fix(early-stop): close the two surfaces #778 left reporting success

[ericleepi314]

#778 made headless report a non-success result subtype when the agent loop cut a run short. It named two gaps. This closes both.

A — the agent-server turn outcome was hardcoded to success

_AgentSession._run_turn emitted subtype="success" regardless of why the loop stopped, so on the TUI / VS Code path a guard-killed turn was indistinguishable from a completed one. Three consumers gate on that field: _maybe_continue_goal fed such a turn to the /goal judge as evidence of progress, _maybe_review_memories learned from it, and the cron loop rearmed on it.

It now derives the subtype from result.terminal.reason through the same EARLY_STOP_SUBTYPES map headless uses, so the two surfaces cannot drift.

The two remaining hardcoded successes in that function are UserPromptSubmit hook blocks, which never run the loop — left deliberately, consistent with excluding hook_stopped/stop_hook_prevented from the map (operator policy working as configured, not the harness cutting a run short).

/goal auto-continues rather than dying

Before the subtype was derived, an early-stopped turn arrived as "success" with [Max tool turns reached] as its text, was judged not-done, and the loop retried. Bailing out would have silently killed loops that used to recover on their own.

So the judge is skipped — there's no output to judge, and asking a model whether [Stopped: …] satisfies the goal only invites a wrong answer — and a synthetic continue is applied. That's also judge_goal's own fail-open verdict, so it's a path the loop already handles.

Deliberately still routed through apply_verdict rather than enqueuing directly: that call ticks turns_used and lets the goal's cap decide. Short-circuiting it would let a turn that keeps stopping early retry forever.

B — a degenerate turn fell through to completed

query.py nudges an assistant turn with no tool calls, no text, nothing in the outbox. Once the nudge budget was spent the whole block was skipped — including the detection — so a still-degenerate turn reached Terminal(reason="completed") and was reported as a clean success with an empty result.

The detection is hoisted into an is_degenerate flag and a new Terminal(reason="empty_response") fires when the budget is spent and it's still degenerate, with an explanation surfaced as response_text.

Two correctness fixes fell out of building it:

• spoke_via_outbox read the session-lifetime outbox, created once per session and never cleared — so one AskUserQuestion / Brief / SendUserMessage anywhere in a session permanently disabled the check for every later prompt, and the silent empty success came straight back. Now scoped to the current query and to SendUserMessage, the only tool whose entry becomes response_text (an AskUserQuestion entry suppressed the check while contributing nothing).
• The two nudge arms shared one budget, so continuation nudges could spend it before the first empty turn arrived — the empty turn then got no retry and the run hard-failed without the round trip that recovers it (3/89 terminal-bench trials). Separate counters now.

The taxonomy change

TerminalReason gains empty_response, which TS has no counterpart for — the arm that detects it is itself port-only (TS's continuation nudge gates on non-empty text, so it never sees the case).

The parity test moves from set-equality against the TS snapshot to "TS is a subset, and every extra is declared in PYTHON_ONLY_TERMINAL_REASONS". Both teeth that matter are kept: a dropped TS reason still fails, an undeclared extra still fails. A companion test requires every declared extra to be in EARLY_STOP_SUBTYPES, so the escape hatch cannot reintroduce a silent success.

Downstream

• VS Code matched subtype === 'error' exactly, so the new subtypes fell through to "Completed (N turns)". Now checks is_error and falls back to msg.result when nothing streamed.
• The TUI already keyed on is_error but printed the explanation twice — once as turn text, once as a red error line — because an early stop carries its explanation in result and sets no separate error field. It now emits a short run stopped early (<subtype>).

Verification

End to end against a local SSE server returning a genuinely empty assistant turn: error_during_execution, is_error: true, num_turns 4 (1 + 3 retries), explanation in result.

Every fix mutation-tested — including against the exact mutation that proved an earlier version of the agent-server test was a placebo (compute the subtype, then discard it; the old test passed, the new behavioural one fails).

Reviewed by the critic subagent across two rounds on this PR. It caught the outbox scoping with a runtime repro, the shared nudge budget, the placebo test, and the VS Code gap.

🤖 Generated with Claude Code