[EdbertChan]

Summary

Reflect turns a chat into lasting lessons for later agents. Before this change, when a person reminded the agent of a rule it had already skipped, Reflect treated that reminder as a new personal preference to capture. The real lesson was that the rule already existed and was not followed. This change adds a small router that picks the lesson bucket before any preference skill is invoked: a restated existing rule is rejected as already covered, and only a true preference with no home becomes preference capture.

Review Claim

A restated existing rule is rejected as already covered and does not open a preference capture; a preference with no existing skill still routes to preference capture.

Review Lane

behavior

Review Unit

engine-runtime

Safety Invariant

Does not change live hooks at install time beyond phrase text for the restated-rule judge; lesson routing is a reflect-step script and skill prose. No merge. No production deploy. Assumptions: Safety Invariant unconfirmed by the user in chat; proceeding with this text.

Slice Rationale

Single claim: the lesson bucket for a restated existing rule. Mechanism, phrase on_hit, skill prose, and tests land together so the gate cannot be skipped by prose alone.

Non-goals

• Does not remove the intervention detector or soften frustration ranking
• Does not change the preference-capture skill itself
• Does not change poteto or pstack reflect
• Does not auto-merge

Architecture

Before

flowchart TD
  audit["intervention flag"]
  auto["Route to preference capture"]
  audit --> auto

Loading

After

flowchart TD
  audit["intervention flag orders pass"]
  gate["lesson_route bucket"]
  audit --> gate
  gate -->|"already covered"| rejected["Rejected"]
  gate -->|"preference no skill"| auto["Route to preference capture"]
  gate -->|"durable new"| accepted["Accepted"]

Loading

Measured

Command: git show origin/main:engine/hooks/llm-judge/phrases/restated-rule.json | python3 -c 'import json,sys; print(json.load(sys.stdin)["on_hit"])'

• base: restated-rule: the person had to restate a rule or skill that already exists. Count it as an intervention; a restated rule the agent should already follow is automate-me evidence.
• head: restated-rule: the person had to restate a rule or skill that already exists. Count it as an intervention signal that orders the reflect pass. Default lesson bucket is already-covered (execution miss on an existing skill) via engine/skills/reflect/scripts/lesson_route.py — do not invent a new preference or call automate-me for this hit alone.

Command: python3 engine/skills/reflect/scripts/lesson_route.py --from-transcript engine/skills/reflect/scripts/tests/fixtures/restated_rule/fires.jsonl

• base: fatal: path 'engine/skills/reflect/scripts/lesson_route.py' exists on disk, but not in 'origin/main' (from git cat-file -e origin/main:engine/skills/reflect/scripts/lesson_route.py; exit 128)
• head: already-covered / action: reject / do not route to automate-me

Command: python3 -m unittest engine.skills.reflect.scripts.tests.test_lesson_route

• base: none: router and its tests are part of this slice; merge-base has no module to import
• head: Ran 10 tests / OK

Test Plan

Test Plan

• python3 engine/skills/reflect/scripts/lesson_route.py --from-transcript engine/skills/reflect/scripts/tests/fixtures/restated_rule/fires.jsonl → already-covered
• preference quote with no skill → automate-me
• skill never opened → skill-not-used
• python3 -m unittest engine.skills.reflect.scripts.tests.test_lesson_route → 10 OK
• grounding tests call the router on fires.jsonl and assert already-covered
• make-pr preflight --commit-only / Review Unit engine-runtime

Revert Plan

Revert Plan

Revert the two commits on this branch. Restores prior restated-rule phrase text and removes the lesson router; Reflect again treats restated rules as preference-capture evidence.

---

Note

Medium Risk
Changes reflect workflow routing and judge phrase guidance; behavior is gated by tests and script output, but mis-routing could skip legitimate preferences or accept wrong lessons if heuristics misfire.

Overview
Adds lesson_route.py so reflect synthesis picks a lesson bucket before Accepted or automate-me routing. Restated existing rules, admitted skips, and skill text that already covers the quote map to already-covered (Rejected), not new preference capture.

intervention-must-automate and frustration signals still order the reflect pass; they no longer imply an automate-me row by themselves. Reflect skill prose, cost-audit.md, and the Frustration lens now require running the router; only bucket automate-me invokes the preference sibling.

The restated-rule llm-judge phrase on_hit text is updated to point at already-covered and lesson_route.py instead of “automate-me evidence.”

The no-comments hook allows trailing text-match-decision-warn: directives so router regex lines stay compliant. Tests and the restated_rule/fires.jsonl fixture assert transcript routing → already-covered.

^{Reviewed by Cursor Bugbot for commit cfbdf1a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: b3cc23cd-c8bc-42eb-a6c2-02dcb6c09638)

[mergify]

Queued — the merge queue status continues in this comment ↓.

[EdbertChan]

This pull request is part of a Mergify stack:

#	Pull Request	Link
1	Route restated existing rules to already-covered, not automate-me	#1324	👈
2	Drop the lesson-router comments that failed the comment gate.	#1330

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: 7e3d7ee2-2373-4ccc-9aea-a0c190c5cecb)

[EdbertChan]

@Mergifyio queue

[mergify]

Merge Queue Status

• ✅ Entered queue — 2026-10-11 19:21 UTC · Rule: admin-bypass · triggered by @EdbertChan with the @mergifyio queue command
• ✅ Checks passed · on draft merge queue: checking #1324 + #1331 + #1334 + #1335 together on main (0d8f2e0) #1338
• ✅ Merged — 2026-10-11 20:22 UTC · at a9995b38b8440b8b933ff7878414b2ffd6e9a9aa · squash

This pull request spent 1 hour 59 seconds in the queue, including 59 minutes 58 seconds running CI.

Required conditions to merge

• check-success = lint
• check-success = test
• check-success = validate