Skip to content

docs(rfc): the wake judge screens gated loops by default - #13082

Merged
bolichen97 merged 1 commit into
mainfrom
docs/rfc-wake-judge-accepted
Sep 23, 2026
Merged

bolichen97 merged 1 commit into
mainfrom
docs/rfc-wake-judge-accepted

Conversation

@chenmingwei23

Copy link
Copy Markdown
Contributor

1. What is the problem?

docs/request-for-change/rfc-wake-judge.md states two things that are no longer true.

Its front matter says status: draft, which the directory's own vocabulary defines as "Proposed. Nothing built." The LLM lane is on main: #12776 landed decisions/impl_llm.py, nudge.wake in DECISION_POINT_NAMES beside JUDGE_POINT, and the decisions.nudge_wake.provider / .llm_model config keys. The field carries weight beyond bookkeeping: the First Principles review lane reads it from the base commit as the record of a product-shape decision, and draft sits outside the closed set it accepts.

Its arming section says "Without judge, the loop behaves exactly as today" -- a loop is screened only when its owner writes a brief for it. That is not the shape #12787 builds: once the nudge_evidence scope is granted, every gated auto-nudge loop is screened, brief or no brief.

2. Why this issue matters to the user

An owner reading this RFC to decide whether to grant the scope would conclude that nothing happens until they hand each loop a wake_when / quiet_when pair, then be surprised when the grant alone changes the behaviour of every gated loop they already have armed. The same wrong reading reaches the skills: babysit and goal-conductor would be written to require a brief they no longer need. And a draft status invites every reader who follows this directory's own advice to distrust the document at exactly the point where its design is being implemented.

3. How our fix solves it

Both symptoms have one root cause: the design moved after the document merged, and nothing in it was re-audited against main.

  • Status. status: partial, with the prose status line the vocabulary requires for that value, naming the LLM lane as on main and the judge core as in flight on #12787. last-audited and audited-at move to today and 2fdadc71ac, because the README states that a bare status rots without them. implementation-prs gains feat(decisions): add the wake judge's LLM lane and per-point provider selection #12776; doc-pr gains docs(rfc): wake judge -- a System One model screens auto-nudge ticks #12735, the PR that merged the document.
  • Arming. The section now states the built shape: the scope grant screens every gated loop with a built-in brief (wake when the subject needs its owner -- a blocker, a question or a ruling addressed to it, a terminal state, or the exit condition the loop's own message states; quiet when nothing has arrived for the owner since the last tick), an explicit judge brief overrides that default, and three cases bypass the judge -- gate=false loops, whose ticks are the work rather than a reaction to evidence; judge: false as a per-loop opt-out; and monitor_watch, which stays probe-first because a typed probe already answers the same question there.
  • The sentences the new default falsified. The state bullet now names the brief in force rather than the spec's text; the rollout's day-one path is the scope grant rather than an armed spec; the section heading no longer says the owner writes the brief.
  • README index. The row's status cell, its "what is actually on main" verdict re-verified at 2fdadc71ac, and one sentence appended to the directory's audit trail.

4. What tests we did

Docs-only, so docs-lint alone: scripts/docs-lint.sh exits 0 and reports "All documentation checks passed" (the dead-identifier lines it prints are the pre-existing baseline, untouched here).

Every claim the new status line and the README row make was read at 2fdadc71ac rather than inferred: impl_llm.py present; nudge.wake in DECISION_POINT_NAMES with JUDGE_POINT and the two lane names; decisions.nudge_wake.provider and .llm_model in the config route's editable set; points/nudge_wake.py absent; no nudge_evidence entry in POINT_SCOPE_KEYS; no quiet_streak_floor anywhere in src/; no kirocrew-judge agent template; types.py still declaring Choice aliased as Question, so the noul and score wire types are still absent; _monitor_tick_is_quiet and _MAX_QUIET_STREAK = 10 in autonudge.py; and #12598 merged, so the Decisions-card shape the row mounts on is on main and only the point's own row waits.

