lsm / neokai / 28248713106

82%

push

github

feat(forge): feed existing lessons and proposals into episode judge prompt (#2146)

* feat(forge): feed existing lessons and proposals into episode judge prompt

The Forge episode judge generates candidate lessons and proposals from
scoped evidence, but had no visibility into items already in the scope.
This caused duplicate lessons (re-deriving active lessons already injected
into tasks) and duplicate proposals (re-proposing work already in the
backlog — a likely contributor to the 10 unconverted proposals).

buildEpisodeInput now fetches up to 10 active lessons and 10 proposed
proposals from the evolution repo and passes them through
EpisodeJudgePromptInput to the judge prompt. The prompt includes a
deduplication instruction so the judge skips duplicates and refines
existing items instead.

* fix(forge): address Codex review — include accepted proposals, candidate lessons, proposal details, drop refine instruction

- Include accepted (not just proposed) proposals in dedup context since
  createTaskFromProposal claims both statuses
- Include candidate (not just active) lessons to prevent duplicate
  candidates before operator review
- Add proposal description and reason to the prompt so the judge can
  match on task semantics, not just title
- Change instruction from "refine or extend" to "omit" since there is
  no update path — createFromEvidence only creates new records
- Expanded test to cover all statuses (active/candidate/proposed/accepted
  included; dismissed excluded)

* fix: accept [bot]-suffixed logins for Codex gate when GitHub reports type User

GitHub inconsistently reports App bots with type "User" in reaction
payloads, which caused the codex_review_bot gate and plan/review approval
hooks to reject valid Codex +1 reactions from chatgpt-codex-connector[bot].

Replace the (.user.type // "") == "Bot" guard with
(.user.login // "") | endswith("[bot]") — the [bot] suffix is a GitHub
convention for App bots that human accounts cannot carry, m... (continued)

11422 of 16451 branches covered (69.43%)

Branch coverage included in aggregate %.

42 of 42 new or added lines in 3 files covered. (100.0%)

96569 of 115290 relevant lines covered (83.76%)

299.04 hits per line