Quality Gate

When To Use

• Running quality checks on egregore work items
• Self-review before creating a PR
• Reviewing another agent's PR in PR-review mode

When NOT To Use

• Manual code reviews outside egregore pipeline
• One-off lint or format checks (use make lint directly)

Orchestrate the QUALITY stage of egregore's pipeline. Each quality step runs convention checks from the codex and invokes mapped skills.

Routing Table

| Step | Conventions | Skills | Modes | |------|------------|--------|-------| | code-review | C1,C2,C3,C4,C5 | pensive:unified-review | self, pr | | unbloat | - | conserve:unbloat | self | | code-refinement | - | pensive:code-refinement | self | | update-tests | - | sanctum:update-tests | self | | update-docs | C5 | sanctum:update-docs, scribe:slop-detector | self |

Inputs

The orchestrator invokes this skill with:

• step: which quality step to run (e.g. "code-review")
• mode: "self-review" or "pr-review"
• work_item_id: the manifest work item ID
• branch: the git branch with changes
• pr_number: (PR-review mode only) the PR number

Self-Review Workflow

When mode is "self-review":

1. Get changed files: git diff --name-only main...HEAD
2. Load conventions from conventions/codex.yml
3. Filter conventions to those mapped to the current step
4. Run convention checks via conventions.py
5. Invoke mapped skills on the changed files
6. Collect all findings
7. Calculate verdict

Auto-Fix Loop

If blocking findings exist:

1. Attempt to fix each finding (skill-dependent)
2. Commit fixes to the work item branch
3. Re-run convention checks
4. If still blocking after 3 attempts, verdict is

"fix-required"

Verdict Calculation

if no findings:
    verdict = "pass"
elif all findings are severity "warning":
    verdict = "pass-with-warnings"
elif blocking findings remain after auto-fix:
    verdict = "fix-required"

Record verdict in manifest decisions:

json
{
  "step": "code-review",
  "chose": "pass-with-warnings",
  "why": "2 warnings (C4: noqa in hooks), 0 blocking"
}

Completion Integrity (opt-in, default OFF)

By default the loop runs indefinitely and autonomously: a fix-required verdict is recorded but does not, on its own, stop the item from advancing. This preserves the historical hands-off posture.

When config.pipeline.completion_integrity is true, the verdict becomes a gate the agent cannot talk its way past:

1. A fix-required verdict is reported to the orchestrator as

a step failure, so the item cannot advance to the ship stage with unresolved blocking findings. It retries in place and, on exhausting max_attempts, is marked failed and the overseer is alerted (never silently completed).

1. Merge is held for human review regardless of auto_merge:

the PR is prepared but left open.

This binds "done" to the verifier (convention checks plus the mapped review skills), not to the agent's own say-so. The principle comes from shipped agent harnesses, stated best by KbWen/agentic-os: an agent "can still cut a corner, it just can't cut this one past a check it doesn't control." The recurring patterns across those harnesses:

| Pattern | What it does | Source | |---------|--------------|--------| | Verifier gate as completion judge | Bind "done" to tests, lints, or result hashes the agent cannot fake | Aider, snarktank/ralph, gaasher | | Completion promise + max-iterations cap | String-matching "done" is insufficient; pair it with an iteration ceiling and manual abort | anthropics ralph-wiggum | | Evidence gates the agent does not control | Credential scan, test-execution validation, phase-log parsing enforced by hooks or CI | KbWen/agentic-os | | Keep-if-better / revert-if-worse ledger | Score each change, revert regressions, log to an append-only ledger for review | gaasher/Agent-Loop-Skills | | Legibility anchor | A persisted current-state or repo-map file that lets a human re-enter the loop | Aider repo map, agentic-os | | Human observer checkpoint | Watch early iterations, monitor diffs, reserve loops for defined problems | anthropics, ghuntley |

Further rationale: .claude/rules/prefer-invariants-over-fallbacks.md (evidence base) and Armin Ronacher, "The Coming Loop" (2026), https://lucumr.pocoo.org/2026/6/23/the-coming-loop/. The flag is off by default; enabling it is a deliberate choice to keep a human as the final judge.

PR-Review Workflow

When mode is "pr-review":

1. Fetch PR diff: gh pr diff <number> --name-only
2. Load conventions and filter to code-review step
3. Run convention checks on changed files
4. Invoke pensive:unified-review on the diff
5. Collect all findings

Posting Reviews

Map findings to GitHub review:

• No findings: gh api POST review with

event "APPROVE"

• Warnings only: POST review with event "COMMENT",

findings as inline comments

• Blocking findings: POST review with event

"REQUEST_CHANGES", blocking findings as inline comments with "must fix" prefix

Comment format per finding:

[egregore:{convention_id}] {message}

Convention: {convention_name}
Severity: {severity}

Quality Config

Work items may have a quality_config field:

json
{
  "skip": ["unbloat"],
  "only": ["code-review", "update-docs"]
}

• skip: list of steps to skip (run all others)
• only: list of steps to run (skip all others)
• If both are set, only takes precedence
• If neither is set, all steps run (default)

Convention Filtering by Step

Not all conventions run on every step. The routing table above defines which conventions apply to which step. The quality gate filters the loaded codex accordingly before running checks.

Exit Criteria

• All applicable convention checks executed
• All mapped skills invoked
• Verdict calculated and recorded in manifest
• For PR-review: GitHub review posted