infrastructure/src/session/ : machinerie de process générique (paramétrable par
la commande = seam d'injection du fake CLI), adapters ClaudeSdkSession/
CodexExecSession avec parsing ISOLÉ par adapter (parse_event), factory
StructuredSessionFactory (routage par structured_adapter), FakeCli scriptable +
harnais de conformité Liskov assert_agent_session_contract.
Incarnation « un run par tour » (send relance claude -p / --resume <id>,
continuité via conversation_id — colle au pivot reprise B).
Tests : 41 contre le FAKE CLI (jamais le vrai claude/codex), workspace vert.
Points en attente des spikes S1/S2 (format réel) — n'impactent que parse_event :
- mapping JSON→ReplyEvent Claude (S1) et Codex (S2) sur schémas SUPPOSÉS ;
- Claude multi-blocs : parse_event ne garde que le 1er bloc (à corriger si Claude
émet plusieurs blocs/message — confirmer S1) ;
- flux sans Final : permissif côté adapter, l'erreur est gérée par send_blocking
(consommateur). À reconfirmer côté UI streaming (D4).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- LOT C2 (§14.5.3) : use cases de configuration des embedders déclaratifs
(List/Save/Delete + DescribeEmbedderEngines : modèles ONNX recommandés,
environnement local détecté, stratégies compilées). UI EmbedderSettings.
- LOT C3 (§14.5.5) : suggestion contextuelle best-effort à l'activation quand la
mémoire dépasse le budget de recall sans embedder configuré (event
EmbedderSuggested, anti-spam 1×/session, « ne plus demander »).
- Contexte projet partagé .ideai/CONTEXT.md (model-agnostic) injecté à tous les
agents/profils au lancement, avant la persona. UI ProjectContextPanel.
Tests : backend workspace vert (0 échec) ; frontend 306/306.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Remplace les StubEmbedder pour les stratégies localServer/api/localOnnx par de
vrais moteurs, chacun derrière une feature cargo off-by-default — la posture
fondatrice « rien d'imposé, zéro dépendance » (défaut none → rappel naïf) reste
byte-for-byte inchangée.
C1a (feature vector-http, reqwest rustls optional):
- HttpEmbedder couvrant localServer (Ollama/llama.cpp) et api (OpenAI/Voyage…),
payload OpenAI-compatible /v1/embeddings, ordre restauré par index, bearer
token lu via env var (jamais en clair), timeout client 30s.
- detect_ollama() pour la détection de l'existant (C3).
C1b (feature vector-onnx, fastembed v5 optional):
- OnnxEmbedder en-process (e5-small, dim 384), init paresseuse + spawn_blocking,
cache modèle sous <app_data>/embedders/onnx — aucun download au build ni au
first-run, uniquement à la demande au 1er embed.
- Catalogue RECOMMENDED_ONNX_MODELS + ONNX_CACHE_SUBDIR + onnx_model_is_cached
exposés (sans feature) pour la config (C2) et la popup (C3).
embedder_from_profile(profile, onnx_cache_dir) dispatche feature-gated ; sans la
feature, retombe sur StubEmbedder (Unsupported) → fallback naïf via
AdaptiveMemoryRecall. Composition root (build_memory_recall) propage le cache dir.
Tests: 10 HTTP + 6 ONNX (dont 2 #[ignore] download réel) + 26 vectoriels, verts
en défaut, --features vector-http et --features vector-onnx.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Base de connaissance persistante par projet, indépendante de tout modèle/CLU
et de git. Cadrage archi en §14.5 (ARCHITECTURE.md), cycle Archi→Dev→Test.
LOT A — étage 1 (.md, source de vérité)
- domaine: entité Memory (+ MemorySlug, MemoryType, MemoryFrontmatter,
MemoryLink, MemoryIndexEntry), liens [[slug]], index MEMORY.md dérivé
- port MemoryStore + MemoryError, adapter FsMemoryStore (.ideai/memory/)
- application: 7 use cases (Create/Update/List/Get/Delete/ReadIndex/
ResolveLinks), From<MemoryError> for AppError
- app-tauri: commandes + DTO, events MemorySaved/MemoryDeleted
- suppression de la variante morte DomainError::MalformedFrontmatter
LOT B — rappel adaptatif (étage 1)
- port MemoryRecall + MemoryQuery, adapter NaiveMemoryRecall (troncature
au budget de tokens, court-circuit budget-0), use case RecallMemory
LOT C — étage 2 vectoriel (structure complète, zéro dépendance lourde)
- port Embedder + EmbedderError, profils déclaratifs EmbedderProfile/
EmbedderStrategy (embedder.json)
- VectorMemoryRecall (cosinus, cache .ideai/memory/.index/ gitignoré)
- AdaptiveMemoryRecall (bascule pure should_use_vector), défaut none
- HashEmbedder (déterministe, tests), StubEmbedder (onnx/server/api)
Tests: 57 binaires verts, build + clippy --workspace sans warning.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Permet de recharger la conversation CLI précédente de chaque cellule à la
réouverture du projet, de façon universelle (indépendant du modèle/CLI).
- profil AgentRuntime: bloc déclaratif optionnel `session { assignFlag, resumeFlag }`
- LeafCell: `conversationId` (persistant, distinct du SessionId PTY) + `agentWasRunning`
- runtime: SessionPlan (None/Assign/Resume) + composition pure des args
- LaunchAgent: décide Assign vs Resume, génère l'UUID, remonte l'id assigné
(persistance par l'appelant via setCellConversation — découplage SRP)
- close: SnapshotRunningAgents fige `agentWasRunning` avant le kill-all
(statut clot/en cours universel, sans parsing CLI)
- SessionInspector: port optionnel best-effort + adapter ClaudeTranscriptInspector
- popup de reprise par cellule (statut + sujet/tokens si dispo), intercalée
avant le Resume auto, jamais sur le chemin reattach
fix(terminals): sérialise les écritures PTY (file FIFO par handle) — corrige
les caractères mélangés/accents dus au réordonnancement des invoke Tauri concurrents
fix(layout): l'opération `move` préservait mal les champs du leaf (perdait `agent`)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Each open/reattach spawned a pump thread feeding pty_bridge[session]. The
broadcast hub added a subscriber per attach without dropping the previous one,
and the stale pump thread kept delivering (its send_output still succeeded
against the new channel) — so every byte was delivered twice (more after
further re-attaches). Accumulated on each tab/layout/agent switch; window
resize made it glaring.
- Broadcast::subscribe() is now single-consumer: a new subscription supersedes
the previous one, ending the old pump thread's stream.
- PtyBridge gains a per-session generation; register() returns it and
unregister_if(session, gen) only removes when still current, so a dying
superseded thread can't tear down the channel that replaced it.
- All three attach sites (open/reattach/agent launch) use unregister_if.
Tests: 6 added (generation guard + single-consumer broadcast). Workspace green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
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>
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>
