Repository navigation
docs(rfc): the wake judge screens gated loops by default - #13082
Conversation
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Verify the owner endorses §3.3's screen-every-gated-loop default: the [DESIGN-REVIEWED] e0b31cb |
Opus 5 Review — ✅ no blocking findingsReviewed Review detailsRollout §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 [OPUS-REVIEWED] e0b31cb Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
|
8774669 to
d30f93c
Compare
|
|
|
66d9a89 to
0b6b411
Compare
647bea2 to
73d55b2
Compare
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.
73d55b2 to
e0b31cb
Compare
|
bolichen97
left a comment
There was a problem hiding this comment.
Approved after full readiness scan: readiness: passed, required check PR Readiness green on e0b31cb, no effective CHANGES_REQUESTED, no unresolved blocking threads.
…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.
1. What is the problem?
docs/request-for-change/rfc-wake-judge.mdstates 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 landeddecisions/impl_llm.py,nudge.wakeinDECISION_POINT_NAMESbesideJUDGE_POINT, and thedecisions.nudge_wake.provider/.llm_modelconfig 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, anddraftsits 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 thenudge_evidencescope 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_whenpair, 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:babysitandgoal-conductorwould be written to require a brief they no longer need. And adraftstatus 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: 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-auditedandaudited-atmove to today and2fdadc71ac, because the README states that a bare status rots without them.implementation-prsgains feat(decisions): add the wake judge's LLM lane and per-point provider selection #12776;doc-prgains docs(rfc): wake judge -- a System One model screens auto-nudge ticks #12735, the PR that merged the document.judgebrief overrides that default, and three cases bypass the judge --gate=falseloops, whose ticks are the work rather than a reaction to evidence;judge: falseas a per-loop opt-out; andmonitor_watch, which stays probe-first because a typed probe already answers the same question there.statebullet 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.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.shexits 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
2fdadc71acrather than inferred:impl_llm.pypresent;nudge.wakeinDECISION_POINT_NAMESwithJUDGE_POINTand the two lane names;decisions.nudge_wake.providerand.llm_modelin the config route's editable set;points/nudge_wake.pyabsent; nonudge_evidenceentry inPOINT_SCOPE_KEYS; noquiet_streak_flooranywhere insrc/; nokirocrew-judgeagent template;types.pystill declaringChoicealiased asQuestion, so thenoulandscorewire types are still absent;_monitor_tick_is_quietand_MAX_QUIET_STREAK = 10inautonudge.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
judgespec is stored and ignored", while open question 2 saysautoresolves 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.rfc-conductor-work-ledgerv4's convention, because a branch under review moves.Implementation: #12787 (judge core, open), #12776 (LLM lane, on main).