upstream tracking: claude-code#39700 (stream-json hangs, never exits) NOT_PLANNED — codify post-result/stuck-after-tool_result watchdog as permanent

[nathanschram]

Summary

The _post_result_idle_watchdog + subcountdown + SIGTERM machinery (#333, #470) and the stuck-after-tool_result detector (#322) exist because Claude Code's --output-format stream-json intermittently hangs — the process stops producing output and never exits.

Surfaced by the v0.35.3 Claude Code workaround-retirement audit (umbrella tracking issue).

Upstream status (checked 2026-05-29)

anthropics/claude-code#39700 — "[BUG] --output-format stream-json intermittently hangs — process stops producing output and never exits" — CLOSED / NOT_PLANNED.

(Sibling: anthropics/claude-code#30333 — ResultMessage never emitted — also NOT_PLANNED.)

Implication: this is NOT a retirement

Upstream declined to fix #39700, so the watchdog/SIGTERM safety net is a permanent mitigation, not a temporary workaround. Actions:

1. Codify it as permanent in code comments + docs so a future cleanup doesn't mistake it for retire-able dead code.
2. Guard the safety net during tracking: unified predicate-based stall detector — supersede the whack-a-mole pattern across releases #527's refactor. tracking: unified predicate-based stall detector — supersede the whack-a-mole pattern across releases #527 ("unified predicate-based stall detector") refactors when Untether warns. Ensure that refactor — and any detector-cleanup it performs — does not drop the actual kill-the-hung-process path (_post_result_idle_watchdog → SIGTERM → SIGKILL). The detector decides messaging; the watchdog decides survival; they must stay decoupled.
3. Consider re-escalating upstream with fresh repros — the channelo/nsd post-result-hang sessions captured during v0.35.3 rc16/rc17 instrumentation (claude.post_result_idle.tick / subprocess.state.* logs) — in case Anthropic reconsiders.

Affected code

• src/untether/runners/claude.py — _post_result_idle_watchdog, post-result subcountdown, _detect_stuck_after_tool_result
• src/untether/runner.py — _classify_jsonl_event (engine-agnostic tool_result classifier)
• Cross-ref: tracking: unified predicate-based stall detector — supersede the whack-a-mole pattern across releases #527 (detector refactor), investigate: Claude CLI session stays alive for ~36 min after final result event (idle-but-alive UX gap) #333 / stall-message backoff once last_event_type=result (post-result idle is benign noise to the user) #470 / Claude Code stuck after MCP tool_result via Untether (upstream claude-code#39700) #322 (the shipped mitigations), background-task lifecycle: clear handle on terminal signal, not first tool_result #374 (bg-handle lifecycle)

Done when

• Watchdog/SIGTERM path documented as a permanent #39700 mitigation.
• tracking: unified predicate-based stall detector — supersede the whack-a-mole pattern across releases #527's design explicitly preserves the SIGTERM safety net (note or acceptance criterion added there).
• Decision recorded on whether to re-file upstream.

---

Part of the v0.35.3 Claude Code workaround-retirement audit (umbrella tracking issue).