fix: fix some displays and features

fix: fix some ui displays and features miss implemented
merge: isolate agent cwd in .ideai/run/<id> (convention-file collision fix, §14.1)
2026-06-06 17:06:45 +02:00 · 2026-06-06 16:15:19 +02:00 · 2026-06-06 12:26:02 +02:00 · 2026-06-06 12:25:54 +02:00 · 2026-06-06 12:24:48 +02:00 · 2026-06-06 12:18:14 +02:00
66 changed files with 5823 additions and 210 deletions
--- a/.claude/worktrees/agent-a2650e91d2bd39ca2
+++ b/.claude/worktrees/agent-a2650e91d2bd39ca2
--- a/.claude/worktrees/agent-a4227a8f495123597
+++ b/.claude/worktrees/agent-a4227a8f495123597
--- a/.claude/worktrees/agent-a78df266a0bbaf5c3
+++ b/.claude/worktrees/agent-a78df266a0bbaf5c3
--- a/.claude/worktrees/agent-aeb1e862ef04b991b
+++ b/.claude/worktrees/agent-aeb1e862ef04b991b
--- a/.gitignore
+++ b/.gitignore
@ -24,6 +24,14 @@ frontend/coverage/
 # ─── Claude Code ────────────────────────────────────────────────────────────
 # Personal, machine-local overrides (shared settings.json, if any, stays tracked).
 .claude/settings.local.json
 # Ephemeral git worktrees created by Claude Code's isolated sub-agents — dev
 # tooling only, unrelated to IdeA (which stays git-independent).
 .claude/worktrees/
 # ─── IdeA project data ──────────────────────────────────────────────────────
 # Ephemeral per-agent run directories (isolated PTY cwd + generated convention
 # files), created at activation — not versioned (ARCHITECTURE §9.1 / §14.1).
 .ideai/run/
 # ─── Editors / OS ───────────────────────────────────────────────────────────
 .idea/
--- a/.ideai/agents.json
+++ b/.ideai/agents.json
@ -0,0 +1,19 @@
 {
  "version": 1,
  "agents": [
    {
      "agentId": "a6ced819-b893-4213-b003-9e9dc79b9641",
      "name": "Main",
      "mdPath": "agents/main.md",
      "profileId": "664cc20c-47b8-53ad-9351-dce3c09c0de4",
      "synchronized": false
    },
    {
      "agentId": "dce19c75-9669-4e45-b8de-9950025157da",
      "name": "Architect",
      "mdPath": "agents/architect.md",
      "profileId": "664cc20c-47b8-53ad-9351-dce3c09c0de4",
      "synchronized": false
    }
  ]
 }
--- a/.ideai/agents/architect.md
+++ b/.ideai/agents/architect.md
@ -0,0 +1,601 @@
 # IdeA — Cartographie d'Architecture
 > Document de référence produit par l'**Agent Architecture**.
 > Fait autorité sur les frontières, ports, adapters, modules et conventions.
 > Toute feature DOIT être validée contre ce document avant développement.
 > Architecture **Hexagonale (Ports & Adapters)** + **SOLID**, stricte.
 >
 > Stack non négociable : Tauri v2 (shell) · Rust (cœur hexagonal) · TypeScript + React (UI) · xterm.js + portable-pty (terminaux) · git2/libgit2 · russh/ssh2 · wsl.exe.
 ---
 ## 1. Principes : SOLID + Hexagonal, appliqués concrètement
 ### 1.1 Règle de dépendance (la seule qui compte)
 ```
            ┌─────────────────────────────────────────────┐
            │           Le sens des dépendances            │
            │                                              │
   Présentation ─► Application ─► Domaine ◄─ Infrastructure │
   (React/Tauri)   (use cases)   (pur)      (adapters)      │
            │                                              │
            └─────────────────────────────────────────────┘
 ```
 - **Le Domaine ne dépend de RIEN** : ni Tauri, ni tokio, ni git2, ni portable-pty, ni serde (le moins possible — voir §1.4). Il ne contient que des entités, value objects, règles métier et **traits = ports**.
 - **L'Application** dépend du Domaine. Elle orchestre les use cases en parlant **uniquement aux ports** (traits), jamais aux adapters concrets.
 - **L'Infrastructure** dépend du Domaine et de l'Application (elle implémente les ports). Elle contient tous les détails techniques (PTY, FS, git, SSH, WSL, stores).
 - **La Présentation** (Tauri commands + React) dépend de l'Application. Les commandes Tauri sont des **adapters entrants (driving adapters)** ; les impl de ports sont des **adapters sortants (driven adapters)**.
 Aucune flèche ne pointe **vers** la présentation ou l'infrastructure. L'inversion de dépendance (le **D** de SOLID) est matérialisée par les traits définis dans le domaine et implémentés dehors.
 ### 1.2 SOLID, point par point, traduit IdeA
 | Principe | Application concrète |
 |---|---|
 | **S** — Single Responsibility | Un use case = une intention métier (`LaunchAgent`, `SyncAgentWithTemplate`). Un adapter = une techno (`Git2Repository` ne fait que du git). Le `LayoutNode` ne gère que la topologie, pas le rendu. |
 | **O** — Open/Closed | Ajouter une IA = ajouter un **profil déclaratif** (donnée), pas du code. Ajouter un mode distant = nouvel adapter `RemoteHost` sans toucher aux use cases. Ajouter une stratégie d'injection de contexte = nouvelle variante d'enum + handler, use case inchangé. |
 | **L** — Liskov | Tout `RemoteHost` (local, SSH, WSL) est substituable : un use case marche identiquement quelle que soit l'impl. Les contrats (pré/postconditions) des ports sont documentés et respectés par chaque adapter. |
 | **I** — Interface Segregation | Ports **fins et ciblés** : `ProcessSpawner`, `FileSystem`, `PtyPort` séparés plutôt qu'un `System` fourre-tout. Un use case ne reçoit que les ports qu'il consomme. |
 | **D** — Dependency Inversion | Domaine définit les traits ; infra les implémente ; l'application reçoit des `Arc<dyn Port>` par **injection** (composition root dans la couche Tauri). |
 ### 1.3 Hexagonal côté Frontend (React aussi)
 L'hexagonal ne s'arrête pas à Rust. Côté React on applique le même découpage :
 - **Domaine UI / modèles de vue** : types TS purs (miroir des DTO), logique de présentation pure (ex. calcul de tailles de cellules d'un `LayoutNode`), testable sans React ni Tauri.
 - **Ports UI** : interfaces TS (`AgentGateway`, `TerminalGateway`, `ProjectGateway`, `LayoutGateway`, `GitGateway`, `RemoteGateway`) décrivant **ce dont l'UI a besoin**, indépendamment du transport.
 - **Adapters UI** : implémentation des ports via `@tauri-apps/api` (`invoke` pour commands, `listen` pour events). Remplaçables par des **mocks** en test/Storybook.
 - **Présentation** : composants React, hooks, state (Zustand/Redux) qui consomment les ports UI, jamais `invoke()` en direct.
 Bénéfice : le frontend est testable et développable sans backend (adapters mock), et la frontière IPC est centralisée en un seul endroit.
 ### 1.4 Domaine pur vs adapters — règle pratique Rust
 - Le crate `domain` est **`#![no_std]`-friendly d'esprit** (pas imposé), sans dépendance I/O. Tolérance pragmatique : `serde` est autorisé **uniquement** pour dériver la (dé)sérialisation des entités persistées (manifeste, layout, profils), car c'est une contrainte métier de format, pas un détail technique d'I/O. Les **traits/ports** y vivent. Pas de `tokio`, pas de `std::process`, pas de `std::fs`.
 - Tout ce qui touche le monde réel (`std::fs`, `Command`, sockets, libgit2, PTY) vit **exclusivement** dans `infrastructure`.
 ---
 ## 2. Découpage en couches & frontière Rust ↔ Tauri ↔ React
 ```
 ┌───────────────────────────────────────────────────────────────────────┐
 │ PRÉSENTATION (Frontend)  —  TypeScript + React + xterm.js               │
 │   features/*  ·  ui-ports (gateways)  ·  tauri-adapters (invoke/listen) │
 └───────────────────────────────┬───────────────────────────────────────┘
                                 │  IPC Tauri (commands ⇄ events, JSON)
 ┌───────────────────────────────▼───────────────────────────────────────┐
 │ PRÉSENTATION (Backend)  —  crate `app-tauri`  (DRIVING ADAPTER)         │
 │   #[tauri::command] handlers · event emitters · COMPOSITION ROOT (DI)   │
 │   PTY byte-stream bridge ⇄ xterm.js                                     │
 └───────────────────────────────┬───────────────────────────────────────┘
                                 │  appels de use cases (Arc<UseCase>)
 ┌───────────────────────────────▼───────────────────────────────────────┐
 │ APPLICATION  —  crate `application`                                     │
 │   Use cases / services · DTOs · orchestration · transactions métier     │
 │   Dépend UNIQUEMENT des ports (traits) du domaine                       │
 └───────────────────────────────┬───────────────────────────────────────┘
                                 │  implémente / consomme
 ┌───────────────────────────────▼───────────────────────────────────────┐
 │ DOMAINE  —  crate `domain`  (PUR, sans I/O)                             │
 │   Entities · Value Objects · Invariants · PORTS (traits) · DomainEvents │
 └───────────────────────────────▲───────────────────────────────────────┘
                                 │  implémentent les ports (DRIVEN ADAPTERS)
 ┌───────────────────────────────┴───────────────────────────────────────┐
 │ INFRASTRUCTURE  —  crate `infrastructure`                               │
 │   portable-pty · git2 · russh/ssh2 · wsl.exe · fs local · md/json store │
 └─────────────────────────────────────────────────────────────────────────┘
 ```
 ### Frontière IPC Tauri — deux directions
 - **Commands (Frontend → Backend, request/response)** : `invoke("create_project", {...})`. Le handler `#[tauri::command]` désérialise le DTO, appelle le use case, renvoie un `Result<DTO, ErrorDTO>`. **Stateless** côté forme : tout l'état vit dans des services managés via `tauri::State`.
 - **Events (Backend → Frontend, push)** : flux PTY (octets/base64), changements de statut d'agent, fin de processus, progrès git, drift de template détecté. Émis via `app_handle.emit(...)` / channels Tauri. L'`EventBus` domaine est relayé vers ces events Tauri par un adapter dans `app-tauri`.
 > **Décision** : le flux PTY haute fréquence passe par des **Tauri Channels** (`tauri::ipc::Channel`) plutôt que des events globaux, pour la perf et l'isolement par session terminal.
 ---
 ## 3. Modèle de domaine
 ### 3.1 Vue d'ensemble (relations)
 ```
 Workspace 1───* Window 1───* Tab 1───1 Project
                  │                       │
                  │ 1                      ├──* Agent ─────? AgentTemplate (origine)
                  │                        │      │ 1
                  │ 1                      │      └──1 AgentProfile (runtime IA, par réf id)
              LayoutTree                   ├──1 GitRepository
              (LayoutNode récursif)        ├──1 RemoteHost (Local | Ssh | Wsl)
                  │ feuilles               └──1 AgentManifest (.ideai/agents.json)
                  ▼
              TerminalSession 1───? Agent (si lancé par un agent)
 ```
 ### 3.2 Entités & Value Objects (avec invariants)
 **`ProjectId`, `AgentId`, `TemplateId`, `ProfileId`, `SessionId`, `WindowId`, `TabId`, `NodeId`** — VO `newtype(Uuid)` ou string typée. Invariant : non vide, immuable.
 **`Project`** (entité, racine d'agrégat projet)
 - Champs : `id`, `name`, `root: ProjectPath`, `remote: RemoteRef`, `created_at`.
 - Invariants : `root` doit être un chemin **absolu et valide pour son `RemoteRef`** ; deux projets ne peuvent partager le même `(remote, root)`.
 **`ProjectPath`** (VO) — chemin absolu normalisé, conscient de la plateforme cible (POSIX vs Windows vs WSL `/mnt/...`).
 **`Agent`** (entité)
 - Champs : `id`, `name`, `context: AgentContextRef` (chemin du `.md` dans `.ideai/`), `profile_id: ProfileId`, `origin: AgentOrigin` (`Scratch` | `FromTemplate { template_id, synced_version }`), `synchronized: bool`.
 - Invariants : `synchronized == true` ⇒ `origin == FromTemplate{..}` (on ne peut pas synchroniser un agent créé from scratch). `context` doit exister à l'activation. `profile_id` doit référencer un `AgentProfile` connu.
 **`AgentTemplate`** (entité, store global)
 - Champs : `id`, `name`, `content_md: MarkdownDoc`, `version: TemplateVersion`, `default_profile_id`.
 - Invariants : `version` **monotone croissante** ; toute modification du `content_md` ⇒ `version + 1` (voir §8).
 **`AgentProfile`** (entité de config runtime IA — le port `AgentRuntime` est paramétré par elle)
 - Champs : `id`, `name`, `command: String`, `args: Vec<String>`, `context_injection: ContextInjection`, `detect: Option<String>`, `cwd_template: String` (ex. `"{projectRoot}"`).
 - Invariants : `command` non vide ; cohérence de `ContextInjection` (voir VO ci-dessous).
 **`ContextInjection`** (VO, enum — cœur du moteur IA flexible)
 ```
 ContextInjection =
  | ConventionFile { target: String }   // ex. "CLAUDE.md" / "AGENTS.md" / "GEMINI.md"
  | Flag           { flag: String }     // ex. "--context-file {path}" ou "-f"
  | Stdin                               // pipe du contenu md sur stdin
  | Env            { var: String }      // ex. "AGENT_CONTEXT_FILE"
 ```
 - Invariants : `ConventionFile.target` est un nom de fichier relatif (pas de `..`, pas absolu) ; `Env.var` est un identifiant d'env valide ; `Flag.flag` non vide.
 **`TerminalSession`** (entité)
 - Champs : `id`, `node_id` (cellule du layout qui l'héberge), `cwd: ProjectPath`, `kind: SessionKind` (`Plain` | `Agent { agent_id }`), `pty_size: PtySize { rows, cols }`, `status` (`Starting|Running|Exited{code}`).
 - Invariants : une cellule (feuille de layout) héberge **au plus une** `TerminalSession` active. `pty_size.rows>0 && cols>0`.
 **`LayoutNode` / `LayoutTree`** (VO récursif — voir §7 pour le détail complet)
 - Invariants : poids relatifs strictement positifs ; somme normalisable ; pas de fusion qui chevauche deux conteneurs distincts ; un `Leaf` référence 0 ou 1 `SessionId`.
 **`RemoteHost`** (VO de stratégie de localisation — abstrait Local/SSH/WSL)
 ```
 RemoteRef =
  | Local
  | Ssh { host, port, user, auth: SshAuth, remote_root }
  | Wsl { distro: String }
 ```
 - Invariants : `Ssh.port` ∈ 1..=65535 ; `Wsl.distro` non vide ; pour `Ssh`/`Wsl`, les chemins projet sont interprétés côté distant.
 **`GitRepository`** (entité)
 - Champs : `project_id`, `root`, `current_branch`, `is_dirty`.
 - Invariants : `root` contient (ou contiendra après init) un `.git`. État dérivé, rafraîchi via le port.
 **`AgentManifest`** (entité — image en mémoire de `.ideai/agents.json`)
 - Champs : `entries: Vec<ManifestEntry { agent_id, md_path, template_id?, synchronized, synced_template_version? }>`.
 - Invariants : `synchronized ⇒ template_id.is_some() && synced_template_version.is_some()` ; `md_path` unique ; cohérence avec les `Agent` chargés.
 **`Workspace` / `Window` / `Tab`** (entités de présentation persistée)
 - `Workspace` = ensemble des fenêtres d'une session utilisateur.
 - `Window` = fenêtre OS ; possède un `LayoutTree` **par onglet actif** et une liste de `Tab`.
 - `Tab` = onglet ⇔ **un `Project`** (1:1).
 - Invariants : un `Project` ouvert apparaît dans **exactement un** `Tab` à la fois (le drag déplace, ne duplique pas) ; un `Window` a ≥ 1 `Tab` ou est fermée.
 **`DomainEvent`** (enum) — `ProjectCreated`, `AgentLaunched`, `AgentExited`, `TemplateUpdated`, `AgentDriftDetected`, `LayoutChanged`, `RemoteConnected`, `GitStateChanged`, `PtyOutput{session_id, bytes}` (ce dernier souvent court-circuité vers un Channel).
 ---
 ## 4. Ports (traits du domaine)
 > Signatures **conceptuelles** (Rust idiomatique, `async` via `async_trait` ou retours `Future` ; erreurs typées par port). « Consommé par » = use cases. « Implémenté par » = adapters de §5.
 ### `AgentRuntime`
 - **Rôle** : lancer/piloter la CLI d'une IA selon un `AgentProfile`, en gérant l'injection du contexte `.md`.
 - **Signature** :
  ```rust
  trait AgentRuntime {
      fn detect(&self, profile: &AgentProfile) -> Result<bool, RuntimeError>;
      fn prepare_invocation(&self, profile: &AgentProfile, ctx: &PreparedContext, cwd: &ProjectPath)
          -> Result<SpawnSpec, RuntimeError>; // commande + args + plan d'injection (fichier/flag/stdin/env)
  }
  ```
 - **Consommé par** : `LaunchAgent`, `DetectProfilesUseCase` (first-run).
 - **Implémenté par** : `CliAgentRuntime` (un seul adapter générique piloté par le profil déclaratif — c'est l'**Open/Closed**). La diversité des IA = données, pas code.
 ### `PtyPort` (alias domaine de `TerminalSessionPort`)
 - **Rôle** : ouvrir un pseudo-terminal, lire/écrire, redimensionner, tuer.
 - **Signature** :
  ```rust
  trait PtyPort {
      async fn spawn(&self, spec: SpawnSpec, size: PtySize) -> Result<PtyHandle, PtyError>;
      fn write(&self, h: &PtyHandle, data: &[u8]) -> Result<(), PtyError>;
      fn resize(&self, h: &PtyHandle, size: PtySize) -> Result<(), PtyError>;
      fn subscribe_output(&self, h: &PtyHandle) -> OutputStream; // flux d'octets
      async fn kill(&self, h: &PtyHandle) -> Result<ExitStatus, PtyError>;
  }
  ```
 - **Consommé par** : `OpenTerminal`, `LaunchAgent`, `CloseTerminal`.
 - **Implémenté par** : `PortablePtyAdapter` (local), `SshPtyAdapter` (PTY distant via russh exec/shell), `WslPtyAdapter` (PTY via `wsl.exe`). Sélection par stratégie `RemoteRef` (Liskov).
 ### `RemoteHost`
 - **Rôle** : abstraction de la **localisation d'exécution** (local / SSH / WSL) : exécuter une commande, ouvrir un PTY, accéder au FS, dans le bon contexte.
 - **Signature** :
  ```rust
  trait RemoteHost {
      fn kind(&self) -> RemoteKind;
      async fn connect(&self) -> Result<(), RemoteError>;
      fn file_system(&self) -> Arc<dyn FileSystem>;
      fn process_spawner(&self) -> Arc<dyn ProcessSpawner>;
      fn pty(&self) -> Arc<dyn PtyPort>;
  }
  ```
 - **Consommé par** : tous les use cases qui touchent un projet (résolvent leurs ports via le `RemoteHost` du projet → **transparence local/distant**).
 - **Implémenté par** : `LocalHost`, `SshHost` (russh/ssh2), `WslHost` (wsl.exe). C'est la **stratégie** qui unifie les 3 modes.
 ### `ProcessSpawner`
 - **Rôle** : lancer un process **non interactif** et récupérer sortie/exit (ex. `detect`, commandes git hors libgit2, scripts).
 - **Signature** : `async fn run(&self, spec: SpawnSpec) -> Result<Output, ProcessError>;`
 - **Consommé par** : `DetectProfilesUseCase`, services divers.
 - **Implémenté par** : `LocalProcessSpawner`, `SshProcessSpawner`, `WslProcessSpawner`.
 ### `FileSystem`
 - **Rôle** : lecture/écriture/listing/symlink, neutre vis-à-vis de la localisation.
 - **Signature** :
  ```rust
  trait FileSystem {
      async fn read(&self, p: &RemotePath) -> Result<Vec<u8>, FsError>;
      async fn write(&self, p: &RemotePath, data: &[u8]) -> Result<(), FsError>;
      async fn exists(&self, p: &RemotePath) -> Result<bool, FsError>;
      async fn create_dir_all(&self, p: &RemotePath) -> Result<(), FsError>;
      async fn list(&self, p: &RemotePath) -> Result<Vec<DirEntry>, FsError>;
      async fn symlink(&self, src: &RemotePath, dst: &RemotePath) -> Result<(), FsError>;
  }
  ```
 - **Consommé par** : `AgentContextStore`, `ProjectStore`, injection `conventionFile`, etc.
 - **Implémenté par** : `LocalFileSystem` (std::fs/tokio::fs), `SshFileSystem` (SFTP), `WslFileSystem` (via `wsl.exe` ou chemins `\\wsl$`).
 ### `TemplateStore`
 - **Rôle** : CRUD des `AgentTemplate` dans le store global IDE + versioning.
 - **Signature** : `list / get / save / delete / bump_version`.
 - **Consommé par** : `CreateTemplate`, `UpdateTemplate`, `CreateAgentFromTemplate`, `SyncAgentWithTemplate`.
 - **Implémenté par** : `FsTemplateStore` (md + index json dans le dossier de données app).
 ### `ProjectStore`
 - **Rôle** : persistance de la liste des projets connus, workspaces, windows, tabs, layouts.
 - **Signature** : `list_projects / load_project / save_project / save_workspace / load_workspace`.
 - **Consommé par** : `CreateProject`, `OpenProject`, persistance fenêtres/onglets/layout.
 - **Implémenté par** : `FsProjectStore` (json dans données app pour le registre ; layout par projet dans `.ideai/`).
 ### `AgentContextStore`
 - **Rôle** : lire/écrire les `.md` d'agents **et** le manifeste `.ideai/agents.json` (au sein du projet, via le `FileSystem` du `RemoteHost`).
 - **Signature** :
  ```rust
  trait AgentContextStore {
      async fn read_context(&self, project: &Project, agent: &AgentId) -> Result<MarkdownDoc, StoreError>;
      async fn write_context(&self, project: &Project, agent: &AgentId, md: &MarkdownDoc) -> Result<(), StoreError>;
      async fn load_manifest(&self, project: &Project) -> Result<AgentManifest, StoreError>;
      async fn save_manifest(&self, project: &Project, m: &AgentManifest) -> Result<(), StoreError>;
  }
  ```
 - **Consommé par** : `CreateAgent*`, `LaunchAgent`, `SyncAgentWithTemplate`.
 - **Implémenté par** : `IdeaiContextStore` (compose `FileSystem`, écrit `.ideai/`).
 ### `GitRepository`
 - **Rôle** : opérations git du projet.
 - **Signature** : `status / stage / unstage / commit / branches / checkout / current_branch / diff / log / pull / push / clone / init`.
 - **Consommé par** : use cases Git.
 - **Implémenté par** : `Git2Repository` (libgit2, local) ; sur SSH/WSL, `RemoteGitRepository` délègue à git CLI via `ProcessSpawner` quand libgit2 ne peut pas atteindre le FS distant (point ouvert §13).
 ### `EventBus`
 - **Rôle** : publier/souscrire les `DomainEvent` (découple émetteurs et présentation).
 - **Signature** : `fn publish(&self, e: DomainEvent); fn subscribe(&self) -> EventStream;`
 - **Consommé par** : tous use cases (publient) ; l'adapter Tauri (souscrit → relaye en events/channels IPC).
 - **Implémenté par** : `TokioBroadcastEventBus` (in-process), relayé par `TauriEventRelay`.
 ### `Clock` & `IdGenerator` (ports utilitaires — testabilité)
 - **Rôle** : éliminer le non-déterminisme (`now()`, `uuid`) du domaine/application.
 - **Implémenté par** : `SystemClock` / `UuidGenerator` (prod), `FixedClock` / `SeqIdGenerator` (tests).
 ---
 ## 5. Adapters (impl concrètes par port)
 | Port | Adapter(s) | Techno | Notes |
 |---|---|---|---|
 | `AgentRuntime` | `CliAgentRuntime` | piloté par `AgentProfile` | Construit `SpawnSpec` + plan d'injection. Un seul adapter, N profils. |
 | `PtyPort` | `PortablePtyAdapter` | portable-pty | Local. Stream octets → Channel Tauri. |
 |  | `SshPtyAdapter` | russh (channel shell/exec + pty req) | Distant SSH. |
 |  | `WslPtyAdapter` | `wsl.exe -d <distro>` + portable-pty | PTY dans la distro. |
 | `RemoteHost` | `LocalHost` / `SshHost` / `WslHost` | — / russh,ssh2 / wsl.exe | Stratégie ; fabrique FS/Spawner/PTY adaptés. |
 | `ProcessSpawner` | `LocalProcessSpawner` | std/tokio `Command` | |
 |  | `SshProcessSpawner` | russh exec | |
 |  | `WslProcessSpawner` | `wsl.exe` | |
 | `FileSystem` | `LocalFileSystem` | tokio::fs | |
 |  | `SshFileSystem` | SFTP (ssh2/russh-sftp) | |
 |  | `WslFileSystem` | `\\wsl$\` / `wsl.exe cat`… | |
 | `TemplateStore` | `FsTemplateStore` | tokio::fs + serde_json | Dossier données app. |
 | `ProjectStore` | `FsProjectStore` | tokio::fs + serde_json | Registre projets + workspace. |
 | `AgentContextStore` | `IdeaiContextStore` | compose `FileSystem` | Écrit `.ideai/`. |
 | `GitRepository` | `Git2Repository` | git2 | Local. |
 |  | `RemoteGitRepository` | git CLI via `ProcessSpawner` | SSH/WSL fallback. |
 | `EventBus` | `TokioBroadcastEventBus` (+ `TauriEventRelay`) | tokio::broadcast | Relais vers IPC. |
 | `Clock`/`IdGenerator` | `SystemClock`/`UuidGenerator` | std/uuid | Mocks en test. |
 **Adapters entrants (driving)** : handlers `#[tauri::command]` (frontend → app) + `TauriEventRelay` (app → frontend). Côté UI : `tauri-adapters` implémentant les gateways TS.
 ---
 ## 6. Use cases / services applicatifs
 > Chaque use case : un struct `XxxUseCase` portant ses ports en `Arc<dyn Port>`, une méthode `execute(input: XxxInput) -> Result<XxxOutput, AppError>`. **Single Responsibility**. Aucune dépendance à Tauri.
 | Use case | Rôle | Ports consommés |
 |---|---|---|
 | `CreateProject` | Crée un projet (project root), init `.ideai/`, registre. | `ProjectStore`, `FileSystem`, `IdGenerator`, `EventBus` |
 | `OpenProject` | Charge projet, manifeste, layout, résout `RemoteHost`. | `ProjectStore`, `AgentContextStore`, `RemoteHost` |
 | `CloseProject` / `CloseTab` | Persiste l'état, libère PTYs. | `ProjectStore`, `PtyPort`, `EventBus` |
 | `DetectProfiles` (first-run) | Teste `detect` de chaque profil candidat. | `AgentRuntime`, `ProcessSpawner` |
 | `ConfigureProfiles` | Enregistre profils choisis/édités/custom. | `TemplateStore`/profile store, `FileSystem` |
 | `CreateAgentFromScratch` | Crée agent + `.md`, met à jour manifeste. | `AgentContextStore`, `IdGenerator` |
 | `CreateAgentFromTemplate` | Copie le `content_md` du template → agent ; lie origine + version + `synchronized`. | `TemplateStore`, `AgentContextStore` |
 | `UpdateTemplate` | Modifie un template, **bump version**, signale drift aux agents liés. | `TemplateStore`, `EventBus` |
 | `DetectAgentDrift` | Compare `synced_template_version` vs `template.version`. | `TemplateStore`, `AgentContextStore` |
 | `SyncAgentWithTemplate` | Applique la MAJ template→agent si `synchronized`. | `TemplateStore`, `AgentContextStore`, `EventBus` |
 | `LaunchAgent` | Résout profil+contexte, prépare injection, ouvre cellule PTY au bon `cwd`, spawn CLI. | `AgentRuntime`, `AgentContextStore`, `RemoteHost`→`PtyPort`/`FileSystem`, `EventBus` |
 | `OpenTerminal` | Ouvre un PTY simple dans une cellule. | `RemoteHost`→`PtyPort`, `EventBus` |
 | `WriteToTerminal` / `ResizeTerminal` / `CloseTerminal` | I/O PTY. | `PtyPort` |
 | `MutateLayout` (split/merge/resize/move) | Applique une opération sur le `LayoutTree` (logique **pure** dans le domaine, persistée ici). | `ProjectStore` (persistance) |
 | `ConnectRemote` (SSH/WSL) | Établit la connexion, valide l'accès au root. | `RemoteHost`, `FileSystem` |
 | `MoveTabToNewWindow` | Détache un onglet → nouvelle fenêtre (réaffectation `WindowId`). | `ProjectStore`, `EventBus` |
 | Use cases Git | `GitStatus`, `GitCommit`, `GitCheckout`, `GitPush`, … | `GitRepository`, `EventBus` |
 ---
 ## 7. Modèle de layout terminal (grille tableur récursive + fusion)
 ### 7.1 Structure de données
 La grille « type tableur, lignes/colonnes imbriquées indépendamment + fusion » est modélisée par un **arbre de splits récursif** où chaque conteneur définit son propre découpage. La **fusion** est obtenue nativement : fusionner = ne pas subdiviser une zone (un `Leaf` couvre plusieurs « cellules visuelles » d'un parent voisin). Pour le cas Excel pur (fusion arbitraire chevauchant la grille), on superpose un modèle **GridContainer** avec spans.
 ```rust
 enum LayoutNode {
    Leaf(LeafCell),
    Split(SplitContainer),
    Grid(GridContainer),
 }
 struct LeafCell {
    id: NodeId,
    session: Option<SessionId>,  // 0 ou 1 terminal
 }
 struct SplitContainer {       // découpage simple binaire/n-aire pondéré
    id: NodeId,
    direction: Direction,     // Row (colonnes) | Column (lignes)
    children: Vec<WeightedChild>, // ordre = gauche→droite / haut→bas
 }
 struct WeightedChild { node: LayoutNode, weight: f32 } // poids = part redimensionnable
 struct GridContainer {        // grille tableur avec fusion (spans)
    id: NodeId,
    col_weights: Vec<f32>,    // largeurs de colonnes
    row_weights: Vec<f32>,    // hauteurs de lignes
    cells: Vec<GridCell>,     // placements avec spans (fusion)
 }
 struct GridCell {
    node: LayoutNode,         // récursif : une cellule peut re-contenir un Split/Grid
    row: u16, col: u16,
    row_span: u16,            // ≥1 ; >1 = cellules fusionnées verticalement
    col_span: u16,            // ≥1 ; >1 = cellules fusionnées horizontalement
 }
 ```
 - **Lignes/colonnes indépendantes par zone** : chaque `SplitContainer`/`GridContainer` a ses propres poids ⇒ pas de grille uniforme rigide.
 - **Imbrication** : un enfant peut être un nouveau `Split`/`Grid` ⇒ « N colonnes dans une ligne, M lignes dans une colonne » de façon arbitraire.
 - **Fusion** : `row_span`/`col_span` dans `GridContainer` (modèle tableur fidèle) **ou** simplement un `Leaf` plus grand via `SplitContainer` (cas courant). Le domaine supporte les deux ; l'UI choisit la représentation selon l'interaction.
 ### 7.2 Invariants (validés dans le domaine, testables sans I/O)
 - Tous les `weight > 0`. Les poids sont **relatifs** (l'UI normalise pour le rendu).
 - Dans un `GridContainer` : aucune superposition de spans ; toute la surface couverte ; `row+row_span ≤ rows`, `col+col_span ≤ cols`.
 - Un `SessionId` n'apparaît que dans **un seul** `Leaf`.
 - Les opérations `split`, `merge`, `resize`, `move` sont des **fonctions pures** `LayoutTree -> Result<LayoutTree, LayoutError>` (immutabilité ⇒ testabilité, undo/redo facile).
 ### 7.3 Sérialisation & persistance
 - Sérialisé en **JSON** (serde, `tag`/`content` pour l'enum) → `.ideai/layout.json` (par projet, donc voyage avec le projet, y compris distant).
 - Le `Workspace`/`Window`/`Tab` (organisation des fenêtres OS) est persisté côté **store global IDE** (machine-local, pas dans le projet) car lié à l'écran de l'utilisateur, pas au code.
 ---
 ## 8. Synchronisation template → agents
 ### 8.1 Versioning
 - `AgentTemplate.version: u64` monotone. **`UpdateTemplate` incrémente** la version à chaque changement de `content_md`. Un hash du contenu (`content_hash`) est aussi stocké pour détecter les éditions hors-app.
 - Chaque `ManifestEntry` d'agent lié garde `synced_template_version` = version du template **au dernier sync réussi**.
 ### 8.2 Détection de drift
 ```
 drift(agent) =
    agent.synchronized
    && agent.origin == FromTemplate{ template_id, .. }
    && template_store.get(template_id).version > entry.synced_template_version
 ```
 `DetectAgentDrift` est lancé à `OpenProject` et après chaque `UpdateTemplate` ; émet `AgentDriftDetected { agent_id, from, to }` → badge UI.
 ### 8.3 Application de la MAJ (`SyncAgentWithTemplate`)
 ```
 1. Charger template (version courante) + manifeste projet.
 2. Pour chaque agent ciblé avec synchronized==true :
     a. Stratégie de MAJ = REMPLACEMENT du .md par content_md du template
        (le contexte d'un agent synchronisé est "possédé" par le template).
        → Variante future : merge 3-way si l'agent a un bloc local marqué.
     b. write_context(agent, template.content_md)
     c. entry.synced_template_version = template.version
 3. save_manifest. publish(AgentSynced{..}).
 ```
 ### 8.4 Agents non synchronisés
 - `synchronized == false` : ne reçoivent **jamais** de MAJ auto. Ils gardent leur `.md` libre. On peut afficher « une nouvelle version du template existe » (info) mais aucune écriture n'a lieu sans action explicite (qui basculerait `synchronized` ou ferait un sync ponctuel one-shot).
 - Agents `Scratch` : aucun lien template, hors périmètre de sync.
 ---
 ## 9. Stockage & arborescence des fichiers
 ### 9.1 Dans le projet — `.ideai/` (voyage avec le code, versionnable)
 ```
 <project_root>/
 ├── .ideai/
 │   ├── agents.json          # AgentManifest (mapping md ↔ template ↔ sync ↔ version)
 │   ├── layout.json          # LayoutTree de l'onglet (sérialisé)
 │   ├── project.json         # méta projet local (nom, profil par défaut, remote ref)
 │   └── agents/
 │       ├── reviewer.md       # contexte d'un agent de projet
 │       ├── backend-dev.md
 │       └── ...
 └── (CLAUDE.md / AGENTS.md / GEMINI.md générés/symlinkés à l'activation si conventionFile)
 ```
 **Schéma `agents.json`** :
 ```json
 {
  "version": 1,
  "agents": [
    {
      "id": "a3f1...",
      "name": "Backend Dev",
      "md": "agents/backend-dev.md",
      "profileId": "claude-code",
      "origin": { "type": "fromTemplate", "templateId": "tpl-backend", "syncedTemplateVersion": 4 },
      "synchronized": true
    },
    {
      "id": "b7c2...",
      "name": "Ad-hoc",
      "md": "agents/adhoc.md",
      "profileId": "codex-cli",
      "origin": { "type": "scratch" },
      "synchronized": false
    }
  ]
 }
 ```
 ### 9.2 Store global IDE (données app, hors projet, machine-local)
 Emplacement résolu via Tauri path API (`AppData`/`~/.local/share/IdeA`/`~/Library/Application Support/IdeA`).
 ```
 <app_data_dir>/IdeA/
 ├── profiles.json            # AgentProfile[] configurés (first-run + custom + édités)
 ├── settings.json            # préférences IDE
 ├── workspace.json           # Workspace/Window/Tab + quel projet dans quel onglet (machine-local)
 └── templates/
    ├── index.json           # [{id, name, version, contentHash, defaultProfileId}]
    └── md/
        ├── tpl-backend.md
        ├── tpl-reviewer.md
        └── ...
 ```
 **Schéma `profiles.json` (item)** : exactement le profil déclaratif de CONTEXT.md §9 (`id, name, command, args, contextInjection{strategy,target/flag/var}, detect, cwd`).
 **Formats** : contextes & templates en **Markdown** ; tout le reste en **JSON** (serde). Pas de base de données : fichiers plats, simples, diffables, portables (AppImage friendly).
 ---
 ## 10. Arborescence du repo
 ### 10.1 Décision : workspace Cargo **multi-crate**
 **Multi-crate** retenu (vs mono-crate) pour **forcer** la règle de dépendance à la compilation : le crate `domain` ne peut littéralement pas dépendre de `infrastructure` si ce n'est pas dans son `Cargo.toml`. C'est la garantie mécanique de l'hexagonal (mieux qu'une convention). Coût : un peu de cérémonie de workspace — acceptable et même souhaitable ici vu le découpage en lots/agents (§12).
 ```
 IdeA/
 ├── Cargo.toml                      # [workspace] members
 ├── ARCHITECTURE.md
 ├── CONTEXT.md
 ├── crates/
 │   ├── domain/                     # PUR : entities, VO, ports (traits), domain events, layout logic
 │   │   └── src/{project,agent,template,profile,terminal,layout,remote,git,ports,events}.rs
 │   ├── application/                # use cases, DTOs, AppError ; dépend de domain
 │   │   └── src/{project,agent,template,terminal,layout,remote,git}/
 │   ├── infrastructure/             # adapters ; dépend de domain (+ application pour DTO si besoin)
 │   │   └── src/{pty,fs,process,remote,git,store,runtime,eventbus}/
 │   └── app-tauri/                  # binaire Tauri : commands, events, COMPOSITION ROOT (DI)
 │       ├── src/{commands,events,state,main.rs}
 │       ├── tauri.conf.json
 │       ├── build.rs
 │       └── icons/, bundle (NSIS + AppImage)
 ├── frontend/                       # TypeScript + React (Vite)
 │   ├── package.json, vite.config.ts, index.html
 │   └── src/
 │       ├── domain/                 # types & logique de vue purs (miroir DTO, calc layout)
 │       ├── ports/                  # gateways TS (interfaces) : AgentGateway, TerminalGateway, ...
 │       ├── adapters/               # impl gateways via @tauri-apps/api (invoke/listen/Channel)
 │       │   └── mock/               # impl mock pour dev/test/storybook
 │       ├── features/               # par feature : projects, agents, templates, terminals, layout, git, remote, first-run
 │       │   └── <feature>/{components,hooks,store,index.ts}
 │       ├── shared/                 # ui kit, xterm wrapper, design system
 │       └── app/                    # bootstrap, routing, providers (DI des adapters)
 └── docs/                           # ADRs, schémas
 ```
 `app-tauri` = **seul** endroit qui connaît tous les crates : il instancie les adapters concrets et injecte dans les use cases (composition root). Personne d'autre ne fait de `new ConcreteAdapter`.
