fix(domain,application,infrastructure,app-tauri,frontend): conversation_id redevient un pivot logique IdeA/log/handoff, reprise provider réservée à la sélection explicite — lifecycle: LaunchAgentInput::conversation_id documenté comme sélection explicite (Some = reprise native si disponible sinon handoff portable, None = conversation provider neuve), plus de copie implicite du pivot logique de la leaf au reboot ; fallback silencieux: provider_resume_fallback résolu via resolve_handoff et câblé jusqu'aux adaptateurs Claude/Codex/OpenCode via AgentSessionFactory::create resume_fallback_preamble + with_resume_fallback_preamble — le préambule reste dormant si la session native répond et n'est injecté que si la session provider s'avère introuvable/périmée (fake stale-then-fresh en test, relance fraîche sans erreur utilisateur) ; app-tauri launch_agent: contrat documenté, l'activation ordinaire d'une cellule doit envoyer None même si la leaf conserve un pivot logique d'avant reboot ; frontend CustomAgentChatView: freshOnOpen réinitialise l'état conversation complet (scrollback, view state, conversationId publié, bannière reprise, pagination historique) quand le runtime structuré n'existe plus au remontage — pas de fausse « Conversation reprise » après reboot ; tests: structured_launch_d3 étendu, structured_session_d0, agent_lifecycle structured, CustomAgentChatView freshOnOpen — #372 (QA verte: cargo check --tests -p domain -p application -p infrastructure, cargo test -p application --test structured_launch_d3, cargo test -p domain --test structured_session_d0, cargo test -p application --test agent_lifecycle structured, npm test -- CustomAgentChatView, npm run typecheck, git diff --check ; fichiers .ideai/* et scripts/ hors périmètre ; aucun push)

This commit is contained in:
2026-09-17 19:43:14 +02:00
parent 0914a77d9f
commit 07488769d3
19 changed files with 483 additions and 57 deletions

View File

@ -25,7 +25,10 @@ use domain::ports::{
use domain::sandbox::{SandboxEnforcer, SandboxPlan};
use domain::SessionId;
use super::process::{run_turn_cancellable, SpawnLine, TurnCancel};
use super::process::{
is_missing_provider_session, resume_fallback_prompt, run_turn_cancellable, SpawnLine,
TurnCancel,
};
/// Résultat du parsing d'une ligne : **zéro ou plusieurs** événements à émettre et/ou
/// un `session_id` capté (init/result). Permet à [`parse_event`] de rester **pure**
@ -300,6 +303,8 @@ pub struct ClaudeSdkSession {
cwd: String,
/// Id de conversation **du moteur** Claude, capté au premier tour, `None` avant.
conversation_id: Mutex<Option<String>>,
/// Portable handoff injected only if the seeded provider session is stale.
resume_fallback_preamble: Option<String>,
/// Plan de sandbox OS **par lancement** (lot LP4-4), porté dans chaque
/// [`SpawnLine`]. `None` ⇒ aucun sandboxing (drain async natif).
sandbox: Option<SandboxPlan>,
@ -330,12 +335,20 @@ impl ClaudeSdkSession {
command: command.into(),
cwd: cwd.into(),
conversation_id: Mutex::new(seed_conversation_id),
resume_fallback_preamble: None,
sandbox,
sandbox_enforcer,
turn_cancel: TurnCancel::new(),
}
}
/// Adds the portable context used by the one-shot stale-resume fallback.
#[must_use]
pub fn with_resume_fallback_preamble(mut self, preamble: Option<&str>) -> Self {
self.resume_fallback_preamble = preamble.map(str::to_owned);
self
}
/// Compose la ligne de commande d'un tour selon l'état de conversation.
///
/// Format RÉEL vérifié 2026-06-09 :
@ -408,6 +421,7 @@ impl ClaudeSdkSession {
prompt: &str,
tap: Option<std::sync::mpsc::Sender<ReplyEvent>>,
) -> Result<ReplyStream, AgentSessionError> {
let had_native_resume = self.conversation_id().is_some();
let spec = self.build_spawn_line(prompt);
let (line_tap, parser_thread) = tap.map_or((None, None), |event_tap| {
let (line_tx, line_rx) = std::sync::mpsc::channel::<String>();
@ -424,14 +438,27 @@ impl ClaudeSdkSession {
});
(Some(line_tx), Some(parser))
});
let raw_result = run_turn_cancellable(
let mut raw_result = run_turn_cancellable(
&spec,
None,
self.sandbox_enforcer.as_ref(),
line_tap,
line_tap.clone(),
&self.turn_cancel,
)
.await;
if had_native_resume && raw_result.as_ref().is_err_and(is_missing_provider_session) {
*self.conversation_id.lock().expect("mutex sain") = None;
let fallback_prompt =
resume_fallback_prompt(self.resume_fallback_preamble.as_deref(), prompt);
raw_result = run_turn_cancellable(
&self.build_spawn_line(&fallback_prompt),
None,
self.sandbox_enforcer.as_ref(),
line_tap,
&self.turn_cancel,
)
.await;
}
if let Some(parser) = parser_thread {
let _ = parser.join();
}

