fix(handler): never silently drop a turn — re-prompt until model commits

[claudiusthebot]

Why

Two failure modes caused turns to silently drop with no user-visible response and no observability beyond "turn completed with empty output":

1. Tool calls without end_turn (NEW — the bug Dylan hit today)

On 2026-05-12 after the PR #151 merge, I ran git pull + talon restart in a Bash tool, the result returned, and I never called end_turn. Dylan had to ask "Why u didn't reply??" — and the handler exited silently because:

const flowViolation =
  trailing.length > 0 &&  // false — no prose was written
  !state.turnTerminated &&
  !isDuplicateOfDelivered(...);

With no trailing prose, flowViolation = false regardless of whether a delivery tool was called. Tool-only turns slipped through the safety net.

2. Single-retry cap

Even when the trailing-prose case DID fire, the handler re-prompted once and silent-dropped on the second violation. Dylan's design rule (2026-05-12 13:52):

automatically dropping should never happen — you can choose to not reply but the system shouldn't drop on its own.

The model has to commit. end_turn() with no args is the explicit "no reply" close — the only valid silent close.

Changes

• New FLOW_VIOLATION_MAX_RETRIES = 3 constant. Re-prompts up to 3 times before accepting a drop (instead of 1). Set high enough that real-world model slip-ups recover; low enough that a pathologically broken model doesn't loop forever burning tokens. After cap exhaustion the drop is logged at ERROR level so observability catches it.

• _retried: boolean → _retryCount: number = 0. Error-recovery branches (session_expired / context_length / model fallback) now check _retryCount === 0 instead of !_retried to preserve "retry once" semantics for transient infra failures.

• flowViolation check detects BOTH failure modes:

  const hadActivity = trailing.length > 0 || state.toolCalls > 0;
  const flowViolation =
    !state.turnTerminated && hadActivity && !isTrailingDuplicate;

  The toolCalls > 0 leg catches the case I hit today. Empty turns with no prose AND no tool calls (rare) still don't trigger — those are unambiguously "nothing to deliver."

• Reminder body rewritten to be more directive — lists all four valid termination paths (end_turn(text=), end_turn(), send(...), react(...)) and explicitly mentions that tool calls alone don't close a turn.

Tests

Four new integration cases in src/__tests__/integration/talon-functional.test.ts:

• Re-prompt fires for trailing prose without delivery (existing case, now verifying the new reminder body text)
• Re-prompt fires for tool-calls-without-end_turn (NEW — the canonical no-delivery case that wasn't caught before; the test I personally wish had existed yesterday)
• Cap exhausts at exactly 3 retries when the model never recovers (would have been 1 in the old code; would loop forever if cap was accidentally removed)
• No re-prompt on clean end_turn() with no args (the explicit "I chose not to reply" close — must be respected without re-prompting)

Infrastructure note: each handleMessage call spawns a fresh stub subprocess that resets turnIndex to 0, so we can verify re-prompts FIRE (via STDIN protocol log inspection) but can't easily test "model recovers via script.turns[1]" without infra surgery. The assertion that matters most — "re-prompt was attempted, not silent-dropped" — is fully covered.

Verification

• npx vitest run — full suite 1852/1865 pass (1 pre-existing package.functional failure unrelated, 12 skipped live-tier)
• 9/9 functional tests pass (5 existing + 4 new)
• tsc --noEmit clean
• prettier --check clean
• npm run lint — 0 errors, 0 warnings on changed files

Real-world impact

Today's failure mode is impossible going forward. Even if a future-me forgets to call end_turn, the handler will surface [FLOW VIOLATION] reminders up to 3 times before giving up — and giving up itself is now an ERROR-level log instead of silent.

🤖 Generated with Claude Code

[dylanneve1]

Superseded by #246. I reviewed the old changes and ported the valid/current fixes onto current main in the consolidated PR, while leaving out stale removed-backend changes and broad churn that no longer applies.