[Bug]: Hardcoded 30-minute idle session reaper + 2-second force-kill causes context loss across all providers on desktop

[skabillium]

Summary

After a break of ~30–60 minutes, conversations lose all provider-side context. The UI still shows the full history, but the agent behaves like a fresh session and it doesn't remember prior turns, plans, or tool-call state.

Root Cause

This is caused by two hardcoded defaults in the server:

1. ProviderSessionReaper.ts kills idle sessions after exactly 30 minutes (DEFAULT_INACTIVITY_THRESHOLD_MS = 30 * 60 * 1000). This is an architecture-level behavior that affects all providers (Codex, OpenCode, Claude, Cursor).

2. CodexSessionRuntime.ts gives the Codex app-server process only 2 seconds (CODEX_APP_SERVER_FORCE_KILL_AFTER = "2 seconds") to shut down before SIGKILL. Codex often needs more time to flush its SQLite state DB to disk. When the process is hard-killed mid-flush, the thread file is corrupted or incomplete, so thread/resume fails and silently falls back to a brand-new thread/start.

Why it affects all providers

• Codex: thread/resume fails because the state DB was not flushed → falls back to thread/start → context is lost.
• OpenCode (ACP): The ACP session ID is not persisted after the reaper kills the session, so T3 calls session/new instead of session/resume on reconnect. See [Bug]: OpenCode (ACP) provider creates new session instead of resuming after app/system restart #2838 for detailed analysis.
• Claude: Similar to OpenCode — the resume cursor can point to a dead CLI session if the process was killed before the first turn was persisted (see [Bug]: Thread becomes permanently unusable if claudeAgent CLI session is killed before writing its first turn (dangling resume_cursor_json at turnCount=0) #2336).

Related issues

• [Bug]: t3code should resume same conversation after idle timeout ? #2378 — "t3code should resume same conversation after idle timeout ?" (OpenCode)
• [Bug]: OpenCode (ACP) provider creates new session instead of resuming after app/system restart #2838 — "OpenCode (ACP) provider creates new session instead of resuming after app/system restart"
• [Bug]: Codex provider drains plan credits while T3 Code is idle in background (Codex equivalent of #2191) #2720 — "Codex provider drains plan credits while T3 Code is idle in background" (reaper not killing properly due to stuck activeTurnId)
• [Bug]: Chat splits into a new session after a pause #2188 — "Chat splits into a new session after a pause" (Codex, closed)
• [Bug]: Loss of context when the app is restarted #2140 — "Loss of context when the app is restarted" (Codex, closed)
• [Bug]: Thread becomes permanently unusable if claudeAgent CLI session is killed before writing its first turn (dangling resume_cursor_json at turnCount=0) #2336 — "Thread becomes permanently unusable if claudeAgent CLI session is killed before writing its first turn"

Proposed fix

1. Make the idle threshold configurable per environment: 3 hours for desktop (where process lifetime is cheap), 30 minutes for web/shared environments.
2. Bump the force-kill grace period for Codex: from 2 seconds to 15 seconds, so the app-server has time to flush its state DB.
3. For ACP providers: persist the ACP session ID in resumeCursor so session/resume can be used on reconnect.

Environment

• Desktop app (macOS, Windows, Linux)
• All providers: Codex, OpenCode, Claude, Cursor