Agent stalls: model announces work, ends turn with end_turn and no tool calls — loop accepts it as done

[BunsDev]

Symptom

The agent announces work and then just stops — the turn ends with stop_reason "end_turn" and nothing happens. Observed repeatedly in one session (BUILD mode, claude-opus-4-8, local):

• Turn A ends with: "Enough exploration — I have the wiring points. Writing the CLI command now. First the self-contained module…" → no tool calls, turn over.
• User: "proceed with implementation" → turn ends with zero output; TUI shows the fix(tui): expand paste placeholders at submit time #149 placeholder (no response — model ended the turn with stop_reason "end_turn").
• User: "continue until completion" → model runs one Bash(cd …) round, then narrates "Now writing the CLI uninstall module." across rounds and ends with "Writing the CLI uninstall module now." → no Write/Edit calls ever happen, turn shows "done".

Session evidence: ~/.coven-code/sessions/d0bca548-…json (assistant messages [7], [9], [11] all end end_turn with announced-but-unexecuted work; [9] is fully empty).

Root cause

run_query_loop treats end_turn as authoritative in both provider paths (crates/query/src/lib.rs — the Anthropic match stop arm and the OpenAI-compat round loop). When the model emits intent-narration text but no tool calls and stops, the loop fires Stop hooks and returns QueryOutcome::EndTurn. The #149 follow-up made the empty variant visible (placeholder text) but not recoverable, and the announce-then-stop variant isn't even surfaced — the turn just reads as if the agent chose to stop mid-task.

There is already a working in-loop recovery precedent: the max_tokens arm injects MAX_TOKENS_RECOVERY_MSG and continues, bounded by MAX_TOKENS_RECOVERY_LIMIT, with should_emit_turn_complete suppressing the intermediate TurnComplete.

Proposed fix

Bounded stall recovery mirroring the max_tokens pattern, in both provider paths:

• When a round ends end_turn with no tool calls and the text is empty or its closing sentences are an imminent-action announcement ("Writing the CLI uninstall module now.", "Let me confirm deps, then write the module.") — inject a continuation nudge as a user message and continue instead of ending the turn.
• Bounded at 2 nudges per user turn; counter resets on any tool round. On exhaustion, fall through to today's behavior (placeholder for the empty case).
• Deliberate handoffs never trigger recovery: questions and wait-for-user phrasing ("Should I…?", "Let me know…", "I'll wait for your review.") end the turn as before.
• Suppress the intermediate TurnComplete while recovering (same as max_tokens recovery).
• Kill switch: COVEN_CODE_DISABLE_STALL_RECOVERY=1.

Worst case for a false positive is one extra bounded round-trip in which the model states the outcome and stops — the nudge text explicitly allows "if the work is complete, state the final outcome".