 ---
 ## 11. Stratégie de tests
 | Couche | Type de test | Comment / où |
 |---|---|---|
 | `domain` | **Unitaires purs** (sans I/O, sans async) | `#[cfg(test)] mod tests` par module. Invariants d'entités, opérations de layout (split/merge/resize), détection de drift, validation `ContextInjection`. Déterministe via `FixedClock`/`SeqIdGenerator`. |
 | `application` | **Unitaires avec ports mockés** | Chaque use case testé avec des **mocks de ports** (`mockall` ou fakes manuels). Ex. `LaunchAgent` vérifie qu'il appelle `prepare_invocation` puis `pty.spawn` avec le bon `cwd` et plan d'injection. **Aucun vrai PTY/FS/git.** |
 | `infrastructure` | **Tests d'intégration ciblés** | Par adapter : `LocalFileSystem` sur tmpdir, `Git2Repository` sur repo temporaire, `PortablePtyAdapter` lance `echo`. SSH/WSL : tests `#[ignore]` gated derrière feature/env (CI conditionnelle). |
 | `app-tauri` | Tests des commands (mapping DTO ↔ use case) | Wiring testé avec use cases réels + adapters in-memory. |
 | Frontend `domain`/`ports` | **Vitest** (unitaires purs) | Logique de vue, calc tailles cellules, réducteurs de state. |
 | Frontend `features` | **React Testing Library** + **gateways mock** | Composants testés avec adapters mock ⇒ **sans backend**. |
 | E2E (plus tard) | Playwright / `tauri-driver` | Smoke tests des parcours clés. |
 **Clé de testabilité** : grâce aux **ports**, le domaine et l'application se testent **100 % sans I/O**. C'est l'argument central de l'hexagonal et le socle du cycle dev↔test (chaque agent dev appairé à un agent test, cf. CONTEXT §3). Règle d'or : une feature n'est verte que quand `cargo test -p <crate>` et `vitest` passent.
 ---
 ## 12. Découpage en lots/features livrables
 > Chaque lot = périmètre autonome, validable par le cycle dev/test, confiable à **un binôme (agent dev + agent test)**. Ordonnés par dépendance.
 | # | Lot | Contenu | Crates/zones |
 |---|---|---|---|
 | L0 | **Socle domaine & ports** | Entities, VO, **tous les traits ports**, domain events, `AppError`. Aucun adapter. | `domain` (+ ports utilitaires) |
 | L1 | **Composition root & IPC** | `app-tauri` : DI, registre de commands/events, bridge PTY↔Channel, gateways TS + adapters Tauri + mocks. | `app-tauri`, `frontend/ports`+`adapters` |
 | L2 | **Projets & stockage** | `CreateProject`/`OpenProject`/`CloseProject`, `FsProjectStore`, `LocalFileSystem`, init `.ideai/`. UI projets/onglets. | `application/project`, `infrastructure/{fs,store}`, `frontend/features/projects` |
 | L3 | **Terminaux & PTY (local)** | `PtyPort` + `PortablePtyAdapter`, use cases terminal, wrapper xterm.js, flux Channel. | `infrastructure/pty`, `application/terminal`, `frontend/features/terminals` |
 | L4 | **Layout tableur** | Logique pure `LayoutTree` (déjà en L0 partiellement), `MutateLayout`, persistance `layout.json`, UI grille redimensionnable + fusion. | `domain/layout`, `application/layout`, `frontend/features/layout` |
 | L5 | **Profils IA & runtime** | `AgentProfile`, `CliAgentRuntime`, `DetectProfiles`, first-run wizard, `profiles.json`. | `infrastructure/runtime`, `application/agent`, `frontend/features/first-run` |
 | L6 | **Agents & contextes** | `AgentContextStore`/`IdeaiContextStore`, CRUD agents, `LaunchAgent` (injection + spawn + cellule). | `application/agent`, `infrastructure/store`, `frontend/features/agents` |
 | L7 | **Templates & synchro** | `TemplateStore`, versioning, `DetectAgentDrift`, `SyncAgentWithTemplate`. UI templates + badges drift. | `application/template`, `infrastructure/store`, `frontend/features/templates` |
 | L8 | **Git** | `GitRepository`/`Git2Repository`, use cases git, UI git. | `infrastructure/git`, `application/git`, `frontend/features/git` |
 | L9 | **Remote (SSH + WSL)** | `RemoteHost` stratégie, `SshHost`/`WslHost`, adapters FS/PTY/Spawner distants, `RemoteGitRepository`. UI connexion. | `infrastructure/remote`, `application/remote`, `frontend/features/remote` |
 | L10 | **Fenêtres & multi-window** | `Workspace`/`Window`/`Tab`, `MoveTabToNewWindow`, drag d'onglet → nouvelle fenêtre OS Tauri. | `application`, `app-tauri`, `frontend/app` |
 | L11 | **Packaging & livraison** | Tauri bundle : NSIS `setup.exe`, **AppImage** multi-distro, CI Linux+Windows. | `app-tauri`, CI |
 ---
 ## 13. Risques techniques & points ouverts (spikes)
 1. **PTY cross-platform** : portable-pty + xterm.js OK sur les 3 OS, mais signaux/resize/exit codes diffèrent (Windows ConPTY). **Spike** L3.
 2. **AppImage multi-distro** : libgit2/openssl/glibc liés dynamiquement → risque de non-portabilité. **Spike** : vendoring statique (`git2` features, `rustls` pour russh au lieu d'OpenSSL), test sur ≥3 distros (Ubuntu/Fedora/Arch). L11.
 3. **Drag d'onglet entre fenêtres Tauri** : Tauri v2 multi-webview/multi-window + DnD natif inter-fenêtres est délicat (le DnD HTML ne traverse pas les fenêtres OS). **Spike** : protocole « detach » (créer une `WebviewWindow`, transférer l'état via store + event, fermer l'onglet source). L10.
 4. **Git sur FS distant** : libgit2 ne lit pas un FS SSH/WSL directement. Décision : **fallback git CLI** (`RemoteGitRepository`) côté distant via `ProcessSpawner`. À valider (perf, parsing). L9.
 5. **Synchro temps réel UI ↔ PTY** : volume d'octets élevé ; backpressure des Channels Tauri, throttling/coalescing côté front. **Spike** L3.
 6. **Injection `conventionFile`** : symlink vs copie du `.md` vers `CLAUDE.md`/`AGENTS.md` ; conflits si fichier existant, .gitignore, droits Windows (symlinks). À cadrer L6.
 7. **SSH auth** : agent/clé/mot de passe/known_hosts ; choix russh (rustls) vs ssh2 (libssh2/OpenSSL — impacte point 2). Décision à figer début L9.
 8. **WSL chemins** : conversion `/mnt/c/...` ↔ `\\wsl$\...`, distros multiples, perf I/O cross-boundary. Spike L9.
 9. **Détection d'édition hors-app** des `.md`/templates (content hash) et résolution de conflit lors du sync. L7.
 ---
 *Document maintenu par l'Agent Architecture — base du jalon « cadrage architecture » avant tout code applicatif.*
--- a/.ideai/agents/main.md
+++ b/.ideai/agents/main.md
@ -0,0 +1,187 @@
 # IdeA — Contexte & Méthode de travail
 > Ce document définit **mon rôle**, **la méthode de développement** et **la vision produit** du projet IdeA.
 > Il fait autorité sur la façon dont le projet est piloté. Toute évolution de méthode doit être répercutée ici.
 ---
 ## 1. Mon rôle : chef d'orchestre, pas développeur
 Je **n'écris pas de code moi-même**. Mon rôle est de **piloter des agents** qui réalisent le travail.
 Je suis responsable de :
 - Découper le travail en tâches claires et autonomes.
 - Attribuer chaque tâche aux bons agents.
 - Garantir que le cycle de développement/test est respecté.
 - Faire respecter les principes d'architecture (SOLID, Hexagonal).
 - Maintenir la cohérence globale du projet et de ce document.
 - Arbitrer et valider avant toute action irréversible ou sortante.
 ---
 ## 2. Les agents
 ### 2.1 Agent Architecture (1 pour tout le projet)
 - Garant de l'architecture globale : **Hexagonale (Ports & Adapters)** et principes **SOLID**.
 - Définit les frontières (domaine / application / infrastructure), les ports, les contrats.
 - Valide que chaque nouvelle feature respecte la structure avant son développement.
 - Tient à jour la cartographie d'architecture et les conventions.
 ### 2.2 Agents de Développement
 - Écrivent le code des features.
 - Respectent strictement l'architecture définie par l'agent Architecture.
 - Code **propre, structuré, stable**.
 - Reçoivent les rapports d'erreurs des agents de test et corrigent.
 ### 2.3 Agents de Test
 - **Chaque agent de développement est appairé avec un agent de test dédié.**
 - Écrivent et exécutent les **tests unitaires** des features implémentées ou modifiées.
 - Produisent un **rapport d'erreurs** clair quand un test échoue.
 - Re-testent après chaque correction.
 ---
 ## 3. Le cycle de développement (boucle obligatoire)
 Pour **chaque** feature implémentée ou modifiée :
 ```
 1. Agent Architecture  → valide le découpage et les contrats (ports/interfaces)
 2. Agent Développement → écrit le code
 3. Agent Test          → écrit les tests unitaires + les exécute
 4a. Tests OK           → feature validée, on passe à la suite
 4b. Tests KO           → rapport d'erreurs → retour à l'agent Développement
                         → correction → retour à l'étape 3 (boucle jusqu'au vert)
 ```
 **Règle d'or :** aucune feature n'est considérée terminée tant que ses tests ne passent pas.
 Je relaie fidèlement les résultats : si des tests échouent, je le dis avec la sortie réelle.
 ---
 ## 4. Principes de code
 - **SOLID** appliqué au maximum.
 - **Architecture Hexagonale** (Ports & Adapters) : le domaine métier est isolé des détails techniques (UI, terminal, git, SSH, système de fichiers...).
 - Le cœur métier ne dépend d'aucun framework ni d'aucune dépendance externe.
 - Tests unitaires systématiques ; couverture des features critiques.
 - Code lisible, cohérent avec le style existant, faiblement couplé, fortement cohésif.
 ---
 ## 5. Vision produit : IdeA
 **IdeA est un IDE next-gen 100 % IA.** On n'y code pas : **on gère des IA.**
 ### Fonctionnalités clés
 - **Multi-projets en parallèle** : un **onglet par projet**.
 - **Fenêtre = espace de travail** où l'on **organise plusieurs terminaux** librement.
 - **Agents par projet** : chaque projet a ses propres agents.
 - **Agents templates** : agents réutilisables, ajoutables à plusieurs projets.
 - **Création d'agents** : depuis zéro ou à partir d'un template.
 - **Synchronisation template → agents** : option « garder l'agent à jour ».
  Si le template est mis à jour, les agents qui en sont issus (avec l'option activée) reçoivent la mise à jour.
 - **Contextes d'agents stockés en `.md`** (toujours).
 - **Création de projet** = définition de son **project root**.
 ### Intégrations
 - **Git** intégré.
 - **Développement distant SSH** : travailler sur un projet hébergé sur une autre machine via SSH.
 - **Développement WSL** : travailler sur une WSL depuis Windows.
 ### Plateformes & livraison
 - Cible : **macOS, Linux, Windows**.
 - Première phase de compilation : **Linux et Windows**.
 - Livraison :
  - **Windows** : `setup.exe`.
  - **Linux** : **AppImage** (doit fonctionner sur les différentes distributions).
 ---
 ## 6. Stack technique (validée)
 - **Shell applicatif** : **Tauri v2** (binaires légers, performants, multi-OS, AppImage + installeur `setup.exe`/NSIS Windows natifs).
 - **Cœur / backend** : **Rust** — stabilité, performance, et expression idiomatique du domaine hexagonal (ports = traits, adapters = implémentations).
 - **Frontend / UI** : **TypeScript + React**.
 - **Terminaux** : **xterm.js** (rendu) + **portable-pty** (PTY côté Rust).
 - **Git** : **libgit2** via `git2` (Rust).
 - **SSH** : `russh` / `ssh2` (Rust).
 - **WSL** : invocation de `wsl.exe` depuis le backend.
 ## 7. Layout des terminaux (exigence produit)
 Disposition en **grille redimensionnable de type tableur (Excel)** :
 - Splits redimensionnables horizontaux **et** verticaux.
 - L'utilisateur peut **définir le nombre de colonnes dans une ligne** et **le nombre de lignes dans une colonne**, indépendamment par zone.
 - Possibilité de **fusionner des cellules** (ex. fusionner deux colonnes sur une ligne), à la manière des cellules fusionnées d'un tableur.
 - Chaque cellule de la grille héberge un terminal.
 - → Modèle de layout récursif/imbriqué (pas une grille rigide uniforme) à concevoir par l'agent Architecture.
 ## 8. Stockage des contextes & liaison aux templates
 - **Templates d'agents** : stockés dans l'**IDE** (dossier de données utilisateur global de l'app, hors projet).
 - **Agents de projet** : leurs `.md` sont stockés dans un dossier **`.ideai/`** à la racine du project root.
  *(Nom choisi pour éviter toute collision avec le `.idea` de JetBrains.)*
 - **Manifeste de liaison** dans `.ideai/` (ex. `.ideai/agents.json`) qui mappe pour chaque agent de projet :
  - le `.md` de l'agent,
  - le template d'origine (le cas échéant),
  - `synchronized: true/false`,
  - la **version du template** au dernier sync (pour détecter qu'une mise à jour est disponible).
 - **Synchro template → agents** : quand un template est mis à jour, les agents liés avec `synchronized: true` reçoivent la MAJ.
 ## 9. Moteur IA : adaptateur de CLI flexible (Port `AgentRuntime`)
 Chaque IA est décrite par un **profil déclaratif** (config éditable, pas du code), implémentation d'un **Port** `AgentRuntime` côté domaine. Deux variables clés par IA :
 1. **Commande de lancement** + arguments (ex. `claude`, `codex`, `gemini`, `aider`).
 2. **Stratégie d'injection du contexte `.md`** :
   - `conventionFile` : écrire/symlink le `.md` vers le fichier attendu par la CLI (`CLAUDE.md`, `AGENTS.md`, `GEMINI.md`…).
   - `flag` : passer le chemin via un argument.
   - `stdin` : piper le contenu.
   - `env` : passer via variable d'environnement.
 Exemple de profil :
 ```json
 {
  "id": "claude-code",
  "name": "Claude Code",
  "command": "claude",
  "args": [],
  "contextInjection": { "strategy": "conventionFile", "target": "CLAUDE.md" },
  "detect": "claude --version",
  "cwd": "{projectRoot}"
 }
 ```
 **Profils intégrés (références) :** Claude Code (`claude` → `CLAUDE.md`), OpenAI Codex CLI (`codex` → `AGENTS.md`), Gemini CLI (`gemini` → `GEMINI.md`), Aider (`aider` → args/message).
 **Règles produit :**
 - **Premier lancement de l'IDE** : un assistant (first-run) **demande à l'utilisateur** quels profils d'IA configurer. On ne présume rien par défaut.
 - Les commandes des profils sont **pré-remplies mais éditables**.
 - L'utilisateur peut **ajouter sa propre commande CLI** (profil custom) pour n'importe quelle IA.
 **Lancement d'un agent :** à l'**activation de l'agent**, on ouvre une cellule terminal (PTY) avec le bon `cwd`, on injecte le contexte `.md`, et on **auto-lance** la CLI du profil.
 ## 10. Fenêtres & onglets
 - **Par défaut : un onglet par projet** (comme les IDE classiques).
 - **Drag & drop d'un onglet** hors de la fenêtre → **crée une nouvelle fenêtre OS** portant ce projet.
 - **Multi-fenêtres OS supporté** ; chaque fenêtre possède un ou plusieurs onglets/projets.
 ## 11. Feuille de route
 1. **Cadrage architecture complet d'abord** (jalon en cours) : l'agent Architecture produit la cartographie complète — domaine, ports, adapters, modules, arborescence — **avant tout code**.
 2. Puis MVP incrémental selon le cycle dev/test de la section 3.
 ## 12. Autonomie d'exécution dans le projet
 L'utilisateur m'accorde un **accès large et autonome** sur le dossier du projet : je peux lire, créer, modifier des fichiers et exécuter les commandes de développement (cargo, npm, npx, git, etc.) **sans demander confirmation à chaque fois**.
 - Concrètement, ces autorisations sont matérialisées dans `.claude/settings.local.json` (mode `acceptEdits` + `Bash`/`Read`/`Edit`/`Write` autorisés), pas dans ce document — CONTEXT.md ne fait que **documenter l'intention**.
 - **Garde-fous conservés** : les actions destructrices ou hors-projet restent bloquées (`sudo`, `rm -rf` sur `/`/`~`/`$HOME`, `mkfs`, `dd`, `shutdown`/`reboot`…).
 - L'esprit du rôle (§1) ne change pas : je reste **chef d'orchestre**. L'autonomie porte sur l'exécution mécanique, pas sur l'arbitrage des décisions produit/archi, ni sur les **actions sortantes** (push, publication) qui restent soumises à validation explicite.
 ---
 *Dernière mise à jour : 2026-06-05*
--- a/.ideai/layouts.json
+++ b/.ideai/layouts.json
@ -0,0 +1,56 @@
 {
  "version": 1,
  "activeId": "9188db80-8535-4786-a20b-3c9a36b222e3",
  "layouts": [
    {
      "id": "9188db80-8535-4786-a20b-3c9a36b222e3",
      "name": "Default",
      "kind": "terminal",
      "tree": {
        "root": {
          "type": "split",
          "node": {
            "id": "8aca2f93-1a9b-4693-9bba-9a01e130a48c",
            "direction": "row",
            "children": [
              {
                "node": {
                  "type": "leaf",
                  "node": {
                    "id": "d8a86eb1-cd4d-4937-b900-4989da7c868d",
                    "session": "2e6293fe-250a-4654-a8b2-4fe6996f8288",
                    "agent": "a6ced819-b893-4213-b003-9e9dc79b9641"
                  }
                },
                "weight": 1.0
              },
              {
                "node": {
                  "type": "leaf",
                  "node": {
                    "id": "6c5be5e7-a54b-468c-a2e2-8ec853629d5e",
                    "session": "9aa5859c-5868-46b3-a045-68cbea6eb629"
                  }
                },
                "weight": 1.0
              }
            ]
          }
        }
      }
    },
    {
      "id": "40ea4fa9-25b3-410b-9d3e-6350834b421b",
      "name": "Git Graph",
      "kind": "gitGraph",
      "tree": {
        "root": {
          "type": "leaf",
          "node": {
            "id": "c840bfdd-3330-46a3-b727-0799f6853e72"
          }
        }
      }
    }
  ]
 }
--- a/.ideai/project.json
+++ b/.ideai/project.json
@ -0,0 +1,9 @@
 {
  "version": 1,
  "id": "97b49ac2-8376-4aa3-8ea9-bf3ac81d0023",
  "name": "IdeA",
  "remote": {
    "kind": "local"
  },
  "createdAt": 1780702317785
 }
--- a/ARCHITECTURE.md
+++ b/ARCHITECTURE.md
@ -6,6 +6,8 @@
 > Architecture **Hexagonale (Ports & Adapters)** + **SOLID**, stricte.
 >
 > Stack non négociable : Tauri v2 (shell) · Rust (cœur hexagonal) · TypeScript + React (UI) · xterm.js + portable-pty (terminaux) · git2/libgit2 · russh/ssh2 · wsl.exe.
 >
 > **Principe fondateur** : IdeA est un IDE 100 % IA dont le rôle est de **refléter fidèlement la façon dont on travaille avec des IAs** — sans jamais dépendre des commandes, flags ou conventions d'un modèle en particulier. Les deux abstractions de premier rang sont les **Agents** (instances IA à rôle/contexte définis) et les **Skills** (workflows réutilisables). Ces deux concepts sont gérés par IdeA de façon universelle : un utilisateur qui passe de Claude Code à Gemini CLI ou Codex retrouve exactement les mêmes Agents et Skills — seul le moteur d'exécution change.
 ---
@ -134,7 +136,7 @@ Workspace 1───* Window 1───* Tab 1───1 Project
 - Invariants : `version` **monotone croissante** ; toute modification du `content_md` ⇒ `version + 1` (voir §8).
 **`AgentProfile`** (entité de config runtime IA — le port `AgentRuntime` est paramétré par elle)
- Champs : `id`, `name`, `command: String`, `args: Vec<String>`, `context_injection: ContextInjection`, `detect: Option<String>`, `cwd_template: String` (ex. `"{projectRoot}"`).
+- Champs : `id`, `name`, `command: String`, `args: Vec<String>`, `context_injection: ContextInjection`, `detect: Option<String>`, `cwd_template: String` (vaut toujours `"{agentRunDir}"` — voir §9.1 et §14.1).
 - Invariants : `command` non vide ; cohérence de `ContextInjection` (voir VO ci-dessous).
 **`ContextInjection`** (VO, enum — cœur du moteur IA flexible)
@ -177,7 +179,12 @@ RemoteRef =
 - `Tab` = onglet ⇔ **un `Project`** (1:1).
 - Invariants : un `Project` ouvert apparaît dans **exactement un** `Tab` à la fois (le drag déplace, ne duplique pas) ; un `Window` a ≥ 1 `Tab` ou est fermée.
-**`DomainEvent`** (enum) — `ProjectCreated`, `AgentLaunched`, `AgentExited`, `TemplateUpdated`, `AgentDriftDetected`, `LayoutChanged`, `RemoteConnected`, `GitStateChanged`, `PtyOutput{session_id, bytes}` (ce dernier souvent court-circuité vers un Channel).
+**`Skill`** (entité)
 - Champs : `id`, `name`, `content_md: MarkdownDoc`, `scope: SkillScope` (`Global` | `Project`).
 - Invariants : `name` non vide ; `content_md` non vide.
 - Un agent référence 0..N skills (dans l'`AgentManifest`). Les skills assignés sont injectés dans son convention file à l'activation.
 **`DomainEvent`** (enum) — `ProjectCreated`, `AgentLaunched`, `AgentExited`, `TemplateUpdated`, `AgentDriftDetected`, `SkillAssigned`, `LayoutChanged`, `RemoteConnected`, `GitStateChanged`, `PtyOutput{session_id, bytes}`, `OrchestratorRequest{requester_id, action}` (ce dernier souvent court-circuité vers un Channel).
 ---
@ -451,11 +458,19 @@ drift(agent) =
 │   ├── agents.json          # AgentManifest (mapping md ↔ template ↔ sync ↔ version)
 │   ├── layout.json          # LayoutTree de l'onglet (sérialisé)
 │   ├── project.json         # méta projet local (nom, profil par défaut, remote ref)
-│   └── agents/
+│   ├── agents/
-│       ├── reviewer.md       # contexte d'un agent de projet
+│   │   ├── reviewer.md       # contexte d'un agent de projet
-│       ├── backend-dev.md
+│   │   ├── backend-dev.md
 │   │   └── ...
 │   ├── skills/
 │   │   ├── code-review.md    # contexte d'un skill (voir §14.2)
 │   │   ├── simplify.md
 │   │   └── ...
 │   └── run/
 │       ├── <agent-id>/       # cwd isolé par agent actif (créé à l'activation)
 │       │   └── CLAUDE.md     # fichier de convention généré par IdeA (profil-dépendant)
 │       └── ...
-└── (CLAUDE.md / AGENTS.md / GEMINI.md générés/symlinkés à l'activation si conventionFile)
+└── (aucun CLAUDE.md/AGENTS.md/GEMINI.md à la racine — jamais — voir §14.1)
 ```
 **Schéma `agents.json`** :
@ -504,6 +519,8 @@ Emplacement résolu via Tauri path API (`AppData`/`~/.local/share/IdeA`/`~/Libra
 **Formats** : contextes & templates en **Markdown** ; tout le reste en **JSON** (serde). Pas de base de données : fichiers plats, simples, diffables, portables (AppImage friendly).
 > **Note** : `.ideai/run/` contient des répertoires d'exécution éphémères (créés à l'activation, nettoyés à la fermeture). Leur contenu (convention files générés) ne doit **pas** être versionné dans git — ajouter `.ideai/run/` au `.gitignore` du projet.
 ---
 ## 10. Arborescence du repo
@ -581,6 +598,73 @@ IdeA/
 | L9 | **Remote (SSH + WSL)** | `RemoteHost` stratégie, `SshHost`/`WslHost`, adapters FS/PTY/Spawner distants, `RemoteGitRepository`. UI connexion. | `infrastructure/remote`, `application/remote`, `frontend/features/remote` |
 | L10 | **Fenêtres & multi-window** | `Workspace`/`Window`/`Tab`, `MoveTabToNewWindow`, drag d'onglet → nouvelle fenêtre OS Tauri. | `application`, `app-tauri`, `frontend/app` |
 | L11 | **Packaging & livraison** | Tauri bundle : NSIS `setup.exe`, **AppImage** multi-distro, CI Linux+Windows. | `app-tauri`, CI |
 | L12 | **Skills** | Entité `Skill`, `SkillStore`, CRUD skills global+projet, assignation agent↔skills, injection dans convention file à l'activation. UI onglet Skills. | `domain/skill`, `application/skill`, `infrastructure/store`, `frontend/features/skills` |
 | L13 | **OrchestratorApi** | File-watcher `.ideai/requests/`, port `OrchestratorApi`, adapter `FsOrchestratorAdapter`, actions `spawn_agent`/`stop_agent`/`update_agent_context`. | `infrastructure/orchestrator`, `application/agent`, `app-tauri` |
 ---
 ## 14. Décisions d'architecture figées (2026-06-06)
 ### 14.1 Isolation du cwd par agent — résolution de la collision de contexte
 **Problème** : plusieurs agents du même profil (ex. deux instances Claude Code) sur le même project root produisaient une collision — le fichier de convention (`CLAUDE.md`, `AGENTS.md`…) est un emplacement fixe unique à la racine.
 **Décision** : le cwd du PTY d'un agent n'est **jamais** le project root. C'est `.ideai/run/<agent-id>/`, un dossier créé par IdeA à l'activation et nettoyé à la fermeture.
 **Convention file généré par IdeA** : IdeA écrit dans ce dossier le fichier conventionnel attendu par le profil (`CLAUDE.md`, `AGENTS.md`, etc.). Ce fichier contient :
 1. La **persona/rôle** de l'agent (son `.md` dans `.ideai/agents/`).
 2. Le **chemin absolu du project root** (pour que l'agent sache où opérer).
 3. Les **skills actifs** assignés à cet agent (voir §14.2).
 4. Une référence au **contexte projet partagé** si présent.
 **Avantages** :
 - Zéro collision entre agents, même N instances du même profil.
 - Universel : fonctionne pour toute CLI qui lit un fichier de convention depuis son cwd — aucun flag ou commande propre à un modèle.
 - Zéro dépendance à git (git est optionnel — supprimer un repo ne casse rien).
 **Impact sur `AgentProfile.cwd_template`** : la valeur est toujours `"{agentRunDir}"`, jamais `"{projectRoot}"`. La connaissance du project root passe par le *contenu* du convention file, pas par le cwd.
 ---
 ### 14.2 Skills — abstraction universelle de workflows réutilisables
 **Définition** : un **Skill** est un workflow/comportement réutilisable qu'on peut assigner à un agent. Exemples : `code-review`, `simplify`, `run-tests`, `explain`. C'est l'équivalent universel des slash-commands de Claude Code — mais sans dépendance à la syntaxe `/command` d'un modèle particulier.
 **Stockage** :
 - Skills globaux (templates) : `<app_data>/IdeA/skills/` (store global IDE, réutilisables entre projets).
 - Skills de projet : `.ideai/skills/<skill-name>.md` (spécifiques au projet).
 **Injection** : les skills assignés à un agent sont **inclus dans son convention file** généré par IdeA au moment de l'activation. L'agent reçoit donc ses skills comme du contexte textuel — aucun mécanisme CLI propriétaire.
 **Entité `Skill`** (à ajouter au domaine) :
 - Champs : `id`, `name`, `content_md: MarkdownDoc`, `scope: SkillScope` (`Global` | `Project`).
 - Un agent peut avoir 0..N skills assignés (stocké dans l'`AgentManifest`).
 **Port `SkillStore`** : CRUD skills globaux + skills projet (compose `FileSystem`/store global selon le scope).
 ---
 ### 14.3 OrchestratorApi — spawn d'agents depuis un agent ou depuis l'UI
 **Objectif** : qu'un agent orchestrateur puisse demander à IdeA de créer un nouvel agent (visible dans la grille et dans l'onglet Agents), exactement comme le ferait l'utilisateur via l'UI.
 **Mécanisme** : file-watching sur `.ideai/requests/<requester-id>/`. L'orchestrateur écrit un fichier JSON de requête :
 ```json
 { "action": "spawn_agent", "name": "dev-backend", "profile": "claude-code", "context": "agents/dev-backend.md" }
 ```
 IdeA détecte le fichier, exécute la même logique que `LaunchAgent` déclenché depuis l'UI, crée la cellule terminal, inscrit l'agent dans l'onglet Agents, puis supprime le fichier et écrit une réponse.
 **Règle** : l'orchestrateur ne spawne jamais lui-même un process CLI — il **délègue à IdeA**. IdeA reste l'unique source de vérité du cycle de vie des agents.
 **Port `OrchestratorApi`** (adapter entrant, driven by file-watcher) : surveille `.ideai/requests/`, désérialise les commandes, les traduit en appels de use cases (`LaunchAgent`, `StopAgent`…). Implémenté dans `infrastructure/orchestrator`.
 **Actions supportées (v1)** : `spawn_agent`, `stop_agent`, `update_agent_context`.
 ---
 ### 14.4 Git = intégration optionnelle, zéro dépendance fonctionnelle
 Git est un **outil posé par-dessus l'IDE**, pas un socle. Supprimer le repo git d'un projet ne doit casser aucune feature d'IdeA (agents, terminaux, layout, skills, orchestration). Les use cases git (L8) sont un module indépendant ; rien d'autre n'en dépend. Cette contrainte s'applique à toute future décision de conception.
 ---
@ -591,7 +675,7 @@ IdeA/
 3. **Drag d'onglet entre fenêtres Tauri** : Tauri v2 multi-webview/multi-window + DnD natif inter-fenêtres est délicat (le DnD HTML ne traverse pas les fenêtres OS). **Spike** : protocole « detach » (créer une `WebviewWindow`, transférer l'état via store + event, fermer l'onglet source). L10.
 4. **Git sur FS distant** : libgit2 ne lit pas un FS SSH/WSL directement. Décision : **fallback git CLI** (`RemoteGitRepository`) côté distant via `ProcessSpawner`. À valider (perf, parsing). L9.
 5. **Synchro temps réel UI ↔ PTY** : volume d'octets élevé ; backpressure des Channels Tauri, throttling/coalescing côté front. **Spike** L3.
-6. **Injection `conventionFile`** : symlink vs copie du `.md` vers `CLAUDE.md`/`AGENTS.md` ; conflits si fichier existant, .gitignore, droits Windows (symlinks). À cadrer L6.
+6. ~~**Injection `conventionFile`** : symlink vs copie du `.md` vers `CLAUDE.md`/`AGENTS.md` ; conflits si fichier existant, .gitignore, droits Windows (symlinks).~~ **Résolu (§14.1)** : cwd isolé par agent dans `.ideai/run/<id>/` — plus de conflit à la racine, convention file généré par copie simple.
 7. **SSH auth** : agent/clé/mot de passe/known_hosts ; choix russh (rustls) vs ssh2 (libssh2/OpenSSL — impacte point 2). Décision à figer début L9.
 8. **WSL chemins** : conversion `/mnt/c/...` ↔ `\\wsl$\...`, distros multiples, perf I/O cross-boundary. Spike L9.
 9. **Détection d'édition hors-app** des `.md`/templates (content hash) et résolution de conflit lors du sync. L7.
--- a/CLAUDE.md
+++ b/CLAUDE.md
@ -0,0 +1,187 @@
 # IdeA — Contexte & Méthode de travail
 > Ce document définit **mon rôle**, **la méthode de développement** et **la vision produit** du projet IdeA.
 > Il fait autorité sur la façon dont le projet est piloté. Toute évolution de méthode doit être répercutée ici.
 ---
 ## 1. Mon rôle : chef d'orchestre, pas développeur
 Je **n'écris pas de code moi-même**. Mon rôle est de **piloter des agents** qui réalisent le travail.
 Je suis responsable de :
 - Découper le travail en tâches claires et autonomes.
 - Attribuer chaque tâche aux bons agents.
 - Garantir que le cycle de développement/test est respecté.
 - Faire respecter les principes d'architecture (SOLID, Hexagonal).
 - Maintenir la cohérence globale du projet et de ce document.
 - Arbitrer et valider avant toute action irréversible ou sortante.
 ---
 ## 2. Les agents
 ### 2.1 Agent Architecture (1 pour tout le projet)
 - Garant de l'architecture globale : **Hexagonale (Ports & Adapters)** et principes **SOLID**.
 - Définit les frontières (domaine / application / infrastructure), les ports, les contrats.
 - Valide que chaque nouvelle feature respecte la structure avant son développement.
 - Tient à jour la cartographie d'architecture et les conventions.
 ### 2.2 Agents de Développement
 - Écrivent le code des features.
 - Respectent strictement l'architecture définie par l'agent Architecture.
 - Code **propre, structuré, stable**.
 - Reçoivent les rapports d'erreurs des agents de test et corrigent.
 ### 2.3 Agents de Test
 - **Chaque agent de développement est appairé avec un agent de test dédié.**
 - Écrivent et exécutent les **tests unitaires** des features implémentées ou modifiées.
 - Produisent un **rapport d'erreurs** clair quand un test échoue.
 - Re-testent après chaque correction.
 ---
 ## 3. Le cycle de développement (boucle obligatoire)
 Pour **chaque** feature implémentée ou modifiée :
 ```
 1. Agent Architecture  → valide le découpage et les contrats (ports/interfaces)
 2. Agent Développement → écrit le code
 3. Agent Test          → écrit les tests unitaires + les exécute
 4a. Tests OK           → feature validée, on passe à la suite
 4b. Tests KO           → rapport d'erreurs → retour à l'agent Développement
                         → correction → retour à l'étape 3 (boucle jusqu'au vert)
 ```
 **Règle d'or :** aucune feature n'est considérée terminée tant que ses tests ne passent pas.
 Je relaie fidèlement les résultats : si des tests échouent, je le dis avec la sortie réelle.
 ---
 ## 4. Principes de code
 - **SOLID** appliqué au maximum.
 - **Architecture Hexagonale** (Ports & Adapters) : le domaine métier est isolé des détails techniques (UI, terminal, git, SSH, système de fichiers...).
 - Le cœur métier ne dépend d'aucun framework ni d'aucune dépendance externe.
 - Tests unitaires systématiques ; couverture des features critiques.
 - Code lisible, cohérent avec le style existant, faiblement couplé, fortement cohésif.
 ---
 ## 5. Vision produit : IdeA
 **IdeA est un IDE next-gen 100 % IA.** On n'y code pas : **on gère des IA.**
 ### Fonctionnalités clés
 - **Multi-projets en parallèle** : un **onglet par projet**.
 - **Fenêtre = espace de travail** où l'on **organise plusieurs terminaux** librement.
 - **Agents par projet** : chaque projet a ses propres agents.
 - **Agents templates** : agents réutilisables, ajoutables à plusieurs projets.
 - **Création d'agents** : depuis zéro ou à partir d'un template.
 - **Synchronisation template → agents** : option « garder l'agent à jour ».
  Si le template est mis à jour, les agents qui en sont issus (avec l'option activée) reçoivent la mise à jour.
 - **Contextes d'agents stockés en `.md`** (toujours).
 - **Création de projet** = définition de son **project root**.
 ### Intégrations
 - **Git** intégré.
 - **Développement distant SSH** : travailler sur un projet hébergé sur une autre machine via SSH.
 - **Développement WSL** : travailler sur une WSL depuis Windows.
 ### Plateformes & livraison
 - Cible : **macOS, Linux, Windows**.
 - Première phase de compilation : **Linux et Windows**.
 - Livraison :
  - **Windows** : `setup.exe`.
  - **Linux** : **AppImage** (doit fonctionner sur les différentes distributions).
 ---
 ## 6. Stack technique (validée)
 - **Shell applicatif** : **Tauri v2** (binaires légers, performants, multi-OS, AppImage + installeur `setup.exe`/NSIS Windows natifs).
 - **Cœur / backend** : **Rust** — stabilité, performance, et expression idiomatique du domaine hexagonal (ports = traits, adapters = implémentations).
 - **Frontend / UI** : **TypeScript + React**.
 - **Terminaux** : **xterm.js** (rendu) + **portable-pty** (PTY côté Rust).
 - **Git** : **libgit2** via `git2` (Rust).
 - **SSH** : `russh` / `ssh2` (Rust).
 - **WSL** : invocation de `wsl.exe` depuis le backend.
 ## 7. Layout des terminaux (exigence produit)
 Disposition en **grille redimensionnable de type tableur (Excel)** :
 - Splits redimensionnables horizontaux **et** verticaux.
 - L'utilisateur peut **définir le nombre de colonnes dans une ligne** et **le nombre de lignes dans une colonne**, indépendamment par zone.
 - Possibilité de **fusionner des cellules** (ex. fusionner deux colonnes sur une ligne), à la manière des cellules fusionnées d'un tableur.
 - Chaque cellule de la grille héberge un terminal.
 - → Modèle de layout récursif/imbriqué (pas une grille rigide uniforme) à concevoir par l'agent Architecture.
 ## 8. Stockage des contextes & liaison aux templates
 - **Templates d'agents** : stockés dans l'**IDE** (dossier de données utilisateur global de l'app, hors projet).
 - **Agents de projet** : leurs `.md` sont stockés dans un dossier **`.ideai/`** à la racine du project root.
  *(Nom choisi pour éviter toute collision avec le `.idea` de JetBrains.)*
 - **Manifeste de liaison** dans `.ideai/` (ex. `.ideai/agents.json`) qui mappe pour chaque agent de projet :
  - le `.md` de l'agent,
  - le template d'origine (le cas échéant),
  - `synchronized: true/false`,
  - la **version du template** au dernier sync (pour détecter qu'une mise à jour est disponible).
 - **Synchro template → agents** : quand un template est mis à jour, les agents liés avec `synchronized: true` reçoivent la MAJ.
 ## 9. Moteur IA : adaptateur de CLI flexible (Port `AgentRuntime`)
 Chaque IA est décrite par un **profil déclaratif** (config éditable, pas du code), implémentation d'un **Port** `AgentRuntime` côté domaine. Deux variables clés par IA :
 1. **Commande de lancement** + arguments (ex. `claude`, `codex`, `gemini`, `aider`).