View File

@ -22,7 +22,10 @@ use domain::ports::{
use domain::sandbox::{SandboxEnforcer, SandboxPlan};
use domain::SessionId;
use super::process::{run_turn_cancellable, SpawnLine, TurnCancel};
use super::process::{
is_missing_provider_session, resume_fallback_prompt, run_turn_cancellable, SpawnLine,
TurnCancel,
};
/// Résultat du parsing d'une ligne Codex : **zéro ou plusieurs** événements et/ou un
/// id de conversation Codex capté. Miroir de `claude::ParsedLine` (vecteur d'events
@ -401,6 +404,8 @@ pub struct CodexExecSession {
env: Vec<(String, String)>,
/// Id de conversation **du moteur** Codex, capté au premier tour, `None` avant.
conversation_id: Mutex<Option<String>>,
/// Portable handoff injected only if the seeded provider session is stale.
resume_fallback_preamble: Option<String>,
/// Plan de sandbox OS **par lancement** (lot LP4-4), porté dans chaque
/// [`SpawnLine`]. `None` ⇒ aucun sandboxing (drain async natif).
sandbox: Option<SandboxPlan>,
@ -503,12 +508,20 @@ impl CodexExecSession {
model_reasoning_effort,
env,
conversation_id: Mutex::new(seed_conversation_id),
resume_fallback_preamble: None,
sandbox,
sandbox_enforcer,
turn_cancel: TurnCancel::new(),
}
}
/// Adds the portable context used by the one-shot stale-resume fallback.
#[must_use]
pub fn with_resume_fallback_preamble(mut self, preamble: Option<&str>) -> Self {
self.resume_fallback_preamble = preamble.map(str::to_owned);
self
}
/// Compose la ligne de commande d'un tour.
///
/// Format RÉEL vérifié 2026-06-10 (codex 0.137.0) :
@ -641,6 +654,7 @@ impl CodexExecSession {
prompt: &str,
tap: Option<std::sync::mpsc::Sender<ReplyEvent>>,
) -> Result<ReplyStream, AgentSessionError> {
let had_native_resume = self.conversation_id().is_some();
let spec = self.build_spawn_line(prompt);
let has_tap = tap.is_some();
let (line_tap, parser_thread) = tap.map_or((None, None), |event_tap| {
@ -670,14 +684,27 @@ impl CodexExecSession {
});
(Some(line_tx), Some(parser))
});
let raw_result = run_turn_cancellable(
let mut raw_result = run_turn_cancellable(
&spec,
None,
self.sandbox_enforcer.as_ref(),
line_tap,
line_tap.clone(),
&self.turn_cancel,
)
.await;
if had_native_resume && raw_result.as_ref().is_err_and(is_missing_provider_session) {
*self.conversation_id.lock().expect("mutex sain") = None;
let fallback_prompt =
resume_fallback_prompt(self.resume_fallback_preamble.as_deref(), prompt);
raw_result = run_turn_cancellable(
&self.build_spawn_line(&fallback_prompt),
None,
self.sandbox_enforcer.as_ref(),
line_tap,
&self.turn_cancel,
)
.await;
}
if let Some(parser) = parser_thread {
let _ = parser.join();
}