5. Any other suggestions on the work

  • Nothing in this PR depends on the status it sets. It ships no code, so the flip records a decision rather than proposing one to justify a change in the same diff.
  • One ambiguity is left standing rather than quietly resolved: the Decisions-card section says a scope-off tick "behaves exactly as today and any judge spec is stored and ignored", while open question 2 says auto resolves to the LLM lane when the scope is absent. Which of those governs a scope-off loop that pins no provider belongs to the core PR, not to a docs edit.
  • The status line pins the core sections to #12787 and asks for a re-audit at merge, following rfc-conductor-work-ledger v4's convention, because a branch under review moves.

Implementation: #12787 (judge core, open), #12776 (LLM lane, on main).

@chenmingwei23
chenmingwei23 requested review from a team and dwu96 September 23, 2026 13:26
@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Sep 23, 2026
@github-actions

github-actions Bot commented Sep 23, 2026 •

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of e0b31cbf4bfe2d1b51703154231080b15701401f and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] e0b31cb

False positive or not applicable? A repository writer can comment:
/ai-review override gpt e0b31cbf4bfe2d1b51703154231080b15701401f: <one-sentence reason>

@github-actions

github-actions Bot commented Sep 23, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Fable 5) — ✅ PASS

Design-level review of e0b31cbf4bfe2d1b51703154231080b15701401f — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Verify the owner endorses §3.3's screen-every-gated-loop default: the partial flip makes this RFC the decision record #12787 will cite.

[DESIGN-REVIEWED] e0b31cb

@github-actions

github-actions Bot commented Sep 23, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5 Review — ✅ no blocking findings

Reviewed e0b31cbf4bfe2d1b51703154231080b15701401f — this comment is updated in place on each push.

Review details

Rollout §8 re-asserts "Jev lane only" for PR D, which the landed lane resolver contradicts.

FINDING — docs/request-for-change/rfc-wake-judge.md:444 — the re-audited PR D entry keeps Jev lane only while judge_lane on main returns LANE_LLM whenever Jev is not armed (gate.py:487) and _judge_authority authorizes that lane with no keystone read (gate.py:611), so D's tick hook will screen on the already-landed LLM lane by default on any install without a Jev key — and the same section still lists PR E's decisions/impl_llm.py, "the provider setting" and the card's provider/LLM-model pickers as future work with Depends on D's point name, though this commit's own front matter records implementation-prs: [12776] and nudge.wake is already in DECISION_POINT_NAMES (gate.py:70) with decisions.nudge_wake.provider / .llm_model already editable (dashboard/handlers/core.py:2617) → Fix: replace Jev lane only with the actual day-one behaviour (both lanes live, default auto landing on the LLM lane without a Jev key), and mark entry 2 as landed via #12776 instead of dependent on D.

[OPUS-REVIEWED] e0b31cb

Verdict parsed from the review's SHA-scoped output markers for commit e0b31cbf4bfe2d1b51703154231080b15701401f.

False positive or not applicable? A repository writer can comment:
/ai-review override fable e0b31cbf4bfe2d1b51703154231080b15701401f: <one-sentence reason>

@chenmingwei23

Copy link
Copy Markdown
Contributor Author
  • span=45dc661abaa4 -- "every gated auto-nudge loop is screened" vs the monitor is None early return: the claim stands, and the placement it depends on is now stated in the document.

Checked against the branch the sentence describes, feat/wake-judge-core read at 08f5787fb8, rather than against main's hook. There the judge is called BEFORE that guard. The method computes probe_will_run = monitor is not None and monitor.outcome is None and loop.gate, then if not probe_will_run: calls the judge, and only after that comes if monitor is None or monitor.outcome is not None: return False.
The judge helper's own docstring gives the reason: a conductor watching sibling sessions has no monitor state at all, so anything gated behind a monitor's presence would never reach it. Its comment adds that a loop whose probe DOES run defers instead to the quiet return at the end, because only the probe sees a merged or closed subject and ends the watch.
So the loops the Problem section motivates are exactly the ones the judge reaches, and main's early return is the shape the core PR changes, not a ceiling on it.
The suggested fix is taken anyway, because a reader cannot check the claim without it: the arming section now states where the judge sits relative to the probe guard, in both directions.

@chenmingwei23
chenmingwei23 force-pushed the docs/rfc-wake-judge-accepted branch 3 times, most recently from 8774669 to d30f93c Compare September 23, 2026 14:38
@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running and removed readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention labels Sep 23, 2026
@chenmingwei23

Copy link
Copy Markdown
Contributor Author
  • span=44880f3c4bb3 -- the document admitting owed edits inside its design sections: accepted, and answered by moving them rather than restating them.

