tmux: keep the fresh-context intent until a turn completes (#352)

[ziomik]

The bug

context_restart launches a fresh REPL, but the fresh-context intent is spent on that single launch. If the fresh REPL wedges before completing its first turn, the inflight watchdog force-restarts it at 600s — well past the 180s FRESH_CONTEXT_RESPAWN_GRACE_SEC window — and the relaunch resolves --continue, which resumes the newest transcript on disk: the pre-restart one. The context window is never actually cleared, so the agent restarts into the same full window and the loop repeats every ~10 minutes.

Observed on engineer (daemon journald, tmux[engineer]):

claude_cmd_built mode=fresh force_fresh=True prior_transcript=True      # restart, binds 5543e594
inflight head aged 613.0s > 600.0s, transcript quiet + REPL not idle
  — REPL stuck; scheduling force_restart
claude_cmd_built mode=continue force_fresh=False prior_transcript=True  # intent already spent
transcript path updated to .../59201989.jsonl                          # the PRE-restart transcript

The decisive detail: the fresh transcript (5543e594.jsonl) never exists on disk. The fresh REPL never completes a turn, so Claude Code never writes the file — which is also why the watchdog sees "transcript quiet" and force-restarts in the first place. So resuming by path alone would not have been enough either: it would point at a file that isn't there.

The fix

Two halves, each insufficient alone.

1. The fresh intent is sticky. Armed on a force-fresh launch (same place the respawn grace is armed), disarmed only by a correlated post-fresh turn completion — the one event that proves the fresh REPL is alive and owns a real transcript. Until then every relaunch, force_restart included, builds mode=fresh. Without this, half 2 is still defeated: the intent is burned by the first force_restart.

2. No blind --continue. --continue means "newest transcript in this project dir", a guess that goes wrong exactly when it matters. When the session has a bound transcript path (SessionStart hook or discovery), resume it explicitly with --resume <path>. When that bound transcript does not exist on disk, launch fresh rather than resume somebody else's conversation. When no bind was ever recorded (e.g. first launch after a daemon restart), the legacy --continue behavior is unchanged.

claude_cmd_built now also logs fresh_intent_pending and resume_path.

Tests

New TestStickyFreshContextIntent — written first, and the three behavioral ones failed with exactly the production symptom ('--continue' in ['claude', '--continue', ...]) before the fix:

• fresh intent survives grace expiry until a turn completes (the reproduction: past the grace, at watchdog time)
• a completed post-fresh turn releases the intent (no permanent-fresh regression)
• a warm relaunch resumes the bound transcript by path
• a relaunch is fresh when the bound transcript does not exist — the decisive case
• an unknown active transcript keeps legacy --continue

Green: 589 passed / 1 skipped across test_tmux_session, test_tmux_transcript, test_tmux_context_autorestart, test_tmux_context_realtime, test_codex_tmux_session, test_codex_tmux_transcript, test_tmux_container_isolation; plus 82 passed across test_daemon and test_agent_status. ruff check clean.

Not in this PR

Raising the watchdog quiet tolerance (600s → ~900s) for wake_reason=context_restart was discussed as a follow-up. Deliberately left out: it changes watchdog timing, and this PR is meant to be reviewable on the launch-mode logic alone.

Deploy note

This touches the session-manager force_restart path. Timing of the production rollout to be agreed before shipping.

🤖 Opened by Engineer