View File

@ -142,6 +142,7 @@ impl AgentSessionFactory for StructuredSessionFactory {
ctx: &PreparedContext,
cwd: &ProjectPath,
session: &SessionPlan,
resume_fallback_preamble: Option<&str>,
requester: Option<&str>,
env: &[(String, String)],
sandbox: Option<&SandboxPlan>,
@ -175,9 +176,10 @@ impl AgentSessionFactory for StructuredSessionFactory {
// injection supplémentaire n'incombe ici en mode structuré (la CLI lit son
// fichier conventionnel — CLAUDE.md / AGENTS.md — depuis le cwd).
let session: Arc<dyn AgentSession> = match adapter {
StructuredAdapter::Claude => Arc::new(ClaudeSdkSession::new(
id, command, cwd, seed, plan, enforcer,
)),
StructuredAdapter::Claude => Arc::new(
ClaudeSdkSession::new(id, command, cwd, seed, plan, enforcer)
.with_resume_fallback_preamble(resume_fallback_preamble),
),
StructuredAdapter::Codex => {
let policy = match structured_policy {
Some(StructuredProviderLaunchPolicy::Codex {
@ -195,32 +197,38 @@ impl AgentSessionFactory for StructuredSessionFactory {
));
}
};
Arc::new(CodexExecSession::new_with_policy_and_overrides(
Arc::new(
CodexExecSession::new_with_policy_and_overrides(
id,
command,
cwd,
seed,
policy.0,
policy.1,
policy.2,
profile.model.clone(),
profile.model_reasoning_effort.clone(),
env.to_vec(),
plan,
enforcer,
)
.with_resume_fallback_preamble(resume_fallback_preamble),
)
}
StructuredAdapter::OpenCode => Arc::new(
OpenCodeSession::new(
id,
command,
cwd,
profile.command.clone(),
profile.args.clone(),
cwd.clone(),
ctx.project_root.clone(),
seed,
policy.0,
policy.1,
policy.2,
profile.model.clone(),
profile.model_reasoning_effort.clone(),
env.to_vec(),
plan,
enforcer,
))
}
StructuredAdapter::OpenCode => Arc::new(OpenCodeSession::new(
id,
profile.command.clone(),
profile.args.clone(),
cwd.clone(),
ctx.project_root.clone(),
seed,
env.to_vec(),
plan,
enforcer,
)?),
)?
.with_resume_fallback_preamble(resume_fallback_preamble),
),
StructuredAdapter::OpenAiCompatible => {
let config = profile.chat_http.clone().ok_or_else(|| {
AgentSessionError::Start(format!(

