Handoff tuning: route exploration to quick tasks, arm the quick grain, confirm quick fan-outs #1016

Goal

Chat is the front desk: cheap to talk to, fast to answer words. Work ships off to tasks. Every decision about when work leaves the chat, what shape it leaves in, and how big a piece is cut should be made at the pareto of cost, wall time and quality — and confirmed where a wrong decision spends real money.

This issue is the tracker. Waves are ordered cheap-and-safe → structural. Wave D records what was considered and decided against so it does not come back.

Wave A — prompt and schema, no new machinery

• A1. Route broad read-only exploration to quick_task. Today the judge text says "a quick task is a small thing: a few files, a few minutes", which steers medium exploration (10–30 files) back into the chat's own context where it costs window and turn time. Add one clause to handoffFacts (internal/session/beltfacts.go): broad read-only sweeps — many files, many sources, "where/how does X work" — go to quick_task with an items list, not into the caller's own context. Add routing examples to quickTaskDescription (internal/session/task_quick.go).
  Acceptance: a words-only broad exploration request produces a quick task with an items list, not a long inline read turn.
• A2. Un-cap the quick grain. The 2026-09-10 measurement (task_quick.go header): a quick task sized at twenty files died at taskNoProgress (6 read-only steps) after 500s and $0.62. propose_task's schema exposes no_progress for exactly this; quick_task's does not. Expose it (same interpolated default), or derive the budget from len(items).
  Acceptance: a 15-item exploration quick task is not stopped at taskNoProgress while still ticking items.
• A3. Trust-the-landing clause. carryontasks_test.go's measured failure: a chat re-read a landed quick task's answer three times. Add one sentence to handoffFacts: a landing note IS the answer — fold it, never re-open the work to check it.
  Acceptance: no continuation is written into a turn waiting on or folding its own quick task's landing.
• A4. Sharpen the quick report contract. prompts/quick.md: findings first, absolute paths, explicit "searched X, no Z" for negative results.
• A5. Prefix budget funding. New bytes in handoffFacts, the quick schema and quick.md must be funded by compressing the same fragments — the lean prefix has ~5 bytes of headroom after the chat-loop-economy wave, and the budget only ratchets down (internal/session/prefixbudget_test.go).

Wave B — structural, one reused mechanism

• B1. Confirm quick fan-outs before admission. propose_task has a human gate (the countdown card). quick_task has none: no card, no audit, starts instantly — and a batch of mis-shaped quick tasks spends money before anybody reads anything. Reuse the division-review shape (reviewDivision, RoleDivision, mastermind tier, callRoleChecked budget): a batch of ≥2 quick tasks, or one whose items read like a division, gets ONE review of all parts together before admission. Refusal is an ordinary tool result; nothing is cancelled, nothing spent.
  Acceptance: a scripted chat emitting 3 overlapping quick tasks gets them merged or amended by one mastermind call; a single small quick task admits unreviewed.
• B2. consult tool. Model-initiated one-shot question to the mastermind tier (callRole), bounded per turn (~2), conversations only (task nodes already have the division-review door). For uncertainty the model can name itself.
  Known limit, stated up front: the models that most need a second opinion ask for it least (the markreader measurement) — so B1, not B2, carries the reliability. B2 is the escape valve.
• B3. Model-conditional worker pages — measured experiment only. Short imperative variant of prompts/worker.md and prompts/quick.md for small tiers. No prefix-cache cost (the prefix is per-model already); the cost is a second page to keep true. Bench before shipping; drop if it does not move outcomes.

Wave C — when plandb attaches natively

• C1. Chat-turn items replace the batch-sentence law and the held-note gate (mandated narration → structured plan rows).
• C2. Re-evaluate RoleMarkReader and RoleHandoff against SELECT from the plan table: both exist to synthesize state a database will then hold. Replace inputs with facts; delete calls only after a bench shows the structured state answers "what is left" as well.

Wave D — decided against (do not re-propose without new measurements)

• Runtime-generated worker personas in a KV store: unbounded prompt variance, cache misses, drift, no measured consumer. build_harness is the door for recurring work shapes.
• Model-summarizing compaction: our mechanical stub+fold with the full record in the store is cheaper and lossless; prose summaries compound summary-of-summary drift.
• Best-of-N task attempts by default: a cost multiplier; maybe later as an option on hard changes, never the default.
• Persona prose on generic tasks: words without mechanism — the shapes-list ruling (beltfacts.go) already measured that a prose routing list teaches matching, not judgement.

Notes

• Every item touching prompt or tool surface carries its internal/manual/chat/ page updates and a change entry in the same PR (the manual law).
• Branched off fix/chat-loop-economy (session, provider: the chat loop's economy — read dedupe, slice folding, floor, billed arms, cache pin, shaped gate, stale-build sweep #1015) and stacked on it until that merges.

Drafted with CodeAF · reviewed and owned by the author