Already-landed open PRs block the review queue: the already-landed gate
correctly stops approval/merge, but no task path could close the
redundant PR. Add the reconciler close path:
- task_capability_map: new reconcile_close_landed_pr (gitea.pr.close)
and reconcile_close_landed_issue (gitea.issue.close) tasks under a
dedicated "reconciler" role. Exact close capabilities only — the
role grants no review, approve, request-changes, or merge authority,
and normal reviewer profiles keep no broad PR-close authority.
- role_session_router: reconciler tasks route wrong_role_stop in
author/reviewer sessions with a dedicated recovery message; matching
reconciler sessions route allowed_current_session.
- review_proofs.assess_reconciler_close_gate: PR close allowed only
when every proof passes — PR live-verified open, candidate head SHA
pinned and matching the live head, target branch freshly fetched
with a full SHA, head an ancestor of the target, exact
gitea.pr.close proven. Non-landed PRs return
NOT_LANDED_CLOSE_BLOCKED; missing close capability returns
RECOVERY_HANDOFF_REQUIRED; incomplete proof fails closed as
GATE_NOT_PROVEN. Linked-issue close is skipped when the issue is
already closed and allowed only for an open issue resolved by the
landed commits with exact gitea.issue.close capability. The gate
also returns the required final-report field list.
17 new tests cover the capability map, wrong-role routing, and every
gate path (landed close, non-landed block, stale target, changed head,
missing capabilities, linked-issue variants, review-mutation denial).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Approving a PR whose full test suite fails is now gated behind complete
baseline proof. New assess_full_suite_failure_approval_gate:
- Full-suite pass: approval allowed, no baseline required.
- Full-suite failure: default action is REQUEST_CHANGES; approval is
blocked unless the #325 baseline validation proof passes (branches/
worktree, pinned baseline/PR SHAs, matching failure signatures,
clean before/after) and the proof also carries the PR and baseline
full-suite commands, evidence that new/changed tests passed, and a
non-empty explanation of why the failures are unrelated to the PR.
- Unknown/unproven full-suite result blocks approval.
- Main-checkout baseline comparison is rejected.
- Vague "same as master" claims without proof are rejected via the
existing pre-existing-failure claim detection.
- Non-approve decisions are never blocked; the gate only reports the
request-changes default.
Tests cover matching baseline failures, mismatched signatures,
incomplete proof (missing commands, new-test evidence, explanation),
main-checkout baseline, unknown suite result, vague claims, and
full-suite pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add assess_linked_issue_consistency to review_proofs: reviewer final
reports must report a "Linked issue status" that matches the issue(s)
the selected PR was live-verified to link this session. Without live
proof the field may only say "not verified in this session"; with
proof, every linked issue must be named, and any other "Issue #N"
mention anywhere in the report fails as stale leakage from a previous
PR (the PR #280 / Issue #260-vs-#275 incident). Missing field fails
closed. Tests cover correct linked issue, stale previous issue in both
the status field and diagnostics, missing proof, unverified wording,
multiple linked issues, and PR-number mentions not being mistaken for
issue mentions.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Merge prgs/master into feat/issue-330-proof-wording-enforcement and keep
both assess_proof_wording (#330) and git ref mutation reporting (#297).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add assess_prior_blocker_skip_proof to validate reviewer reports document
live REQUEST_CHANGES blocker proof for skipped earlier open PRs, reject
stale-memory skips when head changed after blocker, and require
BLOCKER_STATUS_UNVERIFIED when review feedback cannot be fetched.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Reviewer controller handoffs must now report the precise mutation
categories (File edits by reviewer, Worktree/index mutations, Git ref
mutations, MCP/Gitea mutations, Review mutations, Merge mutations,
Cleanup mutations, External-state mutations, Read-only diagnostics)
instead of the ambiguous legacy "Workspace mutations" field.
- assess_controller_handoff(role="review") no longer requires
"Workspace mutations" and fails closed when any legacy
"Workspace mutations:" line is present, regardless of value.
- New HANDOFF_REVIEW_MUTATION_FIELDS makes the nine precise categories
required fields for review-role handoffs, matching the canonical
schemas/review-merge-final-report.md field set.
- Tests cover approval, request-changes, merge, worktree cleanup,
fetch-only, and blocked handoffs, plus rejection of
"Workspace mutations: none" and non-none values.