View File

@ -577,6 +577,7 @@ mod tests {
&cwd(),
&SessionPlan::None,
None,
None,
&[],
None,
None,
@ -601,6 +602,7 @@ mod tests {
&cwd(),
&SessionPlan::None,
None,
None,
&[],
None,
Some(&codex_policy),
@ -634,6 +636,7 @@ mod tests {
&temp_cwd("factory-openai"),
&SessionPlan::None,
None,
None,
&[],
None,
None,
@ -672,6 +675,7 @@ mod tests {
&cwd(),
&SessionPlan::None,
None,
None,
&[],
None,
None,
@ -723,6 +727,7 @@ mod tests {
&cwd(),
&SessionPlan::None,
None,
None,
&[],
None,
Some(&policy),
@ -762,6 +767,7 @@ mod tests {
conversation_id: "repris-42".to_owned(),
},
None,
None,
&[],
None,
None,
@ -916,6 +922,89 @@ mod tests {
(bin.to_string_lossy().into_owned(), argv)
}
/// Fake provider that rejects its first (resume) invocation as stale, then
/// succeeds as a fresh Codex conversation on the retry.
fn make_stale_then_fresh_codex_fake() -> (String, std::path::PathBuf) {
use std::io::Write as _;
use std::sync::atomic::{AtomicU64, Ordering};
static C: AtomicU64 = AtomicU64::new(0);
let n = C.fetch_add(1, Ordering::Relaxed);
let dir = std::env::current_dir()
.expect("cwd")
.join("target")
.join("test-fakes")
.join("session");
std::fs::create_dir_all(&dir).expect("create stale fake dir");
let bin = dir.join(format!("idea-stale-cli-{}-{n}", std::process::id()));
let argv = dir.join(format!("idea-stale-argv-{}-{n}", std::process::id()));
let marker = dir.join(format!("idea-stale-marker-{}-{n}", std::process::id()));
let script = format!(
"#!/bin/sh\nfor a in \"$@\"; do printf '%s\\n' \"$a\" >> '{}'; done\n\
if [ ! -f '{}' ]; then touch '{}'; printf '%s\\n' 'session not found' >&2; exit 1; fi\n\
printf '%s\\n' '{{\"type\":\"thread.started\",\"thread_id\":\"fresh-thread\"}}'\n\
printf '%s\\n' '{{\"type\":\"item.completed\",\"item\":{{\"id\":\"i0\",\"type\":\"agent_message\",\"text\":\"fallback ok\"}}}}'\n",
argv.display(),
marker.display(),
marker.display(),
);
let mut file = std::fs::File::create(&bin).expect("create stale fake");
file.write_all(script.as_bytes()).expect("write stale fake");
file.sync_all().expect("sync stale fake");
#[cfg(unix)]
{
use std::os::unix::fs::PermissionsExt;
let mut permissions = std::fs::metadata(&bin).unwrap().permissions();
permissions.set_mode(0o755);
std::fs::set_permissions(&bin, permissions).unwrap();
}
super::conformance::wait_until_executable(&bin);
(bin.to_string_lossy().into_owned(), argv)
}
#[tokio::test]
async fn stale_codex_resume_retries_fresh_with_handoff_without_surface_error() {
let factory = StructuredSessionFactory::new();
let (command, argv) = make_stale_then_fresh_codex_fake();
let profile = structured_profile(StructuredAdapter::Codex, &command);
let policy = StructuredProviderLaunchPolicy::Codex {
sandbox_mode: "read-only".to_owned(),
writable_roots: Vec::new(),
network_access: false,
};
let session = factory
.start(
&profile,
&prepared_ctx(),
&cwd(),
&SessionPlan::Resume {
conversation_id: "stale-thread".to_owned(),
},
Some("Résumé portable conservé."),
None,
&[],
None,
Some(&policy),
)
.await
.expect("session adapter starts");
let stream = session
.send("Continuer le travail")
.await
.expect("stale provider resume falls back silently");
assert!(stream.into_iter().any(|event| matches!(
event,
ReplyEvent::Final { content } if content == "fallback ok"
)));
assert_eq!(session.conversation_id().as_deref(), Some("fresh-thread"));
let recorded = std::fs::read_to_string(argv).expect("recorded argv");
assert!(recorded.contains("resume\nstale-thread"));
assert!(recorded.contains("Contexte borné de la conversation reprise"));
assert!(recorded.contains("Résumé portable conservé."));
assert!(recorded.contains("Continuer le travail"));
}
// ---- Claude parse_event : plusieurs blocs, robustesse ---------------
/// Messages successifs : chaque message produit ses événements ; ici un par
@ -1848,6 +1937,7 @@ mod tests {
conversation_id: "cx-resume".to_owned(),
},
None,
None,
&[],
None,
Some(&policy),
@ -1874,6 +1964,7 @@ mod tests {
conversation_id: "ignored-by-engine".to_owned(),
},
None,
None,
&[],
None,
None,
@ -1907,6 +1998,7 @@ mod tests {
&cwd(),
&SessionPlan::None,
None,
None,
&[],
None,
None,