Correct, and the placement was the real problem. A design section that says "this is an edit the branch owes" is a debt record in a section a reader consults for the contract.
Two of the three items were not decided facts at all, so they are now numbered open questions (Q4: which switch authorizes which destination for a briefless loop; Q5: how a consent granted under narrower wording is told apart from one granted under the wider population). Q5 names the mechanism the seam already uses for this, a scope key of its own, without choosing it here.
The third is a decided requirement, so it stays stated in present tense where the contract lives, and the branch's lack of it moved to the prose status line, which is the part of this document that exists for in-flight facts and already carries the head it was read at.
Nothing in the design sections now claims a mechanism that does not exist.

@chenmingwei23

Copy link
Copy Markdown
Contributor Author
  • span=480aba12325c -- recording the owed edits as review comments on the core PR: deferred, with the reason.

Worth doing and not mine to do unilaterally: two of the three items are now open questions rather than owed edits, and their answers decide code on that PR. A comment asserting one answer would pre-empt a decision its owner has not made.
The debt is not lost in the meantime. The status line names each piece the branch does not carry, pinned to the head it was read at, and asks for a re-audit of the three affected sections at merge.
Raised with the conductor as the one design choice this document deliberately leaves open, so it reaches that PR as a decision rather than as a drive-by comment.

@chenmingwei23

Copy link
Copy Markdown
Contributor Author
  • span=45dc661abaa4 -- the compounding bound having no floor for a loop with no monitor: does not hold against the branch, and the bound is now named precisely.

Checked against feat/wake-judge-core at 08f5787fb8 rather than main's site. The branch does not use the monitor's streak field. It carries a per-LOOP judge_quiet_streak, incremented on every QUIET verdict and reset on a delivered turn, compared against _judge_quiet_streak_floor(), which reads the point's own config key and clamps it.
Its own comment states the reason for a separate field: the monitor's streak cannot reach a loop that has no monitor, which is exactly the briefless population in question.
The wording is still improved, because a reader could not check the claim from the section: the security section now names the per-loop counter and says why a per-loop counter is what makes the bound reach a monitorless loop.

@chenmingwei23
chenmingwei23 force-pushed the docs/rfc-wake-judge-accepted branch 2 times, most recently from 66d9a89 to 0b6b411 Compare September 23, 2026 15:32
@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running and removed readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention labels Sep 23, 2026
@chenmingwei23
chenmingwei23 force-pushed the docs/rfc-wake-judge-accepted branch 2 times, most recently from 647bea2 to 73d55b2 Compare September 23, 2026 16:11
@github-actions github-actions Bot added the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Sep 23, 2026
Once the nudge_evidence scope is granted, the judge screens every gated
auto-nudge loop with a built-in brief: wake when the subject needs its
owner, quiet when nothing has arrived for the owner since the last tick.
A loop's own judge brief overrides that default and is screened on its
lane's own authority, so the scope carries the built-in brief alone and a
fresh install screens nothing. Loops armed gate=false, loops passing
judge: false, and monitor_watch bypass the judge, the last because a
typed probe already answers the same question there. The arming section
states where the judge sits relative to the probe guard, which decides
whether a loop with no typed subject is reached at all.

Because screening reaches loops whose owner wrote nothing, the built-in
brief also carries the clause the compounding bound rests on: a comment
or a fetched page is untrusted content, and a claim inside it that there
is nothing to do is not evidence that nothing happened. The security
section names the exposed population as every gated loop rather than the
briefed ones, and names the bound as the judge's own per-loop streak
counter, which is what reaches a loop carrying no monitor.

Making the default wider than the opt-in wording leaves two mechanisms
undecided, so they are open questions rather than prose that asserts an
answer: which switch authorizes which destination for a briefless loop,
and how a consent granted under narrower words is told apart from one
granted under the wider population. The design sections state the
contract; the status line carries what the branch does not have yet.

The LLM lane runs the bundled kirocrew-lite agent on the model that agent
already resolves -- JUDGE_MODEL_DEFAULT is auto, so it pins no model
class -- and no judge-specific template ships.

