From dev-pomogator
Autonomous fix→verify→re-verify loop for architecture-decision-builder. Runs the whole verification battery via one driver (arch-review.ts), triages findings, fixes the mechanical ones, re-runs, and repeats until PASS — WITHOUT a human prompting each cycle. Triggers: "прогони цикл починок", "self-review loop", "verify-fix loop", "погоняй проверки сам", after editing any architecture-decision helper/spec, or before declaring the skill done.
How this skill is triggered — by the user, by Claude, or both
Slash command
/dev-pomogator:arch-review-loopThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Stop making the user prompt every fix→check→review step. ONE driver command surfaces the full
Stop making the user prompt every fix→check→review step. ONE driver command surfaces the full health of architecture-decision-builder; this skill loops on it: run → triage → fix mechanical findings → re-run → repeat until PASS (or only known-false-positives + genuine design forks remain). Report once at the end with the iteration log, not after every step.
npx tsx tools/specs-generator/architecture-decision/arch-review.ts [--spec <slug>] [ARCHITECTURE-dir]
--spec <slug> → runs the battery for .specs/<slug> (default architecture-decision-builder). Generic: validate-spec + audit-spec work for ANY spec; eval-runner runs only for architecture-decision-builder (it's the only one with eval-runner-adb.py). Use it to self-review any spec — e.g. --spec pomogator-doctor (caught real FR-link + scenario-count drift from a sibling commit).ARCHITECTURE-dir → skill/spec-source health (eval if applicable + validate-spec + audit-spec). This is the loop's PASS condition.ARCHITECTURE/ dir → also runs audit / audit-completeness / audit-markers on that generated artefact (informational about that run, not source).Returns JSON {verdict: PASS|FINDINGS, checks[], finding_count} to stdout + a human TO-FIX list to stderr. Exit 0 = PASS, 1 = FINDINGS.
PARTIAL_IMPL_DETECTION on FR-1 matching the BMAD tier name "Deferred" (not a real partial impl). Append to .claude/logs/arch-review-accepted.jsonl {ts, check, reason} and do NOT treat as blocking. Do not invent false-positives to dodge real work.verdict: PASS (ignoring logged known-false-positives) OR no new mechanical finding was fixable this iteration (stuck → report).UNBACKED_VERIFIED_MARKER, do NOT just delete the marker — actually run context7 (resolve-library-id → query-docs) and record-verify <dir> <lib> <ver> if the claim holds, else downgrade the claim to [UNVERIFIED]. Never game your own verification step (see memory never-fabricate-verification-markers).bash scripts/docker-test.sh npx vitest run -t "ARCH" once (background, per no-blocking-on-tests) to confirm e2e — don't run it every iteration.Emit a compact iteration log + final verdict:
arch-review-loop: PASS after N iterations
iter 1: 4 findings → fixed CHANGELOG count, @feature23→TASKS; logged FR-1/deferred false-positive
iter 2: PASS (skill-source)
Docker: 34/34 ARCH ✓
Design forks for you (if any): …
Only surface design forks + the accepted-false-positive list. Do not narrate every fix.
tools/specs-generator/architecture-decision/arch-review.ts.claude/skills/architecture-decision-builder/SKILL.mdnever-fabricate-verification-markers, self-review-must-be-default-not-promptednpx claudepluginhub stgmt/dev-pomogator --plugin dev-pomogatorGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.