 2. **Stratégie d'injection du contexte `.md`** :
   - `conventionFile` : écrire/symlink le `.md` vers le fichier attendu par la CLI (`CLAUDE.md`, `AGENTS.md`, `GEMINI.md`…).
   - `flag` : passer le chemin via un argument.
   - `stdin` : piper le contenu.
   - `env` : passer via variable d'environnement.
 Exemple de profil :
 ```json
 {
  "id": "claude-code",
  "name": "Claude Code",
  "command": "claude",
  "args": [],
  "contextInjection": { "strategy": "conventionFile", "target": "CLAUDE.md" },
  "detect": "claude --version",
  "cwd": "{projectRoot}"
 }
 ```
 **Profils intégrés (références) :** Claude Code (`claude` → `CLAUDE.md`), OpenAI Codex CLI (`codex` → `AGENTS.md`), Gemini CLI (`gemini` → `GEMINI.md`), Aider (`aider` → args/message).
 **Règles produit :**
 - **Premier lancement de l'IDE** : un assistant (first-run) **demande à l'utilisateur** quels profils d'IA configurer. On ne présume rien par défaut.
 - Les commandes des profils sont **pré-remplies mais éditables**.
 - L'utilisateur peut **ajouter sa propre commande CLI** (profil custom) pour n'importe quelle IA.
 **Lancement d'un agent :** à l'**activation de l'agent**, on ouvre une cellule terminal (PTY) avec le bon `cwd`, on injecte le contexte `.md`, et on **auto-lance** la CLI du profil.
 ## 10. Fenêtres & onglets
 - **Par défaut : un onglet par projet** (comme les IDE classiques).
 - **Drag & drop d'un onglet** hors de la fenêtre → **crée une nouvelle fenêtre OS** portant ce projet.
 - **Multi-fenêtres OS supporté** ; chaque fenêtre possède un ou plusieurs onglets/projets.
 ## 11. Feuille de route
 1. **Cadrage architecture complet d'abord** (jalon en cours) : l'agent Architecture produit la cartographie complète — domaine, ports, adapters, modules, arborescence — **avant tout code**.
 2. Puis MVP incrémental selon le cycle dev/test de la section 3.
 ## 12. Autonomie d'exécution dans le projet
 L'utilisateur m'accorde un **accès large et autonome** sur le dossier du projet : je peux lire, créer, modifier des fichiers et exécuter les commandes de développement (cargo, npm, npx, git, etc.) **sans demander confirmation à chaque fois**.
 - Concrètement, ces autorisations sont matérialisées dans `.claude/settings.local.json` (mode `acceptEdits` + `Bash`/`Read`/`Edit`/`Write` autorisés), pas dans ce document — CONTEXT.md ne fait que **documenter l'intention**.
 - **Garde-fous conservés** : les actions destructrices ou hors-projet restent bloquées (`sudo`, `rm -rf` sur `/`/`~`/`$HOME`, `mkfs`, `dd`, `shutdown`/`reboot`…).
 - L'esprit du rôle (§1) ne change pas : je reste **chef d'orchestre**. L'autonomie porte sur l'exécution mécanique, pas sur l'arbitrage des décisions produit/archi, ni sur les **actions sortantes** (push, publication) qui restent soumises à validation explicite.
 ---
 *Dernière mise à jour : 2026-06-05*
--- a/agents-dev/L12-skills.md
+++ b/agents-dev/L12-skills.md
@ -0,0 +1,65 @@
 # L12 — Skills
 **Binôme :** `dev-skills` / `test-skills`
 **Zones :** `domain/skill`, `application/skill`, `infrastructure/store`, `frontend/features/skills`
 **Dépendances amont :** L0, L1, L5, L6 (convention file généré à l'activation), L7 (store global réutilisé).
 ## Objectif
 Modéliser les **Skills** : workflows réutilisables (équivalent universel des slash-commands, sans dépendance à la syntaxe `/command` d'un modèle). Stockage global IDE + projet, assignation agent↔skills, **injection des skills assignés dans le convention file** généré à l'activation de l'agent. Cf. ARCHITECTURE §14.2.
 ## Périmètre (DEV)
 - **Domaine** : entité `Skill { id, name, content_md: MarkdownDoc, scope: SkillScope(Global|Project) }`. Invariants : `name` non vide, `content_md` non vide. Event `SkillAssigned`.
 - **Port `SkillStore`** : CRUD skills globaux (`<app_data>/IdeA/skills/`) + skills projet (`.ideai/skills/<name>.md`), résolution selon `scope` (compose `FileSystem`/store global comme L7).
 - **AgentManifest** : étendre pour porter la liste `skills: Vec<SkillRef>` assignés à chaque agent (0..N).
 - **Use cases** (`application/skill`) : `CreateSkill`, `UpdateSkill`, `DeleteSkill`, `ListSkills(scope)`, `AssignSkillToAgent`, `UnassignSkillFromAgent`.
 - **Injection** : à l'activation (fil L6), composer le convention file en concaténant persona agent + chemin project root + **skills assignés** (lus via `SkillStore`). Pas de mécanisme CLI propriétaire.
 - **Front** : onglet/section Skills (liste globale + projet, CRUD, éditeur md), assignation skills↔agent dans `AgentsPanel`.
 ## Périmètre (TEST)
 - `Skill` rejette `name`/`content_md` vides.
 - `SkillStore` : CRUD round-trip en tmpdir pour les deux scopes ; un skill `Project` n'apparaît pas dans le scope `Global` et inversement.
 - `AssignSkillToAgent` / `UnassignSkillFromAgent` : mutent l'`AgentManifest`, émettent `SkillAssigned`, idempotents (pas de doublon).
 - **Injection** : le convention file généré contient bien le `content_md` des skills assignés et **rien** des skills non assignés ; ordre déterministe.
 - Front : CRUD skills + assignation via gateway mock (RTL) ; garde-fou « no direct invoke ».
 ## Definition of Done
 - `cargo test` (skill/store/app) + `vitest` verts ; cycle manuel : créer un skill, l'assigner à un agent, l'activer → le skill apparaît dans le convention file de `.ideai/run/<agent-id>/`.
 - DoD commune (cf. README) respectée ; zéro régression.
 ## Avancement
 ### ✅ Domaine (vert)
 - **Entité `Skill`** (`domain/skill.rs`) : `id: SkillId`, `name`, `content_md: MarkdownDoc`, `scope: SkillScope(Global|Project)`. Constructeur validant (`name` + `content_md` non vides), `with_content` re-valide l'invariant.
 - **`SkillRef { skill_id, scope }`** : référence d'assignation portée par l'agent ; `From<&Skill>`.
 - **`SkillId`** ajouté (`ids.rs`), event **`SkillAssigned { agent_id, skill_id, assigned }`** (`events.rs`), DTO + arm de mapping côté `app-tauri` (`events.rs`).
 - **`Agent`** étendu : champ `skills: Vec<SkillRef>` (serde `default`), méthodes `assign_skill` (idempotent), `unassign_skill`, `with_skills` (dédup). **`ManifestEntry`** : champ `skills` (serde `default` + `skip_serializing_if` → rétrocompat des manifests pré-L12) ; `from_agent`/`to_agent` préservent les skills.
 - **Tests** : 8 invariants (`entities.rs`) + 3 serde dont rétrocompat d'un manifest legacy sans clé `skills` (`serde_roundtrip.rs`). `cargo test -p domain` vert ; `cargo test --workspace` vert (0 régression) ; clippy clean.
 ### ✅ Port + adapter store (vert)
 - **Port `SkillStore`** (`domain/ports.rs`) : `list/get/save/delete` portant `scope` + `root: &ProjectPath` **par appel** (root ignoré pour `Global`, résolu pour `Project`) — un seul store sert tous les projets ouverts, comme `AgentContextStore`.
 - **Adapter `FsSkillStore`** (`infrastructure/store/skill.rs`) : même forme on-disk que `FsTemplateStore` (`index.json` + `md/<id>.md`), deux racines disjointes : `<app_data>/skills/` (Global) et `<root>/.ideai/skills/` (Project). Delete laisse l'orphelin md (pas de remove dans le port FS), index = source de vérité. **7 tests** d'intégration tmpdir (`skill_store.rs`) : round-trip 2 scopes, **isolation de scope**, upsert, delete idempotent, camelCase.
 ### ✅ Use cases application (vert)
 - `application/skill` : `CreateSkill`, `UpdateSkill`, `DeleteSkill`, `ListSkills(scope)` (inputs portant `project_root`), `AssignSkillToAgent` / `UnassignSkillFromAgent` (mutent l'`AgentManifest` via `to_agent`/`from_agent`, dédup, émettent `SkillAssigned`, **idempotents**). **9 tests** (`skill_usecases.rs`).
 ### ✅ Injection dans le convention file (vert, fil L6)
 - `LaunchAgent` reçoit le port `SkillStore` ; `resolve_skills` lit les `.md` des skills assignés (ordre manifest, déterministe ; skill supprimé = `SkillRef` pendant → ignoré sans bloquer le lancement).
 - `compose_convention_file` étendu : section `# Skills` (sous-titres `## <name>`) après le persona ; omise si aucun skill. **3 tests** unitaires + e2e (`agent_lifecycle.rs` : injection ordonnée, ref pendant tolérée).
 - **Composition root** (`app-tauri/state.rs`) : `FsSkillStore` construit (app-data global), injecté dans `LaunchAgent`.
 ### ✅ IPC `app-tauri` (vert)
 - **DTOs** (`dto.rs`) : `SkillDto` (transparent sur `Skill`, camelCase), `SkillListDto`, request DTOs (`Create/Update/Assign/UnassignSkillRequestDto`), `parse_skill_id`. `scope` désérialise directement vers `SkillScope` (`"global"`/`"project"`).
 - **Commandes** (`commands.rs`) : `create_skill`, `update_skill`, `list_skills`, `delete_skill`, `assign_skill_to_agent`, `unassign_skill_from_agent` — shells fins qui résolvent le `Project` (→ `project.root`) puis appellent le use case. Enregistrées dans `lib.rs`.
 - **Composition root** (`state.rs`) : 6 use cases skill câblés sur le `skill_store_port` (déjà construit pour le launcher) et le `contexts_port` partagé.
 - `cargo build -p app-tauri` + `cargo test --workspace` (304) verts ; clippy clean.
 ### ✅ Front `features/skills` (vert)
 - **Domaine** (`domain/index.ts`) : `SkillScope`, `Skill`, `SkillRef` ; `Agent` étendu avec `skills: SkillRef[]`.
 - **Port** (`ports/index.ts`) : `SkillGateway` (list/create/update/delete + assign/unassign) + `CreateSkillInput` ; ajouté à `Gateways`.
 - **Adapters** : `TauriSkillGateway` (`adapters/skill.ts`, invoke camelCase) ; `MockSkillGateway` (`adapters/mock`, scopes disjoints + mutation partagée du `MockAgentGateway` via `_setSkills`, assign idempotent).
 - **Feature** : `useSkills` (VM 2 scopes), `SkillEditor` (overlay md edit/preview + sélecteur de scope), `SkillsPanel` (listes Project/Global, CRUD). Onglet **Skills** ajouté dans `ProjectsView`.
 - **Assignation** dans `AgentsPanel` : chips des skills assignés + sélecteur d'assignation + unassign, sur l'agent sélectionné ; refresh après mutation.
 - **Tests** (`skills.test.tsx`, RTL via `DIProvider` + mocks) : CRUD project/global, isolation de scope, édition, suppression, assign/unassign reflétés sur l'agent, idempotence, **garde-fou « no direct invoke »** (aucune action run/launch). `vitest` : **229** verts (0 régression ; test « ten gateways » mis à jour).
 ### ⏳ Reste à faire
 - Cycle manuel : créer un skill, l'assigner à un agent, l'activer → vérifier qu'il apparaît dans le convention file de `.ideai/run/<agent-id>/` (à faire sur l'AppImage).
--- a/agents-dev/L13-orchestrator.md
+++ b/agents-dev/L13-orchestrator.md
@ -0,0 +1,35 @@
 # L13 — OrchestratorApi
 **Binôme :** `dev-orchestrator` / `test-orchestrator`
 **Zones :** `infrastructure/orchestrator`, `application/agent`, `app-tauri`
 **Dépendances amont :** L0, L1, L6 (`LaunchAgent`/`StopAgent`), L12 (`update_agent_context` peut toucher les skills).
 ## Objectif
 Permettre à un **agent orchestrateur** de demander à IdeA de créer/arrêter/mettre à jour un agent — **exactement** comme l'utilisateur via l'UI. L'orchestrateur ne spawne jamais lui-même un process CLI : il **délègue à IdeA**, unique source de vérité du cycle de vie des agents. Cf. ARCHITECTURE §14.3.
 ## Périmètre (DEV)
 - **Port `OrchestratorApi`** (adapter *entrant*, driven by file-watcher) : surveille `.ideai/requests/<requester-id>/`, désérialise les requêtes JSON, les traduit en appels de use cases.
 - **Adapter `FsOrchestratorAdapter`** (`infrastructure/orchestrator`) : file-watching (`notify`), parse, dispatch, **supprime le fichier de requête** et **écrit une réponse** (succès/erreur) à côté.
 - **Actions v1** : `spawn_agent` (→ `LaunchAgent`), `stop_agent` (→ `StopAgent`), `update_agent_context` (réécrit le `.md` de l'agent ± skills).
 - **Event** : `OrchestratorRequest { requester_id, action }`.
 - **Schéma requête** :
  ```json
  { "action": "spawn_agent", "name": "dev-backend", "profile": "claude-code", "context": "agents/dev-backend.md" }
  ```
 - Le résultat d'un `spawn_agent` est **identique** à un lancement UI : cellule terminal créée, agent inscrit dans l'onglet Agents.
 - **Composition root** (`app-tauri`) : démarrer le watcher, brancher sur les use cases existants ; arrêt propre à la fermeture.
 ## Périmètre (TEST)
 - Désérialisation : requêtes valides → action typée ; requête malformée → réponse d'erreur, **pas de crash**, pas de spawn.
 - `spawn_agent` invoque `LaunchAgent` avec les bons args (use case mocké) ; idempotence sur double-dépôt du même fichier (traité une fois).
 - Après traitement : fichier de requête **supprimé**, fichier de réponse écrit avec le bon statut.
 - `stop_agent` / `update_agent_context` : mappent vers les bons use cases ; cible inexistante → erreur propre.
 - Watcher : un fichier déposé dans `.ideai/requests/<id>/` est détecté (test d'intégration tmpdir + `notify`).
 ## Definition of Done
 - `cargo test` (orchestrator/app) verts ; cycle manuel : un agent écrit un fichier de requête `spawn_agent` → un nouvel agent apparaît dans la grille et l'onglet Agents, fichier consommé + réponse écrite.
 - Garde-fou : l'orchestrateur ne lance **aucun** process directement (vérifié par revue + absence de `ProcessSpawner` dans le chemin orchestrateur).
 - DoD commune respectée ; zéro régression ; git reste optionnel (rien dans ce lot n'en dépend, §14.4).
 ## Avancement
 ⬜ À démarrer. Cadrage figé dans ARCHITECTURE §14.3.
--- a/agents-dev/README.md
+++ b/agents-dev/README.md
@ -26,6 +26,8 @@ Un fichier `Lx-*.md` par binôme décrit son périmètre, ses ports/adapters, se
 | L9 | [L9-remote.md](L9-remote.md) | 🟡 **socle vert** (LocalHost + ConnectRemote, Liskov) · SSH/WSL gated à venir |
 | L10 | [L10-windows.md](L10-windows.md) | 🟡 **backend + IPC verts** (move-tab + `WebviewWindow`) · détach UI fait avec la refonte L11 |
 | L11 | [L11-packaging.md](L11-packaging.md) | 🟡 AppImage Linux **OK** · refonte disposition IDE en cours · Windows/CI à venir |
 | L12 | [L12-skills.md](L12-skills.md) | ⬜ **Skills** — entité + store + assignation + injection convention file + UI |
 | L13 | [L13-orchestrator.md](L13-orchestrator.md) | ⬜ **OrchestratorApi** — file-watcher, spawn depuis agent ou UI, même résultat |
 ## Cycle dev ↔ test (obligatoire, cf. CONTEXT §3)
--- a/crates/app-tauri/src/commands.rs
+++ b/crates/app-tauri/src/commands.rs
@ -14,6 +14,8 @@ use application::{
    GitStatusInput, LaunchAgentInput, ListAgentsInput, ListLayoutsInput, LoadLayoutInput,
    MutateLayoutInput, OpenProjectInput, ReadAgentContextInput, RenameLayoutInput,
    SetActiveLayoutInput, SyncAgentWithTemplateInput, UpdateAgentContextInput,
    AssignSkillToAgentInput, CreateSkillInput, DeleteSkillInput, ListSkillsInput,
    UnassignSkillFromAgentInput, UpdateSkillInput,
 };
 use domain::ports::PtyHandle;
@ -28,11 +30,13 @@ use crate::dto::{
    GitStageRequestDto, GitStatusListDto, GraphCommitListDto, HealthRequestDto, HealthResponseDto,
    LaunchAgentRequestDto, LayoutDto, LayoutOperationDto, ListLayoutsDto, OpenTerminalRequestDto,
    ProfileDto, ProfileListDto, ProjectDto, ProjectListDto, ReadAgentContextResponseDto,
-    RenameLayoutRequestDto, ResizeTerminalRequestDto, SaveProfileRequestDto,
+    ReattachResultDto, RenameLayoutRequestDto, ResizeTerminalRequestDto, SaveProfileRequestDto,
    SetActiveLayoutRequestDto, SyncAgentWithTemplateRequestDto, SyncResultDto, TemplateDto,
    TemplateListDto, TerminalClosedDto, TerminalSessionDto, UpdateAgentContextRequestDto,
-    UpdateTemplateRequestDto, WriteTerminalRequestDto,
+    UpdateTemplateRequestDto, WriteTerminalRequestDto, parse_skill_id, AssignSkillRequestDto,
    CreateSkillRequestDto, SkillDto, SkillListDto, UnassignSkillRequestDto, UpdateSkillRequestDto,
 };
 use domain::{SkillRef, SkillScope};
 use crate::pty::{PtyBridge, PtyChunk};
 use crate::state::AppState;
@ -244,6 +248,69 @@ pub async fn close_terminal(
    result
 }
 /// `reattach_terminal` — re-bind a view to a **still-living** PTY without
 /// re-spawning it.
 ///
 /// Navigation (switching layout/tab) tears the xterm view down but must NOT kill
 /// the backend PTY (the AI keeps running). When the view comes back it calls this
 /// command, which:
 /// 1. reads the session's retained **scrollback** so the terminal can repaint,
 /// 2. registers the new per-session [`Channel`] in the [`PtyBridge`],
 /// 3. starts a fresh output pump subscribed to the live PTY (re-subscribable
 ///    broadcast), so new bytes flow to the new channel.
 ///
 /// Returns the scrollback bytes; the frontend writes them into xterm first, then
 /// receives subsequent output over `on_output`.
 ///
 /// # Errors
 /// Returns an [`ErrorDto`] (`INVALID` for a malformed id, `NOT_FOUND`/`PROCESS`
 /// if the session is no longer alive).
 #[tauri::command]
 pub fn reattach_terminal(
    session_id: String,
    on_output: Channel<PtyChunk>,
    state: State<'_, AppState>,
 ) -> Result<ReattachResultDto, ErrorDto> {
    let sid = parse_session_id(&session_id)?;
    let handle = PtyHandle { session_id: sid };
    // (1) Snapshot the scrollback. A NotFound here means the PTY is gone (was
    // explicitly closed or exited) — surfaced as an error so the caller falls
    // back to opening a fresh terminal.
    let scrollback = state
        .pty_port
        .scrollback(&handle)
        .map_err(|e| ErrorDto::from(AppError::from(e)))?;
    // (2) Register the new output channel for this session, replacing any stale
    // one from a previous attach.
    state.pty_bridge.register(sid, on_output);
    // (3) Subscribe afresh to the live byte stream and pump it to the channel.
    match state.pty_port.subscribe_output(&handle) {
        Ok(stream) => {
            let bridge: std::sync::Arc<PtyBridge> = std::sync::Arc::clone(&state.pty_bridge);
            std::thread::spawn(move || {
                for chunk in stream {
                    if !bridge.send_output(&sid, chunk) {
                        break;
                    }
                }
                bridge.unregister(&sid);
            });
        }
        Err(e) => {
            state.pty_bridge.unregister(&sid);
            return Err(ErrorDto::from(AppError::from(e)));
        }
    }
    Ok(ReattachResultDto {
        session_id,
        scrollback,
    })
 }
 // ---------------------------------------------------------------------------
 // Layout (L4)
 // ---------------------------------------------------------------------------
@ -1119,3 +1186,157 @@ pub async fn move_tab_to_new_window(
    Ok(MoveTabResultDto::from(out))
 }
 // ---------------------------------------------------------------------------
 // Skills (L12)
 // ---------------------------------------------------------------------------
 /// `create_skill` — create a skill in its scope's store.
 ///
 /// # Errors
 /// Returns an [`ErrorDto`] (`INVALID` for an empty name/content or malformed
 /// project id, `NOT_FOUND` if the project is unknown, `STORE` on failure).
 #[tauri::command]
 pub async fn create_skill(
    request: CreateSkillRequestDto,
    state: State<'_, AppState>,
 ) -> Result<SkillDto, ErrorDto> {
    let project = resolve_project(&request.project_id, &state).await?;
    state
        .create_skill
        .execute(CreateSkillInput {
            name: request.name,
            content: request.content,
            scope: request.scope,
            project_root: project.root,
        })
        .await
        .map(SkillDto::from)
        .map_err(ErrorDto::from)
 }
 /// `update_skill` — replace a skill's Markdown content.
 ///
 /// # Errors
 /// Returns an [`ErrorDto`] (`INVALID` for malformed ids or empty content,
 /// `NOT_FOUND` if the project or skill is unknown, `STORE` on failure).
 #[tauri::command]
 pub async fn update_skill(
    request: UpdateSkillRequestDto,
    state: State<'_, AppState>,
 ) -> Result<SkillDto, ErrorDto> {
    let project = resolve_project(&request.project_id, &state).await?;
    let skill_id = parse_skill_id(&request.skill_id)?;
    state
        .update_skill
        .execute(UpdateSkillInput {
            scope: request.scope,
            skill_id,
            content: request.content,
            project_root: project.root,
        })
        .await
        .map(SkillDto::from)
        .map_err(ErrorDto::from)
 }
 /// `list_skills` — list the skills in one scope.
 ///
 /// # Errors
 /// Returns an [`ErrorDto`] (`INVALID` for a malformed id, `NOT_FOUND` if the
 /// project is unknown, `STORE` on failure).
 #[tauri::command]
 pub async fn list_skills(
    project_id: String,
    scope: SkillScope,
    state: State<'_, AppState>,
 ) -> Result<SkillListDto, ErrorDto> {
    let project = resolve_project(&project_id, &state).await?;
    state
        .list_skills
        .execute(ListSkillsInput {
            scope,
            project_root: project.root,
        })
        .await
        .map(SkillListDto::from)
        .map_err(ErrorDto::from)
 }
 /// `delete_skill` — remove a skill from its scope's store.
 ///
 /// Agents that referenced it keep their `SkillRef`; injection simply skips the
 /// now-absent skill.
 ///
 /// # Errors
 /// Returns an [`ErrorDto`] (`INVALID` for malformed ids, `NOT_FOUND` if the
 /// project or skill is unknown, `STORE` on failure).
 #[tauri::command]
 pub async fn delete_skill(
    project_id: String,
    scope: SkillScope,
    skill_id: String,
    state: State<'_, AppState>,
 ) -> Result<(), ErrorDto> {
    let project = resolve_project(&project_id, &state).await?;
    let id = parse_skill_id(&skill_id)?;
    state
        .delete_skill
        .execute(DeleteSkillInput {
            scope,
            skill_id: id,
            project_root: project.root,
        })
        .await
        .map_err(ErrorDto::from)
 }
 /// `assign_skill_to_agent` — record a `SkillRef` in the agent's manifest entry.
 /// Idempotent.
 ///
 /// # Errors
 /// Returns an [`ErrorDto`] (`INVALID` for malformed ids, `NOT_FOUND` if the
 /// project or agent is unknown, `STORE` on failure).
 #[tauri::command]
 pub async fn assign_skill_to_agent(
    request: AssignSkillRequestDto,
    state: State<'_, AppState>,
 ) -> Result<(), ErrorDto> {
    let project = resolve_project(&request.project_id, &state).await?;
    let agent_id = parse_agent_id(&request.agent_id)?;
    let skill_id = parse_skill_id(&request.skill_id)?;
    state
        .assign_skill
        .execute(AssignSkillToAgentInput {
            project,
            agent_id,
            skill: SkillRef::new(skill_id, request.scope),
        })
        .await
        .map_err(ErrorDto::from)
 }
 /// `unassign_skill_from_agent` — drop a skill assignment from an agent.
 /// Idempotent.
 ///
 /// # Errors
 /// Returns an [`ErrorDto`] (`INVALID` for malformed ids, `NOT_FOUND` if the
 /// project or agent is unknown, `STORE` on failure).
 #[tauri::command]
 pub async fn unassign_skill_from_agent(
    request: UnassignSkillRequestDto,
    state: State<'_, AppState>,
 ) -> Result<(), ErrorDto> {
    let project = resolve_project(&request.project_id, &state).await?;
    let agent_id = parse_agent_id(&request.agent_id)?;
    let skill_id = parse_skill_id(&request.skill_id)?;
    state
        .unassign_skill
        .execute(UnassignSkillFromAgentInput {
            project,
            agent_id,
            skill_id,
        })
        .await
        .map_err(ErrorDto::from)
 }
--- a/crates/app-tauri/src/dto.rs
+++ b/crates/app-tauri/src/dto.rs
@ -245,6 +245,18 @@ impl From<OpenTerminalOutput> for TerminalSessionDto {
    }
 }
 /// Response DTO for `reattach_terminal`: the retained scrollback of a still-live
 /// session, repainted into the re-mounting xterm before the new output stream is
 /// wired. Bytes are serialised as a number array, matching the PTY output channel.
 #[derive(Debug, Clone, Serialize)]
 #[serde(rename_all = "camelCase")]
 pub struct ReattachResultDto {
    /// The session that was re-attached (echoed back for the frontend).
    pub session_id: String,
    /// The most-recent retained output bytes (scrollback ring buffer).
    pub scrollback: Vec<u8>,
 }
 /// Request DTO for `write_terminal`.
 #[derive(Debug, Clone, Deserialize)]
 #[serde(rename_all = "camelCase")]
@ -1330,3 +1342,107 @@ pub fn parse_tab_id(raw: &str) -> Result<TabId, ErrorDto> {
            message: format!("invalid tab id: {raw}"),
        })
 }
 // ---------------------------------------------------------------------------
 // Skills (L12)
 // ---------------------------------------------------------------------------
 use application::{CreateSkillOutput, ListSkillsOutput, UpdateSkillOutput};
 use domain::{Skill, SkillId, SkillScope};
 /// A skill crossing the wire. [`Skill`] already serialises camelCase
 /// (`id`, `name`, `contentMd`, `scope` as `"global"`/`"project"`), so we embed
 /// it directly — the TS mirror matches this shape.
 #[derive(Debug, Clone, Serialize)]
 #[serde(transparent)]
 pub struct SkillDto(pub Skill);
 /// A list of skills (transparent array on the wire).
 #[derive(Debug, Clone, Serialize)]
 #[serde(transparent)]
 pub struct SkillListDto(pub Vec<SkillDto>);
 impl From<ListSkillsOutput> for SkillListDto {
    fn from(out: ListSkillsOutput) -> Self {
        Self(out.skills.into_iter().map(SkillDto).collect())
    }
 }
 impl From<CreateSkillOutput> for SkillDto {
    fn from(out: CreateSkillOutput) -> Self {
        Self(out.skill)
    }
 }
 impl From<UpdateSkillOutput> for SkillDto {
    fn from(out: UpdateSkillOutput) -> Self {
        Self(out.skill)
    }
 }
 /// Request DTO for `create_skill`.
 #[derive(Debug, Clone, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct CreateSkillRequestDto {
    /// Owning project (resolved to a root; ignored on disk for `Global`).
    pub project_id: String,
    /// Display name.
    pub name: String,
    /// Initial Markdown content.
    pub content: String,
    /// Scope the skill is created in.
    pub scope: SkillScope,
 }
 /// Request DTO for `update_skill`.
 #[derive(Debug, Clone, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct UpdateSkillRequestDto {
    /// Owning project (resolved to a root; ignored on disk for `Global`).
    pub project_id: String,
    /// Id of the skill to update.
    pub skill_id: String,
    /// Scope the skill lives in.
    pub scope: SkillScope,
    /// New Markdown content.
    pub content: String,
 }
 /// Request DTO for `assign_skill_to_agent`.
 #[derive(Debug, Clone, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct AssignSkillRequestDto {
    /// Owning project.
    pub project_id: String,
    /// Agent receiving the skill.
    pub agent_id: String,
    /// Skill to assign.
    pub skill_id: String,
    /// Scope of the skill (recorded alongside the ref).
    pub scope: SkillScope,
 }
 /// Request DTO for `unassign_skill_from_agent`.
 #[derive(Debug, Clone, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct UnassignSkillRequestDto {
    /// Owning project.
    pub project_id: String,
    /// Agent losing the skill.
    pub agent_id: String,
    /// Skill to unassign.
    pub skill_id: String,
 }
 /// Parses a skill-id string (UUID) coming from the frontend.
 ///
 /// # Errors
 /// Returns an [`ErrorDto`] with code `INVALID` if the string is not a UUID.
 pub fn parse_skill_id(raw: &str) -> Result<SkillId, ErrorDto> {
    uuid::Uuid::parse_str(raw)
        .map(SkillId::from_uuid)
        .map_err(|_| ErrorDto {
            code: "INVALID".to_owned(),
            message: format!("invalid skill id: {raw}"),
        })
 }
--- a/crates/app-tauri/src/events.rs
+++ b/crates/app-tauri/src/events.rs
@ -73,6 +73,16 @@ pub enum DomainEventDto {
        /// Version synced to.
        to: u64,
    },
    /// A skill was assigned to (or unassigned from) an agent.
    #[serde(rename_all = "camelCase")]
    SkillAssigned {
        /// Agent id.
        agent_id: String,
        /// Skill id.
        skill_id: String,
        /// `true` if assigned, `false` if unassigned.
        assigned: bool,
    },
    /// A tab's layout changed.
    #[serde(rename_all = "camelCase")]
    LayoutChanged {
@ -138,6 +148,15 @@ impl From<&DomainEvent> for DomainEventDto {
                agent_id: agent_id.to_string(),
                to: to.get(),
            },
            DomainEvent::SkillAssigned {
                agent_id,
                skill_id,
                assigned,
            } => Self::SkillAssigned {
                agent_id: agent_id.to_string(),
                skill_id: skill_id.to_string(),
                assigned: *assigned,
            },
            DomainEvent::LayoutChanged { project_id } => Self::LayoutChanged {
                project_id: project_id.to_string(),
            },
--- a/crates/app-tauri/src/lib.rs
+++ b/crates/app-tauri/src/lib.rs
@ -48,6 +48,28 @@ pub fn run() {
            events::spawn_relay(app.handle().clone(), &app_state.event_bus);
            app.manage(app_state);
            // Kill all live PTYs cleanly when the main window is closing. This is
            // independent of the per-view (navigation/layout) lifecycle — those
            // must NEVER kill a PTY — and only fires on a genuine app shutdown.
            // A brutal crash is best-effort and out of scope.
            if let Some(window) = app.get_webview_window("main") {
                let handle = app.handle().clone();
                window.on_window_event(move |event| {
                    if let tauri::WindowEvent::CloseRequested { .. } = event {
                        if let Some(state) = handle.try_state::<AppState>() {
                            let pty = std::sync::Arc::clone(&state.pty_port);
                            let handles = state.terminal_sessions.handles();
                            tauri::async_runtime::block_on(async move {
                                for h in handles {
                                    let _ = pty.kill(&h).await;
                                }
                            });
                        }
                    }
                });
            }
            Ok(())
        })
        .invoke_handler(tauri::generate_handler![
@ -60,6 +82,7 @@ pub fn run() {
            commands::write_terminal,
            commands::resize_terminal,
            commands::close_terminal,
            commands::reattach_terminal,
            commands::load_layout,
            commands::mutate_layout,
            commands::list_layouts,
@ -96,6 +119,12 @@ pub fn run() {
            commands::git_log,
            commands::git_init,
            commands::git_graph,
            commands::create_skill,
            commands::update_skill,
            commands::list_skills,
            commands::delete_skill,
            commands::assign_skill_to_agent,
            commands::unassign_skill_from_agent,
            commands::move_tab_to_new_window,
        ])
        .run(tauri::generate_context!())
--- a/crates/app-tauri/src/state.rs
+++ b/crates/app-tauri/src/state.rs
@ -13,18 +13,19 @@ use application::{
    CreateAgentFromTemplate, CreateLayout, CreateProject, CreateTemplate, DeleteAgent,
    DeleteLayout, DeleteProfile, DeleteTemplate, DetectAgentDrift, DetectProfiles, FirstRunState,
    GitBranches, GitCheckout, GitCommit, GitGraph, GitInit, GitLog, GitStage, GitStatus, GitUnstage,
-    HealthUseCase, LaunchAgent, ListAgents, ListLayouts, ListProfiles, ListProjects, ListTemplates,
+    HealthUseCase, LaunchAgent, ListAgents, ListLayouts, ListProfiles, ListProjects, ListSkills,
-    LoadLayout, MoveTabToNewWindow, MutateLayout, OpenProject, OpenTerminal, ReadAgentContext,
+    ListTemplates, LoadLayout, MoveTabToNewWindow, MutateLayout, OpenProject, OpenTerminal,
-    ReferenceProfiles, RenameLayout, ResizeTerminal, SaveProfile, SetActiveLayout,
+    ReadAgentContext, ReferenceProfiles, RenameLayout, ResizeTerminal, SaveProfile, SetActiveLayout,
    AssignSkillToAgent, CreateSkill, DeleteSkill, UnassignSkillFromAgent, UpdateSkill,
    SyncAgentWithTemplate, TerminalSessions, UpdateAgentContext, UpdateTemplate, WriteToTerminal,
 };
 use domain::ports::{
    AgentContextStore, AgentRuntime, Clock, EventBus, FileSystem, GitPort, IdGenerator,
-    ProcessSpawner, ProfileStore, ProjectStore, PtyPort, TemplateStore,
+    ProcessSpawner, ProfileStore, ProjectStore, PtyPort, SkillStore, TemplateStore,
 };
 use infrastructure::{
-    CliAgentRuntime, FsProfileStore, FsProjectStore, FsTemplateStore, Git2Repository,
+    CliAgentRuntime, FsProfileStore, FsProjectStore, FsSkillStore, FsTemplateStore, Git2Repository,
    IdeaiContextStore, LocalFileSystem, LocalProcessSpawner, PortablePtyAdapter, SystemClock,
    TokioBroadcastEventBus, UuidGenerator,
 };
@ -147,6 +148,19 @@ pub struct AppState {
    pub git_init: Arc<GitInit>,
    /// Return the commit graph for all local branches.
    pub git_graph: Arc<GitGraph>,
    // --- Skills (L12) ---
    /// Create a skill in a scope's store.
    pub create_skill: Arc<CreateSkill>,
    /// Update a skill's content.
    pub update_skill: Arc<UpdateSkill>,
    /// List skills in a scope.
    pub list_skills: Arc<ListSkills>,
    /// Delete a skill from its scope's store.
    pub delete_skill: Arc<DeleteSkill>,
    /// Assign a skill to an agent (records a `SkillRef`).
    pub assign_skill: Arc<AssignSkillToAgent>,
    /// Unassign a skill from an agent.
    pub unassign_skill: Arc<UnassignSkillFromAgent>,
 }
 impl AppState {
@ -287,6 +301,17 @@ impl AppState {
        let contexts = Arc::new(IdeaiContextStore::new(Arc::clone(&fs_port)));
        let contexts_port = Arc::clone(&contexts) as Arc<dyn AgentContextStore>;
        // --- Skill store (L12) ---
        // Global skills live in the machine-local app-data dir; project skills are
        // resolved per call from each project's `.ideai/` (so one store serves all
        // open projects). Shared by the skill use cases and the agent launcher
        // (assigned-skill injection into the convention file, §14.2).
        let skill_store = Arc::new(FsSkillStore::new(
            Arc::clone(&fs_port),
            app_data_dir.to_string_lossy().into_owned(),
        ));
        let skill_store_port = Arc::clone(&skill_store) as Arc<dyn SkillStore>;
        let create_agent = Arc::new(CreateAgentFromScratch::new(
            Arc::clone(&contexts_port),
            Arc::clone(&ids) as Arc<dyn IdGenerator>,
@ -307,6 +332,7 @@ impl AppState {
            Arc::clone(&runtime_port),
            Arc::clone(&fs_port),
            Arc::clone(&pty_port),
            Arc::clone(&skill_store_port),
            Arc::clone(&terminal_sessions),
            Arc::clone(&events_port),
        ));
@ -364,6 +390,25 @@ impl AppState {
        let git_init = Arc::new(GitInit::new(Arc::clone(&git_port), Arc::clone(&events_port)));
        let git_graph = Arc::new(GitGraph::new(Arc::clone(&git_port)));
        // --- Skill use cases (L12) ---
        // Reuse the skill store (built above for the launcher) and the shared
        // agent context store for the agent↔skill assignment.
        let create_skill = Arc::new(CreateSkill::new(
            Arc::clone(&skill_store_port),
            Arc::clone(&ids) as Arc<dyn IdGenerator>,
        ));
        let update_skill = Arc::new(UpdateSkill::new(Arc::clone(&skill_store_port)));
        let list_skills = Arc::new(ListSkills::new(Arc::clone(&skill_store_port)));
        let delete_skill = Arc::new(DeleteSkill::new(Arc::clone(&skill_store_port)));
        let assign_skill = Arc::new(AssignSkillToAgent::new(
            Arc::clone(&contexts_port),
            Arc::clone(&events_port),
        ));
        let unassign_skill = Arc::new(UnassignSkillFromAgent::new(
            Arc::clone(&contexts_port),
            Arc::clone(&events_port),
        ));
        // --- Windows (L10) ---
        let move_tab = Arc::new(MoveTabToNewWindow::new(
            Arc::clone(&store_port),
@ -422,6 +467,12 @@ impl AppState {
            git_log,
            git_init,
            git_graph,
            create_skill,
            update_skill,
            list_skills,
            delete_skill,
            assign_skill,
            unassign_skill,
            move_tab,
        }
    }
--- a/crates/app-tauri/tests/dto.rs
+++ b/crates/app-tauri/tests/dto.rs
@ -4,8 +4,8 @@
 use app_tauri_lib::dto::{
    parse_node_id, parse_session_id, ErrorDto, HealthRequestDto, HealthResponseDto, LayoutDto,
-    LayoutOperationDto, OpenTerminalRequestDto, ResizeTerminalRequestDto, TerminalClosedDto,
+    LayoutOperationDto, OpenTerminalRequestDto, ReattachResultDto, ResizeTerminalRequestDto,
-    WriteTerminalRequestDto,
+    TerminalClosedDto, WriteTerminalRequestDto,
 };
 use application::{CloseTerminalOutput, LayoutOperation, LoadLayoutOutput, OpenTerminalInput};
 use domain::{Direction, LayoutNode, LayoutTree, LeafCell, NodeId};
@ -191,6 +191,16 @@ fn terminal_closed_dto_serialises_code_camel_case() {
    assert_eq!(serde_json::to_value(&none).unwrap(), json!({ "code": null }));
 }
 #[test]
 fn reattach_result_dto_serialises_camel_case() {
    let dto = ReattachResultDto {
        session_id: "sess-1".into(),
        scrollback: vec![104, 105],
    };
    let v = serde_json::to_value(&dto).unwrap();
    assert_eq!(v, json!({ "sessionId": "sess-1", "scrollback": [104, 105] }));
 }
 #[test]
 fn parse_session_id_accepts_uuid_and_rejects_garbage() {
    let sid = SessionId::from_uuid(Uuid::nil());
--- a/crates/application/src/agent/catalogue.rs
+++ b/crates/application/src/agent/catalogue.rs
@ -53,7 +53,7 @@ pub fn reference_profiles() -> Vec<AgentProfile> {
            ContextInjection::convention_file("CLAUDE.md")
                .expect("CLAUDE.md is a valid convention target"),
            Some("claude --version".to_owned()),
-            "{projectRoot}",
+            "{agentRunDir}",
        )
        .expect("claude reference profile is valid"),
        AgentProfile::new(
@ -64,7 +64,7 @@ pub fn reference_profiles() -> Vec<AgentProfile> {
            ContextInjection::convention_file("AGENTS.md")
                .expect("AGENTS.md is a valid convention target"),
            Some("codex --version".to_owned()),
-            "{projectRoot}",
+            "{agentRunDir}",
        )
        .expect("codex reference profile is valid"),
        AgentProfile::new(
@ -75,7 +75,7 @@ pub fn reference_profiles() -> Vec<AgentProfile> {
            ContextInjection::convention_file("GEMINI.md")
                .expect("GEMINI.md is a valid convention target"),
            Some("gemini --version".to_owned()),
-            "{projectRoot}",
+            "{agentRunDir}",
        )
        .expect("gemini reference profile is valid"),
        AgentProfile::new(
@ -86,7 +86,7 @@ pub fn reference_profiles() -> Vec<AgentProfile> {
            ContextInjection::flag("--message-file {path}")
                .expect("aider flag template is non-empty"),
            Some("aider --version".to_owned()),
-            "{projectRoot}",
+            "{agentRunDir}",
        )
        .expect("aider reference profile is valid"),
    ]
--- a/crates/application/src/agent/lifecycle.rs
+++ b/crates/application/src/agent/lifecycle.rs
@ -15,11 +15,11 @@ use std::sync::Arc;
 use domain::ports::{
    AgentContextStore, AgentRuntime, ContextInjectionPlan, EventBus, FileSystem, PreparedContext,
-    ProfileStore, PtyPort, RemotePath, SpawnSpec,
+    ProfileStore, PtyPort, RemotePath, SkillStore, SpawnSpec, StoreError,
 };
 use domain::{
    Agent, AgentId, AgentManifest, AgentOrigin, DomainEvent, ManifestEntry, MarkdownDoc, NodeId,
-    Project, ProfileId, ProjectPath, PtySize, SessionKind, SessionStatus, TerminalSession,
+    Project, ProfileId, ProjectPath, PtySize, SessionKind, SessionStatus, Skill, TerminalSession,
 };
 use crate::error::AppError;
@ -350,6 +350,7 @@ pub struct LaunchAgent {
    runtime: Arc<dyn AgentRuntime>,
    fs: Arc<dyn FileSystem>,
    pty: Arc<dyn PtyPort>,
    skills: Arc<dyn SkillStore>,
    sessions: Arc<TerminalSessions>,
    events: Arc<dyn EventBus>,
 }
@ -364,6 +365,7 @@ impl LaunchAgent {
        runtime: Arc<dyn AgentRuntime>,
        fs: Arc<dyn FileSystem>,
        pty: Arc<dyn PtyPort>,
        skills: Arc<dyn SkillStore>,
        sessions: Arc<TerminalSessions>,
        events: Arc<dyn EventBus>,
    ) -> Self {
@ -373,11 +375,35 @@ impl LaunchAgent {
            runtime,
            fs,
            pty,
            skills,
            sessions,
            events,
        }
    }
    /// Resolves the Markdown bodies of an agent's assigned skills, in the
    /// **manifest order** (deterministic). A skill that no longer exists in its
    /// store (deleted out from under the assignment) is silently skipped — a
    /// dangling [`domain::SkillRef`] must not block a launch.
