[captcrock]

What

A reasoning model on a tier with max_output_tokens: 4096 could spend the whole cap thinking. The stream ended finish_reason: length with no text and no tool call, and the turn failed as empty_completion. The chat told the user to start a new session even though the context was about a tenth full. Nothing retried, so the next turns failed the same way. The window had no copy for the code, so a toast also called it an unrecognised error.

Daemon:

• stream_reply withdraws a reply that thought to the cap (assistant_retract, reason output_cap_thinking) and asks once more on its own counter. The new max_tokens is min(4× cap, max(32768, cap), room left in the window), sent only when that is at least twice the cap. provider_retry.max_retries: 0 turns this off. Later attempts of the same call keep the raised cap.
• If the raised request is refused before any output (a 402 balance reservation, a 400, an overflow), the first result is kept so the real cause gets reported. A refusal of a later stall or guard retry is reported as that call's own error.
• A turn that still has no answer fails as output_cap_thinking. The chat line names the config entry whose max_output_tokens applied, as presets.<preset>.<tier>, a backup's fallbacks[n], or the computer-use worker's preset.
• A non-cap empty reply stays empty_completion, with accurate copy.
• The "a new chat may help" sentence appears only when the prompt is at the compaction budget.
• A Stop that arrives just as an empty reply comes back ends the turn as stopped, not failed.

UI:

• Copy for both codes: a banner for output_cap_thinking, cleared once a later turn answers, live or replayed; a toast for empty_completion.
• Activity title and preview for the new retract reason.
• A reason-aware Working line.
• The OS notification uses the copy's title.

Class B entry in DECISIONS.md. The protocol change is additive: one new error code and one new retract reason.

Tests

• core/tests/test_thinking_cap_retry.py, 45 tests. With the retry switched off, 10 of them fail. Each fix that came out of review has its own test, and each of those fails without its fix.
• UI: notifications.test.ts, chat-retract.test.ts, os-notify.test.ts, and the new chat-pane-retry.test.ts, which mounts the real ChatPane. The new behaviour tests fail on the old code.
• Core:
  • 8450 passed and 1 failed, test_worktree_daemon::test_a_failed_checkout_moves_the_session_back_to_its_folder. It passes alone three times out of three; it's a timing flake under load. The baseline run on fe2810c3 had a different flake, test_workflow::test_stop_reaches_a_childs_shell, which also passes alone.
  • ruff, ruff format and mypy --strict are clean.
• UI: vitest 2737 passed, svelte-check and tsc clean, vite build succeeds, Playwright walk 26 passed, and the protocol fixtures are unchanged.
• Live against OpenRouter (deepseek/deepseek-v4.1-flash-1017), with this branch's daemon in a throwaway data dir and a tiny max_output_tokens. Three runs:
  • Cap 512, harder prompt: the first reply spent the cap thinking and was withdrawn with output_cap_thinking. The retry at 2,048 spent that too, and the turn failed with output_cap_thinking and the new text naming presets.<preset>.brain.
  • Cap 128, easier prompt: the first reply was withdrawn the same way. The retry at 512 answered correctly.
  • Cap 512, easier prompt: answered the first time, with no retry.

Not in this PR

These hit the same failure and are listed as follow-ups in the story: the delegate child, which still says "Worker finished with no summary.", the progress check, the validator review, distill, the Settings model probe, OpenRouter reasoning params, usage reasoning_tokens, and machine-aware copy for bridged sessions.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HUthoCrBcxKu4GVj429dwD