bug(sessions): the in-feed Thinking counter counts through laptop sleep and can stick for hours #963

Summary

After the laptop sleeps, an agent pane's in-feed work indicator can read Thinking · 20h03m. The counter is WorkIndicator painting Date.now() - turnStartedAt (useElapsedSeconds), and nothing in the session pipeline knows the machine slept. Reported by the user on 2026-09-12 against the Thinking counter specifically.

Expected

• A turn whose stream died while the machine was asleep does not keep showing a live counter after wake.
• A turn that genuinely continues after wake does not count the sleep as working time.
• A prompt sent after wake starts its own clock.

Actual (reproduced against the real adapters and reducers)

The reproduction drives the real ClaudeProxyAdapter / CodexResponsesAdapter and the renderer's real foldSemanticEvent + reduceStreamPhase, in the order useIpcSubscriptions runs them, with only the wall clock simulated. Timings come from a real recording (below).

Mechanism, by file:

• src/renderer/src/session-runtime/semantic/streamPhaseMachine.ts:63-72: turnStartedAt is set only while null and cleared only on an idle phase or a clean turn_completed.
• packages/claude-code-headless/src/proxy/ClaudeProxyAdapter.ts:1009-1022: stale-flow reap only on a new flow's first chunk; with the proxy on, screen idle is suppressed.
• packages/codex-headless/src/proxy/CodexResponsesAdapter.ts:983-1021: watchdog seals the turn but publishes no idle phase.
• src/main/index.ts:672-673: powerMonitor suspend/resume feeds only the performance monitor; no session code reacts to sleep.
• All activity clocks use Date.now(); the run journal's monotonicMs (performance.now()) also advanced through sleep in real recordings, so it cannot detect sleep either.

Evidence (real, on disk)

• Turn across a confirmed clamshell sleep, 2026-08-31 → 09-01: Claude transcript (prompt 20:57:09 PDT, last assistant 22:45:08, turn_duration 39,888,429 ms = 11.08 h written after wake), the matching Codex rollout (turn_aborted: wall span 11.62 h vs duration_ms 1.48 h), the Agent Code run journal (app alive through the sleep; wall and monotonic deltas equal), and /var/log/powermanagement sleep 23:43:59 → wake 08:01:40.
• Unclean shutdown, 2026-08-29 → 08-30: 20.7 h gap in a Claude transcript with a tool_use that never got a result; next app run reports app.prior_unclean_shutdown. (A cold relaunch starts every runtime empty, so this shape cannot itself produce the counter; sleep with the app running is the reproduced cause.)

Impact

Every agent pane after a laptop sleep; Dispatch rows, Close Old Agents and Close Idle Orchestration Agents read the same phase, so a stuck phase also shows the agent as working and keeps it from being judged idle.

Environment

macOS (darwin 25.6), Agent Code at d12cd347, Claude and Codex with the proxy enabled.

Reproduction test and plan: branch feat/agent-working-time, src/renderer/src/session-runtime/semantic/turnClockAcrossSleep.test.ts and docs/decomposition/agent-working-time.md.