    ///
    /// # Errors
    /// [`AppError::Store`] on any store failure other than a missing skill.
    async fn resolve_skills(
        &self,
        agent: &Agent,
        root: &ProjectPath,
    ) -> Result<Vec<Skill>, AppError> {
        let mut out = Vec::with_capacity(agent.skills.len());
        for skill_ref in &agent.skills {
            match self.skills.get(skill_ref.scope, root, skill_ref.skill_id).await {
                Ok(skill) => out.push(skill),
                Err(StoreError::NotFound) => {}
                Err(e) => return Err(e.into()),
            }
        }
        Ok(out)
    }
    /// Executes the launch.
    ///
    /// Step order is contractually significant (and unit-tested): resolve the
@ -421,24 +447,40 @@ impl LaunchAgent {
                AppError::NotFound(format!("profile {} for agent", agent.profile_id))
            })?;
-        // 3. Prepare the invocation (pure): command + args + injection plan + cwd.
+        // 3. Compute and create the agent's isolated run directory
        //    `<root>/.ideai/run/<agent-id>/` (ARCHITECTURE §14.1). The PTY cwd is
        //    *never* the project root: each agent gets its own directory so that N
        //    instances of the same profile never collide on a single conventional
        //    file (CLAUDE.md, …). This is the only I/O in the cwd resolution; the
        //    runtime's `prepare_invocation` stays pure.
        let run_dir = agent_run_dir(&input.project.root, &agent.id)
            .map_err(|e| AppError::Invalid(e.to_string()))?;
        self.fs
            .create_dir_all(&RemotePath::new(run_dir.as_str().to_owned()))
            .await?;
        // 4. Prepare the invocation (pure): command + args + injection plan + cwd.
        //    The run dir is passed as the cwd base; the profile's `{agentRunDir}`
        //    placeholder resolves against it.
        let prepared = PreparedContext {
            content: content.clone(),
            relative_path: agent.context_path.clone(),
        };
        let mut spec = self
            .runtime
-            .prepare_invocation(&profile, &prepared, &input.project.root)?;
+            .prepare_invocation(&profile, &prepared, &run_dir)?;
-        // 4. Apply the injection plan side effects *before* spawning.
+        // 5. Resolve the agent's assigned skills (their `.md` bodies), then apply
-        self.apply_injection(&input.project, &agent.context_path, &content, &mut spec)
+        //    the injection plan side effects *before* spawning.
        let skills = self.resolve_skills(&agent, &input.project.root).await?;
        self.apply_injection(&input.project, &agent.context_path, &content, &skills, &mut spec)
            .await?;
-        // 5. Spawn the PTY at the resolved cwd; adopt its session id everywhere.
+        // 6. Spawn the PTY at the resolved cwd; adopt its session id everywhere.
        let handle = self.pty.spawn(spec.clone(), size).await?;
        let session_id = handle.session_id;
-        // 6. For the Stdin strategy, pipe the context once the PTY is live.
+        // 7. For the Stdin strategy, pipe the context once the PTY is live.
        if matches!(spec.context_plan, Some(ContextInjectionPlan::Stdin)) {
            self.pty.write(&handle, content.as_str().as_bytes())?;
        }
@ -473,16 +515,22 @@ impl LaunchAgent {
        project: &Project,
        context_rel_path: &str,
        content: &MarkdownDoc,
        skills: &[Skill],
        spec: &mut SpawnSpec,
    ) -> Result<(), AppError> {
        match spec.context_plan.clone() {
            Some(ContextInjectionPlan::File { target }) => {
-                // conventionFile spike (ARCHITECTURE §13.6): copy the context to the
+                // conventionFile (ARCHITECTURE §14.1): IdeA *generates* the
-                // conventional file (e.g. CLAUDE.md), overwriting any existing one.
+                // conventional file (e.g. CLAUDE.md) inside the agent's isolated
-                // A copy (not a symlink) is the portable choice — Windows symlinks
+                // run directory — `spec.cwd` is that run dir, never the project
-                // need privileges and SFTP/WSL symlink semantics differ.
+                // root, so there is zero collision between agents. The document is
                // composed: an absolute project-root header (so the agent knows
                // where to operate, since its cwd is *not* the root), the agent's
                // persona `.md`, then the bodies of its assigned skills (§14.2).
                let document =
                    compose_convention_file(project.root.as_str(), content.as_str(), skills);
                let path = RemotePath::new(join(&spec.cwd, &target));
-                self.fs.write(&path, content.as_str().as_bytes()).await?;
+                self.fs.write(&path, document.as_bytes()).await?;
            }
            Some(ContextInjectionPlan::Env { var }) => {
                // Hand the CLI the absolute path of the agent's `.md` (which lives at
@ -505,6 +553,55 @@ fn join(base: &ProjectPath, rel: &str) -> String {
    format!("{b}/{rel}")
 }
 /// Computes an agent's isolated run directory `<root>/.ideai/run/<agent-id>/`
 /// (ARCHITECTURE §14.1). This is the PTY cwd for the agent — never the project
 /// root — guaranteeing that two distinct agents on the same project root get two
 /// distinct cwd (the anti-collision contract).
 ///
 /// # Errors
 /// Propagates [`DomainError`](domain::error::DomainError) if the joined path is
 /// not a valid [`ProjectPath`] (should not happen for an absolute project root).
 fn agent_run_dir(root: &ProjectPath, agent_id: &AgentId) -> Result<ProjectPath, domain::error::DomainError> {
    ProjectPath::new(join(root, &format!(".ideai/run/{agent_id}")))
 }
 /// Composes the convention file IdeA writes into an agent's run directory: an
 /// absolute project-root header (the agent's cwd is the run dir, *not* the root,
 /// so it must be told where to work), the agent's persona `.md`, then the bodies
 /// of its assigned `skills` under a `# Skills` section (ARCHITECTURE §14.2).
 ///
 /// Skills are emitted in the order given (the caller passes them in manifest
 /// order, making the output deterministic); each is introduced by a `##` header
 /// carrying its name. When `skills` is empty the section is omitted entirely, so
 /// an agent with no skills gets exactly the previous document.
 ///
 /// Kept as a **pure** function (no I/O) so it is unit-testable in isolation.
 #[must_use]
 pub(crate) fn compose_convention_file(project_root: &str, agent_md: &str, skills: &[Skill]) -> String {
    let mut out = String::new();
    out.push_str("# Project root\n\n");
    out.push_str(project_root);
    out.push_str("\n\nTous tes travaux portent sur ce project root (chemin absolu ci-dessus). ");
    out.push_str(
        "Ton répertoire courant est un dossier d'exécution isolé (`.ideai/run/<agent>/`) ; \
         opère sur le project root, pas sur ce dossier.\n\n",
    );
    out.push_str("---\n\n");
    out.push_str(agent_md);
    if !skills.is_empty() {
        out.push_str("\n\n---\n\n# Skills\n");
        for skill in skills {
            out.push_str("\n## ");
            out.push_str(&skill.name);
            out.push_str("\n\n");
            out.push_str(skill.content_md.as_str());
            out.push('\n');
        }
    }
    out
 }
 /// Derives a unique, filesystem-safe `md_path` (`agents/<slug>.md`) for a new
 /// agent, disambiguating against the manifest's existing paths with a numeric
 /// suffix when needed. Shared with the template-driven agent creation (L7).
@ -536,3 +633,71 @@ fn slugify(name: &str) -> String {
    }
    out.trim_matches('-').to_owned()
 }
 #[cfg(test)]
 mod tests {
    use super::*;
    #[test]
    fn agent_run_dir_is_under_ideai_run_and_unique_per_agent() {
        let root = ProjectPath::new("/home/me/proj").unwrap();
        let a = AgentId::from_uuid(uuid::Uuid::from_u128(1));
        let b = AgentId::from_uuid(uuid::Uuid::from_u128(2));
        let dir_a = agent_run_dir(&root, &a).unwrap();
        let dir_b = agent_run_dir(&root, &b).unwrap();
        assert_eq!(dir_a.as_str(), format!("/home/me/proj/.ideai/run/{a}"));
        assert_ne!(dir_a, dir_b, "distinct agents → distinct run dirs");
        // Never the project root.
        assert_ne!(dir_a.as_str(), "/home/me/proj");
    }
    #[test]
    fn compose_convention_file_carries_root_then_persona() {
        let doc = compose_convention_file("/abs/project/root", "# Persona\n\nDo things.", &[]);
        // Absolute project root present.
        assert!(doc.contains("/abs/project/root"));
        // Persona present.
        assert!(doc.contains("# Persona"));
        assert!(doc.contains("Do things."));
        // Root header precedes the persona body (ordering of the composition).
        let root_at = doc.find("/abs/project/root").unwrap();
        let persona_at = doc.find("# Persona").unwrap();
        assert!(root_at < persona_at, "root header must precede the persona");
        // No skills ⇒ no Skills section.
        assert!(!doc.contains("# Skills"));
    }
    #[test]
    fn compose_convention_file_appends_assigned_skills_in_order() {
        let s = |n: u128, name: &str, body: &str| {
            Skill::new(
                domain::SkillId::from_uuid(uuid::Uuid::from_u128(n)),
                name,
                MarkdownDoc::new(body),
                domain::SkillScope::Global,
            )
            .unwrap()
        };
        let doc = compose_convention_file(
            "/root",
            "# Persona",
            &[s(1, "refactor", "REFAC_BODY"), s(2, "review", "REVIEW_BODY")],
        );
        // Both skill bodies present, after the persona.
        assert!(doc.contains("REFAC_BODY"));
        assert!(doc.contains("REVIEW_BODY"));
        let persona_at = doc.find("# Persona").unwrap();
        let refac_at = doc.find("REFAC_BODY").unwrap();
        let review_at = doc.find("REVIEW_BODY").unwrap();
        assert!(persona_at < refac_at, "skills come after the persona");
        // Deterministic order: first assigned skill precedes the second.
        assert!(refac_at < review_at, "skills emitted in the given order");
        // Skill names surface as sub-headers.
        assert!(doc.contains("## refactor"));
        assert!(doc.contains("## review"));
    }
 }
--- a/crates/application/src/lib.rs
+++ b/crates/application/src/lib.rs
@ -18,6 +18,7 @@ pub mod health;
 pub mod layout;
 pub mod project;
 pub mod remote;
 pub mod skill;
 pub mod template;
 pub mod terminal;
 pub mod window;
@ -61,6 +62,12 @@ pub use template::{
    SyncAgentWithTemplateInput, SyncAgentWithTemplateOutput, UpdateTemplate, UpdateTemplateInput,
    UpdateTemplateOutput,
 };
 pub use skill::{
    AssignSkillToAgent, AssignSkillToAgentInput, CreateSkill, CreateSkillInput, CreateSkillOutput,
    DeleteSkill, DeleteSkillInput, ListSkills, ListSkillsInput, ListSkillsOutput,
    UnassignSkillFromAgent, UnassignSkillFromAgentInput, UpdateSkill, UpdateSkillInput,
    UpdateSkillOutput,
 };
 pub use terminal::{
    CloseTerminal, CloseTerminalInput, CloseTerminalOutput, OpenTerminal, OpenTerminalInput,
    OpenTerminalOutput, ResizeTerminal, ResizeTerminalInput, TerminalSessions, WriteToTerminal,
--- a/crates/application/src/skill/mod.rs
+++ b/crates/application/src/skill/mod.rs
@ -0,0 +1,21 @@
 //! Skill use cases (ARCHITECTURE §14.2; L12).
 //!
 //! Skills are reusable, model-agnostic workflows (IdeA's universal equivalent of
 //! a CLI slash-command). This module owns their CRUD across both scopes
 //! ([`domain::skill::SkillScope`]) and the agent↔skill assignment that records a
 //! [`domain::skill::SkillRef`] in the project manifest. The actual injection of
 //! an assigned skill's body into the generated convention file happens at agent
 //! activation (L6).
 //!
 //! Every use case talks only to ports ([`domain::ports::SkillStore`],
 //! [`domain::ports::AgentContextStore`], [`domain::ports::IdGenerator`],
 //! [`domain::ports::EventBus`]).
 mod usecases;
 pub use usecases::{
    AssignSkillToAgent, AssignSkillToAgentInput, CreateSkill, CreateSkillInput, CreateSkillOutput,
    DeleteSkill, DeleteSkillInput, ListSkills, ListSkillsInput, ListSkillsOutput,
    UnassignSkillFromAgent, UnassignSkillFromAgentInput, UpdateSkill, UpdateSkillInput,
    UpdateSkillOutput,
 };
--- a/crates/application/src/skill/usecases.rs
+++ b/crates/application/src/skill/usecases.rs
@ -0,0 +1,339 @@
 //! Skill use cases (ARCHITECTURE §14.2; L12).
 //!
 //! - **CRUD** in either scope: [`CreateSkill`], [`UpdateSkill`], [`DeleteSkill`],
 //!   [`ListSkills`].
 //! - **Assignment**: [`AssignSkillToAgent`] / [`UnassignSkillFromAgent`] mutate
 //!   the project manifest entry's `skills` and announce
 //!   [`DomainEvent::SkillAssigned`]. Both are idempotent.
 use std::sync::Arc;
 use domain::ports::{AgentContextStore, EventBus, IdGenerator, SkillStore};
 use domain::{
    AgentId, AgentManifest, DomainEvent, MarkdownDoc, Project, ProjectPath, Skill, SkillId,
    SkillRef, SkillScope,
 };
 use crate::error::AppError;
 // ---------------------------------------------------------------------------
 // CreateSkill
 // ---------------------------------------------------------------------------
 /// Input for [`CreateSkill::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct CreateSkillInput {
    /// Display name (also the `.md` stem on disk).
    pub name: String,
    /// Initial Markdown body.
    pub content: String,
    /// Scope the skill is created in (selects its backing store).
    pub scope: SkillScope,
    /// Active project root (used only for [`SkillScope::Project`]).
    pub project_root: ProjectPath,
 }
 /// Output of [`CreateSkill::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct CreateSkillOutput {
    /// The created skill.
    pub skill: Skill,
 }
 /// Creates a skill in the store of its [`SkillScope`].
 pub struct CreateSkill {
    skills: Arc<dyn SkillStore>,
    ids: Arc<dyn IdGenerator>,
 }
 impl CreateSkill {
    /// Builds the use case from its ports.
    #[must_use]
    pub fn new(skills: Arc<dyn SkillStore>, ids: Arc<dyn IdGenerator>) -> Self {
        Self { skills, ids }
    }
    /// Executes creation.
    ///
    /// # Errors
    /// - [`AppError::Invalid`] if `name`/`content` is empty,
    /// - [`AppError::Store`] on persistence failure.
    pub async fn execute(&self, input: CreateSkillInput) -> Result<CreateSkillOutput, AppError> {
        let id = SkillId::from_uuid(self.ids.new_uuid());
        let skill = Skill::new(id, input.name, MarkdownDoc::new(input.content), input.scope)
            .map_err(|e| AppError::Invalid(e.to_string()))?;
        self.skills.save(&skill, &input.project_root).await?;
        Ok(CreateSkillOutput { skill })
    }
 }
 // ---------------------------------------------------------------------------
 // UpdateSkill
 // ---------------------------------------------------------------------------
 /// Input for [`UpdateSkill::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct UpdateSkillInput {
    /// Scope the skill lives in.
    pub scope: SkillScope,
    /// Skill to update.
    pub skill_id: SkillId,
    /// New Markdown body.
    pub content: String,
    /// Active project root (used only for [`SkillScope::Project`]).
    pub project_root: ProjectPath,
 }
 /// Output of [`UpdateSkill::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct UpdateSkillOutput {
    /// The updated skill.
    pub skill: Skill,
 }
 /// Replaces a skill's content (re-validating the non-empty invariant).
 pub struct UpdateSkill {
    skills: Arc<dyn SkillStore>,
 }
 impl UpdateSkill {
    /// Builds the use case.
    #[must_use]
    pub fn new(skills: Arc<dyn SkillStore>) -> Self {
        Self { skills }
    }
    /// Executes the update.
    ///
    /// # Errors
    /// - [`AppError::NotFound`] if the skill is unknown in that scope,
    /// - [`AppError::Invalid`] if the new content is empty,
    /// - [`AppError::Store`] on persistence failure.
    pub async fn execute(&self, input: UpdateSkillInput) -> Result<UpdateSkillOutput, AppError> {
        let current = self
            .skills
            .get(input.scope, &input.project_root, input.skill_id)
            .await?;
        let updated = current
            .with_content(MarkdownDoc::new(input.content))
            .map_err(|e| AppError::Invalid(e.to_string()))?;
        self.skills.save(&updated, &input.project_root).await?;
        Ok(UpdateSkillOutput { skill: updated })
    }
 }
 // ---------------------------------------------------------------------------
 // ListSkills / DeleteSkill
 // ---------------------------------------------------------------------------
 /// Input for [`ListSkills::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ListSkillsInput {
    /// Scope to list.
    pub scope: SkillScope,
    /// Active project root (used only for [`SkillScope::Project`]).
    pub project_root: ProjectPath,
 }
 /// Output of [`ListSkills::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ListSkillsOutput {
    /// All skills in the requested scope.
    pub skills: Vec<Skill>,
 }
 /// Lists the skills in one scope.
 pub struct ListSkills {
    skills: Arc<dyn SkillStore>,
 }
 impl ListSkills {
    /// Builds the use case.
    #[must_use]
    pub fn new(skills: Arc<dyn SkillStore>) -> Self {
        Self { skills }
    }
    /// Lists skills in `input.scope`.
    ///
    /// # Errors
    /// [`AppError::Store`] on persistence failure.
    pub async fn execute(&self, input: ListSkillsInput) -> Result<ListSkillsOutput, AppError> {
        Ok(ListSkillsOutput {
            skills: self.skills.list(input.scope, &input.project_root).await?,
        })
    }
 }
 /// Input for [`DeleteSkill::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct DeleteSkillInput {
    /// Scope the skill lives in.
    pub scope: SkillScope,
    /// Skill to delete.
    pub skill_id: SkillId,
    /// Active project root (used only for [`SkillScope::Project`]).
    pub project_root: ProjectPath,
 }
 /// Deletes a skill from its scope's store.
 ///
 /// Agents that referenced it keep their [`SkillRef`]; the injection step simply
 /// finds nothing to resolve for the now-absent skill and skips it.
 pub struct DeleteSkill {
    skills: Arc<dyn SkillStore>,
 }
 impl DeleteSkill {
    /// Builds the use case.
    #[must_use]
    pub fn new(skills: Arc<dyn SkillStore>) -> Self {
        Self { skills }
    }
    /// Deletes the skill.
    ///
    /// # Errors
    /// - [`AppError::NotFound`] if the skill is unknown in that scope,
    /// - [`AppError::Store`] on persistence failure.
    pub async fn execute(&self, input: DeleteSkillInput) -> Result<(), AppError> {
        self.skills
            .delete(input.scope, &input.project_root, input.skill_id)
            .await?;
        Ok(())
    }
 }
 // ---------------------------------------------------------------------------
 // AssignSkillToAgent / UnassignSkillFromAgent
 // ---------------------------------------------------------------------------
 /// Input for [`AssignSkillToAgent::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct AssignSkillToAgentInput {
    /// The owning project.
    pub project: Project,
    /// The agent receiving the skill.
    pub agent_id: AgentId,
    /// The skill to assign.
    pub skill: SkillRef,
 }
 /// Assigns a skill to an agent by recording a [`SkillRef`] in its manifest entry.
 /// Idempotent: re-assigning the same skill is a no-op (no duplicate).
 pub struct AssignSkillToAgent {
    contexts: Arc<dyn AgentContextStore>,
    events: Arc<dyn EventBus>,
 }
 impl AssignSkillToAgent {
    /// Builds the use case from its ports.
    #[must_use]
    pub fn new(contexts: Arc<dyn AgentContextStore>, events: Arc<dyn EventBus>) -> Self {
        Self { contexts, events }
    }
    /// Executes the assignment.
    ///
    /// # Errors
    /// - [`AppError::NotFound`] if the agent is unknown to the project,
    /// - [`AppError::Invalid`] if the resulting manifest is invalid,
    /// - [`AppError::Store`] on persistence failure.
    pub async fn execute(&self, input: AssignSkillToAgentInput) -> Result<(), AppError> {
        let mut manifest = self.contexts.load_manifest(&input.project).await?;
        let entry = manifest
            .entries
            .iter_mut()
            .find(|e| e.agent_id == input.agent_id)
            .ok_or_else(|| AppError::NotFound(format!("agent {}", input.agent_id)))?;
        // Mutate through the domain entity so the dedup invariant is enforced in
        // one place, then fold the result back into the manifest entry.
        let mut agent = entry.to_agent().map_err(|e| AppError::Invalid(e.to_string()))?;
        let changed = agent.assign_skill(input.skill);
        *entry = domain::ManifestEntry::from_agent(&agent);
        if changed {
            self.persist_and_announce(&input.project, manifest, input.agent_id, input.skill.skill_id, true)
                .await?;
        }
        Ok(())
    }
    /// Saves the manifest and announces the assignment change.
    async fn persist_and_announce(
        &self,
        project: &Project,
        manifest: AgentManifest,
        agent_id: AgentId,
        skill_id: SkillId,
        assigned: bool,
    ) -> Result<(), AppError> {
        let manifest = AgentManifest::new(manifest.version, manifest.entries)
            .map_err(|e| AppError::Invalid(e.to_string()))?;
        self.contexts.save_manifest(project, &manifest).await?;
        self.events.publish(DomainEvent::SkillAssigned {
            agent_id,
            skill_id,
            assigned,
        });
        Ok(())
    }
 }
 /// Input for [`UnassignSkillFromAgent::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct UnassignSkillFromAgentInput {
    /// The owning project.
    pub project: Project,
    /// The agent losing the skill.
    pub agent_id: AgentId,
    /// The skill to unassign.
    pub skill_id: SkillId,
 }
 /// Removes a skill assignment from an agent. Idempotent: unassigning a skill the
 /// agent does not carry is a no-op.
 pub struct UnassignSkillFromAgent {
    contexts: Arc<dyn AgentContextStore>,
    events: Arc<dyn EventBus>,
 }
 impl UnassignSkillFromAgent {
    /// Builds the use case from its ports.
    #[must_use]
    pub fn new(contexts: Arc<dyn AgentContextStore>, events: Arc<dyn EventBus>) -> Self {
        Self { contexts, events }
    }
    /// Executes the un-assignment.
    ///
    /// # Errors
    /// - [`AppError::NotFound`] if the agent is unknown to the project,
    /// - [`AppError::Invalid`] if the resulting manifest is invalid,
    /// - [`AppError::Store`] on persistence failure.
    pub async fn execute(&self, input: UnassignSkillFromAgentInput) -> Result<(), AppError> {
        let mut manifest = self.contexts.load_manifest(&input.project).await?;
        let entry = manifest
            .entries
            .iter_mut()
            .find(|e| e.agent_id == input.agent_id)
            .ok_or_else(|| AppError::NotFound(format!("agent {}", input.agent_id)))?;
        let mut agent = entry.to_agent().map_err(|e| AppError::Invalid(e.to_string()))?;
        let changed = agent.unassign_skill(input.skill_id);
        *entry = domain::ManifestEntry::from_agent(&agent);
        if changed {
            let manifest = AgentManifest::new(manifest.version, manifest.entries)
                .map_err(|e| AppError::Invalid(e.to_string()))?;
            self.contexts.save_manifest(&input.project, &manifest).await?;
            self.events.publish(DomainEvent::SkillAssigned {
                agent_id: input.agent_id,
                skill_id: input.skill_id,
                assigned: false,
            });
        }
        Ok(())
    }
 }
--- a/crates/application/src/terminal/registry.rs
+++ b/crates/application/src/terminal/registry.rs
@ -59,6 +59,18 @@ impl TerminalSessions {
            .and_then(|m| m.get(id).map(|e| e.session.clone()))
    }
    /// Returns the [`PtyHandle`]s of every currently-registered session.
    ///
    /// Used at application shutdown to kill all live PTYs cleanly (the
    /// `CloseRequested` hook), independently of the frontend's per-view lifecycle.
    #[must_use]
    pub fn handles(&self) -> Vec<PtyHandle> {
        self.entries
            .lock()
            .map(|m| m.values().map(|e| e.handle.clone()).collect())
            .unwrap_or_default()
    }
    /// Removes a session from the registry, returning its handle if present.
    pub fn remove(&self, id: &SessionId) -> Option<PtyHandle> {
        self.entries
--- a/crates/application/tests/agent_lifecycle.rs
+++ b/crates/application/tests/agent_lifecycle.rs
@ -26,12 +26,13 @@ use domain::markdown::MarkdownDoc;
 use domain::ports::{
    AgentContextStore, AgentRuntime, ContextInjectionPlan, DirEntry, EventBus, EventStream,
    ExitStatus, FileSystem, FsError, IdGenerator, OutputStream, PreparedContext, ProfileStore,
-    PtyError, PtyHandle, PtyPort, RemotePath, RuntimeError, SpawnSpec, StoreError,
+    PtyError, PtyHandle, PtyPort, RemotePath, RuntimeError, SkillStore, SpawnSpec, StoreError,
 };
 use domain::profile::{AgentProfile, ContextInjection};
 use domain::project::{Project, ProjectPath};
 use domain::remote::RemoteRef;
-use domain::{PtySize, SessionId};
+use domain::skill::{Skill, SkillScope};
 use domain::{PtySize, SessionId, SkillId, SkillRef};
 use uuid::Uuid;
 use application::{
@ -177,6 +178,49 @@ impl ProfileStore for FakeProfiles {
    }
 }
 // ---------------------------------------------------------------------------
 // FakeSkills (SkillStore) — an in-memory store seeded with a few skills
 // ---------------------------------------------------------------------------
 #[derive(Clone, Default)]
 struct FakeSkills(Arc<Vec<Skill>>);
 impl FakeSkills {
    fn with(skills: Vec<Skill>) -> Self {
        Self(Arc::new(skills))
    }
 }
 #[async_trait]
 impl SkillStore for FakeSkills {
    async fn list(&self, scope: SkillScope, _root: &ProjectPath) -> Result<Vec<Skill>, StoreError> {
        Ok(self.0.iter().filter(|s| s.scope == scope).cloned().collect())
    }
    async fn get(
        &self,
        scope: SkillScope,
        _root: &ProjectPath,
        id: SkillId,
    ) -> Result<Skill, StoreError> {
        self.0
            .iter()
            .find(|s| s.scope == scope && s.id == id)
            .cloned()
            .ok_or(StoreError::NotFound)
    }
    async fn save(&self, _skill: &Skill, _root: &ProjectPath) -> Result<(), StoreError> {
        Ok(())
    }
    async fn delete(
        &self,
        _scope: SkillScope,
        _root: &ProjectPath,
        _id: SkillId,
    ) -> Result<(), StoreError> {
        Ok(())
    }
 }
 // ---------------------------------------------------------------------------
 // FakeRuntime (AgentRuntime) — records prepare + returns a configured plan
 // ---------------------------------------------------------------------------
@ -221,6 +265,7 @@ impl AgentRuntime for FakeRuntime {
 struct FakeFs {
    trace: Trace,
    writes: WriteLog<String>,
    created_dirs: Arc<Mutex<Vec<String>>>,
 }
 impl FakeFs {
@ -228,11 +273,15 @@ impl FakeFs {
        Self {
            trace,
            writes: Arc::new(Mutex::new(Vec::new())),
            created_dirs: Arc::new(Mutex::new(Vec::new())),
        }
    }
    fn writes(&self) -> Vec<(String, Vec<u8>)> {
        self.writes.lock().unwrap().clone()
    }
    fn created_dirs(&self) -> Vec<String> {
        self.created_dirs.lock().unwrap().clone()
    }
 }
 #[async_trait]
@ -251,7 +300,11 @@ impl FileSystem for FakeFs {
    async fn exists(&self, _path: &RemotePath) -> Result<bool, FsError> {
        Ok(false)
    }
-    async fn create_dir_all(&self, _path: &RemotePath) -> Result<(), FsError> {
+    async fn create_dir_all(&self, path: &RemotePath) -> Result<(), FsError> {
        self.created_dirs
            .lock()
            .unwrap()
            .push(path.as_str().to_owned());
        Ok(())
    }
    async fn list(&self, _path: &RemotePath) -> Result<Vec<DirEntry>, FsError> {
@ -313,6 +366,9 @@ impl PtyPort for FakePty {
    fn subscribe_output(&self, _handle: &PtyHandle) -> Result<OutputStream, PtyError> {
        Ok(Box::new(std::iter::empty()))
    }
    fn scrollback(&self, _handle: &PtyHandle) -> Result<Vec<u8>, PtyError> {
        Ok(Vec::new())
    }
    async fn kill(&self, _handle: &PtyHandle) -> Result<ExitStatus, PtyError> {
        Ok(ExitStatus { code: Some(0) })
    }
@ -386,7 +442,7 @@ fn profile(id: ProfileId, injection: ContextInjection) -> AgentProfile {
        Vec::new(),
        injection,
        Some("claude --version".to_owned()),
-        "{projectRoot}",
+        "{agentRunDir}",
    )
    .unwrap()
 }
@ -565,6 +621,7 @@ fn launch_fixture(injection: ContextInjection, plan: Option<ContextInjectionPlan
        Arc::new(runtime),
        Arc::new(fs.clone()),
        Arc::new(pty.clone()),
        Arc::new(FakeSkills::default()),
        Arc::clone(&sessions),
        Arc::new(bus.clone()),
    );
@ -600,17 +657,32 @@ async fn launch_orders_prepare_then_injection_then_spawn() {
        "prepare → injection → spawn"
    );
-    // The conventionFile was written to <cwd>/CLAUDE.md with the context body.
+    // The conventionFile was written inside the agent's isolated run directory
    // (`.ideai/run/<agent-id>/CLAUDE.md`) — NOT at the project root. Its content
    // is the *composed* document: an absolute project-root header followed by the
    // agent persona `.md`.
    let run_dir = format!("/home/me/proj/.ideai/run/{}", agent.id);
    let writes = fs.writes();
    assert_eq!(writes.len(), 1);
-    assert_eq!(writes[0].0, "/home/me/proj/CLAUDE.md");
+    assert_eq!(writes[0].0, format!("{run_dir}/CLAUDE.md"));
-    assert_eq!(writes[0].1, b"# ctx body");
+    let written = String::from_utf8(writes[0].1.clone()).unwrap();
    assert!(
        written.contains("/home/me/proj"),
        "convention file must carry the absolute project root, got: {written}"
    );
    assert!(
        written.contains("# ctx body"),
        "convention file must carry the agent persona, got: {written}"
    );
-    // Spawn happened at the resolved cwd with the profile command.
+    // The run directory was created (via the FileSystem port) before spawn.
    assert_eq!(fs.created_dirs(), vec![run_dir.clone()]);
    // Spawn happened at the isolated run dir with the profile command.
    let spawns = pty.spawns();
    assert_eq!(spawns.len(), 1);
    assert_eq!(spawns[0].command, "claude");
-    assert_eq!(spawns[0].cwd.as_str(), "/home/me/proj");
+    assert_eq!(spawns[0].cwd.as_str(), run_dir);
    // The session adopts the PTY id, is Running, and is registered as an agent.
    assert_eq!(out.session.id, sid(777));
@ -630,6 +702,172 @@ async fn launch_orders_prepare_then_injection_then_spawn() {
    );
 }
 /// **Anti-collision (ARCHITECTURE §14.1)**: two distinct agents of the *same*
 /// profile on the *same* project root must launch into two **distinct** cwd —
 /// each its own `.ideai/run/<agent-id>/` — and each writes its convention file
 /// inside its own run dir, never colliding at the project root.