- Author/create-issue/inventory roles keep their existing contracts.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add assess_workspace_mutation_consistency to review_proofs: final
reports fail validation when they claim "Workspace mutations: none"
while worktree mutations (reset --hard, checkout, clean, worktree
add/remove, stash, merge, rebase, ...) occurred — either observed in
the workflow command log or self-reported in the report's own
"Worktree mutations" field. Observed worktree and git ref mutations
must be reported under their precise category fields; reports pass
when they say "File edits by reviewer: none" alongside a non-none
"Worktree mutations" entry. Tests cover reset, checkout, worktree add,
clean, fetch-only, and no-op cases.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add assess_non_mergeable_skip_proof to validate reviewer reports document
head SHA, mergeability result, conflict proof command, conflicting files,
and head-change status for each skipped non-mergeable PR, or classify
MERGEABILITY_UNVERIFIED when proof cannot be retrieved.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add assess_reviewer_baseline_validation_proof to block test-suite runs in
the main checkout and require complete baseline worktree proof when reports
claim pre-existing master failures.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add assess_queue_status_report and assess_proof_wording to reject
contradictory queue-status-only reports: passed gates without a selected
PR, prior diagnostic conflict proof, pagination assumptions, and
blocker contradictions. Wire into build_final_report and document the
queue-status-only schema path.
Closes#339
Add assess_proof_wording to reject unsupported phrases such as
inventory complete, live proof, same as master, and file edits none
unless matching evidence is present. Wire into build_final_report and
add doc-contract tests.
Closes#330
Resolve task_capability_map.py conflict by keeping all three capability
entries: commit_files and gitea_commit_files (gitea.repo.commit, #262,
from master) and reconcile_merged_cleanups (gitea.read, #269, this
branch). Capability boundaries unchanged — reconciliation reporting stays
read-only; no close/merge capability broadened.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
route_task_session now calls assess_infra_stop() live instead of
check_mid_merge(). Update TestMCPHealth.test_route_task_session_blocks_during_merge
to mock the infra-stop assessment path and assert infra_stop_assessment
is returned when blocked.
Part of #285 / PR #286 review fix.
Extends the single-terminal-decision machinery (#211) to session-level
hard-stop semantics:
- terminal_review_hard_stop_reasons: after a live terminal verdict, only
the merge sequence for the same approved PR may continue; REQUEST_CHANGES
stops the run; operator-approved corrections re-open the review path
only, never a cross-PR merge.
- gitea_merge_pr: new Gate 2b consults the hard-stop before any API call,
so a run can never verdict one PR and merge another.
- gitea_mark_final_review_decision: hard-stop check with explicit stop
guidance naming the consumed terminal mutation; plus duplicate
REQUEST_CHANGES suppression — an unresolved request-changes at the
current head SHA refuses a second request-changes (fail closed when
feedback cannot be verified).
- tests/test_terminal_review_hard_stop.py: 13 tests covering the
acceptance criteria; existing TestSubmitPrReview setups stub the new
feedback fetch.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Reviewer reports treated 'git fetch prgs master' as read-only
diagnostics while claiming 'Git/worktree mutations: None'. Fetch edits
no files but updates remote-tracking refs, so hiding it under
diagnostics understates what the run mutated.
Add git_ref_mutating_commands() classifying fetch, pull, push, merge,
update-ref, remote update, branch -d/-D, and reset --hard as git ref
mutations, and assess_git_ref_mutation_report() which fails a report
when ref-updating commands ran but the report lacks a non-none 'Git ref
mutations' entry, omits 'fetched <remote>/<branch>' for a targeted
fetch, still claims 'Git/worktree mutations: None', or lists the
command under read-only diagnostics.
Tests cover fetch-only review, missing ledger entry, missing fetch
target, contradictory none-claim, read-only misclassification, worktree
creation (not a ref mutation), merge, cleanup branch deletion,
no-mutation blocked handoff, and dict command-log entries.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Return wrapped {prs, pagination} from gitea_list_prs with has_more,
inventory_complete, and page traversal via api_fetch_page. Route
gitea_review_pr inventory through gitea_list_prs so trust gates can
prove pagination finality. Add assess_list_prs_pagination_proof verifier.
Closes#340
Extend review_proofs with assess_create_issue_final_report to enforce
workflow-source proof, create-issue handoff schema fields, and rejection
of work-issue/review-merge cross-mode handoff fields.
Closes#337
Introduce assess_mutation_ledger_report so reviewer final reports cannot
claim no file edits when the action log records performed mutations.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Consolidate all four canonical workflow files and router SKILL.md:
- review-merge-pr.md (from #334 canonical prompt)
- reconcile-landed-pr.md (reconciliation gates + handoff schema)
- create-issue.md (canonical issue-creation prompt)
- work-issue.md (canonical author/coder prompt)
SKILL.md is now a short router with mode isolation and universal rules.
Adds final-report schemas for all modes and 13 doc-contract tests.
Supersedes partial split work in PR #335; full #333 acceptance in one PR.
