Merge pull request 'Harden LLM role/task alignment and PR review feedback discovery' (#169) from feat/issue-167-review-feedback-discovery into master

2026-07-05 12:17:43 -05:00
parent 0b6733a7e8 a4060c59cc
commit 29f04d003a
3 changed files with 757 additions and 0 deletions
@@ -398,6 +398,69 @@ touching anything.
  files, detected secret, or any production/deploy behavior — **stop, report the
  blocker, and take no mutating action.** Fail closed; never work around a gate.

+## Task/role alignment (#167)
+
+The **requested task** decides what a session may do — not the credential it
+happens to hold. Resolve the task first with
+`gitea_resolve_task_capability(task=...)`: it returns `stop_required` and
+`task_role_guidance` alongside the permission decision. An LLM asked to
+review must never degrade into author work just because it is connected as
+an author.
+
+| Requested task | Required identity/profile | Allowed | Forbidden | Stop when |
+|---|---|---|---|---|
+| Review PR (`review_pr`) | reviewer (e.g. `sysadmin` / `prgs-reviewer`) | read, gated review verdicts | commits, pushes, file edits, author comments, merge without eligibility | active profile is an author profile — stop immediately; do **not** switch to author-side fixes unless the operator explicitly re-tasks |
+| Address PR change requests (`address_pr_change_requests`) | author (e.g. `jcwalker3` / `prgs-author`) | commit/push fixes to the PR branch, PR comment summarizing fixes | review verdicts, approve, request-changes, merge | active profile lacks branch push |
+| Merge PR (`merge_pr`) | reviewer/merger | gated merge after eligibility + approval | merging own PR, merging without pinned head match | active profile is an author profile, or any merge gate fails |
+| Comment on issue discussion (`comment_issue`) | any profile with `gitea.issue.comment` | issue thread comments | review verdicts, closing via comment | permission missing (`gitea.pr.comment` does **not** imply it) |
+| Comment on PR (`comment_pr`) | any profile with `gitea.pr.comment` | PR thread comments | review verdicts | permission missing |
+| Author implementation (`create_branch`/`push_branch`/`create_pr`) | author | branch, commit, push, open PR | self-review, self-merge | profile lacks the author permissions |
+
+If the task is review/merge and the session is an author profile, the only
+correct outputs are: the read-only PR queue inventory (#164), the structured
+permission report (#142), and a stop. Ask the operator to reconnect to the
+reviewer namespace; a credential or profile swap in the same session never
+cures same-session authorship.
+
+## Review feedback discovery (MCP-native)
+
+Formal review verdicts (APPROVED / REQUEST_CHANGES / COMMENT) live on the
+review endpoints, **not** in the issue-comment thread. Never infer review
+state from issue comments — use `gitea_get_pr_review_feedback(pr_number=...)`
+(read-only, requires `gitea.read`). It reports:
+
+- every submitted review: reviewer, verdict, redacted body, timestamp, and
+  the head SHA it reviewed;
+- `latest_review_state_by_reviewer` (PENDING drafts never count);
+- `has_blocking_change_requests` / `approval_visible` (dismissed reviews do
+  not block);
+- `current_head_sha` vs `latest_reviewed_head_sha`, `review_feedback_stale`,
+  and `author_pushed_after_request_changes` — so a reviewer can see whether
+  feedback predates new commits, and an author can see whether fixes have
+  been pushed since the REQUEST_CHANGES.
+
+A permission block returns `feedback_not_attempted: true` with a structured
+permission report — distinct from a successful "no reviews yet" result, so a
+blocked lookup is never misread as "no feedback exists".
+
+## Validation reporting discipline (#167)
+
+Validation results in handoffs and PR bodies must state exactly what ran and
+what happened. Build the validation section with
+`build_validation_report(...)` (in `mcp_server.py`) or follow its contract by
+hand — every command is one of:
+
+- `passed` — the exact command ran and succeeded;
+- `failed` — must include the exact command **and its output**; never
+  paraphrase ("shell-invocation quirks" is not a status);
+- `skipped` — deliberately not run; name the reason and any targeted check
+  that replaced it;
+- `not-run` — was required but never executed; say so plainly.
+
+Never imply full-suite success unless the full-suite command itself passed
+(`full_suite_passed: true`). A report that hides a failed or skipped check
+is worse than a failing report.
+
 ## Controller Handoff (required, every task)

 Every task — implementation, review, merge, triage, documentation,