القائمة
Orchestrate iterative code review and remediation (Marge loop) over the feature branch diff until all findings are resolved.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Orchestrate iterative spec clarification and remediation (Homer loop) on spec.md until all findings are resolved.
Orchestrate iterative cross-artifact analysis and remediation (Lisa loop) on spec.md, plan.md, and tasks.md until all findings are resolved.
Orchestrate the full SpecKit pipeline (reconcile, specify, homer, phase, plan, tasks, lisa, split, ralph, marge) from feature description to reviewed implementation.
Post a PR review with inline comments for findings requiring human judgment — one-way doors, concurrency risks, architectural decisions, and project-specific patterns.
Analyze a feature branch diff against baseline and project-specific review packs; optionally remediate the auto-fixable findings in severity order.
Challenge and refine a vague idea into a well-formed feature description through adversarial questioning. Produces output ready for /speckit-specify or /speckit-pipeline.
| name | speckit-marge-review |
| description | Orchestrate iterative code review and remediation (Marge loop) over the feature branch diff until all findings are resolved. |
$ARGUMENTS
You MUST consider the user input before proceeding (if not empty).
Verify that the single-pass review skill exists. Marge's review phase delegates to it:
if test -f ".claude/skills/speckit-review/SKILL.md"; then echo "speckit-review: EXISTS"; else echo "speckit-review: MISSING"; fi
if test -f ".claude/agents/findings-ledger.md"; then echo "findings-ledger: EXISTS"; else echo "findings-ledger: MISSING"; fi
If either is MISSING, display this error and STOP:
ERROR: Required file(s) not found.
Missing: <the missing path(s)>
Marge invokes /speckit-review during its review phase to generate findings, and
follows the shared ledger protocol in .claude/agents/findings-ledger.md for
reappearance detection. Run setup.sh to (re)install:
.claude/skills/speckit-review/SKILL.md
.claude/agents/findings-ledger.md
After resolving FEATURE_DIR (the orchestrator handles this), confirm there is a diff to review:
git diff --quiet $(git merge-base HEAD origin/main 2>/dev/null || git merge-base HEAD main)...HEAD
If the command exits 0 (no diff), abort: "No changes detected between the feature branch and main. Nothing to review — run /speckit-ralph-implement first."
Marge uses two gates: an optional fast scoped gate (.specify/quality-gates-fast.sh) per iteration (invoked from .claude/agents/marge.md) and the full gate (.specify/quality-gates.sh) once after the loop terminates.
Validate that the full gate file exists and contains executable content. Run via Bash tool:
test -f .specify/quality-gates.sh && grep -v '^\s*#' .specify/quality-gates.sh | grep -v '^\s*$' | head -1
If the file is missing or contains only comments/whitespace (the command produces no output), STOP with this error:
ERROR: Quality gates file is missing or empty.
Expected: .specify/quality-gates.sh with executable commands.
The full quality gates file is required for Marge to validate review fixes.
Create or update .specify/quality-gates.sh with your project's quality gate
commands (e.g., npm test && npm run lint). The file must exit 0 for gates to pass.
Also check for the optional fast gate:
test -f .specify/quality-gates-fast.sh && echo "EXISTS" || echo "MISSING"
If the fast gate exists, note this — the marge agent will use it per iteration. If missing, the agent falls back to the full gate per iteration.
Set the following LOOP_CONFIG values for this execution:
Read and follow the instructions in .claude/agents/loop-orchestrator.md, using the LOOP_CONFIG values above. Pass $ARGUMENTS through for argument parsing.
Run only when the loop exited via the success path (<promise>ALL_FINDINGS_RESOLVED</promise> observed). Skip on max-iterations, stuck, or failure exits.
Run via Bash tool:
bash .specify/quality-gates.sh
If it exits non-zero, treat the loop as incomplete: the fast gate missed a regression in files outside the per-iteration scope. Set the completion status to failure with reason "end-of-loop full quality gates failed", surface the failing output in the report, and suggest rerunning marge or fixing the issue manually before re-running this command.
If it exits zero, proceed to review report verification.
After the loop completes, confirm <FEATURE_DIR>/review-report.md exists and was written by the final sub agent — the persisted review report is the audit trail of what was found and fixed, and the ledger the next Marge run uses for reappearance detection. If the file is absent after the loop completes, surface this as a failure and instruct the user to re-run /speckit-marge-review.
Every iteration of this loop spawns a Marge sub agent (per .claude/agents/marge.md) that MUST overwrite <FEATURE_DIR>/review-report.md before exiting. This command does NOT write the file directly. The authoritative schema and status lifecycle live in the shared ledger protocol .claude/agents/findings-ledger.md (sourced from claude-agents/findings-ledger.md), which claude-agents/marge.md configures with REPORT_FILE: <FEATURE_DIR>/review-report.md. The report is the audit trail of what was found and fixed, and the ledger the next Marge run loads for reappearance detection — do not restate its schema here.
The orchestrator MUST treat absence of <FEATURE_DIR>/review-report.md after the loop completes as a Marge failure — without it the next Marge run has no ledger for reappearance detection.
/speckit-marge-review — Auto-detect spec dir from current branch, use default max iterations (10)/speckit-marge-review specs/003-fix-pipeline-delegation — Run for specific spec dir/speckit-marge-review 5 — Auto-detect spec dir, limit to 5 iterations/speckit-marge-review specs/003-fix-pipeline-delegation 5 — Specific spec dir with 5 max iterations