View File

@ -23,7 +23,7 @@ use domain::ports::{
use domain::sandbox::{SandboxEnforcer, SandboxPlan};
use domain::SessionId;
use super::process::{TurnCancel, TurnLease};
use super::process::{is_missing_provider_session, resume_fallback_prompt, TurnCancel, TurnLease};
/// Un événement OpenCode parsé depuis stdout JSONL.
#[derive(Debug, Clone, PartialEq, Eq)]
@ -555,6 +555,8 @@ pub struct OpenCodeSession {
project_root: String,
env: Vec<(String, String)>,
engine_session_id: Mutex<Option<String>>,
/// Portable handoff injected only if the seeded provider session is stale.
resume_fallback_preamble: Option<String>,
turn_cancel: TurnCancel,
sandbox: Option<SandboxPlan>,
sandbox_enforcer: Option<Arc<dyn SandboxEnforcer>>,
@ -585,12 +587,20 @@ impl OpenCodeSession {
project_root: project_root.into(),
env,
engine_session_id: Mutex::new(seed),
resume_fallback_preamble: None,
turn_cancel: TurnCancel::new(),
sandbox,
sandbox_enforcer,
})
}
/// Adds the portable context used by the one-shot stale-resume fallback.
#[must_use]
pub fn with_resume_fallback_preamble(mut self, preamble: Option<&str>) -> Self {
self.resume_fallback_preamble = preamble.map(str::to_owned);
self
}
fn build_args(&self, prompt: &str, engine_session_id: Option<&str>) -> Vec<String> {
let mut args = self.prefix_args.clone();
args.extend(["run".to_owned(), "--format".to_owned(), "json".to_owned()]);
@ -663,9 +673,16 @@ impl OpenCodeSession {
Ok(events) => events,
Err(err)
if initial_session.is_some()
&& matches!(err, AgentSessionError::Decode(ref message) if opencode_missing_final(message)) =>
&& (matches!(err, AgentSessionError::Decode(ref message) if opencode_missing_final(message))
|| is_missing_provider_session(&err)) =>
{
self.run_once(prompt, tap, None, &turn).await?
*self
.engine_session_id
.lock()
.unwrap_or_else(std::sync::PoisonError::into_inner) = None;
let fallback_prompt =
resume_fallback_prompt(self.resume_fallback_preamble.as_deref(), prompt);
self.run_once(&fallback_prompt, tap, None, &turn).await?
}
Err(err) => return Err(err),
};

View File

@ -722,11 +722,59 @@ fn process_exit_message(command: &str, code: Option<i32>, stderr: &[u8]) -> Stri
}
}
/// Whether a provider rejected a native resume because its remote/local session
/// no longer exists. Kept deliberately narrow: authentication, permissions and
/// generic process failures must remain visible instead of being retried.
pub(super) fn is_missing_provider_session(error: &AgentSessionError) -> bool {
let AgentSessionError::Io(message) = error else {
return false;
};
let message = message.to_ascii_lowercase();
[
"session not found",
"thread not found",
"conversation not found",
"no session found",
"no thread found",
"no conversation found",
"unknown session",
"unknown thread",
"no rollout found with id",
]
.iter()
.any(|needle| message.contains(needle))
}
pub(super) fn resume_fallback_prompt(preamble: Option<&str>, prompt: &str) -> String {
preamble.map_or_else(
|| prompt.to_owned(),
|summary| format!(
"Contexte borné de la conversation reprise :\n\n{summary}\n\n---\n\nNouveau message utilisateur :\n{prompt}"
),
)
}
#[cfg(test)]
mod tests {
use super::*;
use std::time::Instant;
#[test]
fn missing_provider_session_classifier_is_narrow() {
assert!(is_missing_provider_session(&AgentSessionError::Io(
"codex exited with code 1: session not found".to_owned()
)));
assert!(is_missing_provider_session(&AgentSessionError::Io(
"no rollout found with id abc".to_owned()
)));
assert!(!is_missing_provider_session(&AgentSessionError::Io(
"permission denied".to_owned()
)));
assert!(!is_missing_provider_session(&AgentSessionError::Start(
"session not found".to_owned()
)));
}
#[tokio::test]
async fn run_turn_cancellable_kills_current_child_without_waiting_for_natural_exit() {
let spec = SpawnLine {

View File

@ -476,6 +476,7 @@ async fn structured_sandboxed_turn_preserves_conversation_id() {
&cwd,
&SessionPlan::None,
None,
None,
&[],
Some(&plan),
None,

View File

@ -496,6 +496,7 @@ impl AgentSessionFactory for BlockingReplyFactory {
_ctx: &PreparedContext,
_cwd: &ProjectPath,
_session: &SessionPlan,
_resume_fallback_preamble: Option<&str>,
_requester: Option<&str>,
_env: &[(String, String)],
_sandbox: Option<&domain::sandbox::SandboxPlan>,