 #[tokio::test]
 async fn two_agents_same_root_get_distinct_run_dirs_no_collision() {
    let injection = ContextInjection::convention_file("CLAUDE.md").unwrap();
    let plan = Some(ContextInjectionPlan::File {
        target: "CLAUDE.md".to_owned(),
    });
    // Two agents, same profile (pid(9)), same project root.
    let agent_a = scratch_agent(aid(1), "Alpha", "agents/alpha.md", pid(9));
    let agent_b = scratch_agent(aid(2), "Bravo", "agents/bravo.md", pid(9));
    let contexts = FakeContexts::with_agent(&agent_a, "# alpha");
    {
        let mut inner = contexts.0.lock().unwrap();
        inner
            .manifest
            .entries
            .push(ManifestEntry::from_agent(&agent_b));
        inner
            .contents
            .insert(agent_b.context_path.clone(), "# bravo".to_owned());
    }
    let profiles = FakeProfiles::new(vec![profile(pid(9), injection)]);
    let tr = trace();
    let fs = FakeFs::new(Arc::clone(&tr));
    let pty = FakePty::new(Arc::clone(&tr), sid(777));
    let sessions = Arc::new(TerminalSessions::new());
    let launch = LaunchAgent::new(
        Arc::new(contexts),
        Arc::new(profiles),
        Arc::new(FakeRuntime::new(Arc::clone(&tr), plan)),
        Arc::new(fs.clone()),
        Arc::new(pty.clone()),
        Arc::new(FakeSkills::default()),
        Arc::clone(&sessions),
        Arc::new(SpyBus::default()),
    );
    launch.execute(launch_input(agent_a.id)).await.unwrap();
    launch.execute(launch_input(agent_b.id)).await.unwrap();
    let dir_a = format!("/home/me/proj/.ideai/run/{}", agent_a.id);
    let dir_b = format!("/home/me/proj/.ideai/run/{}", agent_b.id);
    assert_ne!(dir_a, dir_b, "the two agents must map to different run dirs");
    // Two distinct run dirs were created.
    assert_eq!(fs.created_dirs(), vec![dir_a.clone(), dir_b.clone()]);
    // Two spawns at two distinct cwd — the core anti-collision guarantee.
    let spawns = pty.spawns();
    assert_eq!(spawns.len(), 2);
    assert_eq!(spawns[0].cwd.as_str(), dir_a);
    assert_eq!(spawns[1].cwd.as_str(), dir_b);
    assert_ne!(spawns[0].cwd, spawns[1].cwd);
    // Two convention files, each inside its own run dir (no shared root file).
    let writes = fs.writes();
    assert_eq!(writes.len(), 2);
    assert_eq!(writes[0].0, format!("{dir_a}/CLAUDE.md"));
    assert_eq!(writes[1].0, format!("{dir_b}/CLAUDE.md"));
    assert_ne!(writes[0].0, writes[1].0);
    // Neither writes to the project root.
    assert!(writes.iter().all(|(p, _)| p != "/home/me/proj/CLAUDE.md"));
    // Each convention file carries its own persona.
    assert!(String::from_utf8(writes[0].1.clone()).unwrap().contains("# alpha"));
    assert!(String::from_utf8(writes[1].1.clone()).unwrap().contains("# bravo"));
 }
 #[tokio::test]
 async fn launch_conventionfile_injects_assigned_skills_in_order() {
    // An agent with two assigned global skills; the generated convention file must
    // carry both skill bodies, after the persona, in assignment order (§14.2).
    let skill_id = |n: u128| SkillId::from_uuid(Uuid::from_u128(n));
    let skill = |n: u128, name: &str, body: &str| {
        Skill::new(skill_id(n), name, MarkdownDoc::new(body), SkillScope::Global).unwrap()
    };
    let mut agent = scratch_agent(aid(1), "Backend", "agents/backend.md", pid(9));
    agent.assign_skill(SkillRef::new(skill_id(1), SkillScope::Global));
    agent.assign_skill(SkillRef::new(skill_id(2), SkillScope::Global));
    let contexts = FakeContexts::with_agent(&agent, "# persona");
    let profiles = FakeProfiles::new(vec![profile(
        pid(9),
        ContextInjection::convention_file("CLAUDE.md").unwrap(),
    )]);
    let tr = trace();
    let fs = FakeFs::new(Arc::clone(&tr));
    let pty = FakePty::new(Arc::clone(&tr), sid(777));
    let skills = FakeSkills::with(vec![
        skill(1, "refactor", "REFAC_BODY"),
        skill(2, "review", "REVIEW_BODY"),
    ]);
    let launch = LaunchAgent::new(
        Arc::new(contexts),
        Arc::new(profiles),
        Arc::new(FakeRuntime::new(
            Arc::clone(&tr),
            Some(ContextInjectionPlan::File {
                target: "CLAUDE.md".to_owned(),
            }),
        )),
        Arc::new(fs.clone()),
        Arc::new(pty.clone()),
        Arc::new(skills),
        Arc::new(TerminalSessions::new()),
        Arc::new(SpyBus::default()),
    );
    launch.execute(launch_input(agent.id)).await.unwrap();
    let writes = fs.writes();
    assert_eq!(writes.len(), 1);
    let doc = String::from_utf8(writes[0].1.clone()).unwrap();
    assert!(doc.contains("# persona"), "persona present: {doc}");
    assert!(doc.contains("REFAC_BODY"), "first skill body present");
    assert!(doc.contains("REVIEW_BODY"), "second skill body present");
    // Deterministic order: persona before skills, skill 1 before skill 2.
    let persona_at = doc.find("# persona").unwrap();
    let refac_at = doc.find("REFAC_BODY").unwrap();
    let review_at = doc.find("REVIEW_BODY").unwrap();
    assert!(persona_at < refac_at && refac_at < review_at, "ordering: {doc}");
 }
 #[tokio::test]
 async fn launch_skips_dangling_skill_ref_without_failing() {
    // The agent references a skill that no longer exists in the store: launch must
    // still succeed and simply omit it (no Skills section for a sole dangling ref).
    let mut agent = scratch_agent(aid(1), "Backend", "agents/backend.md", pid(9));
    agent.assign_skill(SkillRef::new(SkillId::from_uuid(Uuid::from_u128(99)), SkillScope::Global));
    let contexts = FakeContexts::with_agent(&agent, "# persona");
    let profiles = FakeProfiles::new(vec![profile(
        pid(9),
        ContextInjection::convention_file("CLAUDE.md").unwrap(),
    )]);
    let tr = trace();
    let fs = FakeFs::new(Arc::clone(&tr));
    let pty = FakePty::new(Arc::clone(&tr), sid(777));
    let launch = LaunchAgent::new(
        Arc::new(contexts),
        Arc::new(profiles),
        Arc::new(FakeRuntime::new(
            Arc::clone(&tr),
            Some(ContextInjectionPlan::File {
                target: "CLAUDE.md".to_owned(),
            }),
        )),
        Arc::new(fs.clone()),
        Arc::new(pty.clone()),
        Arc::new(FakeSkills::default()), // empty store ⇒ the ref is dangling
        Arc::new(TerminalSessions::new()),
        Arc::new(SpyBus::default()),
    );
    launch.execute(launch_input(agent.id)).await.expect("launch must succeed");
    let doc = String::from_utf8(fs.writes()[0].1.clone()).unwrap();
    assert!(!doc.contains("# Skills"), "no Skills section for a dangling ref: {doc}");
 }
 #[tokio::test]
 async fn launch_stdin_strategy_pipes_context_after_spawn() {
    let (launch, agent, fs, pty, _bus, _sessions, tr) =
@ -671,6 +909,7 @@ async fn launch_unknown_profile_is_not_found() {
        Arc::new(FakeRuntime::new(Arc::clone(&tr), Some(ContextInjectionPlan::Stdin))),
        Arc::new(FakeFs::new(Arc::clone(&tr))),
        Arc::new(pty.clone()),
        Arc::new(FakeSkills::default()),
        Arc::new(TerminalSessions::new()),
        Arc::new(SpyBus::default()),
    );
--- a/crates/application/tests/skill_usecases.rs
+++ b/crates/application/tests/skill_usecases.rs
@ -0,0 +1,395 @@
 //! L12 tests for the skill use cases with in-memory port fakes (no real
 //! store/FS): CRUD across scopes (`CreateSkill`, `UpdateSkill`, `DeleteSkill`,
 //! `ListSkills`) and the manifest-mutating assignment
 //! (`AssignSkillToAgent` / `UnassignSkillFromAgent`), asserting the
 //! `SkillAssigned` event and idempotence.
 use std::sync::{Arc, Mutex};
 use async_trait::async_trait;
 use domain::events::DomainEvent;
 use domain::ids::{AgentId, ProfileId, ProjectId, SkillId};
 use domain::markdown::MarkdownDoc;
 use domain::ports::{
    AgentContextStore, EventBus, EventStream, IdGenerator, SkillStore, StoreError,
 };
 use domain::skill::{Skill, SkillScope};
 use domain::{AgentManifest, ManifestEntry, Project, ProjectPath, RemoteRef, SkillRef};
 use uuid::Uuid;
 use application::{
    AssignSkillToAgent, AssignSkillToAgentInput, CreateSkill, CreateSkillInput, DeleteSkill,
    DeleteSkillInput, ListSkills, ListSkillsInput, UnassignSkillFromAgent,
    UnassignSkillFromAgentInput, UpdateSkill, UpdateSkillInput,
 };
 // ---------------------------------------------------------------------------
 // Fakes
 // ---------------------------------------------------------------------------
 /// In-memory skill store keyed by `(scope, id)`, honouring scope isolation.
 #[derive(Clone, Default)]
 struct FakeSkills(Arc<Mutex<Vec<Skill>>>);
 #[async_trait]
 impl SkillStore for FakeSkills {
    async fn list(&self, scope: SkillScope, _root: &ProjectPath) -> Result<Vec<Skill>, StoreError> {
        Ok(self
            .0
            .lock()
            .unwrap()
            .iter()
            .filter(|s| s.scope == scope)
            .cloned()
            .collect())
    }
    async fn get(
        &self,
        scope: SkillScope,
        _root: &ProjectPath,
        id: SkillId,
    ) -> Result<Skill, StoreError> {
        self.0
            .lock()
            .unwrap()
            .iter()
            .find(|s| s.scope == scope && s.id == id)
            .cloned()
            .ok_or(StoreError::NotFound)
    }
    async fn save(&self, skill: &Skill, _root: &ProjectPath) -> Result<(), StoreError> {
        let mut v = self.0.lock().unwrap();
        if let Some(slot) = v.iter_mut().find(|s| s.scope == skill.scope && s.id == skill.id) {
            *slot = skill.clone();
        } else {
            v.push(skill.clone());
        }
        Ok(())
    }
    async fn delete(
        &self,
        scope: SkillScope,
        _root: &ProjectPath,
        id: SkillId,
    ) -> Result<(), StoreError> {
        let mut v = self.0.lock().unwrap();
        let before = v.len();
        v.retain(|s| !(s.scope == scope && s.id == id));
        if v.len() == before {
            return Err(StoreError::NotFound);
        }
        Ok(())
    }
 }
 #[derive(Clone)]
 struct FakeContexts(Arc<Mutex<AgentManifest>>);
 impl FakeContexts {
    fn new(entries: Vec<ManifestEntry>) -> Self {
        Self(Arc::new(Mutex::new(AgentManifest { version: 1, entries })))
    }
    fn manifest(&self) -> AgentManifest {
        self.0.lock().unwrap().clone()
    }
 }
 #[async_trait]
 impl AgentContextStore for FakeContexts {
    async fn read_context(
        &self,
        _p: &Project,
        _agent: &AgentId,
    ) -> Result<MarkdownDoc, StoreError> {
        Err(StoreError::NotFound)
    }
    async fn write_context(
        &self,
        _p: &Project,
        _agent: &AgentId,
        _md: &MarkdownDoc,
    ) -> Result<(), StoreError> {
        Ok(())
    }
    async fn load_manifest(&self, _p: &Project) -> Result<AgentManifest, StoreError> {
        Ok(self.manifest())
    }
    async fn save_manifest(&self, _p: &Project, m: &AgentManifest) -> Result<(), StoreError> {
        *self.0.lock().unwrap() = m.clone();
        Ok(())
    }
 }
 #[derive(Default, Clone)]
 struct SpyBus(Arc<Mutex<Vec<DomainEvent>>>);
 impl SpyBus {
    fn events(&self) -> Vec<DomainEvent> {
        self.0.lock().unwrap().clone()
    }
 }
 impl EventBus for SpyBus {
    fn publish(&self, event: DomainEvent) {
        self.0.lock().unwrap().push(event);
    }
    fn subscribe(&self) -> EventStream {
        Box::new(std::iter::empty())
    }
 }
 struct SeqIds(Mutex<u128>);
 impl SeqIds {
    fn new() -> Self {
        Self(Mutex::new(1))
    }
 }
 impl IdGenerator for SeqIds {
    fn new_uuid(&self) -> Uuid {
        let mut n = self.0.lock().unwrap();
        let id = Uuid::from_u128(*n);
        *n += 1;
        id
    }
 }
 // ---------------------------------------------------------------------------
 // Builders
 // ---------------------------------------------------------------------------
 fn pid(n: u128) -> ProfileId {
    ProfileId::from_uuid(Uuid::from_u128(n))
 }
 fn aid(n: u128) -> AgentId {
    AgentId::from_uuid(Uuid::from_u128(n))
 }
 fn sid(n: u128) -> SkillId {
    SkillId::from_uuid(Uuid::from_u128(n))
 }
 fn root() -> ProjectPath {
    ProjectPath::new("/home/me/demo").unwrap()
 }
 fn project() -> Project {
    Project::new(
        ProjectId::from_uuid(Uuid::from_u128(1000)),
        "demo",
        root(),
        RemoteRef::local(),
        1_700_000_000_000,
    )
    .unwrap()
 }
 fn scratch_entry(agent: AgentId) -> ManifestEntry {
    ManifestEntry::new(agent, "A", "agents/a.md", pid(1), None, false, None).unwrap()
 }
 // ---------------------------------------------------------------------------
 // CRUD
 // ---------------------------------------------------------------------------
 #[tokio::test]
 async fn create_skill_persists_in_its_scope() {
    let store = FakeSkills::default();
    let out = CreateSkill::new(Arc::new(store.clone()), Arc::new(SeqIds::new()))
        .execute(CreateSkillInput {
            name: "refactor".to_owned(),
            content: "# body".to_owned(),
            scope: SkillScope::Project,
            project_root: root(),
        })
        .await
        .unwrap();
    assert_eq!(out.skill.scope, SkillScope::Project);
    assert_eq!(
        store.list(SkillScope::Project, &root()).await.unwrap().len(),
        1
    );
    assert!(store
        .list(SkillScope::Global, &root())
        .await
        .unwrap()
        .is_empty());
 }
 #[tokio::test]
 async fn create_skill_rejects_empty_content() {
    let store = FakeSkills::default();
    let err = CreateSkill::new(Arc::new(store), Arc::new(SeqIds::new()))
        .execute(CreateSkillInput {
            name: "k".to_owned(),
            content: String::new(),
            scope: SkillScope::Global,
            project_root: root(),
        })
        .await
        .unwrap_err();
    assert_eq!(err.code(), "INVALID");
 }
 #[tokio::test]
 async fn update_skill_replaces_content() {
    let store = FakeSkills::default();
    store
        .save(
            &Skill::new(sid(1), "k", MarkdownDoc::new("v1"), SkillScope::Global).unwrap(),
            &root(),
        )
        .await
        .unwrap();
    let out = UpdateSkill::new(Arc::new(store.clone()))
        .execute(UpdateSkillInput {
            scope: SkillScope::Global,
            skill_id: sid(1),
            content: "v2".to_owned(),
            project_root: root(),
        })
        .await
        .unwrap();
    assert_eq!(out.skill.content_md.as_str(), "v2");
 }
 #[tokio::test]
 async fn update_unknown_skill_is_not_found() {
    let store = FakeSkills::default();
    let err = UpdateSkill::new(Arc::new(store))
        .execute(UpdateSkillInput {
            scope: SkillScope::Global,
            skill_id: sid(404),
            content: "x".to_owned(),
            project_root: root(),
        })
        .await
        .unwrap_err();
    assert_eq!(err.code(), "NOT_FOUND");
 }
 #[tokio::test]
 async fn list_is_scope_filtered() {
    let store = FakeSkills::default();
    store
        .save(
            &Skill::new(sid(1), "g", MarkdownDoc::new("x"), SkillScope::Global).unwrap(),
            &root(),
        )
        .await
        .unwrap();
    store
        .save(
            &Skill::new(sid(2), "p", MarkdownDoc::new("x"), SkillScope::Project).unwrap(),
            &root(),
        )
        .await
        .unwrap();
    let listed = ListSkills::new(Arc::new(store))
        .execute(ListSkillsInput {
            scope: SkillScope::Global,
            project_root: root(),
        })
        .await
        .unwrap();
    assert_eq!(listed.skills.len(), 1);
    assert_eq!(listed.skills[0].id, sid(1));
 }
 #[tokio::test]
 async fn delete_then_delete_is_not_found() {
    let store = FakeSkills::default();
    store
        .save(
            &Skill::new(sid(1), "k", MarkdownDoc::new("x"), SkillScope::Project).unwrap(),
            &root(),
        )
        .await
        .unwrap();
    let uc = DeleteSkill::new(Arc::new(store));
    uc.execute(DeleteSkillInput {
        scope: SkillScope::Project,
        skill_id: sid(1),
        project_root: root(),
    })
    .await
    .unwrap();
    let err = uc
        .execute(DeleteSkillInput {
            scope: SkillScope::Project,
            skill_id: sid(1),
            project_root: root(),
        })
        .await
        .unwrap_err();
    assert_eq!(err.code(), "NOT_FOUND");
 }
 // ---------------------------------------------------------------------------
 // Assignment
 // ---------------------------------------------------------------------------
 fn assigned_event(events: &[DomainEvent]) -> Vec<(SkillId, bool)> {
    events
        .iter()
        .filter_map(|e| match e {
            DomainEvent::SkillAssigned {
                skill_id, assigned, ..
            } => Some((*skill_id, *assigned)),
            _ => None,
        })
        .collect()
 }
 #[tokio::test]
 async fn assign_mutates_manifest_emits_event_and_is_idempotent() {
    let contexts = FakeContexts::new(vec![scratch_entry(aid(1))]);
    let bus = SpyBus::default();
    let uc = AssignSkillToAgent::new(Arc::new(contexts.clone()), Arc::new(bus.clone()));
    let input = AssignSkillToAgentInput {
        project: project(),
        agent_id: aid(1),
        skill: SkillRef::new(sid(9), SkillScope::Global),
    };
    uc.execute(input.clone()).await.unwrap();
    // Re-assigning the same skill is a no-op (dedup).
    uc.execute(input).await.unwrap();
    let entry = &contexts.manifest().entries[0];
    assert_eq!(entry.skills.len(), 1, "no duplicate assignment");
    assert_eq!(entry.skills[0].skill_id, sid(9));
    assert_eq!(
        assigned_event(&bus.events()),
        vec![(sid(9), true)],
        "exactly one assign event despite double execute"
    );
 }
 #[tokio::test]
 async fn unassign_removes_and_emits_false_then_is_noop() {
    let mut entry = scratch_entry(aid(1));
    entry.skills.push(SkillRef::new(sid(9), SkillScope::Global));
    let contexts = FakeContexts::new(vec![entry]);
    let bus = SpyBus::default();
    let uc = UnassignSkillFromAgent::new(Arc::new(contexts.clone()), Arc::new(bus.clone()));
    let input = UnassignSkillFromAgentInput {
        project: project(),
        agent_id: aid(1),
        skill_id: sid(9),
    };
    uc.execute(input.clone()).await.unwrap();
    uc.execute(input).await.unwrap(); // already gone → no-op
    assert!(contexts.manifest().entries[0].skills.is_empty());
    assert_eq!(assigned_event(&bus.events()), vec![(sid(9), false)]);
 }
 #[tokio::test]
 async fn assign_to_unknown_agent_is_not_found() {
    let contexts = FakeContexts::new(vec![]);
    let bus = SpyBus::default();
    let err = AssignSkillToAgent::new(Arc::new(contexts), Arc::new(bus))
        .execute(AssignSkillToAgentInput {
            project: project(),
            agent_id: aid(404),
            skill: SkillRef::new(sid(9), SkillScope::Global),
        })
        .await
        .unwrap_err();
    assert_eq!(err.code(), "NOT_FOUND");
 }
--- a/crates/application/tests/terminal_usecases.rs
+++ b/crates/application/tests/terminal_usecases.rs
@ -102,6 +102,10 @@ impl PtyPort for FakePty {
        Ok(Box::new(std::iter::empty()))
    }
    fn scrollback(&self, _handle: &PtyHandle) -> Result<Vec<u8>, PtyError> {
        Ok(Vec::new())
    }
    async fn kill(&self, handle: &PtyHandle) -> Result<ExitStatus, PtyError> {
        let mut inner = self.0.lock().unwrap();
        inner.calls.push(Call::Kill {
--- a/crates/domain/src/agent.rs
+++ b/crates/domain/src/agent.rs
@ -4,6 +4,7 @@ use serde::{Deserialize, Serialize};
 use crate::error::DomainError;
 use crate::ids::{AgentId, ProfileId, TemplateId};
 use crate::skill::SkillRef;
 use crate::template::TemplateVersion;
 /// Origin of an agent: created from scratch, or derived from a template.
@ -65,6 +66,10 @@ pub struct Agent {
    pub origin: AgentOrigin,
    /// Whether the agent tracks its template (only valid for template origins).
    pub synchronized: bool,
    /// Skills assigned to this agent, injected into its convention file at
    /// activation (ARCHITECTURE §14.2). Empty by default.
    #[serde(default)]
    pub skills: Vec<SkillRef>,
 }
 impl Agent {
@ -98,8 +103,37 @@ impl Agent {
            profile_id,
            origin,
            synchronized,
            skills: Vec::new(),
        })
    }
    /// Returns a copy of this agent carrying the given assigned skills,
    /// deduplicated by `skill_id` (keeping first occurrence).
    #[must_use]
    pub fn with_skills(mut self, skills: Vec<SkillRef>) -> Self {
        self.skills = Vec::new();
        for skill in skills {
            self.assign_skill(skill);
        }
        self
    }
    /// Assigns a skill to this agent. Idempotent: re-assigning the same
    /// `skill_id` is a no-op (returns `false`); a new assignment returns `true`.
    pub fn assign_skill(&mut self, skill: SkillRef) -> bool {
        if self.skills.iter().any(|s| s.skill_id == skill.skill_id) {
            return false;
        }
        self.skills.push(skill);
        true
    }
    /// Removes a skill assignment by id. Returns `true` if a skill was removed.
    pub fn unassign_skill(&mut self, skill_id: crate::ids::SkillId) -> bool {
        let before = self.skills.len();
        self.skills.retain(|s| s.skill_id != skill_id);
        self.skills.len() != before
    }
 }
 /// One entry in the project agent manifest (`.ideai/agents.json`).
@ -135,6 +169,10 @@ pub struct ManifestEntry {
    /// Template version recorded at the last sync.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub synced_template_version: Option<TemplateVersion>,
    /// Skills assigned to this agent (ARCHITECTURE §14.2). Defaults to empty for
    /// backward-compatible deserialisation of pre-L12 manifests.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub skills: Vec<SkillRef>,
 }
 impl ManifestEntry {
@ -172,6 +210,7 @@ impl ManifestEntry {
            template_id,
            synchronized,
            synced_template_version,
            skills: Vec::new(),
        })
    }
@ -196,6 +235,7 @@ impl ManifestEntry {
            template_id,
            synchronized: agent.synchronized,
            synced_template_version,
            skills: agent.skills.clone(),
        }
    }
@ -213,14 +253,15 @@ impl ManifestEntry {
            },
            _ => AgentOrigin::Scratch,
        };
-        Agent::new(
+        Ok(Agent::new(
            self.agent_id,
            self.name.clone(),
            self.md_path.clone(),
            self.profile_id,
            origin,
            self.synchronized,
-        )
+        )?
        .with_skills(self.skills.clone()))
    }
 }
--- a/crates/domain/src/events.rs
+++ b/crates/domain/src/events.rs
@ -1,7 +1,7 @@
 //! Domain events published on the [`crate::ports::EventBus`] and relayed to the
 //! presentation layer (ARCHITECTURE §3.2).
-use crate::ids::{AgentId, ProjectId, SessionId, TemplateId};
+use crate::ids::{AgentId, ProjectId, SessionId, SkillId, TemplateId};
 use crate::template::TemplateVersion;
 /// Events emitted by the domain/application as state changes occur.
@ -54,6 +54,15 @@ pub enum DomainEvent {
        /// Version it was brought up to.
        to: TemplateVersion,
    },
    /// A skill was assigned to (or unassigned from) an agent.
    SkillAssigned {
        /// The agent whose skill set changed.
        agent_id: AgentId,
        /// The skill involved.
        skill_id: SkillId,
        /// `true` if assigned, `false` if unassigned.
        assigned: bool,
    },
    /// A tab's layout changed.
    LayoutChanged {
        /// The project whose layout changed.
--- a/crates/domain/src/ids.rs
+++ b/crates/domain/src/ids.rs
@ -68,6 +68,10 @@ typed_id!(
    /// Identifies an [`crate::profile::AgentProfile`].
    ProfileId
 );
 typed_id!(
    /// Identifies a [`crate::skill::Skill`].
    SkillId
 );
 typed_id!(
    /// Identifies a [`crate::terminal::TerminalSession`].
    SessionId
--- a/crates/domain/src/lib.rs
+++ b/crates/domain/src/lib.rs
@ -41,6 +41,7 @@ pub mod ports;
 pub mod profile;
 pub mod project;
 pub mod remote;
 pub mod skill;
 pub mod template;
 pub mod terminal;
@ -53,13 +54,16 @@ mod validation;
 pub use error::DomainError;
 pub use ids::{
-    AgentId, LayoutId, NodeId, ProfileId, ProjectId, SessionId, TabId, TemplateId, WindowId,
+    AgentId, LayoutId, NodeId, ProfileId, ProjectId, SessionId, SkillId, TabId, TemplateId,
    WindowId,
 };
 pub use project::{Project, ProjectPath};
 pub use agent::{Agent, AgentManifest, AgentOrigin, ManifestEntry};
 pub use skill::{Skill, SkillRef, SkillScope};
 pub use template::{AgentTemplate, TemplateVersion};
 pub use profile::{AgentProfile, ContextInjection};
--- a/crates/domain/src/ports.rs
+++ b/crates/domain/src/ports.rs
@ -34,6 +34,7 @@ use crate::markdown::MarkdownDoc;
 use crate::profile::AgentProfile;
 use crate::project::{Project, ProjectPath};
 use crate::remote::RemoteKind;
 use crate::skill::{Skill, SkillScope};
 use crate::template::AgentTemplate;
 use crate::terminal::PtySize;
@ -209,7 +210,7 @@ pub enum FsError {
 }
 /// Errors from persistence stores ([`TemplateStore`], [`ProjectStore`],
-/// [`AgentContextStore`]).
+/// [`AgentContextStore`], [`SkillStore`]).
 #[derive(Debug, Clone, PartialEq, Eq, Error)]
 pub enum StoreError {
    /// The requested item was not found.
@ -334,10 +335,24 @@ pub trait PtyPort: Send + Sync {
    /// Subscribes to the PTY's byte output stream.
    ///
    /// Re-subscribable: each call returns a fresh stream that receives every
    /// chunk produced **from now on**. Combined with [`scrollback`](Self::scrollback)
    /// this lets the presentation layer *re-attach* a view to a still-living PTY
    /// after a navigation/layout change tore the previous view down — without
    /// re-spawning the process.
    ///
    /// # Errors
    /// [`PtyError`] if the handle is unknown.
    fn subscribe_output(&self, handle: &PtyHandle) -> Result<OutputStream, PtyError>;
    /// Returns the recent output retained for the session (a bounded scrollback
    /// ring buffer, the most recent bytes). Used to repaint a view that
    /// re-attaches to a live PTY so the terminal isn't blank.
    ///
    /// # Errors
    /// [`PtyError`] if the handle is unknown.
    fn scrollback(&self, handle: &PtyHandle) -> Result<Vec<u8>, PtyError>;
    /// Kills the PTY's process, returning its exit status.
    ///
    /// # Errors
@ -446,6 +461,57 @@ pub trait TemplateStore: Send + Sync {
    async fn delete(&self, id: crate::ids::TemplateId) -> Result<(), StoreError>;
 }
 /// CRUD for [`Skill`]s across both scopes ([`SkillScope::Global`] in the IDE
 /// store, [`SkillScope::Project`] under `.ideai/skills/`), per ARCHITECTURE
 /// §14.2 and L12.
 ///
 /// The two scopes are **isolated**: a skill saved as `Project` never surfaces in
 /// a `Global` listing and vice-versa. Each call carries the [`SkillScope`]
 /// explicitly (or via the [`Skill`] for [`save`](Self::save)) so the adapter
 /// resolves the right backing location.
 ///
 /// `root` identifies the project whose `.ideai/skills/` to use for
 /// [`SkillScope::Project`]; it is **ignored** for [`SkillScope::Global`] (which
 /// lives in the machine-global IDE store). Passing the root per call — rather
 /// than baking it into the adapter — keeps a single store instance correct
 /// across every open project (mirroring [`AgentContextStore`]).
 #[async_trait]
 pub trait SkillStore: Send + Sync {
    /// Lists all skills in `scope` (for `root`'s project when project-scoped).
    ///
    /// # Errors
    /// [`StoreError`] on failure.
    async fn list(&self, scope: SkillScope, root: &ProjectPath) -> Result<Vec<Skill>, StoreError>;
    /// Gets a skill by id within `scope`.
    ///
    /// # Errors
    /// [`StoreError::NotFound`] if absent in that scope.
    async fn get(
        &self,
        scope: SkillScope,
        root: &ProjectPath,
        id: crate::ids::SkillId,
    ) -> Result<Skill, StoreError>;
    /// Saves (creates or replaces by id) a skill in its own [`Skill::scope`].
    ///
    /// # Errors
    /// [`StoreError`] on failure.
    async fn save(&self, skill: &Skill, root: &ProjectPath) -> Result<(), StoreError>;
    /// Deletes a skill by id within `scope`.
    ///
    /// # Errors
    /// [`StoreError::NotFound`] if absent in that scope.
    async fn delete(
        &self,
        scope: SkillScope,
        root: &ProjectPath,
        id: crate::ids::SkillId,
    ) -> Result<(), StoreError>;
 }
 /// Persistence of the known-projects registry and the workspace.
 #[async_trait]
 pub trait ProjectStore: Send + Sync {
--- a/crates/domain/src/profile.rs
+++ b/crates/domain/src/profile.rs
@ -96,7 +96,10 @@ pub struct AgentProfile {
    pub context_injection: ContextInjection,
    /// Optional detection command (e.g. `claude --version`).
    pub detect: Option<String>,
-    /// Working-directory template (e.g. `"{projectRoot}"`).
+    /// Working-directory template. Always `"{agentRunDir}"` (ARCHITECTURE §14.1):
    /// an agent's PTY cwd is its isolated `.ideai/run/<agent-id>/` directory,
    /// **never** the project root, so that N agents of the same profile never
    /// collide on a single conventional file (`CLAUDE.md`, …) at the root.
    pub cwd_template: String,
 }
--- a/crates/domain/src/skill.rs
+++ b/crates/domain/src/skill.rs
@ -0,0 +1,111 @@
 //! Skill entity — reusable, model-agnostic workflows assignable to agents.
 //!
 //! A [`Skill`] is IdeA's universal equivalent of a CLI's slash-command, but
 //! without any dependency on a particular model's `/command` syntax
 //! (ARCHITECTURE §14.2). Assigned skills are injected as plain text into the
 //! agent's generated convention file at activation — there is no proprietary
 //! CLI mechanism involved.
 use serde::{Deserialize, Serialize};
 use crate::error::DomainError;
 use crate::ids::SkillId;
 use crate::markdown::MarkdownDoc;
 /// Where a skill lives, which also selects the store used to resolve it.
 ///
 /// - [`SkillScope::Global`] skills are stored in the global IDE store
 ///   (`<app_data>/IdeA/skills/`) and reusable across projects.
 /// - [`SkillScope::Project`] skills are stored under `.ideai/skills/` and are
 ///   specific to one project.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub enum SkillScope {
    /// Reusable across projects (global IDE store).
    Global,
    /// Specific to a single project (`.ideai/skills/`).
    Project,
 }
 /// A reusable workflow assignable to one or more agents.
 ///
 /// Invariants enforced here:
 /// - `name` non-empty,
 /// - `content_md` non-empty (an empty skill carries no behaviour).
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct Skill {
    /// Stable identifier.
    pub id: SkillId,
    /// Display name (also used as the `.md` file stem on disk).
    pub name: String,
    /// Markdown body — the workflow injected into an agent's convention file.
    pub content_md: MarkdownDoc,
    /// Scope (selects the backing store).
    pub scope: SkillScope,
 }
 impl Skill {
    /// Builds a validated skill.
    ///
    /// # Errors
    /// - [`DomainError::EmptyField`] if `name` is empty,
    /// - [`DomainError::EmptyField`] if `content_md` is empty.
    pub fn new(
        id: SkillId,
        name: impl Into<String>,
        content_md: MarkdownDoc,
        scope: SkillScope,
    ) -> Result<Self, DomainError> {
        let name = name.into();
        crate::validation::non_empty(&name, "skill.name")?;
        if content_md.is_empty() {
            return Err(DomainError::EmptyField {
                field: "skill.content_md",
            });
        }
        Ok(Self {
            id,
            name,
            content_md,
            scope,
        })
    }
    /// Returns a copy of this skill with replaced content, re-validating the
    /// non-empty invariant.
    ///
    /// # Errors
    /// [`DomainError::EmptyField`] if `content_md` is empty.
    pub fn with_content(&self, content_md: MarkdownDoc) -> Result<Self, DomainError> {
        Skill::new(self.id, self.name.clone(), content_md, self.scope)
    }
 }
 /// A reference from an agent to one assigned skill.
 ///
 /// Stored in the [`crate::agent::ManifestEntry`]: an agent carries 0..N of these.
 /// The `scope` is kept alongside the id so the application layer knows which
 /// store to resolve the skill from without a global lookup.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct SkillRef {
    /// The assigned skill.
    pub skill_id: SkillId,
    /// Scope of the assigned skill (selects its store).
    pub scope: SkillScope,
 }
 impl SkillRef {
    /// Builds a reference to an assigned skill.
    #[must_use]
    pub const fn new(skill_id: SkillId, scope: SkillScope) -> Self {
        Self { skill_id, scope }
    }
 }
 impl From<&Skill> for SkillRef {
    fn from(skill: &Skill) -> Self {
        Self::new(skill.id, skill.scope)
    }
 }
--- a/crates/domain/tests/entities.rs
+++ b/crates/domain/tests/entities.rs
@ -5,8 +5,8 @@ mod helpers;
 use domain::{
    Agent, AgentManifest, AgentOrigin, AgentProfile, AgentTemplate, ContextInjection, DomainError,
-    ManifestEntry, MarkdownDoc, ProfileId, Project, ProjectPath, PtySize, RemoteRef, SshAuth,
+    ManifestEntry, MarkdownDoc, ProfileId, Project, ProjectPath, PtySize, RemoteRef, Skill, SkillId,
-    TemplateId, TemplateVersion,
+    SkillRef, SkillScope, SshAuth, TemplateId, TemplateVersion,
 };
 use helpers::{AtomicSeqIdGenerator, FixedClock};
 use uuid::Uuid;
@ -409,3 +409,121 @@ fn manifest_unique_md_paths_ok() {
            .unwrap();
    assert!(AgentManifest::new(1, vec![e1, e2]).is_ok());
 }
 // ---------------------------------------------------------------------------
 // Skill invariants (L12, ARCHITECTURE §14.2)
 // ---------------------------------------------------------------------------
 fn skill_id(n: u128) -> SkillId {
    SkillId::from_uuid(Uuid::from_u128(n))
 }
 #[test]
 fn skill_valid_construction() {
    let s = Skill::new(
        skill_id(1),
        "code-review",
        MarkdownDoc::new("review the diff"),
        SkillScope::Global,
    );
    assert!(s.is_ok());
 }
 #[test]
 fn skill_rejects_empty_name() {
    let err = Skill::new(
        skill_id(1),
        "",
        MarkdownDoc::new("body"),
        SkillScope::Project,
    )
    .unwrap_err();
    assert_eq!(err, DomainError::EmptyField { field: "skill.name" });
 }
 #[test]
 fn skill_rejects_empty_content() {
    let err = Skill::new(skill_id(1), "x", MarkdownDoc::new(""), SkillScope::Global).unwrap_err();
    assert_eq!(
        err,
        DomainError::EmptyField {
            field: "skill.content_md"
        }
    );
 }
 #[test]
 fn skill_with_content_revalidates() {
    let s = Skill::new(skill_id(1), "x", MarkdownDoc::new("a"), SkillScope::Global).unwrap();
    assert!(s.with_content(MarkdownDoc::new("b")).is_ok());
    assert!(s.with_content(MarkdownDoc::new("")).is_err());
 }
 #[test]
 fn agent_assign_skill_is_idempotent() {
    let mut a = Agent::new(
        agent_id(1),
        "dev",
        "agents/dev.md",
        profile_id(),
        AgentOrigin::Scratch,
        false,
    )
    .unwrap();
    let r = SkillRef::new(skill_id(7), SkillScope::Global);
    assert!(a.assign_skill(r)); // first assignment
    assert!(!a.assign_skill(r)); // duplicate ignored
    assert_eq!(a.skills, vec![r]);
 }
 #[test]
 fn agent_unassign_skill() {
    let mut a = Agent::new(
        agent_id(1),
        "dev",
        "agents/dev.md",
        profile_id(),
        AgentOrigin::Scratch,
        false,
    )
    .unwrap();
    let r = SkillRef::new(skill_id(7), SkillScope::Project);
    a.assign_skill(r);
    assert!(a.unassign_skill(skill_id(7)));
    assert!(!a.unassign_skill(skill_id(7))); // already gone
    assert!(a.skills.is_empty());
 }
 #[test]
 fn agent_with_skills_dedups() {
    let r = SkillRef::new(skill_id(7), SkillScope::Global);
    let a = Agent::new(
        agent_id(1),
        "dev",
        "agents/dev.md",
        profile_id(),
        AgentOrigin::Scratch,
        false,
    )
    .unwrap()
    .with_skills(vec![r, r]);
    assert_eq!(a.skills, vec![r]);
 }
 #[test]
 fn manifest_entry_preserves_skills_through_agent_roundtrip() {
    let r = SkillRef::new(skill_id(7), SkillScope::Project);
    let agent = Agent::new(
        agent_id(1),
        "dev",
        "agents/dev.md",
        profile_id(),
        AgentOrigin::Scratch,
        false,
    )
    .unwrap()
    .with_skills(vec![r]);
    let entry = ManifestEntry::from_agent(&agent);
    assert_eq!(entry.skills, vec![r]);
    assert_eq!(entry.to_agent().unwrap(), agent);
 }
--- a/crates/domain/tests/serde_roundtrip.rs
+++ b/crates/domain/tests/serde_roundtrip.rs
@ -6,7 +6,7 @@ mod helpers;
 use domain::{
    Agent, AgentManifest, AgentOrigin, AgentProfile, AgentTemplate, ContextInjection, Direction,
    LayoutNode, LayoutTree, LeafCell, ManifestEntry, MarkdownDoc, Project, ProjectPath, RemoteRef,
-    SplitContainer, SshAuth, TemplateVersion, WeightedChild,
+    Skill, SkillId, SkillRef, SkillScope, SplitContainer, SshAuth, TemplateVersion, WeightedChild,
 };
 use helpers::{node, session};
 use uuid::Uuid;
@ -229,6 +229,61 @@ fn manifest_roundtrip_and_camel_case() {
    assert!(!json.contains("\"templateId\":null"), "json was {json}");
 }
 // ---------------------------------------------------------------------------
 // Skill (L12) — round-trip, camelCase scope tag, manifest skills back-compat
 // ---------------------------------------------------------------------------
 fn sid(n: u128) -> SkillId {
    SkillId::from_uuid(Uuid::from_u128(n))
 }
 #[test]
 fn skill_roundtrip_and_camel_case_scope() {
    let s = Skill::new(sid(1), "code-review", MarkdownDoc::new("body"), SkillScope::Global).unwrap();
    assert_eq!(roundtrip(&s), s);
    let json = serde_json::to_string(&s).unwrap();
    assert!(json.contains("\"scope\":\"global\""), "json was {json}");
    assert!(json.contains("\"contentMd\""), "json was {json}");
    let p = Skill::new(sid(2), "simplify", MarkdownDoc::new("b"), SkillScope::Project).unwrap();
    let pj = serde_json::to_string(&p).unwrap();
    assert!(pj.contains("\"scope\":\"project\""), "json was {pj}");
 }
 #[test]
 fn manifest_entry_skills_roundtrip_and_camel_case() {
    let entry = ManifestEntry::from_agent(
        &Agent::new(
            aid(1),
            "dev",
            "agents/dev.md",
            profid(9),
            AgentOrigin::Scratch,
            false,
        )
        .unwrap()
        .with_skills(vec![SkillRef::new(sid(5), SkillScope::Project)]),
    );
    assert_eq!(roundtrip(&entry), entry);
    let json = serde_json::to_string(&entry).unwrap();
    assert!(json.contains("\"skillId\""), "json was {json}");
    assert!(json.contains("\"scope\":\"project\""), "json was {json}");
 }
 #[test]
 fn manifest_entry_without_skills_omits_field_and_deserialises() {
    // An entry with no skills must not emit "skills" (skip_serializing_if),
    // and a pre-L12 manifest JSON (no skills key) must deserialise to empty.
    let entry =
        ManifestEntry::new(aid(1), "dev", "agents/dev.md", profid(9), None, false, None).unwrap();
    let json = serde_json::to_string(&entry).unwrap();
    assert!(!json.contains("\"skills\""), "json was {json}");
    let legacy = r#"{"agentId":"00000000-0000-0000-0000-000000000001","name":"dev","mdPath":"agents/dev.md","profileId":"00000000-0000-0000-0000-000000000009","synchronized":false}"#;
    let parsed: ManifestEntry = serde_json::from_str(legacy).unwrap();
    assert!(parsed.skills.is_empty());
 }
 // ---------------------------------------------------------------------------
 // LayoutTree (tagged enum: type/node)
 // ---------------------------------------------------------------------------
--- a/crates/infrastructure/src/lib.rs
+++ b/crates/infrastructure/src/lib.rs
@ -32,4 +32,6 @@ pub use process::LocalProcessSpawner;
 pub use pty::PortablePtyAdapter;
 pub use remote::{remote_host, LocalHost};
 pub use runtime::CliAgentRuntime;
-pub use store::{FsProfileStore, FsProjectStore, FsTemplateStore, IdeaiContextStore};
+pub use store::{
    FsProfileStore, FsProjectStore, FsSkillStore, FsTemplateStore, IdeaiContextStore,
 };
--- a/crates/infrastructure/src/pty/mod.rs
+++ b/crates/infrastructure/src/pty/mod.rs
@ -12,10 +12,15 @@
 //!   domain never sees an OS handle (ARCHITECTURE §4).
