Review Swarm — Multi-Agent Parallel Review
Start specialized review helpers in parallel on one change, validate their findings, and merge them into one prioritized report (P1 must fix before merge, P2 should fix, P3 suggestion). Narrow reviewers, each out to disprove the change, miss less than one broad one. Done when the report is presented and the offer to act on it is answered or defaulted.
Announce at start: "Starting review swarm — dispatching specialized reviewers in parallel."
Working folder. Blueprint working files live under .agent-blueprint/ in the project root. Before the first write there, make sure .agent-blueprint/.gitignore exists and lists run/, team/, review-runs/, cache/ and .gitignore, so run state and the ignore file itself stay out of commits while plans and notes stay tracked.
Provenance record. Before any other work, run this as written from the project root: mkdir -p .agent-blueprint/run/provenance && date -u +'{"skill": "ab-review-swarm", "version": "4.2.1", "started_at": "%Y-%m-%dT%H:%M:%SZ", "helper_steps": []}' > .agent-blueprint/run/provenance/ab-review-swarm.json. It replaces any older record, and the Working folder step above has already made sure of .agent-blueprint/.gitignore. Each Helper step adds {step, path: helper|inline} to helper_steps. The record tells a run, and the smoke test, which skill ran and how; it is not a security control.
Step 1: Determine Scope
- Files or a path in the request (such as
src/auth/): only those. --pr, a PR or a branch: diff from the merge base,git diff $(git merge-base origin/main HEAD)..HEAD(same as three-dotorigin/main...HEAD). Never use a two-dot range against bareorigin/main: it shows main's newer files as phantom deletions.- Otherwise: uncommitted changes (
git diffplus untracked files fromgit ls-files --others --exclude-standard), or, if none, the last commit (git diff HEAD~1).
If the range is empty or the base is not an ancestor of HEAD, stop and report the range instead of reviewing nothing.
No-commit mode. If the user says not to commit, AGENT_BLUEPRINT_GIT_WRITABLE is 0, or .git is read-only, make no commits: leave changes in the working tree and add the intended message to .agent-blueprint/run/commit-msg.md. With AGENT_BLUEPRINT_GIT_WRITABLE=0, the ship runner commits after the session, so it is not for work the user wants uncommitted. Reviews use the working tree and untracked files against the merge base instead of a commit range.
Here: review git diff $(git merge-base origin/main HEAD) plus the untracked files, whatever range was named, since the work has no commits yet.
Step 2: Select Reviewers
A review-agents list in the YAML frontmatter of blueprint.local.md (project root) is the dispatch list. Otherwise follow references/reviewer-selection.md: always code-reviewer, code-simplicity-reviewer and test-coverage-reviewer, plus each of security-sentinel, performance-oracle, convention-enforcer, frontend-reviewer, architecture-strategist, data-integrity-guardian and schema-drift-detector whose signals the diff shows. Log each start or skip with its reason. --full starts every reviewer.
Step 3: Prepare Review Context
Create .agent-blueprint/review-runs/{run_id}/, where run_id is review-YYYYMMDD-HHMMSS or a short UUID. Build each reviewer's inputs per references/dispatch-notes.md § Reviewer inputs: the rubric in references/review-calibration.md, the two-output contract in references/output-contract.md, and the artifact with the contract it must meet, never the author's claims that the work is correct, which anchor a reviewer toward agreement (§ Input hygiene there).
Step 4: Dispatch All Helpers in Parallel
Helper step. Start a helper (subagent) for this step if you can, with the prompt file named below (its absolute path when the helper can read it, else its full text) and the listed inputs; leave its model and effort at the session's. If you cannot start one, follow the prompt file yourself. Either way, return its Output section, and note which path ran in the run's provenance record if there is one.
Prompts: references/agents/<reviewer>.md for each selected reviewer. Inputs for each: the scope, run_id={run_id}, the diff/files, the Step 3 context and its focus; code-reviewer also gets the plan and standards.
Start them all at once, since they are independent; under a host cap on concurrent helpers, see references/dispatch-notes.md § Helper limits.
Step 5: Collect, Validate, Synthesize
5a: Validation (above 5 findings)
The findings-validator re-checks each finding against the code and the diff (references/dispatch-notes.md § Validation details). Skip it at five or fewer, where its overhead exceeds the benefit.
Helper step. Start a helper (subagent) for this step if you can, with the prompt file named below (its absolute path when the helper can read it, else its full text) and the listed inputs; leave its model and effort at the session's. If you cannot start one, follow the prompt file yourself. Either way, return its Output section, and note which path ran in the run's provenance record if there is one.
Prompt: references/agents/findings-validator.md. Inputs: the merged finding list from the compact returns, numbered F1, F2, … in list order (its verdicts come back by that id), the diff, and run_id={run_id}.
Drop rejected findings. Keep each unresolved one (a protected subject neither confirmed nor refuted), marked unresolved: it reaches the report as advisory with a human owner, because a wrong dismissal there costs more than a false alarm.
5b: Synthesis
Helper step. Start a helper (subagent) for this step if you can, with the prompt file named below (its absolute path when the helper can read it, else its full text) and the listed inputs; leave its model and effort at the session's. If you cannot start one, follow the prompt file yourself. Either way, return its Output section, and note which path ran in the run's provenance record if there is one.
Prompt: references/agents/findings-synthesizer.md. Inputs: the validated finding list, run_id={run_id}, and the artifacts in .agent-blueprint/review-runs/{run_id}/.
Step 6: Present Results
Present the report with its P1, P2 and P3 counts, then offer to act on the actionable findings (present, gated_auto, advisory). With none, the report ends the run.
Asking the user. Ask with your question tool if you have one, offering at most three options; otherwise ask in plain text with a numbered list. In a headless or unattended run nobody will answer: take the default named below, say so in your output, and log it in the run state's decisions if there is a run state.
Five or fewer: fix them now with the ab-resolve-in-parallel skill (independent fixes in parallel), or leave them. More than five: decide each in turn following references/walkthrough.md, since a bulk list does not fit per-item decisions at that volume, or leave them. Default when nobody answers: fix nothing and return the report for the calling pipeline or the user to act on.