Closes#334
- Add workflows/review-merge-pr.md with reviewer-only gates and steps
- Add schemas/review-merge-final-report.md with mutation ledger fields
- Turn SKILL.md into task-mode router; replace inlined §F/§G with links
- Point review/merge templates at extracted workflow
- Add doc-contract tests for split structure
Part of #333.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Workflow reports included the authenticated user's personal email even
when username/profile identity was sufficient (observed: 'Identity:
jcwalker3 / [email protected]' in a reconciliation handoff).
Add format_identity_summary() producing the canonical no-email identity
line ('<username> / <profile>' with optional role/remote, reducing an
email passed in the username slot to its local part), and
assess_email_disclosure() flagging personal email addresses in report
text unless the report itself justifies the disclosure (tool proof,
explicit request, or identity disambiguation).
Tests cover no-email summaries for author and reviewer identities,
role/remote suffixes, email-in-username reduction, clean reports,
single and multiple unjustified disclosures, and justified
disambiguation.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Resolve _GUIDE_RULES conflict by keeping both operator guide keys:
shell_spawn_hard_stop (#258, landed on master via PR #281) and
subagent_delegation (#266, this branch).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Introduce assess_request_changes_approval_proof in review_proofs.py so
review workflows can fail closed when approving an unchanged head after a
prior REQUEST_CHANGES unless explicit override proof is provided.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Adds subagent_gate.py following the author_proofs/review_proofs pure-policy
pattern:
- assess_subagent_delegation: deterministic write tasks (issue claim,
branch creation, code edits, commits, PR creation, review, merge,
cleanup) are blocked for subagents unless explicitly allowed with a
recorded justification and the full inherited gate context (issue lock,
branch/worktree path, identity/profile, allowed tool class, command
deny list, validation ledger requirement, final report schema);
read-only delegation stays allowed; unknown task types fail closed.
- validate_subagent_report: subagent output is accepted only when it
carries the same proof fields as a parent workflow final report.
- Operator guide: new subagent_delegation rule in _GUIDE_RULES so MCP
clients discover the policy.
- tests/test_subagent_gate.py: 13 tests covering blocked write
delegation, justification requirement, missing/blank inherited
context, allowed read-only delegation, unknown tasks, and invalid
subagent final reports.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Introduce validation_ledger with command ledger format, full-suite claim
checks, and structured final-report verification so sessions cannot
overclaim validation or misreport review/merge state.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add assess_infra_stop() to scan git state and conflict markers on each
reviewer capability check, return the exact conflict path in responses,
and clear stale capability-stop terminal state once the worktree is clean.
Closes#285
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
test_skip_scan_walk_root_skips_sibling_worktrees_only derived its
fixture paths from __file__, so the branches/ sibling-worktree skip
assertion only held when the suite ran from a branches/<slug> worktree
whose orchestration checkout has a real branches/ directory —
skip_python_scan_walk_root skips a branches/ walk root only when
<project_root>/branches exists on disk. From the main checkout or any
plain clone the test failed with 'AssertionError: False is not true'.
Build the <main>/branches/<worktree> layout in a TemporaryDirectory and
assert against those paths instead. All four behavioural assertions are
preserved and no production code changes.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Document default tool-call and wall-time budgets for deterministic MCP
work, forbid subagent commit delegation when gitea_commit_files is
available, and pin the rules in runbooks, portable skill, operator guide,
and doc-contract tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Implement gitea_reconcile_merged_cleanups with dry-run reporting for merged
PR remote branches and local branches/ worktrees, plus confirmation-gated
execute mode. Safety gates cover protected branches, open PR references,
active issue locks, dirty worktrees, and delete_branch capability.
Closes#269
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Allow issue lock from base-equivalent branches/ worktrees instead of
requiring the literal master/main branch name. Runtime preflight and
mark_issue/lock_issue now inspect the declared active task workspace so
dirty control-checkout state does not block clean task worktrees.
Closes#275
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Allow issue lock from base-equivalent branches/ worktrees instead of
requiring the literal master/main branch name. Runtime preflight and
mark_issue/lock_issue now inspect the declared active task workspace so
dirty control-checkout state does not block clean task worktrees.
Closes#275
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Document MCP-native gitea_commit_files as the only approved commit path when
gitea.repo.commit is allowed; ban WebFetch, Playwright, and manual base64
workarounds in runbooks and safety model. Add doc tests.
Closes#260.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add runbook section, root-level .gitignore patterns, and detection helper
for _encode_/_emit_/_inline_ throwaway scripts. Surface non-blocking
warnings from assess_preflight_status and gitea_lock_issue.
Closes#261
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add runbook guidance for deleting throwaway _encode/_emit/_inline commit
helpers after MCP commit attempts, ignore those patterns in .gitignore,
and cross-link the worktree-cleanup template.
Closes#261.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
sysadminandClaude Opus 4.8