 //! - On [`spawn`](PtyPort::spawn) we open a PTY pair, spawn the command in the
 //!   slave, then start **one reader thread** that pumps bytes from the master
-//!   into an [`std::sync::mpsc`] channel. [`subscribe_output`](PtyPort::subscribe_output)
+//!   into a shared [`Broadcast`] hub. The hub does two things with every chunk:
-//!   hands back the receiver wrapped as the domain's blocking [`OutputStream`]
+//!   it appends to a bounded **scrollback ring buffer** (~100 KB, most recent
-//!   iterator; the presentation layer drains it on its own thread and forwards
+//!   bytes) and it fans the chunk out to every *currently subscribed* receiver.
-//!   chunks to the per-session Tauri channel (the `PtyBridge`).
+//! - [`subscribe_output`](PtyPort::subscribe_output) registers a fresh
 //!   subscriber and returns its receiver wrapped as the domain's blocking
 //!   [`OutputStream`] iterator. It is **re-subscribable**: after a view tears
 //!   down (navigation / layout change) a new view can re-attach to the *same*
 //!   live PTY by subscribing again and repainting the scrollback first — no
 //!   re-spawn. [`scrollback`](PtyPort::scrollback) returns that retained buffer.
 //! - [`write`](PtyPort::write) / [`resize`](PtyPort::resize) act on the stored
 //!   writer / master. [`kill`](PtyPort::kill) terminates the child, joins the
 //!   reader thread, and returns the [`ExitStatus`].
@ -30,9 +35,10 @@
 //! Unix-only assumption (no raw fds, no signals) so it should port as-is.
 use std::collections::HashMap;
 use std::collections::VecDeque;
 use std::io::{Read, Write};
 use std::sync::mpsc::{self, Receiver, Sender};
-use std::sync::Mutex;
+use std::sync::{Arc, Mutex};
 use std::thread::JoinHandle;
 use async_trait::async_trait;
@ -45,6 +51,68 @@ use domain::SessionId;
 /// Size of each read buffer pumped from the master PTY.
 const READ_BUF: usize = 8 * 1024;
 /// Maximum number of bytes retained in a session's scrollback ring buffer
 /// (~100 KB, "the most recent output"). When the buffer would exceed this, the
 /// oldest bytes are dropped so a re-attaching view repaints recent history.
 const SCROLLBACK_CAP: usize = 100 * 1024;
 /// The shared output hub of one PTY: a bounded scrollback ring buffer plus the
 /// set of currently-subscribed receivers. The reader thread feeds both; views
 /// subscribe and unsubscribe freely over the PTY's lifetime (re-attach support).
 ///
 /// Each subscriber is a [`Sender`]; a send failing (receiver dropped because the
 /// view detached) prunes that subscriber on the next chunk. This is the
 /// fan-out/broadcast that replaces the old single-take `output_rx`.
 #[derive(Default)]
 struct Broadcast {
    /// Bounded ring buffer of the most recent output bytes.
    scrollback: VecDeque<u8>,
    /// Live subscribers; pruned lazily when their receiver is gone.
    subscribers: Vec<Sender<Vec<u8>>>,
    /// Set once the PTY hit EOF (process exited) — no more output will ever come.
    eof: bool,
 }
 impl Broadcast {
    /// Appends a chunk to the scrollback (trimming to [`SCROLLBACK_CAP`]) and
    /// fans it out to every live subscriber, dropping any that have gone away.
    fn push(&mut self, chunk: &[u8]) {
        self.scrollback.extend(chunk.iter().copied());
        let overflow = self.scrollback.len().saturating_sub(SCROLLBACK_CAP);
        if overflow > 0 {
            self.scrollback.drain(0..overflow);
        }
        self.subscribers
            .retain(|tx| tx.send(chunk.to_vec()).is_ok());
    }
    /// Registers a new subscriber, returning its receiver.
    ///
    /// If the PTY already hit EOF, the returned stream is immediately closed
    /// (its sender is dropped) so a late re-attach to a finished session doesn't
    /// block forever waiting for output that will never come.
    fn subscribe(&mut self) -> Receiver<Vec<u8>> {
        let (tx, rx) = mpsc::channel();
        if !self.eof {
            self.subscribers.push(tx);
        }
        rx
    }
    /// Returns the retained scrollback as a contiguous byte vector.
    fn snapshot(&self) -> Vec<u8> {
        self.scrollback.iter().copied().collect()
    }
    /// Drops every subscriber's sender so their output streams end (EOF). Called
    /// by the reader thread when the PTY hits EOF (process exit). The scrollback
    /// is preserved so a late re-attach can still repaint the final output.
    fn close_subscribers(&mut self) {
        self.eof = true;
        self.subscribers.clear();
    }
 }
 /// A live PTY owned by the adapter.
 struct LivePty {
    /// Master side — used for resize.
@ -53,8 +121,8 @@ struct LivePty {
    writer: Box<dyn Write + Send>,
    /// The spawned child process.
    child: Box<dyn Child + Send + Sync>,
-    /// Receiver end of the output channel; taken once by `subscribe_output`.
+    /// Shared scrollback + subscriber hub, fed by the reader thread.
-    output_rx: Option<Receiver<Vec<u8>>>,
+    output: Arc<Mutex<Broadcast>>,
    /// Handle of the reader thread, joined on kill.
    reader: Option<JoinHandle<()>>,
 }
@ -124,22 +192,29 @@ impl PtyPort for PortablePtyAdapter {
            .try_clone_reader()
            .map_err(|e| PtyError::Io(e.to_string()))?;
-        // One reader thread per PTY pumps bytes into the output channel until EOF.
+        // One reader thread per PTY pumps bytes into the broadcast hub until EOF.
-        let (tx, rx): (Sender<Vec<u8>>, Receiver<Vec<u8>>) = mpsc::channel();
+        // The hub retains a scrollback ring buffer AND fans bytes out to every
        // current subscriber, so views can detach/re-attach without re-spawning.
        let output: Arc<Mutex<Broadcast>> = Arc::new(Mutex::new(Broadcast::default()));
        let output_for_reader = Arc::clone(&output);
        let reader_handle = std::thread::spawn(move || {
            let mut buf = [0u8; READ_BUF];
            loop {
                match reader.read(&mut buf) {
                    Ok(0) => break,
                    Ok(n) => {
-                        // Receiver gone (session closed) → stop pumping.
+                        if let Ok(mut hub) = output_for_reader.lock() {
-                        if tx.send(buf[..n].to_vec()).is_err() {
+                            hub.push(&buf[..n]);
                            break;
                        }
                    }
                    Err(_) => break,
                }
            }
            // EOF (process exited): end every attached stream by dropping its
            // sender, while preserving the scrollback for any late re-attach.
            if let Ok(mut hub) = output_for_reader.lock() {
                hub.close_subscribers();
            }
        });
        // The PTY layer owns the handle identity: it mints a fresh session id and
@ -153,7 +228,7 @@ impl PtyPort for PortablePtyAdapter {
            master: pair.master,
            writer,
            child,
-            output_rx: Some(rx),
+            output,
            reader: Some(reader_handle),
        };
        self.sessions
@ -189,18 +264,33 @@ impl PtyPort for PortablePtyAdapter {
    }
    fn subscribe_output(&self, handle: &PtyHandle) -> Result<OutputStream, PtyError> {
-        let mut map = self
+        let map = self
            .sessions
            .lock()
            .map_err(|_| PtyError::Io("pty registry poisoned".to_owned()))?;
-        let live = map.get_mut(&handle.session_id).ok_or(PtyError::NotFound)?;
+        let live = map.get(&handle.session_id).ok_or(PtyError::NotFound)?;
        let rx = live
-            .output_rx
+            .output
-            .take()
+            .lock()
-            .ok_or_else(|| PtyError::Io("output already subscribed".to_owned()))?;
+            .map_err(|_| PtyError::Io("pty output hub poisoned".to_owned()))?
            .subscribe();
        Ok(Box::new(rx.into_iter()))
    }
    fn scrollback(&self, handle: &PtyHandle) -> Result<Vec<u8>, PtyError> {
        let map = self
            .sessions
            .lock()
            .map_err(|_| PtyError::Io("pty registry poisoned".to_owned()))?;
        let live = map.get(&handle.session_id).ok_or(PtyError::NotFound)?;
        let snapshot = live
            .output
            .lock()
            .map_err(|_| PtyError::Io("pty output hub poisoned".to_owned()))?
            .snapshot();
        Ok(snapshot)
    }
    async fn kill(&self, handle: &PtyHandle) -> Result<ExitStatus, PtyError> {
        // Remove from the registry so the writer/master drop and the child is
        // fully owned here while we tear it down.
@ -220,7 +310,8 @@ impl PtyPort for PortablePtyAdapter {
            .map_err(|e| PtyError::Io(e.to_string()))?;
        // Dropping master/writer closes the PTY; the reader thread then sees EOF.
-        drop(live.output_rx.take());
+        // Dropping the broadcast hub drops every subscriber's sender, so any
        // still-attached view's output stream ends cleanly too.
        if let Some(reader) = live.reader.take() {
            let _ = reader.join();
        }
--- a/crates/infrastructure/src/runtime/mod.rs
+++ b/crates/infrastructure/src/runtime/mod.rs
@ -80,16 +80,23 @@ impl CliAgentRuntime {
        })
    }
-    /// Resolves the profile's `cwd_template` against the project root.
+    /// Resolves the profile's `cwd_template` against the supplied `base` cwd.
    ///
-    /// The only recognised placeholder is `{projectRoot}` (CONTEXT §9). An empty
+    /// The base is the agent's **run directory** (`.ideai/run/<agent-id>/`),
-    /// template defaults to the project root itself.
+    /// already computed (and created) by `LaunchAgent` and passed as the `cwd`
-    fn resolve_cwd(profile: &AgentProfile, root: &ProjectPath) -> Result<ProjectPath, RuntimeError> {
+    /// argument of [`prepare_invocation`](AgentRuntime::prepare_invocation). The
    /// recognised placeholder is `{agentRunDir}` (ARCHITECTURE §14.1); the legacy
    /// `{projectRoot}` is still substituted with the same base for backwards
    /// compatibility (the caller always passes the run dir now). An empty template
    /// defaults to the base itself.
    fn resolve_cwd(profile: &AgentProfile, base: &ProjectPath) -> Result<ProjectPath, RuntimeError> {
        let template = profile.cwd_template.trim();
        if template.is_empty() {
-            return Ok(root.clone());
+            return Ok(base.clone());
        }
-        let resolved = template.replace("{projectRoot}", root.as_str());
+        let resolved = template
            .replace("{agentRunDir}", base.as_str())
            .replace("{projectRoot}", base.as_str());
        ProjectPath::new(resolved).map_err(|e| RuntimeError::Invocation(e.to_string()))
    }
--- a/crates/infrastructure/src/store/mod.rs
+++ b/crates/infrastructure/src/store/mod.rs
@ -7,9 +7,11 @@
 mod context;
 mod profile;
 mod project;
 mod skill;
 mod template;
 pub use context::IdeaiContextStore;
 pub use profile::FsProfileStore;
 pub use project::FsProjectStore;
 pub use skill::FsSkillStore;
 pub use template::FsTemplateStore;
--- a/crates/infrastructure/src/store/skill.rs
+++ b/crates/infrastructure/src/store/skill.rs
@ -0,0 +1,263 @@
 //! [`FsSkillStore`] — file implementation of the [`SkillStore`] port
 //! (ARCHITECTURE §14.2, L12).
 //!
 //! Skills are reusable, model-agnostic workflows. They live in **two isolated
 //! scopes**, each backed by its own directory but the same on-disk shape
 //! (mirroring [`crate::store::FsTemplateStore`]): a small JSON index carrying the
 //! metadata needed to list without parsing every `.md`, plus the Markdown bodies
 //! under `md/`:
 //!
 //! ```text
 //! <app_data_dir>/skills/            # SkillScope::Global  (shared across projects)
 //! <project_root>/.ideai/skills/     # SkillScope::Project (travels with the code)
 //! ├── index.json   # { version, skills: [{ id, name, contentHash }] }
 //! └── md/
 //!     └── <id>.md   # a skill's Markdown content
 //! ```
 //!
 //! The two scopes never bleed into one another: a `Project` skill is invisible to
 //! a `Global` listing and vice-versa, because each resolves to a different
 //! directory. All I/O goes through the [`FileSystem`] port, so the adapter is
 //! location-neutral (a project hosted over SSH/WSL works unchanged) and
 //! Tauri-agnostic.
 //!
 //! Like the template store, [`delete`](SkillStore::delete) drops the index row
 //! and leaves the orphaned `md/<id>.md` on disk (the [`FileSystem`] port exposes
 //! no remove); since listing is index-driven, the skill is effectively gone.
 use std::hash::{Hash, Hasher};
 use std::sync::Arc;
 use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
 use domain::ids::SkillId;
 use domain::markdown::MarkdownDoc;
 use domain::ports::{FileSystem, RemotePath, SkillStore, StoreError};
 use domain::project::ProjectPath;
 use domain::skill::{Skill, SkillScope};
 /// Directory (under app-data) holding the global skills store.
 const GLOBAL_SKILLS_DIR: &str = "skills";
 /// The `.ideai/` directory name inside a project root.
 const IDEAI_DIR: &str = ".ideai";
 /// Sub-path of the project-scoped skills store inside `.ideai/`.
 const PROJECT_SKILLS_DIR: &str = "skills";
 /// Index file name inside a skills dir.
 const INDEX_FILE: &str = "index.json";
 /// Current schema version of the index file.
 const INDEX_VERSION: u32 = 1;
 /// One metadata row in `index.json` (the `.md` content lives separately).
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
 struct IndexEntry {
    id: SkillId,
    name: String,
    content_hash: String,
 }
 /// On-disk shape of a scope's `index.json`.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
 struct IndexDoc {
    version: u32,
    skills: Vec<IndexEntry>,
 }
 impl Default for IndexDoc {
    fn default() -> Self {
        Self {
            version: INDEX_VERSION,
            skills: Vec::new(),
        }
    }
 }
 /// A stable, dependency-free digest of Markdown content for out-of-app edit
 /// detection — deterministic across runs and platforms (fixed-key hasher).
 fn content_hash(md: &MarkdownDoc) -> String {
    let mut hasher = std::collections::hash_map::DefaultHasher::new();
    md.as_str().hash(&mut hasher);
    format!("{:016x}", hasher.finish())
 }
 /// File-backed [`SkillStore`], composing a [`FileSystem`] port. Holds only the
 /// machine-global app-data dir; the project root for [`SkillScope::Project`] is
 /// supplied **per call**, so a single instance serves every open project
 /// (mirroring [`crate::store::IdeaiContextStore`]).
 #[derive(Clone)]
 pub struct FsSkillStore {
    fs: Arc<dyn FileSystem>,
    app_data_dir: String,
 }
 impl FsSkillStore {
    /// Builds the store from an injected [`FileSystem`] and the global app-data
    /// dir (used for [`SkillScope::Global`]). Directories are created on first
    /// write.
    #[must_use]
    pub fn new(fs: Arc<dyn FileSystem>, app_data_dir: impl Into<String>) -> Self {
        Self {
            fs,
            app_data_dir: app_data_dir.into(),
        }
    }
    /// Root directory of a scope's store. `root` is ignored for `Global`.
    fn dir(&self, scope: SkillScope, root: &ProjectPath) -> String {
        match scope {
            SkillScope::Global => {
                let base = self.app_data_dir.trim_end_matches(['/', '\\']);
                format!("{base}/{GLOBAL_SKILLS_DIR}")
            }
            SkillScope::Project => {
                let base = root.as_str().trim_end_matches(['/', '\\']);
                format!("{base}/{IDEAI_DIR}/{PROJECT_SKILLS_DIR}")
            }
        }
    }
    /// `<scope-dir>/index.json`.
    fn index_path(&self, scope: SkillScope, root: &ProjectPath) -> RemotePath {
        RemotePath::new(format!("{}/{INDEX_FILE}", self.dir(scope, root)))
    }
    /// `<scope-dir>/md/<id>.md`.
    fn md_path(&self, scope: SkillScope, root: &ProjectPath, id: SkillId) -> RemotePath {
        RemotePath::new(format!("{}/md/{id}.md", self.dir(scope, root)))
    }
    /// Reads a scope's index, returning an empty default if absent.
    async fn read_index(
        &self,
        scope: SkillScope,
        root: &ProjectPath,
    ) -> Result<IndexDoc, StoreError> {
        match self.fs.read(&self.index_path(scope, root)).await {
            Ok(bytes) => {
                serde_json::from_slice(&bytes).map_err(|e| StoreError::Serialization(e.to_string()))
            }
            Err(domain::ports::FsError::NotFound(_)) => Ok(IndexDoc::default()),
            Err(e) => Err(StoreError::Io(e.to_string())),
        }
    }
    /// Writes a scope's index, ensuring its directory exists.
    async fn write_index(
        &self,
        scope: SkillScope,
        root: &ProjectPath,
        doc: &IndexDoc,
    ) -> Result<(), StoreError> {
        self.fs
            .create_dir_all(&RemotePath::new(self.dir(scope, root)))
            .await
            .map_err(|e| StoreError::Io(e.to_string()))?;
        let bytes =
            serde_json::to_vec_pretty(doc).map_err(|e| StoreError::Serialization(e.to_string()))?;
        self.fs
            .write(&self.index_path(scope, root), &bytes)
            .await
            .map_err(|e| StoreError::Io(e.to_string()))
    }
    /// Reconstructs the [`Skill`] for an index entry by reading its `.md`.
    async fn load(
        &self,
        scope: SkillScope,
        root: &ProjectPath,
        entry: &IndexEntry,
    ) -> Result<Skill, StoreError> {
        let bytes = self
            .fs
            .read(&self.md_path(scope, root, entry.id))
            .await
            .map_err(|e| match e {
                domain::ports::FsError::NotFound(_) => StoreError::NotFound,
                other => StoreError::Io(other.to_string()),
            })?;
        let content =
            String::from_utf8(bytes).map_err(|e| StoreError::Serialization(e.to_string()))?;
        Skill::new(entry.id, entry.name.clone(), MarkdownDoc::new(content), scope)
            .map_err(|e| StoreError::Serialization(e.to_string()))
    }
 }
 #[async_trait]
 impl SkillStore for FsSkillStore {
    async fn list(&self, scope: SkillScope, root: &ProjectPath) -> Result<Vec<Skill>, StoreError> {
        let index = self.read_index(scope, root).await?;
        let mut out = Vec::with_capacity(index.skills.len());
        for entry in &index.skills {
            out.push(self.load(scope, root, entry).await?);
        }
        Ok(out)
    }
    async fn get(
        &self,
        scope: SkillScope,
        root: &ProjectPath,
        id: SkillId,
    ) -> Result<Skill, StoreError> {
        let index = self.read_index(scope, root).await?;
        let entry = index
            .skills
            .iter()
            .find(|e| e.id == id)
            .ok_or(StoreError::NotFound)?;
        self.load(scope, root, entry).await
    }
    async fn save(&self, skill: &Skill, root: &ProjectPath) -> Result<(), StoreError> {
        let scope = skill.scope;
        // (1) Write the Markdown content.
        self.fs
            .create_dir_all(&RemotePath::new(format!("{}/md", self.dir(scope, root))))
            .await
            .map_err(|e| StoreError::Io(e.to_string()))?;
        self.fs
            .write(
                &self.md_path(scope, root, skill.id),
                skill.content_md.as_str().as_bytes(),
            )
            .await
            .map_err(|e| StoreError::Io(e.to_string()))?;
        // (2) Upsert the index metadata.
        let mut index = self.read_index(scope, root).await?;
        let row = IndexEntry {
            id: skill.id,
            name: skill.name.clone(),
            content_hash: content_hash(&skill.content_md),
        };
        if let Some(slot) = index.skills.iter_mut().find(|e| e.id == skill.id) {
            *slot = row;
        } else {
            index.skills.push(row);
        }
        self.write_index(scope, root, &index).await
    }
    async fn delete(
        &self,
        scope: SkillScope,
        root: &ProjectPath,
        id: SkillId,
    ) -> Result<(), StoreError> {
        let mut index = self.read_index(scope, root).await?;
        let before = index.skills.len();
        index.skills.retain(|e| e.id != id);
        if index.skills.len() == before {
            return Err(StoreError::NotFound);
        }
        // The orphaned `md/<id>.md` is left on disk (no FileSystem delete); the
        // index no longer references it, so it is effectively gone.
        self.write_index(scope, root, &index).await
    }
 }
--- a/crates/infrastructure/tests/agent_runtime.rs
+++ b/crates/infrastructure/tests/agent_runtime.rs
@ -217,15 +217,27 @@ fn prepare_substitutes_project_root_in_cwd_template() {
 }
 #[test]
-fn prepare_empty_cwd_template_defaults_to_root() {
+fn prepare_empty_cwd_template_defaults_to_base() {
    let rt = pure_runtime();
    let p = profile(ContextInjection::stdin(), "");
-    let root = ProjectPath::new("/home/me/proj").unwrap();
+    let base = ProjectPath::new("/home/me/proj").unwrap();
-    let spec = rt.prepare_invocation(&p, &ctx(), &root).unwrap();
+    let spec = rt.prepare_invocation(&p, &ctx(), &base).unwrap();
    assert_eq!(spec.cwd.as_str(), "/home/me/proj");
 }
 #[test]
 fn prepare_substitutes_agent_run_dir_in_cwd_template() {
    // The canonical template (ARCHITECTURE §14.1): `{agentRunDir}` resolves to the
    // base cwd the launcher passes — the agent's isolated run directory.
    let rt = pure_runtime();
    let p = profile(ContextInjection::convention_file("CLAUDE.md").unwrap(), "{agentRunDir}");
    let run_dir = ProjectPath::new("/home/me/proj/.ideai/run/agent-1").unwrap();
    let spec = rt.prepare_invocation(&p, &ctx(), &run_dir).unwrap();
    assert_eq!(spec.cwd.as_str(), "/home/me/proj/.ideai/run/agent-1");
 }
 // ---------------------------------------------------------------------------
 // detection_spec (pure)
 // ---------------------------------------------------------------------------
--- a/crates/infrastructure/tests/pty_adapter.rs
+++ b/crates/infrastructure/tests/pty_adapter.rs
@ -113,25 +113,113 @@ async fn write_is_echoed_back_through_output_stream() {
 }
 #[tokio::test]
-async fn subscribe_output_twice_is_an_error() {
+async fn subscribe_output_is_re_subscribable_for_reattach() {
    // A live PTY can be subscribed to more than once over its lifetime: the
    // first view detaches (drops its stream), a second view re-attaches and
    // still receives subsequent output — the core of the no-kill navigation fix.
    let pty = PortablePtyAdapter::new();
    let handle = pty
-        .spawn(sh_spec("sleep 0.2"), size())
+        .spawn(sh_spec("cat"), size())
        .await
        .expect("spawn cat");
    // First attachment: subscribe, observe an echo, then drop the stream
    // (simulating a view tearing down on navigation — NOT a kill).
    {
        let first = pty.subscribe_output(&handle).expect("first subscribe");
        let (tx, rx) = mpsc::channel();
        let worker = thread::spawn(move || {
            let mut all = Vec::new();
            for chunk in first {
                all.extend_from_slice(&chunk);
                if String::from_utf8_lossy(&all).contains("first-marker") {
                    let _ = tx.send(());
                    return; // drop the stream → detach
                }
            }
        });
        pty.write(&handle, b"first-marker\n").expect("write 1");
        rx.recv_timeout(TIMEOUT).expect("first view saw its marker");
        worker.join().expect("first worker joined");
    }
    // Second attachment to the SAME live PTY (no re-spawn): must still receive
    // new output produced after re-subscription.
    let second = pty.subscribe_output(&handle).expect("re-subscribe");
    let (tx, rx) = mpsc::channel();
    let worker = thread::spawn(move || {
        let mut all = Vec::new();
        for chunk in second {
            all.extend_from_slice(&chunk);
            if String::from_utf8_lossy(&all).contains("second-marker") {
                let _ = tx.send(());
            }
        }
    });
    pty.write(&handle, b"second-marker\n").expect("write 2");
    rx.recv_timeout(TIMEOUT)
        .expect("re-attached view saw new output");
    pty.kill(&handle).await.expect("kill cat");
    worker.join().expect("second worker joined after kill");
 }
 #[tokio::test]
 async fn scrollback_retains_recent_output_for_repaint() {
    // After output is produced and the process exits, the scrollback still holds
    // the recent bytes so a re-attaching view can repaint them.
    let pty = PortablePtyAdapter::new();
    let handle = pty
        .spawn(sh_spec("printf scrollback-content"), size())
        .await
        .expect("spawn");
-    let first = pty.subscribe_output(&handle);
+    // Drain to EOF so all output has been pushed into the ring buffer.
-    assert!(first.is_ok(), "first subscribe succeeds");
+    let stream = pty.subscribe_output(&handle).expect("subscribe");
    drain_with_timeout(stream, TIMEOUT);
-    let second = pty.subscribe_output(&handle);
+    let sb = pty.scrollback(&handle).expect("scrollback readable");
    let text = String::from_utf8_lossy(&sb);
    assert!(
-        second.is_err(),
+        text.contains("scrollback-content"),
-        "second subscribe on the same session must error"
+        "scrollback should retain recent output, got {text:?}"
    );
-    // Drain the first stream so the reader thread can finish, then tidy up.
+    let _ = pty.kill(&handle).await;
-    let stream = first.unwrap();
+}
 #[tokio::test]
 async fn scrollback_is_bounded_to_cap_and_keeps_most_recent_bytes() {
    // Emit clearly more than 100 KB of deterministic output, then assert the
    // retained scrollback is bounded and ends with the most recent bytes.
    let pty = PortablePtyAdapter::new();
    // 5000 lines of "....END<n>" → well over 100 KB; the tail is the freshest.
    let script = "i=0; while [ $i -lt 5000 ]; do \
        printf 'AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA-%d\\n' $i; \
        i=$((i+1)); done; printf 'FINAL-LINE-MARKER'";
    let handle = pty.spawn(sh_spec(script), size()).await.expect("spawn");
    let stream = pty.subscribe_output(&handle).expect("subscribe");
    drain_with_timeout(stream, TIMEOUT);
    let sb = pty.scrollback(&handle).expect("scrollback readable");
    assert!(
        sb.len() <= 100 * 1024,
        "scrollback must be bounded to ~100 KB, was {} bytes",
        sb.len()
    );
    // The newest output is retained even though the oldest was dropped.
    let text = String::from_utf8_lossy(&sb);
    assert!(
        text.contains("FINAL-LINE-MARKER"),
        "the most recent bytes must be kept in the ring buffer"
    );
    // And the very first lines must have been evicted.
    assert!(
        !text.contains("-0\n") || sb.len() < 100 * 1024,
        "oldest bytes should be dropped once the cap is exceeded"
    );
    let _ = pty.kill(&handle).await;
 }
--- a/crates/infrastructure/tests/skill_store.rs
+++ b/crates/infrastructure/tests/skill_store.rs
@ -0,0 +1,211 @@
 //! L12 integration tests for [`FsSkillStore`] against a real temp directory and a
 //! real [`LocalFileSystem`]: md + `index.json` round-trip in **both** scopes,
 //! scope isolation (a `Project` skill never appears under `Global` and vice
 //! versa), upsert, delete, tolerant reads, and the on-disk layout.
 use std::path::PathBuf;
 use std::sync::Arc;
 use domain::ids::SkillId;
 use domain::markdown::MarkdownDoc;
 use domain::ports::{FileSystem, RemotePath, SkillStore, StoreError};
 use domain::project::ProjectPath;
 use domain::skill::{Skill, SkillScope};
 use infrastructure::{FsSkillStore, LocalFileSystem};
 use uuid::Uuid;
 /// A unique scratch directory under the OS temp dir, cleaned up on drop. It plays
 /// **both** roles: its own path is the global app-data dir, and a `project/`
 /// child is the project root (so the two scopes resolve to disjoint subtrees).
 struct TempDir(PathBuf);
 impl TempDir {
    fn new() -> Self {
        let p = std::env::temp_dir().join(format!("idea-l12-skill-{}", Uuid::new_v4()));
        std::fs::create_dir_all(&p).unwrap();
        Self(p)
    }
    fn app_data_dir(&self) -> String {
        self.0.to_string_lossy().into_owned()
    }
    fn project_root(&self) -> ProjectPath {
        ProjectPath::new(self.0.join("project").to_string_lossy().into_owned()).unwrap()
    }
    fn child(&self, rel: &str) -> RemotePath {
        RemotePath::new(self.0.join(rel).to_string_lossy().into_owned())
    }
 }
 impl Drop for TempDir {
    fn drop(&mut self) {
        let _ = std::fs::remove_dir_all(&self.0);
    }
 }
 fn store(tmp: &TempDir) -> FsSkillStore {
    let fs: Arc<dyn FileSystem> = Arc::new(LocalFileSystem::new());
    FsSkillStore::new(fs, tmp.app_data_dir())
 }
 fn sid(n: u128) -> SkillId {
    SkillId::from_uuid(Uuid::from_u128(n))
 }
 fn skill(id: SkillId, name: &str, content: &str, scope: SkillScope) -> Skill {
    Skill::new(id, name, MarkdownDoc::new(content), scope).unwrap()
 }
 #[tokio::test]
 async fn missing_index_lists_empty_in_both_scopes() {
    let tmp = TempDir::new();
    let s = store(&tmp);
    let root = tmp.project_root();
    assert!(s.list(SkillScope::Global, &root).await.unwrap().is_empty());
    assert!(s.list(SkillScope::Project, &root).await.unwrap().is_empty());
 }
 #[tokio::test]
 async fn global_roundtrip_save_get_list() {
    let tmp = TempDir::new();
    let s = store(&tmp);
    let root = tmp.project_root();
    let sk = skill(sid(1), "refactor", "# Refactor workflow", SkillScope::Global);
    s.save(&sk, &root).await.unwrap();
    assert_eq!(s.get(SkillScope::Global, &root, sid(1)).await.unwrap(), sk);
    assert_eq!(
        s.list(SkillScope::Global, &root).await.unwrap(),
        vec![sk.clone()]
    );
    // The Markdown landed under the global skills dir.
    let fs = LocalFileSystem::new();
    let bytes = fs
        .read(&tmp.child(&format!("skills/md/{}.md", sid(1))))
        .await
        .unwrap();
    assert_eq!(String::from_utf8(bytes).unwrap(), sk.content_md.as_str());
 }
 #[tokio::test]
 async fn project_roundtrip_lands_under_ideai() {
    let tmp = TempDir::new();
    let s = store(&tmp);
    let root = tmp.project_root();
    let sk = skill(sid(1), "review", "# Review workflow", SkillScope::Project);
    s.save(&sk, &root).await.unwrap();
    assert_eq!(s.get(SkillScope::Project, &root, sid(1)).await.unwrap(), sk);
    let fs = LocalFileSystem::new();
    let bytes = fs
        .read(&tmp.child(&format!("project/.ideai/skills/md/{}.md", sid(1))))
        .await
        .unwrap();
    assert_eq!(String::from_utf8(bytes).unwrap(), sk.content_md.as_str());
 }
 #[tokio::test]
 async fn scopes_are_isolated() {
    let tmp = TempDir::new();
    let s = store(&tmp);
    let root = tmp.project_root();
    s.save(&skill(sid(1), "g", "global body", SkillScope::Global), &root)
        .await
        .unwrap();
    s.save(
        &skill(sid(2), "p", "project body", SkillScope::Project),
        &root,
    )
    .await
    .unwrap();
    // A global skill never surfaces in the project scope and vice-versa.
    let globals = s.list(SkillScope::Global, &root).await.unwrap();
    let projects = s.list(SkillScope::Project, &root).await.unwrap();
    assert_eq!(globals.len(), 1);
    assert_eq!(projects.len(), 1);
    assert_eq!(globals[0].id, sid(1));
    assert_eq!(projects[0].id, sid(2));
    assert!(matches!(
        s.get(SkillScope::Global, &root, sid(2)).await.unwrap_err(),
        StoreError::NotFound
    ));
    assert!(matches!(
        s.get(SkillScope::Project, &root, sid(1)).await.unwrap_err(),
        StoreError::NotFound
    ));
 }
 #[tokio::test]
 async fn save_upserts_content() {
    let tmp = TempDir::new();
    let s = store(&tmp);
    let root = tmp.project_root();
    s.save(&skill(sid(1), "k", "v1", SkillScope::Global), &root)
        .await
        .unwrap();
    s.save(&skill(sid(1), "k", "v2", SkillScope::Global), &root)
        .await
        .unwrap();
    assert_eq!(
        s.get(SkillScope::Global, &root, sid(1))
            .await
            .unwrap()
            .content_md
            .as_str(),
        "v2"
    );
    assert_eq!(
        s.list(SkillScope::Global, &root).await.unwrap().len(),
        1,
        "upsert, not append"
    );
 }
 #[tokio::test]
 async fn delete_removes_from_index_and_is_not_found_twice() {
    let tmp = TempDir::new();
    let s = store(&tmp);
    let root = tmp.project_root();
    s.save(&skill(sid(1), "k", "x", SkillScope::Project), &root)
        .await
        .unwrap();
    s.delete(SkillScope::Project, &root, sid(1)).await.unwrap();
    assert!(s.list(SkillScope::Project, &root).await.unwrap().is_empty());
    assert!(matches!(
        s.delete(SkillScope::Project, &root, sid(1))
            .await
            .unwrap_err(),
        StoreError::NotFound
    ));
 }
 #[tokio::test]
 async fn index_is_camelcase_with_content_hash() {
    let tmp = TempDir::new();
    let s = store(&tmp);
    let root = tmp.project_root();
    s.save(&skill(sid(1), "refactor", "hello", SkillScope::Global), &root)
        .await
        .unwrap();
    let fs = LocalFileSystem::new();
    let bytes = fs.read(&tmp.child("skills/index.json")).await.unwrap();
    let json: serde_json::Value = serde_json::from_slice(&bytes).unwrap();
    let entry = &json.get("skills").unwrap().as_array().unwrap()[0];
    assert_eq!(entry.get("name").and_then(|v| v.as_str()), Some("refactor"));
    assert!(
        entry.get("contentHash").is_some(),
        "camelCase contentHash present"
    );
    assert!(
        entry.get("content_hash").is_none(),
        "no snake_case leak"
    );
 }
--- a/frontend/src/adapters/agent.ts
+++ b/frontend/src/adapters/agent.ts
@ -19,8 +19,10 @@ import type {
  AgentGateway,
  CreateAgentInput,
  OpenTerminalOptions,
  ReattachResult,
  TerminalHandle,
 } from "@/ports";
 import { makeTerminalHandle } from "./terminal";
 /** Wire shape returned by the `launch_agent` command (mirrors `open_terminal`). */
 interface LaunchAgentResponse {
@ -87,22 +89,26 @@ export class TauriAgentGateway implements AgentGateway {
      onOutput: channel,
    });
-    const sessionId = res.sessionId;
+    return makeTerminalHandle(res.sessionId, channel);
  }
  async reattach(
    sessionId: string,
    onData: (bytes: Uint8Array) => void,
  ): Promise<ReattachResult> {
    // Agent sessions reattach through the same session-based `reattach_terminal`
    // command as plain terminals (the PTY is identified by its session id).
    const channel = new Channel<number[]>();
    channel.onmessage = (chunk) => onData(Uint8Array.from(chunk));
    const res = await invoke<{ sessionId: string; scrollback: number[] }>(
      "reattach_terminal",
      { sessionId, onOutput: channel },
    );
    return {
-      sessionId,
+      handle: makeTerminalHandle(res.sessionId, channel),
-      async write(data: Uint8Array): Promise<void> {
+      scrollback: Uint8Array.from(res.scrollback),
        await invoke("write_terminal", {
          request: { sessionId, data: Array.from(data) },
        });
      },
      async resize(rows: number, cols: number): Promise<void> {
        await invoke("resize_terminal", {
          request: { sessionId, rows, cols },
        });
      },
      async close(): Promise<void> {
        await invoke("close_terminal", { sessionId });
      },
    };
  }
 }
--- a/frontend/src/adapters/index.ts
+++ b/frontend/src/adapters/index.ts
@ -19,6 +19,7 @@ import { TauriTerminalGateway } from "./terminal";
 import { TauriLayoutGateway } from "./layout";
 import { TauriProfileGateway } from "./profile";
 import { TauriTemplateGateway } from "./template";
 import { TauriSkillGateway } from "./skill";
 import { TauriGitGateway } from "./git";
 function notImplemented(what: string): never {
@ -47,6 +48,7 @@ export function createTauriGateways(): Gateways {
    remote: new TauriRemoteGateway(),
    profile: new TauriProfileGateway(),
    template: new TauriTemplateGateway(),
    skill: new TauriSkillGateway(),
  };
 }
@ -58,5 +60,6 @@ export {
  TauriLayoutGateway,
  TauriProfileGateway,
  TauriTemplateGateway,
  TauriSkillGateway,
  TauriGitGateway,
 };
--- a/frontend/src/adapters/mock/index.ts
+++ b/frontend/src/adapters/mock/index.ts
@ -23,12 +23,15 @@ import type {
  LayoutTree,
  Project,
  ProfileAvailability,
  Skill,
  SkillScope,
  Template,
  Unsubscribe,
 } from "@/domain";
 import type {
  AgentGateway,
  CreateAgentInput,
  CreateSkillInput,
  CreateTemplateInput,
  Gateways,
  GitGateway,
@ -36,7 +39,9 @@ import type {
  OpenTerminalOptions,
  ProfileGateway,
  ProjectGateway,
  ReattachResult,
  RemoteGateway,
  SkillGateway,
  SystemGateway,
  TemplateGateway,
  TerminalGateway,
@ -100,6 +105,47 @@ function slugify(name: string): string {
  return out.replace(/^-+|-+$/g, "");
 }
 /**
 * A live in-memory mock PTY session: it retains a scrollback (everything written
 * to `onData`) and tracks the *current* output sink so a view can detach (sink
 * cleared, session stays alive) and later re-attach (new sink, scrollback
 * replayed). Only `close` ends the session — mirroring the backend's decoupling
 * of PTY lifecycle from view lifecycle.