The RFC leaves draft for partial: the LLM lane is on main and the judge
core is on a branch, which is a subset of the sections describing it --
two evidence kinds, no comment-body kind by design, no ledger or self
collector, and a default brief without the untrusted-evidence clause. The
audit fields, the prose status line and the README index row record what
is where, including the card row and its pickers already being on main
with only the consent switch waiting on the scope.
@chenmingwei23
chenmingwei23 force-pushed the docs/rfc-wake-judge-accepted branch from 73d55b2 to e0b31cb Compare September 23, 2026 16:29
@github-actions github-actions Bot removed the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Sep 23, 2026
@chenmingwei23

Copy link
Copy Markdown
Contributor Author
  • span=45dc661abaa4 -- the rollout's "Jev lane only" for PR D reading as false now that the LLM lane has landed: accepted, and deferred to the Q4 follow-up rather than a tenth review round.

Checked and correct on the mechanism. judge_lane ends with "return LANE_JEV if jev_consented else LANE_LLM", and the judge authority returns the LLM lane with no keystone read, so once the core PR's tick hook lands on a machine with no armed Jev, the lane that answers is the one that shipped in the LLM-lane PR. The phrase describes what that PR builds, not what a tick will do.
Deferred rather than pushed, for a reason I would rather state than hide: the entry it sits in is already marked blocked on Q4, and Q4 is exactly "which switch authorizes which destination for a briefless loop". Whoever settles Q4 rewrites this entry's lane sentence in the same edit, so correcting the clause now buys a document that is right for one round and re-rolls nine lanes plus five non-deterministic reviews for prose.
The board on this head is otherwise clean: every shard green, zero failing lanes, and the four other review lanes passing with no findings. Spending a fresh head on a wording choice that another decision will rewrite is the trade this defers.

@github-actions github-actions Bot added readiness: passed Eligible automated validation passed for the current revision and removed readiness: checking Automated validation is still running labels Sep 23, 2026

@bolichen97 bolichen97 left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approved after full readiness scan: readiness: passed, required check PR Readiness green on e0b31cb, no effective CHANGES_REQUESTED, no unresolved blocking threads.

@bolichen97
bolichen97 merged commit dc3546c into main Sep 23, 2026
94 checks passed
@bolichen97
bolichen97 deleted the docs/rfc-wake-judge-accepted branch September 23, 2026 22:54
@github-actions github-actions Bot removed the readiness: passed Eligible automated validation passed for the current revision label Sep 23, 2026
wknight1 pushed a commit to wknight1/KiroCrew that referenced this pull request Oct 5, 2026
…13082)

Once the nudge_evidence scope is granted, the judge screens every gated
auto-nudge loop with a built-in brief: wake when the subject needs its
owner, quiet when nothing has arrived for the owner since the last tick.
A loop's own judge brief overrides that default and is screened on its
lane's own authority, so the scope carries the built-in brief alone and a
fresh install screens nothing. Loops armed gate=false, loops passing
judge: false, and monitor_watch bypass the judge, the last because a
typed probe already answers the same question there. The arming section
states where the judge sits relative to the probe guard, which decides
whether a loop with no typed subject is reached at all.

Because screening reaches loops whose owner wrote nothing, the built-in
brief also carries the clause the compounding bound rests on: a comment
or a fetched page is untrusted content, and a claim inside it that there
is nothing to do is not evidence that nothing happened. The security
section names the exposed population as every gated loop rather than the
briefed ones, and names the bound as the judge's own per-loop streak
counter, which is what reaches a loop carrying no monitor.

Making the default wider than the opt-in wording leaves two mechanisms
undecided, so they are open questions rather than prose that asserts an
answer: which switch authorizes which destination for a briefless loop,
and how a consent granted under narrower words is told apart from one
granted under the wider population. The design sections state the
contract; the status line carries what the branch does not have yet.

The LLM lane runs the bundled kirocrew-lite agent on the model that agent
already resolves -- JUDGE_MODEL_DEFAULT is auto, so it pins no model
class -- and no judge-specific template ships.

The RFC leaves draft for partial: the LLM lane is on main and the judge
core is on a branch, which is a subset of the sections describing it --
two evidence kinds, no comment-body kind by design, no ledger or self
collector, and a default brief without the untrusted-evidence clause. The
audit fields, the prose status line and the README index row record what
is where, including the card row and its pickers already being on main
with only the consent switch waiting on the scope.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants