grill-before-distill

name	grill-before-distill
description	Standard memory admission mode for harness-mem (grill-me role). Runs automatically on distill — depth scales by risk, not a forced heavy loop every time. Three modes: deep interrogation, light checklist, lookback for stale truth. No writes. More than a review-stage helper: admission plus lookback.
allowed-tools	["Read","Grep"]

grill-before-distill (grill-me 准入)

What this is

标准准入模式 — not a mandatory heavy re-interrogation on every distill. It is a question router plus pressure-test verdict for the main memory chain.

Complements tools/session-distill/references/distillation-rules.md; does not replace /hm:review.

Phase	Job
grill-me / grill-before-distill	Admission (depth by risk)
session-distill	Extract candidates
/hm:review + govern_memory(action="decide")	Durable confirmation

Runs automatically when distill or memory write is in play — user does not say 「先拷问」. Pick the lightest sufficient mode; escalate when risk rises.

admit / narrow / defer / reject are main-chain actions, not dream actions:

Action	Main-chain meaning
admit	Candidate is allowed to continue to govern_memory(action="suggest").
narrow	Candidate is useful but wording/scope must be narrowed before govern_memory(action="suggest").
defer	Do not write yet; keep as pending/note and attach the missing evidence or question.
reject	Do not write a candidate; treat as noise, local-only, duplicate, or misleading.

Dream is different: dream/lookback handles already confirmed or maintained truth and may suggest keep, narrow, supersede, or reject for later review/undo. Admission actions only decide whether the write path continues.

Three modes (pick by risk)

Mode A — Deep grill-style interrogation (必须)

When:

• User explicit rule: 「记住」「以后都这样」「写成项目规则」
• govern_memory(action="suggest", arguments={kind:"rule", ...}) or repo-wide AI behavior change
• High blast-radius: architecture fact, security, release policy, cross-team default
• Ambiguous fuzzy conclusion on a draft claim after packet read (not before prepare_session_distill)

How: route one unresolved question at a time until the action is clear (see checklist below). Answer from evidence first; ask the user only when evidence and local context cannot decide.

Mode B — Light checklist pass (默认)

When:

• Ordinary /hm:distill / session-distill candidates
• Low–medium risk govern_memory(action="suggest"), handoff, or relation with clear evidence
• Packet-backed facts with observation ids already attached

How: run the 5-point checklist inline in one pass (no multi-turn loop unless an item fails). Tag risk: low; admit if all pass.

[ ] reusable beyond this session?
[ ] kind + scope clear?
[ ] evidence id / path present?
[ ] not session noise (distillation-rules reject list)?
[ ] wording not overbroad?

Fail any → route the unresolved question, then either narrow, escalate to Mode A for that item only, defer, or reject.

Mode C — Lookback (已确认记忆回看，grill-me 主力)

When:

• Dream maintenance, /hm:review spot-check, wake surfaced stale truth
• User asks「这条还对吗」「会不会误导」
• Supersede / conflict signals on confirmed memory

How: adversarial lookback — failure modes, expiry, misleading scope, missing supersede. May use search_memory and, in future, a source-backed external evidence tool such as smart-search to re-verify. Output: keep / narrow / supersede / reject for the maintenance/review path (no silent truth overwrite).

Default routing

Situation	Mode
/hm:distill bulk candidates	B per item; A if rule/high-impact
User「记住这个」	A
Autopilot boundary suggest	B; A if rule candidate
Confirmed memory audit	C
Pure code work, no memory intent	skip

Question routing

grill-before-distill asks the question and routes it to the right answer source. It does not default to asking the user.

Question type	First answer source	Escalate to
Evidence: "what proves this?"	packet, observation id, search_memory, files, tests, docs	answer-memory-evidence
Scope/kind: "fact, rule, handoff, relation, or local note?"	packet + distillation-rules	user only if intent is ambiguous
Freshness: "could this be stale?"	current repo files, tests, docs, timeline/search	external source-backed evidence in future
Misleading risk: "would future Agents over-apply this?"	grill-style self-critique	ask-memory-boundary if product/architecture boundary matters
Architecture/product boundary	existing docs, roadmap, code ownership	ask-memory-boundary
User preference or policy intent	user-stated instruction	ask user

Answer before asking. Ask the user only when the missing information is a preference, intent, or product decision that evidence cannot settle.

Checklist (Mode A — one unresolved question per turn)

Order	Question	If no →
1	Reusable?	defer to note / reject
2	Kind?	keep asking
3	Scope?	keep asking
4	Evidence plan?	route to evidence lookup
5	Expiry?	tag risk

Main-chain action: admit | narrow | defer | reject

Admission note (no MCP writes):

mode: deep | light | lookback
action: admit | narrow | defer | reject
kind + scope: (if admit/narrow)
routing: agent | answer-memory-evidence | ask-memory-boundary | user
evidence_plan: internal | external | both | none
risk: low | medium | high

Main chain position

模糊结论 / session-end
  -> prepare_session_distill
  -> 读 packet + draft candidate claims
  -> grill-me 准入（A/B 按风险；对 draft claim，不是空会话先拷问）
  -> govern_memory(action="suggest")（admit；narrow 后可写；defer/reject 不写）
  -> 内部 search_memory / 代码检索；外部来源证据（smart-search 为参考候选）
  -> auto_review preview + /hm:review
  -> govern_memory(action="decide", decision="confirm") -> confirmed truth

已确认记忆维护 / dream / 抽查 -> Mode C lookback

Boundaries

Does	Does not
Scale admission depth by risk	Force heavy loop on every candidate
Run automatically on distill path	Write candidates or confirmed truth
Mode C for stale truth	Replace /hm:review gate

Non-goals

• Not only compare-bundle review-stage pressure testing; primary role is admission plus lookback
• Not Trellis journal or second truth source
• Not a new MCP tool

References

• docs/memory-adoption.md
• tools/session-distill/references/distillation-rules.md
• tools/session-distill/SKILL.md