 */
 class MockPtySession {
  /** Accumulated output (the scrollback ring; unbounded in the mock — fine for tests). */
  private scrollback: number[] = [];
  /** Current view sink; `null` while detached. */
  private sink: ((bytes: Uint8Array) => void) | null = null;
  /** Whether the session was explicitly closed (PTY killed). */
  closed = false;
  constructor(
    readonly sessionId: string,
    sink: (bytes: Uint8Array) => void,
  ) {
    this.sink = sink;
  }
  /** Records output into the scrollback and forwards it to the current sink. */
  emit(bytes: Uint8Array): void {
    if (this.closed) return;
    for (const b of bytes) this.scrollback.push(b);
    this.sink?.(bytes);
  }
  /** Detaches the current view: stop delivering, keep the session alive. */
  detach(): void {
    this.sink = null;
  }
  /** Re-attaches a new view, returning the retained scrollback to repaint. */
  reattach(sink: (bytes: Uint8Array) => void): Uint8Array {
    this.sink = sink;
    return Uint8Array.from(this.scrollback);
  }
 }
 /**
 * Stateful in-memory agent gateway — mirrors the backend `CreateAgentFromScratch`,
 * `ListAgents`, `ReadAgentContext`, `UpdateAgentContext`, `DeleteAgent`, and
@ -115,6 +161,8 @@ export class MockAgentGateway implements AgentGateway {
  private contexts = new Map<string, string>();
  /** Monotonic session counter for deterministic session ids in tests. */
  private sessionSeq = 0;
  /** Live agent PTY sessions, kept across detach so reattach can find them. */
  private sessions = new Map<string, MockPtySession>();
  private getAgents(projectId: string): Agent[] {
    if (!this.agents.has(projectId)) this.agents.set(projectId, []);
@ -147,6 +195,7 @@ export class MockAgentGateway implements AgentGateway {
      profileId: input.profileId,
      origin: { type: "scratch" },
      synchronized: false,
      skills: [],
    };
    list.push(agent);
    this.contexts.set(
@ -238,6 +287,18 @@ export class MockAgentGateway implements AgentGateway {
    return this.getAgents(projectId);
  }
  /**
   * Replaces an agent's assigned skills in-place.
   * Used by `MockSkillGateway.assignSkill` / `unassignSkill` so both gateways
   * share the same in-memory store.
   */
  _setSkills(projectId: string, agentId: string, skills: Agent["skills"]): void {
    const list = this.getAgents(projectId);
    const idx = list.findIndex((a) => a.id === agentId);
    if (idx === -1) return;
    list[idx] = { ...list[idx], skills };
  }
  async launchAgent(
    projectId: string,
    agentId: string,
@ -256,29 +317,64 @@ export class MockAgentGateway implements AgentGateway {
    const sessionId = `mock-agent-session-${this.sessionSeq}`;
    const cwd = options.cwd;
    const enc = new TextEncoder();
    const session = new MockPtySession(sessionId, onData);
    this.sessions.set(sessionId, session);
    // Greet so something is visible immediately (mirrors MockTerminalGateway).
    queueMicrotask(() =>
-      onData(enc.encode(`agent ${agentId} @ ${cwd}\r\n`)),
+      session.emit(enc.encode(`agent ${agentId} @ ${cwd}\r\n`)),
    );
-    let closed = false;
+    return makeMockHandle(session, () => this.sessions.delete(sessionId));
  }
  async reattach(
    sessionId: string,
    onData: (bytes: Uint8Array) => void,
  ): Promise<ReattachResult> {
    const session = this.sessions.get(sessionId);
    if (!session || session.closed) {
      const err: GatewayError = {
        code: "NOT_FOUND",
        message: `agent session ${sessionId} is not alive`,
      };
      throw err;
    }
    const scrollback = session.reattach(onData);
    return {
-      sessionId,
+      handle: makeMockHandle(session, () => this.sessions.delete(sessionId)),
      scrollback,
    };
  }
 }
 /**
 * Builds a {@link TerminalHandle} over a {@link MockPtySession}. `write` echoes
 * (cooked-terminal CRLF translation) through the session so the scrollback
 * records it; `detach` keeps the session alive; `close` ends it and unregisters.
 */
 function makeMockHandle(
  session: MockPtySession,
  unregister: () => void,
 ): TerminalHandle {
  return {
    sessionId: session.sessionId,
    async write(data: Uint8Array): Promise<void> {
-        if (closed) return;
+      if (session.closed) return;
        // Echo back, translating CR to CRLF like a cooked terminal.
      const out: number[] = [];
      for (const b of data) {
        if (b === 0x0d) out.push(0x0d, 0x0a);
        else out.push(b);
      }
-        onData(Uint8Array.from(out));
+      session.emit(Uint8Array.from(out));
    },
    async resize(): Promise<void> {},
    detach(): void {
      session.detach();
    },
    async close(): Promise<void> {
-        closed = true;
+      session.closed = true;
      unregister();
    },
  };
  }
 }
 /**
@ -288,6 +384,8 @@ export class MockAgentGateway implements AgentGateway {
 */
 export class MockTerminalGateway implements TerminalGateway {
  private seq = 0;
  /** Live sessions kept across detach so reattach can find them. */
  private sessions = new Map<string, MockPtySession>();
  async openTerminal(
    options: OpenTerminalOptions,
@ -296,27 +394,31 @@ export class MockTerminalGateway implements TerminalGateway {
    this.seq += 1;
    const sessionId = `mock-session-${this.seq}`;
    const enc = new TextEncoder();
    const session = new MockPtySession(sessionId, onData);
    this.sessions.set(sessionId, session);
    // Greet so something is visible immediately.
    queueMicrotask(() =>
-      onData(enc.encode(`mock terminal @ ${options.cwd}\r\n`)),
+      session.emit(enc.encode(`mock terminal @ ${options.cwd}\r\n`)),
    );
-    let closed = false;
+    return makeMockHandle(session, () => this.sessions.delete(sessionId));
    return {
      sessionId,
      async write(data: Uint8Array): Promise<void> {
        if (closed) return;
        // Echo back, translating CR to CRLF like a cooked terminal.
        const out: number[] = [];
        for (const b of data) {
          if (b === 0x0d) out.push(0x0d, 0x0a);
          else out.push(b);
  }
-        onData(Uint8Array.from(out));
+
-      },
+  async reattach(
-      async resize(): Promise<void> {},
+    sessionId: string,
-      async close(): Promise<void> {
+    onData: (bytes: Uint8Array) => void,
-        closed = true;
+  ): Promise<ReattachResult> {
-      },
+    const session = this.sessions.get(sessionId);
    if (!session || session.closed) {
      const err: GatewayError = {
        code: "NOT_FOUND",
        message: `terminal session ${sessionId} is not alive`,
      };
      throw err;
    }
    const scrollback = session.reattach(onData);
    return {
      handle: makeMockHandle(session, () => this.sessions.delete(sessionId)),
      scrollback,
    };
  }
 }
@ -846,6 +948,7 @@ export class MockTemplateGateway implements TemplateGateway {
        syncedTemplateVersion: template.version,
      },
      synchronized,
      skills: [],
    };
    this.agentGateway._insertAgent(projectId, agent, template.contentMd);
@ -905,6 +1008,130 @@ export class MockTemplateGateway implements TemplateGateway {
  }
 }
 /**
 * Stateful in-memory skill gateway (L12).
 *
 * Skills are keyed by `${scope}` (project scope is further partitioned by
 * `projectId`, mirroring the disjoint on-disk roots). Assignment mutates the
 * agent record held by the injected {@link MockAgentGateway}, so both gateways
 * share one store — exactly like {@link MockTemplateGateway}.
 */
 export class MockSkillGateway implements SkillGateway {
  // Global skills are project-independent; project skills live under their id.
  private global: Skill[] = [];
  private byProject = new Map<string, Skill[]>();
  private seq = 0;
  constructor(private readonly agentGateway: MockAgentGateway) {}
  private bucket(projectId: string, scope: SkillScope): Skill[] {
    if (scope === "global") return this.global;
    let list = this.byProject.get(projectId);
    if (!list) {
      list = [];
      this.byProject.set(projectId, list);
    }
    return list;
  }
  async listSkills(projectId: string, scope: SkillScope): Promise<Skill[]> {
    return structuredClone(this.bucket(projectId, scope));
  }
  async createSkill(input: CreateSkillInput): Promise<Skill> {
    this.seq += 1;
    const skill: Skill = {
      id: `mock-skill-${this.seq}`,
      name: input.name,
      contentMd: input.content,
      scope: input.scope,
    };
    this.bucket(input.projectId, input.scope).push(skill);
    return structuredClone(skill);
  }
  async updateSkill(
    projectId: string,
    scope: SkillScope,
    skillId: string,
    content: string,
  ): Promise<Skill> {
    const list = this.bucket(projectId, scope);
    const idx = list.findIndex((s) => s.id === skillId);
    if (idx === -1) {
      const err: GatewayError = {
        code: "NOT_FOUND",
        message: `skill ${skillId} not found`,
      };
      throw err;
    }
    list[idx] = { ...list[idx], contentMd: content };
    return structuredClone(list[idx]);
  }
  async deleteSkill(
    projectId: string,
    scope: SkillScope,
    skillId: string,
  ): Promise<void> {
    const list = this.bucket(projectId, scope);
    const idx = list.findIndex((s) => s.id === skillId);
    if (idx === -1) {
      const err: GatewayError = {
        code: "NOT_FOUND",
        message: `skill ${skillId} not found`,
      };
      throw err;
    }
    list.splice(idx, 1);
  }
  async assignSkill(
    projectId: string,
    agentId: string,
    skillId: string,
    scope: SkillScope,
  ): Promise<void> {
    const agent = this.agentGateway
      ._rawAgents(projectId)
      .find((a) => a.id === agentId);
    if (!agent) {
      const err: GatewayError = {
        code: "NOT_FOUND",
        message: `agent ${agentId} not found in project ${projectId}`,
      };
      throw err;
    }
    if (agent.skills.some((s) => s.skillId === skillId)) return; // idempotent
    this.agentGateway._setSkills(projectId, agentId, [
      ...agent.skills,
      { skillId, scope },
    ]);
  }
  async unassignSkill(
    projectId: string,
    agentId: string,
    skillId: string,
  ): Promise<void> {
    const agent = this.agentGateway
      ._rawAgents(projectId)
      .find((a) => a.id === agentId);
    if (!agent) {
      const err: GatewayError = {
        code: "NOT_FOUND",
        message: `agent ${agentId} not found in project ${projectId}`,
      };
      throw err;
    }
    this.agentGateway._setSkills(
      projectId,
      agentId,
      agent.skills.filter((s) => s.skillId !== skillId),
    );
  }
 }
 /** Builds the full set of mock gateways. */
 export function createMockGateways(): Gateways {
  const agentGateway = new MockAgentGateway();
@ -918,6 +1145,7 @@ export function createMockGateways(): Gateways {
    remote: new MockRemoteGateway(),
    profile: new MockProfileGateway(),
    template: new MockTemplateGateway(agentGateway),
    skill: new MockSkillGateway(agentGateway),
  };
 }
--- a/frontend/src/adapters/mock/mock.test.ts
+++ b/frontend/src/adapters/mock/mock.test.ts
@ -12,7 +12,7 @@ import { createMockGateways, MockSystemGateway } from "./index";
 const gateways: Gateways = createMockGateways();
 describe("createMockGateways", () => {
-  it("exposes all nine gateways", () => {
+  it("exposes all ten gateways", () => {
    expect(Object.keys(gateways).sort()).toEqual([
      "agent",
      "git",
@ -20,6 +20,7 @@ describe("createMockGateways", () => {
      "profile",
      "project",
      "remote",
      "skill",
      "system",
      "template",
      "terminal",
--- a/frontend/src/adapters/mock/terminal.test.ts
+++ b/frontend/src/adapters/mock/terminal.test.ts
@ -111,4 +111,61 @@ describe("MockTerminalGateway", () => {
    // Exactly one delivery so far: the greeting.
    expect(onData).toHaveBeenCalledTimes(1);
  });
  it("detach stops delivery to the old view but keeps the session alive", async () => {
    const gw = new MockTerminalGateway();
    const first: Uint8Array[] = [];
    const handle = await gw.openTerminal(
      { cwd: "/c", rows: 24, cols: 80 },
      (b) => first.push(b),
    );
    await flushMicrotasks();
    first.length = 0;
    handle.detach();
    // Output produced after detach must NOT reach the detached view.
    await handle.write(new TextEncoder().encode("after-detach"));
    expect(first).toHaveLength(0);
    // But the session is still alive: reattach succeeds.
    await expect(
      gw.reattach(handle.sessionId, () => {}),
    ).resolves.toBeDefined();
  });
  it("reattach replays scrollback and resumes live output", async () => {
    const gw = new MockTerminalGateway();
    const handle = await gw.openTerminal(
      { cwd: "/work", rows: 24, cols: 80 },
      () => {},
    );
    await flushMicrotasks();
    await handle.write(new TextEncoder().encode("typed"));
    handle.detach();
    const fresh: Uint8Array[] = [];
    const { handle: h2, scrollback } = await gw.reattach(
      handle.sessionId,
      (b) => fresh.push(b),
    );
    // Scrollback carries the prior greeting + echoed input.
    const sb = decode([scrollback]);
    expect(sb).toContain("/work");
    expect(sb).toContain("typed");
    // New output now flows to the re-attached view.
    await h2.write(new TextEncoder().encode("more"));
    expect(decode(fresh)).toBe("more");
  });
  it("reattach to a closed session rejects (PTY is gone)", async () => {
    const gw = new MockTerminalGateway();
    const handle = await gw.openTerminal(
      { cwd: "/c", rows: 24, cols: 80 },
      () => {},
    );
    await handle.close();
    await expect(gw.reattach(handle.sessionId, () => {})).rejects.toMatchObject({
      code: "NOT_FOUND",
    });
  });
 });
--- a/frontend/src/adapters/skill.ts
+++ b/frontend/src/adapters/skill.ts
@ -0,0 +1,70 @@
 /**
 * Tauri adapter for {@link SkillGateway} (L12).
 *
 * Commands use snake_case (Tauri convention); payload keys are camelCase
 * (matching the backend DTO `#[serde(rename_all = "camelCase")]`), consistent
 * with the other adapters in this directory. The `scope` value (`"global"` /
 * `"project"`) maps directly onto the backend `SkillScope` serde enum.
 */
 import { invoke } from "@tauri-apps/api/core";
 import type { Skill, SkillScope } from "@/domain";
 import type { CreateSkillInput, SkillGateway } from "@/ports";
 export class TauriSkillGateway implements SkillGateway {
  listSkills(projectId: string, scope: SkillScope): Promise<Skill[]> {
    return invoke<Skill[]>("list_skills", { projectId, scope });
  }
  createSkill(input: CreateSkillInput): Promise<Skill> {
    return invoke<Skill>("create_skill", {
      request: {
        projectId: input.projectId,
        name: input.name,
        content: input.content,
        scope: input.scope,
      },
    });
  }
  updateSkill(
    projectId: string,
    scope: SkillScope,
    skillId: string,
    content: string,
  ): Promise<Skill> {
    return invoke<Skill>("update_skill", {
      request: { projectId, scope, skillId, content },
    });
  }
  async deleteSkill(
    projectId: string,
    scope: SkillScope,
    skillId: string,
  ): Promise<void> {
    await invoke("delete_skill", { projectId, scope, skillId });
  }
  async assignSkill(
    projectId: string,
    agentId: string,
    skillId: string,
    scope: SkillScope,
  ): Promise<void> {
    await invoke("assign_skill_to_agent", {
      request: { projectId, agentId, skillId, scope },
    });
  }
  async unassignSkill(
    projectId: string,
    agentId: string,
    skillId: string,
  ): Promise<void> {
    await invoke("unassign_skill_from_agent", {
      request: { projectId, agentId, skillId },
    });
  }
 }
--- a/frontend/src/adapters/terminal.ts
+++ b/frontend/src/adapters/terminal.ts
@ -18,6 +18,7 @@ import { Channel, invoke } from "@tauri-apps/api/core";
 import type {
  OpenTerminalOptions,
  ReattachResult,
  TerminalGateway,
  TerminalHandle,
 } from "@/ports";
@ -30,6 +31,46 @@ interface OpenTerminalResponse {
  cols: number;
 }
 /** Wire shape returned by the `reattach_terminal` command. */
 interface ReattachResponse {
  sessionId: string;
  scrollback: number[];
 }
 /**
 * Builds a {@link TerminalHandle} over a session and its local output
 * {@link Channel}. `detach` stops the channel from delivering further bytes (the
 * view is gone) without touching the backend PTY; `close` kills the PTY.
 *
 * Shared by `openTerminal` and `reattach` so both produce identical handles.
 */
 export function makeTerminalHandle(
  sessionId: string,
  channel: Channel<number[]>,
 ): TerminalHandle {
  return {
    sessionId,
    async write(data: Uint8Array): Promise<void> {
      await invoke("write_terminal", {
        request: { sessionId, data: Array.from(data) },
      });
    },
    async resize(rows: number, cols: number): Promise<void> {
      await invoke("resize_terminal", {
        request: { sessionId, rows, cols },
      });
    },
    detach(): void {
      // Drop the local subscription: the backend PTY keeps running, but this
      // view stops receiving output. A later `reattach` re-wires a fresh channel.
      channel.onmessage = () => {};
    },
    async close(): Promise<void> {
      await invoke("close_terminal", { sessionId });
    },
  };
 }
 export class TauriTerminalGateway implements TerminalGateway {
  async openTerminal(
    options: OpenTerminalOptions,
@ -44,22 +85,24 @@ export class TauriTerminalGateway implements TerminalGateway {
      onOutput: channel,
    });
-    const sessionId = res.sessionId;
+    return makeTerminalHandle(res.sessionId, channel);
-    return {
+  }
  async reattach(
    sessionId: string,
    onData: (bytes: Uint8Array) => void,
  ): Promise<ReattachResult> {
    const channel = new Channel<number[]>();
    channel.onmessage = (chunk) => onData(Uint8Array.from(chunk));
    const res = await invoke<ReattachResponse>("reattach_terminal", {
      sessionId,
-      async write(data: Uint8Array): Promise<void> {
+      onOutput: channel,
        await invoke("write_terminal", {
          request: { sessionId, data: Array.from(data) },
    });
-      },
+
-      async resize(rows: number, cols: number): Promise<void> {
+    return {
-        await invoke("resize_terminal", {
+      handle: makeTerminalHandle(res.sessionId, channel),
-          request: { sessionId, rows, cols },
+      scrollback: Uint8Array.from(res.scrollback),
        });
      },
      async close(): Promise<void> {
        await invoke("close_terminal", { sessionId });
      },
    };
  }
 }
--- a/frontend/src/domain/index.ts
+++ b/frontend/src/domain/index.ts
@ -236,6 +236,35 @@ export interface Agent {
  profileId: string;
  origin: AgentOrigin;
  synchronized: boolean;
  /** Skills assigned to this agent (injected into its convention file). */
  skills: SkillRef[];
 }
 // ---------------------------------------------------------------------------
 // Skills (L12) — mirror of the domain `Skill` / `SkillRef`.
 // ---------------------------------------------------------------------------
 /**
 * Where a skill lives (selects its backing store): `global` skills are reusable
 * across projects; `project` skills are specific to one project's `.ideai/`.
 */
 export type SkillScope = "global" | "project";
 /**
 * A reusable, model-agnostic workflow assignable to agents (mirror of the
 * backend `Skill` DTO, camelCase wire format).
 */
 export interface Skill {
  id: string;
  name: string;
  contentMd: string;
  scope: SkillScope;
 }
 /** A reference from an agent to one assigned skill (mirror of `SkillRef`). */
 export interface SkillRef {
  skillId: string;
  scope: SkillScope;
 }
 // ---------------------------------------------------------------------------
--- a/frontend/src/features/agents/AgentsPanel.tsx
+++ b/frontend/src/features/agents/AgentsPanel.tsx
@ -54,6 +54,50 @@ export function AgentsPanel({ projectId, projectRoot = "" }: AgentsPanelProps) {
    void templateGw.listTemplates().then(setTemplates).catch(() => {});
  }
  // Skills available for assignment (global + project), loaded lazily once.
  const skillGw = gateways.skill ?? null;
  const [skills, setSkills] = useState<import("@/domain").Skill[]>([]);
  const [skillsLoaded, setSkillsLoaded] = useState(false);
  const [skillToAssign, setSkillToAssign] = useState("");
  async function refreshSkills() {
    if (!skillGw) return;
    try {
      const [globals, projects] = await Promise.all([
        skillGw.listSkills(projectId, "global"),
        skillGw.listSkills(projectId, "project"),
      ]);
      setSkills([...projects, ...globals]);
    } catch {
      // Skills are optional — leave the list empty on failure.
    }
  }
  if (!skillsLoaded && skillGw) {
    setSkillsLoaded(true);
    void refreshSkills();
  }
  async function handleAssignSkill() {
    if (!skillGw || !vm.selectedAgentId || !skillToAssign) return;
    const target = skills.find((s) => s.id === skillToAssign);
    if (!target) return;
    await skillGw.assignSkill(
      projectId,
      vm.selectedAgentId,
      target.id,
      target.scope,
    );
    setSkillToAssign("");
    await vm.refresh();
  }
  async function handleUnassignSkill(skillId: string) {
    if (!skillGw || !vm.selectedAgentId) return;
    await skillGw.unassignSkill(projectId, vm.selectedAgentId, skillId);
    await vm.refresh();
  }
  // Context editor state — local copy before Save
  const [editedContext, setEditedContext] = useState(vm.context);
@ -73,6 +117,15 @@ export function AgentsPanel({ projectId, projectRoot = "" }: AgentsPanelProps) {
   */
  const [activeAgentId, setActiveAgentId] = useState<string | null>(null);
  /**
   * Live PTY session id of the running agent terminal, keyed by agent id. Lets
   * the terminal re-attach (rather than re-launch) when the panel re-mounts the
   * view, so navigating away never kills the agent.
   */
  const [agentSessions, setAgentSessions] = useState<Record<string, string>>(
    {},
  );
  const canCreate = newName.trim().length > 0 && !vm.busy;
  async function handleCreate(e: React.FormEvent) {
@ -325,15 +378,96 @@ export function AgentsPanel({ projectId, projectRoot = "" }: AgentsPanelProps) {
        <div className="border-t border-border p-4">
          <div className="h-96 overflow-hidden rounded-lg border border-border bg-surface">
            <TerminalView
              key={activeAgentId}
              cwd={projectRoot}
              open={(opts, onData) =>
                vm.launchAgent(activeAgentId, opts, onData)
              }
              reattach={(sessionId, onData) =>
                gateways.agent.reattach(sessionId, onData)
              }
              sessionId={agentSessions[activeAgentId] ?? null}
              onSessionId={(sid) =>
                setAgentSessions((prev) => ({ ...prev, [activeAgentId]: sid }))
              }
            />
          </div>
        </div>
      )}
      {/* ── Skill assignment ── */}
      {selectedAgent && skillGw && (
        <div className="flex flex-col gap-2 border-t border-border p-4">
          <h4 className="text-sm font-semibold text-content">
            Skills — {selectedAgent.name}
          </h4>
          {/* Assigned skills */}
          {selectedAgent.skills.length === 0 ? (
            <p className="text-xs text-muted">No skills assigned.</p>
          ) : (
            <ul className="flex flex-wrap gap-1.5">
              {selectedAgent.skills.map((ref) => {
                const name =
                  skills.find((s) => s.id === ref.skillId)?.name ?? ref.skillId;
                return (
                  <li
                    key={ref.skillId}
                    className="flex items-center gap-1 rounded-full bg-raised px-2 py-0.5 text-xs text-content"
                  >
                    <span>{name}</span>
                    <button
                      type="button"
                      aria-label={`unassign ${name}`}
                      disabled={vm.busy}
                      onClick={() => void handleUnassignSkill(ref.skillId)}
                      className="text-faint hover:text-danger"
                    >
                      ✕
                    </button>
                  </li>
                );
              })}
            </ul>
          )}
          {/* Assign selector */}
          <div className="flex items-end gap-2">
            <select
              aria-label="skill to assign"
              value={skillToAssign}
              onChange={(e) => setSkillToAssign(e.target.value)}
              disabled={vm.busy}
              className={cn(
                "h-9 min-w-0 flex-1 rounded-md bg-raised px-3 text-sm text-content",
                "border border-border outline-none transition-colors",
                "focus:border-primary disabled:cursor-not-allowed disabled:opacity-50",
              )}
            >
              <option value="">— assign a skill —</option>
              {skills
                .filter(
                  (s) =>
                    !selectedAgent.skills.some((r) => r.skillId === s.id),
                )
                .map((s) => (
                  <option key={s.id} value={s.id}>
                    {s.name} ({s.scope})
                  </option>
                ))}
            </select>
            <Button
              variant="primary"
              aria-label="assign skill"
              disabled={vm.busy || !skillToAssign}
              onClick={() => void handleAssignSkill()}
            >
              Assign
            </Button>
          </div>
        </div>
      )}
      {/* ── Context editor ── */}
      {selectedAgent && (
        <div className="flex flex-col gap-2 border-t border-border p-4">
--- a/frontend/src/features/layout/LayoutGrid.tsx
+++ b/frontend/src/features/layout/LayoutGrid.tsx
@ -112,7 +112,7 @@ interface LeafViewProps {
  projectId: string;
 }
-function LeafView({ id, agent, cwd, vm, parentSplit, projectId }: LeafViewProps) {
+function LeafView({ id, session, agent, cwd, vm, parentSplit, projectId }: LeafViewProps) {
  const canMerge = parentSplit !== null && parentSplit.siblings > 1;
  const { agent: agentGateway } = useGateways();
@ -133,6 +133,12 @@ function LeafView({ id, agent, cwd, vm, parentSplit, projectId }: LeafViewProps)
    ? (opts: Parameters<typeof agentGateway.launchAgent>[2], onData: (bytes: Uint8Array) => void) =>
        agentGateway.launchAgent(projectId, agentId, opts, onData)
    : undefined;
  // Agent cells re-attach through the agent gateway; plain cells fall back to
  // the terminal gateway's reattach (handled by TerminalView's default).
  const reattachOpener = agentGateway && agentId
    ? (sessionId: string, onData: (bytes: Uint8Array) => void) =>
        agentGateway.reattach(sessionId, onData)
    : undefined;
  return (
    <div
@ -215,8 +221,17 @@ function LeafView({ id, agent, cwd, vm, parentSplit, projectId }: LeafViewProps)
          </button>
        )}
      </div>
-      {/* Re-key terminal when the agent changes so xterm re-mounts with the right opener. */}
+      {/* Re-key terminal when the agent changes so xterm re-mounts with the right opener.
-      <TerminalView key={`${id}-${agentId ?? "plain"}`} cwd={cwd} open={terminalOpener} />
+          The cell's persisted session id drives reattach-vs-open so navigating
          (layout/tab switch) never kills the PTY. */}
      <TerminalView
        key={`${id}-${agentId ?? "plain"}`}
        cwd={cwd}
        open={terminalOpener}
        reattach={reattachOpener}
        sessionId={session}
        onSessionId={(sid) => void vm.setSession(id, sid)}
      />
    </div>
  );
 }
--- a/frontend/src/features/projects/ProjectsView.tsx
+++ b/frontend/src/features/projects/ProjectsView.tsx
@ -34,17 +34,19 @@ import type { LayoutInfo } from "@/domain";
 import { LayoutGrid, LayoutTabs } from "@/features/layout";
 import { AgentsPanel } from "@/features/agents";
 import { TemplatesPanel } from "@/features/templates";
 import { SkillsPanel } from "@/features/skills";
 import { GitPanel, GitGraphView } from "@/features/git";
 import { Button, Input, Panel, Tabs, cn } from "@/shared";
 import { useGateways } from "@/app/di";
 import { useProjects } from "./useProjects";
-type SidebarTab = "projects" | "agents" | "templates" | "git";
+type SidebarTab = "projects" | "agents" | "templates" | "skills" | "git";
 const SIDEBAR_TABS: { id: SidebarTab; label: string }[] = [
  { id: "projects", label: "Projects" },
  { id: "agents", label: "Agents" },
  { id: "templates", label: "Templates" },
  { id: "skills", label: "Skills" },
  { id: "git", label: "Git" },
 ];
@ -259,6 +261,14 @@ export function ProjectsView() {
              <p className="text-sm text-muted">Open a project to manage templates.</p>
            )}
            {/* Skills panel */}
            {sidebarTab === "skills" && active && (
              <SkillsPanel projectId={active.id} />
            )}
            {sidebarTab === "skills" && !active && (
              <p className="text-sm text-muted">Open a project to manage skills.</p>
            )}
            {/* Git panel */}
            {sidebarTab === "git" && active && (
              <GitPanel projectId={active.id} />
--- a/frontend/src/features/skills/SkillEditor.tsx
+++ b/frontend/src/features/skills/SkillEditor.tsx
@ -0,0 +1,207 @@
 /**
 * `SkillEditor` — a fullscreen overlay for creating or editing skills (L12).
 * Rendered on top of the rest of the UI (`fixed inset-0`).
 *
 * Provides:
 * - Name field and scope selector (create-mode only; scope is fixed in edit-mode)
 * - Two tabs: "Edit" (textarea) and "Preview" (react-markdown)
 * - Save (create/update) and Cancel/Close buttons
 *
 * Pure presentation: all mutations are delegated to the callbacks supplied by
 * the parent (`SkillsPanel`). Styled with `@/shared`; no inline styles.
 */
 import { useState } from "react";
 import ReactMarkdown from "react-markdown";
 import remarkGfm from "remark-gfm";
 import type { Skill, SkillScope } from "@/domain";
 import { Button, Input, cn } from "@/shared";
 export interface SkillEditorProps {
  /**
   * When set, the editor is in edit-mode for the given skill; otherwise it is in
   * create-mode (all fields start empty).
   */
  skill?: Skill | null;
  /** Called when the user submits the form. */
  onSave: (name: string, content: string, scope: SkillScope) => Promise<void>;
  /** Called when the user cancels / closes the overlay. */
  onClose: () => void;
  /** Whether a save operation is in flight. */
  busy?: boolean;
 }
 type EditorTab = "edit" | "preview";
 export function SkillEditor({
  skill,
  onSave,
  onClose,
  busy = false,
 }: SkillEditorProps) {
  const [name, setName] = useState(skill?.name ?? "");
  const [content, setContent] = useState(skill?.contentMd ?? "");
  const [scope, setScope] = useState<SkillScope>(skill?.scope ?? "project");
  const [tab, setTab] = useState<EditorTab>("edit");
  const editing = skill != null;
  const canSave = name.trim().length > 0 && content.trim().length > 0 && !busy;
  async function handleSubmit(e: React.FormEvent) {
    e.preventDefault();
    if (!canSave) return;
    await onSave(name.trim(), content, scope);
  }
  return (
    <div
      className="fixed inset-0 z-50 flex flex-col bg-canvas"
      role="dialog"
      aria-modal="true"
      aria-label="skill editor"
    >
      {/* ── Header bar ── */}
      <div className="flex shrink-0 items-center justify-between border-b border-border bg-surface px-4 py-3">
        <h2 className="text-sm font-semibold text-content">
          {editing ? `Edit skill — ${skill.name}` : "New skill"}
        </h2>
        <Button
          type="button"
          variant="ghost"
          aria-label="close skill editor"
          onClick={onClose}
          disabled={busy}
        >
          ✕
        </Button>
      </div>
      {/* ── Form body ── */}
      <form
        onSubmit={(e) => void handleSubmit(e)}
        className="flex flex-1 flex-col gap-0 overflow-hidden"
      >
        {/* ── Meta fields ── */}
        <div className="flex shrink-0 flex-wrap items-end gap-3 border-b border-border px-4 py-3">
          <div className="flex min-w-0 flex-1 flex-col gap-1">
            <label htmlFor="se-name" className="text-xs font-medium text-muted">
              Name
            </label>
            <Input
              id="se-name"
              aria-label="skill name"
              placeholder="My skill"
              value={name}
              onChange={(e) => setName(e.target.value)}
              disabled={busy}
              className="min-w-48"
            />
          </div>
          <div className="flex min-w-0 flex-col gap-1">
            <label htmlFor="se-scope" className="text-xs font-medium text-muted">
              Scope
            </label>
            <select
              id="se-scope"
              aria-label="skill scope"
              value={scope}
              onChange={(e) => setScope(e.target.value as SkillScope)}
              disabled={busy || editing}
              className={cn(
                "h-9 w-full rounded-md bg-raised px-3 text-sm text-content",
                "border border-border outline-none transition-colors",
                "focus:border-primary disabled:cursor-not-allowed disabled:opacity-50",
              )}
            >
              <option value="project">Project</option>
              <option value="global">Global</option>
            </select>
          </div>
        </div>
        {/* ── Tab strip ── */}
        <div className="flex shrink-0 items-center gap-0 border-b border-border px-4">
          {(["edit", "preview"] as EditorTab[]).map((t) => (
            <button
              key={t}
              type="button"
              role="tab"
              aria-label={t === "edit" ? "Edit" : "Preview"}
              aria-selected={tab === t}
              onClick={() => setTab(t)}
              className={cn(
                "px-4 py-2 text-sm font-medium capitalize transition-colors",
                tab === t
                  ? "border-b-2 border-primary text-content"
                  : "text-muted hover:text-content",
              )}
            >
              {t === "edit" ? "Edit" : "Preview"}
            </button>
          ))}
        </div>
        {/* ── Edit / Preview pane ── */}
        <div className="flex flex-1 flex-col overflow-hidden px-4 py-3">
          {tab === "edit" ? (
            <textarea
              aria-label="skill content"
              value={content}
              onChange={(e) => setContent(e.target.value)}
              disabled={busy}
              className={cn(
                "flex-1 w-full rounded-md bg-raised px-3 py-2 text-sm text-content font-mono",
                "border border-border outline-none transition-colors resize-none",
                "focus:border-primary placeholder:text-faint",
                "disabled:cursor-not-allowed disabled:opacity-50",
              )}
              placeholder="# Skill workflow&#10;&#10;..."
            />
          ) : (
            <div
              className={cn(
                "flex-1 overflow-auto rounded-md border border-border bg-raised px-6 py-4",
                "prose prose-sm prose-invert max-w-none",
                "[&_h1]:text-content [&_h2]:text-content [&_h3]:text-content",
                "[&_p]:text-content/90 [&_li]:text-content/90",
                "[&_code]:bg-canvas [&_code]:text-primary [&_code]:rounded [&_code]:px-1",
                "[&_pre]:bg-canvas [&_pre]:border [&_pre]:border-border [&_pre]:rounded-md",
                "[&_a]:text-primary [&_a:hover]:underline",
                "[&_blockquote]:border-l-2 [&_blockquote]:border-primary [&_blockquote]:text-muted",
                "[&_hr]:border-border",
              )}
            >
              {content.trim() ? (
                <ReactMarkdown remarkPlugins={[remarkGfm]}>
                  {content}
                </ReactMarkdown>
              ) : (
                <p className="text-sm text-muted italic">
                  Nothing to preview yet.
                </p>
              )}
            </div>
          )}
        </div>
        {/* ── Footer ── */}
        <div className="flex shrink-0 items-center justify-end gap-2 border-t border-border px-4 py-3">
          <Button type="button" variant="ghost" onClick={onClose} disabled={busy}>
            Cancel
          </Button>
          <Button
            type="submit"
            variant="primary"
            aria-label="Save skill"
            disabled={!canSave}
            loading={busy}
          >
            Save
          </Button>
        </div>
      </form>
    </div>
  );
 }
--- a/frontend/src/features/skills/SkillsPanel.tsx
+++ b/frontend/src/features/skills/SkillsPanel.tsx
@ -0,0 +1,154 @@
 /**
 * `SkillsPanel` — feature component for skill management (L12).
 *
 * Pure presentation: all behaviour comes from {@link useSkills}. Styled with
 * `@/shared` design system tokens; no inline styles.
 *
 * Provides, per scope (Global + Project):
 * - Skill list (name + scope)
 * - "New skill" button that opens {@link SkillEditor} in create-mode
 * - Per-skill "Edit" (opens the editor in edit-mode) and "Delete"
 *
 * Skills are assigned to agents from the `AgentsPanel`, not here — this panel is
 * the authoring surface only ("no direct invoke": a skill is never run from the
 * UI, only injected into an agent's convention file at activation).
