Give the agenda planner a token budget that survives thinking

[EllAchE]

Finishes what #221 started. The AI reviewer and the agenda planner had the same defect; #221 fixed the reviewer, this fixes the planner.

The bug

max_tokens is a single ceiling over reasoning and visible text. claude-sonnet-5 reasons by default when the request omits thinking. lib/ai/agenda.ts passed max_tokens: 4096 — sized for the placements alone — so a long thinking pass could spend the budget and leave the JSON cut off.

This one is more exposed than the reviewer was. A programme is exactly the kind of answer that earns a long think, and the response body grows with the queue, so both halves are large.

It failed more quietly than the reviewer did

The reviewer at least said something was wrong. Here, extractJson returns no placements from a fragment — exactly as it does from prose — and the caller then returns status: 'ok' with an empty proposal. The organizer sees a planner that apparently had nothing to suggest, when in fact it had plenty and ran out of room saying it. No error, no log line, nothing to investigate.

The fix

• Ceiling raised to 16,000, matching the reviewer. Still under the size where a non-streaming request risks an HTTP timeout, and an unused ceiling is free — billing counts tokens generated, not tokens allowed.
• proposalText reads stop_reason and withholds a cut-off draft, so the caller returns status: 'error' naming the real cause rather than an empty board. The message suggests a shorter queue, which is the half of the budget the organizer actually controls.

The two alternatives stay rejected for the same reasons as in #221: budget_tokens is rejected outright by this model, and turning thinking off is documented to let reasoning prose leak into the text block — which is the string extractJson reads.

Nothing about the trust boundary changed. Every placement still goes through validateProposal and the same detectConflicts the board uses.

Tests

New lib/ai/agenda.test.ts (6 cases) — the file had none. Covers text blocks joined, non-text blocks ignored, stop_reason: "max_tokens" withholding the draft, and the point of the guard: that truncation is not detectable from the text, since a budget can run out on the token after a syntactically complete object.

bun run test (1758 passed / 163 files), bun run lint, bun run typecheck all clean.

Note for #173

#173 also touches lib/ai/agenda.ts. This change is confined to the constant, one new exported function, and the block that reads the response — but whichever merges second should check MAX_OUTPUT_TOKENS survived, and lib/ai/agenda-optimizer.ts deserves the same look before it ships.

🤖 Generated with Claude Code