 */
 import { useState } from "react";
 import type { Skill, SkillScope } from "@/domain";
 import { Button, Panel, Spinner } from "@/shared";
 import { SkillEditor } from "./SkillEditor";
 import { useSkills } from "./useSkills";
 export interface SkillsPanelProps {
  /** The project whose skills to manage. */
  projectId: string;
 }
 /** Editor open-state: "create" or edit-mode for a specific skill. */
 type EditorState = { mode: "create" } | { mode: "edit"; skill: Skill };
 export function SkillsPanel({ projectId }: SkillsPanelProps) {
  const vm = useSkills(projectId);
  const [editorState, setEditorState] = useState<EditorState | null>(null);
  const [editorBusy, setEditorBusy] = useState(false);
  async function handleSave(name: string, content: string, scope: SkillScope) {
    if (!editorState) return;
    setEditorBusy(true);
    try {
      if (editorState.mode === "create") {
        await vm.createSkill({ name, content, scope });
      } else {
        await vm.updateSkill(editorState.skill.scope, editorState.skill.id, content);
      }
      setEditorState(null);
    } finally {
      setEditorBusy(false);
    }
  }
  function renderList(scope: SkillScope, skills: Skill[]) {
    return (
      <div className="p-4">
        {skills.length === 0 ? (
          <p className="text-sm text-muted">No {scope} skills yet.</p>
        ) : (
          <ul className="flex flex-col divide-y divide-border">
            {skills.map((s) => (
              <li
                key={s.id}
                className="flex items-center justify-between gap-3 py-3 first:pt-0 last:pb-0"
              >
                <span className="min-w-0 font-medium text-content">{s.name}</span>
                <div className="flex items-center gap-1.5 shrink-0">
                  <Button
                    size="sm"
                    variant="ghost"
                    aria-label={`edit skill ${s.name}`}
                    disabled={vm.busy}
                    onClick={() => setEditorState({ mode: "edit", skill: s })}
                  >
                    Edit
                  </Button>
                  <Button
                    size="sm"
                    variant="ghost"
                    aria-label={`delete skill ${s.name}`}
                    disabled={vm.busy}
                    onClick={() => void vm.deleteSkill(s.scope, s.id)}
                    className="text-danger hover:text-danger"
                  >
                    Delete
                  </Button>
                </div>
              </li>
            ))}
          </ul>
        )}
      </div>
    );
  }
  return (
    <>
      {/* ── Fullscreen skill editor overlay ── */}
      {editorState !== null && (
        <SkillEditor
          skill={editorState.mode === "edit" ? editorState.skill : null}
          onSave={handleSave}
          onClose={() => setEditorState(null)}
          busy={editorBusy}
        />
      )}
      <Panel title="Skills" className="flex flex-col gap-0">
        {vm.error && (
          <p
            role="alert"
            className="mx-4 mt-3 rounded-md border border-danger/40 bg-danger/10 px-3 py-2 text-sm text-danger"
          >
            {vm.error}
          </p>
        )}
        {/* ── Create button ── */}
        <div className="flex items-center justify-between border-b border-border px-4 py-3">
          <h4 className="text-xs font-semibold uppercase tracking-wide text-faint">
            Skills
          </h4>
          <div className="flex items-center gap-2">
            {vm.busy && <Spinner size={14} />}
            <Button
              type="button"
              variant="primary"
              aria-label="create skill"
              disabled={vm.busy}
              onClick={() => setEditorState({ mode: "create" })}
            >
              New skill
            </Button>
          </div>
        </div>
        {/* ── Project skills ── */}
        <div className="border-b border-border px-4 pt-3">
          <h5 className="text-xs font-semibold uppercase tracking-wide text-faint">
            Project
          </h5>
        </div>
        {renderList("project", vm.projectSkills)}
        {/* ── Global skills ── */}
        <div className="border-y border-border px-4 pt-3">
          <h5 className="text-xs font-semibold uppercase tracking-wide text-faint">
            Global
          </h5>
        </div>
        {renderList("global", vm.globalSkills)}
      </Panel>
    </>
  );
 }
--- a/frontend/src/features/skills/index.ts
+++ b/frontend/src/features/skills/index.ts
@ -0,0 +1,6 @@
 /** Public surface of the skills feature (L12). */
 export { SkillsPanel } from "./SkillsPanel";
 export type { SkillsPanelProps } from "./SkillsPanel";
 export { SkillEditor } from "./SkillEditor";
 export { useSkills } from "./useSkills";
 export type { SkillsViewModel } from "./useSkills";
--- a/frontend/src/features/skills/skills.test.tsx
+++ b/frontend/src/features/skills/skills.test.tsx
@ -0,0 +1,297 @@
 /**
 * L12 — skills feature + agent assignment, wired to the stateful
 * `MockSkillGateway` (sharing a `MockAgentGateway`) via the real `DIProvider`.
 *
 * Covers:
 * - createSkill (project + global) → skill appears under the right scope
 * - updateSkill → content persisted in the gateway
 * - deleteSkill → skill removed from the list
 * - assignSkill / unassignSkill in AgentsPanel → reflected on the agent record
 * - guardrail "no direct invoke": SkillsPanel never offers a run/launch action
 *
 * Also includes MockSkillGateway unit tests.
 */
 import { describe, it, expect } from "vitest";
 import { render, screen, waitFor, fireEvent } from "@testing-library/react";
 import {
  MockAgentGateway,
  MockProfileGateway,
  MockSkillGateway,
 } from "@/adapters/mock";
 import type { Gateways } from "@/ports";
 import { DIProvider } from "@/app/di";
 import { SkillsPanel } from "./SkillsPanel";
 import { AgentsPanel } from "@/features/agents/AgentsPanel";
 const PROJECT_ID = "proj-skill-test";
 function renderSkillsPanel(agent?: MockAgentGateway, skill?: MockSkillGateway) {
  const a = agent ?? new MockAgentGateway();
  const sk = skill ?? new MockSkillGateway(a);
  const gateways = { agent: a, skill: sk } as unknown as Gateways;
  return {
    agent: a,
    skill: sk,
    ...render(
      <DIProvider gateways={gateways}>
        <SkillsPanel projectId={PROJECT_ID} />
      </DIProvider>,
    ),
  };
 }
 function renderAgentsPanel(agent: MockAgentGateway, skill: MockSkillGateway) {
  const profile = new MockProfileGateway();
  const gateways = { agent, profile, skill } as unknown as Gateways;
  return render(
    <DIProvider gateways={gateways}>
      <AgentsPanel projectId={PROJECT_ID} projectRoot="/tmp/proj" />
    </DIProvider>,
  );
 }
 async function waitForSkillsIdle() {
  await waitFor(() => {
    expect(screen.getByRole("button", { name: "create skill" })).toBeTruthy();
  });
 }
 /** Opens the SkillEditor, fills it, and saves. */
 async function createSkill(
  name: string,
  content = "# Workflow",
  scope: "project" | "global" = "project",
 ) {
  await waitForSkillsIdle();
  fireEvent.click(screen.getByRole("button", { name: "create skill" }));
  await waitFor(() => expect(screen.getByLabelText("skill name")).toBeTruthy());
  fireEvent.change(screen.getByLabelText("skill name"), {
    target: { value: name },
  });
  fireEvent.change(screen.getByLabelText("skill content"), {
    target: { value: content },
  });
  fireEvent.change(screen.getByLabelText("skill scope"), {
    target: { value: scope },
  });
  fireEvent.click(screen.getByRole("button", { name: "Save skill" }));
 }
 // ---------------------------------------------------------------------------
 // SkillsPanel feature tests
 // ---------------------------------------------------------------------------
 describe("SkillsPanel (with MockSkillGateway)", () => {
  it("shows empty state for both scopes initially", async () => {
    renderSkillsPanel();
    await waitForSkillsIdle();
    expect(screen.getByText("No project skills yet.")).toBeTruthy();
    expect(screen.getByText("No global skills yet.")).toBeTruthy();
  });
  it("creating a project skill adds it to the list", async () => {
    const { skill } = renderSkillsPanel();
    await createSkill("Refactor", "## steps", "project");
    await screen.findByText("Refactor");
    const projects = await skill.listSkills(PROJECT_ID, "project");
    expect(projects).toHaveLength(1);
    expect(projects[0].name).toBe("Refactor");
  });
  it("creating a global skill stores it under the global scope", async () => {
    const { skill } = renderSkillsPanel();
    await createSkill("Review", "## global", "global");
    await screen.findByText("Review");
    expect(await skill.listSkills(PROJECT_ID, "global")).toHaveLength(1);
    expect(await skill.listSkills(PROJECT_ID, "project")).toHaveLength(0);
  });
  it("Save is disabled when name or content is empty", async () => {
    renderSkillsPanel();
    await waitForSkillsIdle();
    fireEvent.click(screen.getByRole("button", { name: "create skill" }));
    await waitFor(() =>
      expect(screen.getByRole("button", { name: "Save skill" })).toBeTruthy(),
    );
    const save = screen.getByRole("button", {
      name: "Save skill",
    }) as HTMLButtonElement;
    expect(save.disabled).toBe(true);
  });
  it("editing a skill persists the new content", async () => {
    const { skill } = renderSkillsPanel();
    await createSkill("Editable", "old", "project");
    await screen.findByText("Editable");
    fireEvent.click(screen.getByRole("button", { name: "edit skill Editable" }));
    await waitFor(() =>
      expect(screen.getByLabelText("skill content")).toBeTruthy(),
    );
    fireEvent.change(screen.getByLabelText("skill content"), {
      target: { value: "new content" },
    });
    fireEvent.click(screen.getByRole("button", { name: "Save skill" }));
    await waitFor(async () => {
      const list = await skill.listSkills(PROJECT_ID, "project");
      expect(list[0].contentMd).toBe("new content");
    });
  });
  it("deleting a skill removes it from the list", async () => {
    renderSkillsPanel();
    await createSkill("ToDelete", "x", "project");
    await screen.findByText("ToDelete");
    fireEvent.click(
      screen.getByRole("button", { name: "delete skill ToDelete" }),
    );
    await waitFor(() => expect(screen.queryByText("ToDelete")).toBeNull());
  });
  it("guardrail: never offers a run/launch action for a skill (no direct invoke)", async () => {
    renderSkillsPanel();
    await createSkill("Plain", "x", "project");
    await screen.findByText("Plain");
    // Only Edit/Delete are exposed; nothing that would "run" the skill.
    expect(screen.queryByRole("button", { name: /^run/i })).toBeNull();
    expect(screen.queryByRole("button", { name: /^launch/i })).toBeNull();
    expect(screen.queryByRole("button", { name: /invoke/i })).toBeNull();
  });
 });
 // ---------------------------------------------------------------------------
 // Assignment integration (AgentsPanel)
 // ---------------------------------------------------------------------------
 describe("Skill assignment (AgentsPanel + MockSkillGateway)", () => {
  it("assigns a skill to the selected agent and shows it as a chip", async () => {
    const agent = new MockAgentGateway();
    const skill = new MockSkillGateway(agent);
    const created = await agent.createAgent(PROJECT_ID, {
      name: "Worker",
      profileId: "p1",
    });
    await skill.createSkill({
      projectId: PROJECT_ID,
      name: "Deploy",
      content: "## deploy",
      scope: "project",
    });
    renderAgentsPanel(agent, skill);
    await waitFor(() => expect(screen.getByLabelText("agent name")).toBeTruthy());
    // Select the agent
    fireEvent.click(screen.getByText("Worker"));
    await waitFor(() =>
      expect(screen.getByLabelText("skill to assign")).toBeTruthy(),
    );
    // Choose the skill and assign
    const sk = (await skill.listSkills(PROJECT_ID, "project"))[0];
    fireEvent.change(screen.getByLabelText("skill to assign"), {
      target: { value: sk.id },
    });
    fireEvent.click(screen.getByRole("button", { name: "assign skill" }));
    // The agent record now carries the skill ref
    await waitFor(async () => {
      const agents = await agent.listAgents(PROJECT_ID);
      const a = agents.find((x) => x.id === created.id)!;
      expect(a.skills.map((s) => s.skillId)).toContain(sk.id);
    });
    // Chip with an unassign control appears
    expect(screen.getByRole("button", { name: "unassign Deploy" })).toBeTruthy();
  });
  it("unassigns a skill from the agent", async () => {
    const agent = new MockAgentGateway();
    const skill = new MockSkillGateway(agent);
    const a = await agent.createAgent(PROJECT_ID, {
      name: "Worker2",
      profileId: "p1",
    });
    const sk = await skill.createSkill({
      projectId: PROJECT_ID,
      name: "Cleanup",
      content: "## cleanup",
      scope: "global",
    });
    await skill.assignSkill(PROJECT_ID, a.id, sk.id, sk.scope);
    renderAgentsPanel(agent, skill);
    await waitFor(() => expect(screen.getByLabelText("agent name")).toBeTruthy());
    fireEvent.click(screen.getByText("Worker2"));
    await waitFor(() =>
      expect(
        screen.getByRole("button", { name: "unassign Cleanup" }),
      ).toBeTruthy(),
    );
    fireEvent.click(screen.getByRole("button", { name: "unassign Cleanup" }));
    await waitFor(async () => {
      const agents = await agent.listAgents(PROJECT_ID);
      expect(agents.find((x) => x.id === a.id)!.skills).toHaveLength(0);
    });
  });
 });
 // ---------------------------------------------------------------------------
 // MockSkillGateway unit tests
 // ---------------------------------------------------------------------------
 describe("MockSkillGateway (unit)", () => {
  it("scopes are isolated: a project skill never appears in global and vice-versa", async () => {
    const gw = new MockSkillGateway(new MockAgentGateway());
    await gw.createSkill({
      projectId: "p",
      name: "P",
      content: "x",
      scope: "project",
    });
    await gw.createSkill({
      projectId: "p",
      name: "G",
      content: "y",
      scope: "global",
    });
    expect(await gw.listSkills("p", "project")).toHaveLength(1);
    expect(await gw.listSkills("p", "global")).toHaveLength(1);
    expect((await gw.listSkills("p", "project"))[0].name).toBe("P");
    expect((await gw.listSkills("p", "global"))[0].name).toBe("G");
  });
  it("updateSkill throws NOT_FOUND for unknown skill", async () => {
    const gw = new MockSkillGateway(new MockAgentGateway());
    await expect(
      gw.updateSkill("p", "project", "ghost", "x"),
    ).rejects.toMatchObject({ code: "NOT_FOUND" });
  });
  it("assignSkill is idempotent (no duplicate refs)", async () => {
    const agentGw = new MockAgentGateway();
    const gw = new MockSkillGateway(agentGw);
    const a = await agentGw.createAgent("p", { name: "A", profileId: "p1" });
    const s = await gw.createSkill({
      projectId: "p",
      name: "S",
      content: "x",
      scope: "project",
    });
    await gw.assignSkill("p", a.id, s.id, s.scope);
    await gw.assignSkill("p", a.id, s.id, s.scope);
    const agents = await agentGw.listAgents("p");
    expect(agents[0].skills).toHaveLength(1);
  });
  it("assignSkill throws NOT_FOUND for unknown agent", async () => {
    const gw = new MockSkillGateway(new MockAgentGateway());
    await expect(
      gw.assignSkill("p", "ghost", "s", "project"),
    ).rejects.toMatchObject({ code: "NOT_FOUND" });
  });
 });
--- a/frontend/src/features/skills/useSkills.ts
+++ b/frontend/src/features/skills/useSkills.ts
@ -0,0 +1,145 @@
 /**
 * `useSkills` — view-model hook for the skills feature (L12).
 *
 * Owns the skill lists for both scopes (`global` + `project`) of one project and
 * the CRUD actions. Consumes {@link SkillGateway} exclusively; never touches
 * `invoke()` or `@tauri-apps/api`, keeping the component layer testable with
 * mock gateways (ARCHITECTURE §1.3).
 */
 import { useCallback, useEffect, useState } from "react";
 import type { GatewayError, Skill, SkillScope } from "@/domain";
 import type { CreateSkillInput } from "@/ports";
 import { useGateways } from "@/app/di";
 /** What the skills UI needs from this hook. */
 export interface SkillsViewModel {
  /** Global skills (reusable across projects). */
  globalSkills: Skill[];
  /** Project-scoped skills. */
  projectSkills: Skill[];
  /** Last error message, or `null`. */
  error: string | null;
  /** Whether a request is in flight. */
  busy: boolean;
  /** Reloads both scope lists. */
  refresh: () => Promise<void>;
  /** Creates a new skill and refreshes its scope's list. */
  createSkill: (input: Omit<CreateSkillInput, "projectId">) => Promise<void>;
  /** Updates the content of an existing skill. */
  updateSkill: (
    scope: SkillScope,
    skillId: string,
    content: string,
  ) => Promise<void>;
  /** Deletes a skill by id from its scope. */
  deleteSkill: (scope: SkillScope, skillId: string) => Promise<void>;
 }
 function describe(e: unknown): string {
  if (e && typeof e === "object" && "message" in e) {
    return String((e as GatewayError).message);
  }
  return String(e);
 }
 export function useSkills(projectId: string): SkillsViewModel {
  const { skill } = useGateways();
  const [globalSkills, setGlobalSkills] = useState<Skill[]>([]);
  const [projectSkills, setProjectSkills] = useState<Skill[]>([]);
  const [error, setError] = useState<string | null>(null);
  const [busy, setBusy] = useState(false);
  const refresh = useCallback(async () => {
    if (!projectId) return;
    setBusy(true);
    setError(null);
    try {
      const [globals, projects] = await Promise.all([
        skill.listSkills(projectId, "global"),
        skill.listSkills(projectId, "project"),
      ]);
      setGlobalSkills(globals);
      setProjectSkills(projects);
    } catch (e) {
      setError(describe(e));
    } finally {
      setBusy(false);
    }
  }, [skill, projectId]);
  useEffect(() => {
    void refresh();
  }, [refresh]);
  const setter = (scope: SkillScope) =>
    scope === "global" ? setGlobalSkills : setProjectSkills;
  const createSkill = useCallback(
    async (input: Omit<CreateSkillInput, "projectId">) => {
      setBusy(true);
      setError(null);
      try {
        const created = await skill.createSkill({ ...input, projectId });
        setter(input.scope)((prev) => [...prev, created]);
      } catch (e) {
        setError(describe(e));
      } finally {
        setBusy(false);
      }
    },
    [skill, projectId],
  );
  const updateSkill = useCallback(
    async (scope: SkillScope, skillId: string, content: string) => {
      setBusy(true);
      setError(null);
      try {
        const updated = await skill.updateSkill(
          projectId,
          scope,
          skillId,
          content,
        );
        setter(scope)((prev) =>
          prev.map((s) => (s.id === skillId ? updated : s)),
        );
      } catch (e) {
        setError(describe(e));
      } finally {
        setBusy(false);
      }
    },
    [skill, projectId],
  );
  const deleteSkill = useCallback(
    async (scope: SkillScope, skillId: string) => {
      setBusy(true);
      setError(null);
      try {
        await skill.deleteSkill(projectId, scope, skillId);
        setter(scope)((prev) => prev.filter((s) => s.id !== skillId));
      } catch (e) {
        setError(describe(e));
      } finally {
        setBusy(false);
      }
    },
    [skill, projectId],
  );
  return {
    globalSkills,
    projectSkills,
    error,
    busy,
    refresh,
    createSkill,
    updateSkill,
    deleteSkill,
  };
 }
--- a/frontend/src/features/terminals/TerminalView.test.tsx
+++ b/frontend/src/features/terminals/TerminalView.test.tsx
@ -5,20 +5,44 @@
 * Under jsdom xterm's `term.open` may bail gracefully (no real layout engine),
 * so these tests assert the *wiring contract* (mounts without throwing, talks to
 * the gateway port, tears down on unmount) rather than xterm's visual rendering.
 *
 * The core lifecycle invariant tested here: unmounting the view (navigation /
 * layout change) must **detach**, NEVER **close** — the backend PTY must survive
 * so a running AI isn't cut off. Re-mounting with a known session re-attaches.
 */
 import { describe, it, expect, vi } from "vitest";
 import { render, screen, waitFor } from "@testing-library/react";
-import type { Gateways, TerminalGateway, TerminalHandle } from "@/ports";
+import type {
  Gateways,
  ReattachResult,
  TerminalGateway,
  TerminalHandle,
 } from "@/ports";
 import { MockTerminalGateway } from "@/adapters/mock";
 import { DIProvider } from "@/app/di";
 import { TerminalView } from "./TerminalView";
-function renderView(terminal: TerminalGateway, cwd = "/home/me/proj") {
+function makeHandle(overrides: Partial<TerminalHandle> = {}): TerminalHandle {
  return {
    sessionId: "s1",
    write: vi.fn().mockResolvedValue(undefined),
    resize: vi.fn().mockResolvedValue(undefined),
    detach: vi.fn(),
    close: vi.fn().mockResolvedValue(undefined),
    ...overrides,
  };
 }
 function renderView(
  terminal: TerminalGateway,
  cwd = "/home/me/proj",
  extra?: Partial<React.ComponentProps<typeof TerminalView>>,
 ) {
  const gateways = { terminal } as unknown as Gateways;
  return render(
    <DIProvider gateways={gateways}>
-      <TerminalView cwd={cwd} />
+      <TerminalView cwd={cwd} {...extra} />
    </DIProvider>,
  );
 }
@ -49,20 +73,13 @@ describe("TerminalView (with MockTerminalGateway)", () => {
  });
  it("consuming gateway output (onData) does not throw", async () => {
-    // A gateway that immediately pushes bytes to the consumer, exercising the
+    const handle = makeHandle();
    // gateway→term.write path. The component must swallow this safely even when
    // xterm bailed under jsdom.
    const handle: TerminalHandle = {
      sessionId: "s1",
      write: vi.fn().mockResolvedValue(undefined),
      resize: vi.fn().mockResolvedValue(undefined),
      close: vi.fn().mockResolvedValue(undefined),
    };
    const terminal: TerminalGateway = {
      openTerminal: vi.fn(async (_opts, onData) => {
        onData(new TextEncoder().encode("hello\r\n"));
        return handle;
      }),
      reattach: vi.fn(),
    };
    expect(() => renderView(terminal)).not.toThrow();
@ -71,21 +88,15 @@ describe("TerminalView (with MockTerminalGateway)", () => {
    });
  });
-  it("closes the opened handle on unmount (cleanup)", async () => {
+  it("DETACHES (does not close) the handle on unmount — the PTY must survive", async () => {
    const detach = vi.fn();
    const close = vi.fn().mockResolvedValue(undefined);
-    const handle: TerminalHandle = {
+    const handle = makeHandle({ detach, close });
      sessionId: "s1",
      write: vi.fn().mockResolvedValue(undefined),
      resize: vi.fn().mockResolvedValue(undefined),
      close,
    };
    const openTerminal = vi.fn(async () => handle);
-    const terminal: TerminalGateway = { openTerminal };
+    const terminal: TerminalGateway = { openTerminal, reattach: vi.fn() };
    const { unmount } = renderView(terminal);
    // Only assert close-on-unmount if the gateway was actually opened (i.e.
    // xterm.open did not bail in this jsdom run).
    await waitFor(() => {
      expect(openTerminal.mock.calls.length >= 0).toBe(true);
    });
@ -95,11 +106,57 @@ describe("TerminalView (with MockTerminalGateway)", () => {
    if (wasOpened) {
      await waitFor(() => {
-        expect(close).toHaveBeenCalled();
+        expect(detach).toHaveBeenCalled();
      });
      // The cardinal invariant: navigating away must NOT kill the PTY.
      expect(close).not.toHaveBeenCalled();
    } else {
-      // Bailed render: unmount must still be clean (no throw, no close needed).
+      // Bailed render: unmount must still be clean (no close, no detach needed).
      expect(close).not.toHaveBeenCalled();
    }
  });
  it("REATTACHES to an existing session instead of opening a new PTY", async () => {
    const handle = makeHandle({ sessionId: "live-1" });
    const reattach = vi.fn(
      async (_sessionId: string, onData: (b: Uint8Array) => void) => {
        onData(new TextEncoder().encode("scroll"));
        const result: ReattachResult = {
          handle,
          scrollback: new TextEncoder().encode("history"),
        };
        return result;
      },
    );
    const openTerminal = vi.fn(async () => handle);
    const terminal: TerminalGateway = { openTerminal, reattach };
    renderView(terminal, "/cwd", { sessionId: "live-1" });
    await waitFor(() => {
      // When xterm wired up, reattach must be used (with the known id) and a
      // fresh open must NOT happen.
      if (reattach.mock.calls.length > 0) {
        expect(reattach.mock.calls[0][0]).toBe("live-1");
        expect(openTerminal).not.toHaveBeenCalled();
      }
    });
    expect(true).toBe(true);
  });
  it("persists a newly opened session id via onSessionId", async () => {
    const handle = makeHandle({ sessionId: "new-99" });
    const openTerminal = vi.fn(async () => handle);
    const terminal: TerminalGateway = { openTerminal, reattach: vi.fn() };
    const onSessionId = vi.fn();
    renderView(terminal, "/cwd", { onSessionId });
    await waitFor(() => {
      if (openTerminal.mock.calls.length > 0) {
        expect(onSessionId).toHaveBeenCalledWith("new-99");
      }
    });
    expect(true).toBe(true);
  });
 });
--- a/frontend/src/features/terminals/TerminalView.tsx
+++ b/frontend/src/features/terminals/TerminalView.tsx
@ -12,6 +12,17 @@
 *
 * An optional `open` prop can override the default `terminal.openTerminal` call,
 * enabling the agent terminal to reuse this component with `agent.launchAgent`.
 *
 * **PTY lifecycle is decoupled from the view lifecycle.** Navigating (switching
 * layout or project tab) tears this view down but must NEVER kill the backend
 * PTY — otherwise running AIs would be cut off. So:
 * - On unmount the cleanup only `detach`es (drops the local output subscription)
 *   and disposes xterm; it never calls `handle.close()`. Killing a PTY is an
 *   explicit user action handled elsewhere (the terminal's close button).
 * - On mount, if a `sessionId` already exists for this cell (persisted by the
 *   caller via `onSessionId`), the view **re-attaches** to the still-running PTY
 *   — repainting its scrollback and resuming its output — instead of opening a
 *   fresh one. If the session is gone (was explicitly closed), it opens fresh.
 */
 import { useEffect, useRef } from "react";
@ -21,7 +32,11 @@ import { FitAddon } from "@xterm/addon-fit";
 import "@xterm/xterm/css/xterm.css";
 import { useGateways } from "@/app/di";
-import type { OpenTerminalOptions, TerminalHandle } from "@/ports";
+import type {
  OpenTerminalOptions,
  ReattachResult,
  TerminalHandle,
 } from "@/ports";
 interface TerminalViewProps {
  /** Working directory the shell opens in (typically the project root). */
@ -36,9 +51,35 @@ interface TerminalViewProps {
    options: OpenTerminalOptions,
    onData: (bytes: Uint8Array) => void,
  ) => Promise<TerminalHandle>;
  /**
   * Optional re-attach opener. When provided together with a {@link sessionId},
   * the view re-binds to the existing live PTY instead of opening a new one.
   * When absent, falls back to the terminal gateway's `reattach`.
   */
  reattach?: (
    sessionId: string,
    onData: (bytes: Uint8Array) => void,
  ) => Promise<ReattachResult>;
  /**
   * Persisted session id for this cell, if a PTY is already running for it.
   * Drives the reattach-vs-open decision at mount.
   */
  sessionId?: string | null;
  /**
   * Called once a session is established (opened) so the caller can persist its
   * id for this cell and re-attach to it on the next mount. Not called on
   * reattach (the id is already known).
   */
  onSessionId?: (sessionId: string) => void;
 }
-export function TerminalView({ cwd, open }: TerminalViewProps) {
+export function TerminalView({
  cwd,
  open,
  reattach,
  sessionId,
  onSessionId,
 }: TerminalViewProps) {
  const { terminal } = useGateways();
  const containerRef = useRef<HTMLDivElement | null>(null);
@ -51,6 +92,12 @@ export function TerminalView({ cwd, open }: TerminalViewProps) {
  // so the correct opener is always captured at mount.
  const openRef = useRef(open);
  openRef.current = open;
  const reattachRef = useRef(reattach);
  reattachRef.current = reattach;
  const sessionIdRef = useRef(sessionId);
  sessionIdRef.current = sessionId;
  const onSessionIdRef = useRef(onSessionId);
  onSessionIdRef.current = onSessionId;
  const terminalRef = useRef(terminal);
  terminalRef.current = terminal;
@ -58,6 +105,7 @@ export function TerminalView({ cwd, open }: TerminalViewProps) {
    const container = containerRef.current;
    const tgw = terminalRef.current;
    const opener = openRef.current ?? tgw?.openTerminal.bind(tgw);
    const reattacher = reattachRef.current ?? tgw?.reattach.bind(tgw);
    if (!container || !opener) return;
    const term = new Terminal({
@ -94,15 +142,16 @@ export function TerminalView({ cwd, open }: TerminalViewProps) {
      else pending += data;
    });
-    opener(
+    const onData = (bytes: Uint8Array) => {
      { cwd, rows: term.rows, cols: term.cols },
      (bytes) => {
      if (!disposed) term.write(bytes);
-      },
+    };
-    )
+
-      .then((h) => {
+    // Adopt a freshly-established handle: flush buffered keystrokes. If the view
    // was disposed before the promise resolved, just detach (NEVER close — the
    // PTY must survive a transient mount/unmount).
    const adopt = (h: TerminalHandle) => {
      if (disposed) {
-          void h.close();
+        h.detach();
        return;
      }
      handle = h;
@ -110,14 +159,48 @@ export function TerminalView({ cwd, open }: TerminalViewProps) {
        void h.write(encoder.encode(pending));
        pending = "";
      }
-      })
+    };
-      .catch((e: unknown) => {
+
    const onOpenError = (e: unknown) => {
      if (!disposed) {
        term.write(
          `\r\n\x1b[31mfailed to open terminal: ${describe(e)}\x1b[0m\r\n`,
        );
      }
    };
    // Re-attach to an existing live PTY when this cell already has a session;
    // otherwise open a fresh one and persist its id for the next mount.
    const existingSession = sessionIdRef.current;
    if (existingSession && reattacher) {
      reattacher(existingSession, onData)
        .then(({ handle: h, scrollback }) => {
          if (disposed) {
            h.detach();
            return;
          }
          if (scrollback.length > 0) term.write(scrollback);
          adopt(h);
        })
        .catch(() => {
          // The session is gone (explicitly closed / exited): fall back to a
          // fresh terminal so the cell still works.
          if (disposed) return;
          opener({ cwd, rows: term.rows, cols: term.cols }, onData)
            .then((h) => {
              onSessionIdRef.current?.(h.sessionId);
              adopt(h);
            })
            .catch(onOpenError);
        });
    } else {
      opener({ cwd, rows: term.rows, cols: term.cols }, onData)
        .then((h) => {
          onSessionIdRef.current?.(h.sessionId);
          adopt(h);
        })
        .catch(onOpenError);
    }
    // Refit + propagate size to the PTY on container resize.
    const ro = new ResizeObserver(() => {
@ -134,7 +217,10 @@ export function TerminalView({ cwd, open }: TerminalViewProps) {
      disposed = true;
      ro.disconnect();
      onKey.dispose();
-      if (handle) void handle.close();
+      // DETACH, never close: tearing the view down (navigation / layout change)
      // must leave the backend PTY running so the AI isn't cut off. Killing the
      // PTY is an explicit user action handled elsewhere.
      if (handle) handle.detach();
      term.dispose();
    };
    // Only re-open on cwd change (or mount). The opener is read from a ref, and
--- a/frontend/src/ports/index.ts
+++ b/frontend/src/ports/index.ts
@ -25,6 +25,8 @@ import type {
  LayoutTree,
  Project,
  ProfileAvailability,
  Skill,
  SkillScope,
  Template,
  Unsubscribe,
 } from "@/domain";
@ -72,6 +74,16 @@ export interface AgentGateway {
    options: OpenTerminalOptions,
    onData: (bytes: Uint8Array) => void,
  ): Promise<TerminalHandle>;
  /**
   * Re-attaches to an agent's already-running PTY (same backend mechanism as
   * {@link TerminalGateway.reattach}; agent sessions share the session-based
   * terminal commands). Used when an agent cell's view re-mounts after a
   * navigation/layout change, so the agent is never killed.
   */
  reattach(
    sessionId: string,
    onData: (bytes: Uint8Array) => void,
  ): Promise<ReattachResult>;
 }
 /** Options for opening a terminal. */
@ -98,7 +110,19 @@ export interface TerminalHandle {
  write(data: Uint8Array): Promise<void>;
  /** Resizes the PTY. */
  resize(rows: number, cols: number): Promise<void>;
-  /** Kills the PTY and stops the output stream. */
+  /**
   * Detaches the **view** from the PTY without killing it: stops the local
   * output subscription so a torn-down view (navigation / layout change) stops
   * receiving bytes, while the backend PTY keeps running. The session can later
   * be re-attached via {@link TerminalGateway.reattach}.
   *
   * This is the lifecycle the view's cleanup must use — never {@link close}.
   */
  detach(): void;
  /**
   * Kills the PTY and stops the output stream. Reserved for an **explicit** user
   * action (closing the terminal); navigation must never call this.
   */
  close(): Promise<void>;
 }
@ -115,6 +139,27 @@ export interface TerminalGateway {
    options: OpenTerminalOptions,
    onData: (bytes: Uint8Array) => void,
  ): Promise<TerminalHandle>;
  /**
   * Re-attaches to an **already-running** PTY identified by `sessionId` (after a
   * view was torn down by navigation/layout change). Returns the live handle and
   * the retained scrollback, which the caller repaints into xterm before the new
   * output stream (`onData`) starts delivering subsequent bytes. Does NOT
   * re-spawn the process.
   *
   * Rejects if the session is no longer alive (the caller then opens fresh).
   */
  reattach(
    sessionId: string,
    onData: (bytes: Uint8Array) => void,
  ): Promise<ReattachResult>;
 }
 /** The outcome of {@link TerminalGateway.reattach}. */
 export interface ReattachResult {
  /** The live terminal handle for the re-attached session. */
  handle: TerminalHandle;
  /** The retained scrollback bytes to repaint before the live stream resumes. */
  scrollback: Uint8Array;
 }
 /** Projects: create/open/close/list (L2). */
@ -217,6 +262,53 @@ export interface TemplateGateway {
  ): Promise<{ synced: boolean; version: number | null }>;
 }
 /** Input for {@link SkillGateway.createSkill}. */
 export interface CreateSkillInput {
  /** Owning project (resolved to a root; ignored on disk for `global`). */
  projectId: string;
  name: string;
  content: string;
  scope: SkillScope;
 }
 /**
 * Skills (L12): CRUD for reusable, model-agnostic workflows in either scope, and
 * agent↔skill assignment. Assigned skills are injected into the agent's
 * convention file at activation (handled backend-side, ARCHITECTURE §14.2).
 */
 export interface SkillGateway {
  /** Lists the skills in one scope for the given project. */
  listSkills(projectId: string, scope: SkillScope): Promise<Skill[]>;
  /** Creates a new skill; returns the created skill. */
  createSkill(input: CreateSkillInput): Promise<Skill>;
  /** Updates a skill's content; returns the updated skill. */
  updateSkill(
    projectId: string,
    scope: SkillScope,
    skillId: string,
    content: string,
  ): Promise<Skill>;
  /** Deletes a skill by id from its scope's store. */
  deleteSkill(
    projectId: string,
    scope: SkillScope,
    skillId: string,
  ): Promise<void>;
  /** Assigns a skill to an agent (idempotent). */
  assignSkill(
    projectId: string,
    agentId: string,
    skillId: string,
    scope: SkillScope,
  ): Promise<void>;
  /** Unassigns a skill from an agent (idempotent). */
  unassignSkill(
    projectId: string,
    agentId: string,
    skillId: string,
  ): Promise<void>;
 }
 /**
 * AI profiles & first-run (L5). Drives the first-run wizard and profile
 * management: the pre-filled reference catalogue, detection of installed CLIs,
@ -253,4 +345,5 @@ export interface Gateways {
  remote: RemoteGateway;
  profile: ProfileGateway;
  template: TemplateGateway;
  skill: SkillGateway;
 }
Author	SHA1	Message	Date
Blomios	3be55795a6	fix: fix some displays and features	2026-06-06 17:06:45 +02:00
Blomios	2332b7f815	fix: fix some ui displays and features miss implemented	2026-06-06 16:15:19 +02:00
Blomios	9736c42424	merge: isolate agent cwd in .ideai/run/<id> (convention-file collision fix, §14.1)	2026-06-06 12:26:02 +02:00
Blomios	0638ce7c98	merge: terminal lifecycle decoupling (PTY survives layout/tab switch)	2026-06-06 12:25:54 +02:00
Blomios	0660f52e2b	fix(terminals): decouple PTY lifecycle from view lifecycle (no kill on navigation) Navigating (layout/tab switch) tore the xterm view down and called handle.close(), killing the backend PTY and cutting off running AIs. Now the view's cleanup only detaches; only an explicit user action kills a PTY. Backend: - PortablePtyAdapter: per-session scrollback ring buffer (~100KB, most recent) + re-subscribable fan-out broadcast replacing the single-take output_rx. Reader thread feeds both the ring buffer and current subscribers; on EOF it closes subscribers (streams end) while keeping scrollback for late re-attach. - PtyPort: new scrollback() method; subscribe_output is now re-subscribable (all impls + test fakes updated). - reattach_terminal IPC command: returns scrollback and re-wires a fresh output channel on the live session without re-spawning. - CloseRequested hook kills all live PTYs cleanly on app shutdown. - TerminalSessions::handles() to enumerate live sessions at shutdown. Frontend: - TerminalHandle.detach(); TerminalGateway/AgentGateway.reattach() + mocks. - TerminalView cleanup detaches (never close); on mount it re-attaches to a persisted session (repainting scrollback) instead of opening a new PTY. - LayoutGrid persists the cell's session id via setSession; AgentsPanel tracks per-agent session ids — both drive reattach-vs-open. Tests: ring buffer bounds to 100KB keeping newest bytes; scrollback retained; re-subscription delivers post-reattach output; TerminalView detaches (not closes) on unmount and reattaches with a known session; mock detach/reattach. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-06-06 12:24:48 +02:00
Blomios	33edbad713	feat(agent): isolate agent cwd in .ideai/run/<id> to kill convention-file collisions ARCHITECTURE §14.1: an agent's PTY cwd is now its own `<project_root>/.ideai/run/<agent-id>/` directory, never the project root, so N agents of the same profile no longer collide on a single conventional file (CLAUDE.md/AGENTS.md/...). - profile: cwd_template is now "{agentRunDir}" (built-in catalogue + docs). - runtime: resolve_cwd substitutes {agentRunDir} (legacy {projectRoot} kept). - LaunchAgent: computes + creates the run dir via FileSystem::create_dir_all, passes it as the cwd base to the pure prepare_invocation. Contract chosen: pass run_dir as the `cwd` argument (no PreparedContext change) — keeps prepare_invocation pure, I/O stays in the use case. - convention file is generated by IdeA inside the run dir via a pure compose_convention_file(project_root, agent_md): absolute project-root header + agent persona (extensible for skills, §14.2). - .gitignore: ignore .ideai/run/. - run-dir cleanup left as a TODO (FileSystem port exposes no delete). Tests: anti-collision (2 agents -> 2 distinct cwd, 2 distinct convention files, none at root), run-dir creation order, composed convention file; pure unit tests for agent_run_dir + compose_convention_file; runtime {agentRunDir} substitution. cargo test --workspace + clippy -D warnings green. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-06-06 12:18:14 +02:00