qa-swarm

Claude Code plugin marketplace: qa-swarm (multi-agent code review + TDD fixes), code-atlas (architecture index + queryable graph), plan-runner (parallel plan execution), jupiter (spec consolidation)

Scan a codebase and generate an architecture index loaded at session start — directory map, key files, tech stack, patterns, build commands — plus a semantic dependency graph queryable deterministically via /code-atlas:query (impact analysis, blast radius, risk filters)

Consolidate scattered specs into a canonical master-spec tree with a navigable index, and append stubs for undocumented public surface found in code

Take a Markdown implementation plan, run it through a parallel agent swarm with per-wave verification, and generate a bug-fix plan for re-runs; TDD red-green mode on by default (--no-tdd to disable) with per-task red/green gate evidence; auto-detects Claude Code Agent Teams for token-lean orchestration with a subagent fallback. Tallies best-effort token usage for every dispatched subagent (analyzer, dev agents, verifiers, aggregator) into manifest.json token_usage with a grand total and honest coverage counters, surfaced in dashboards, an end-of-run per-phase Token Report with top consumers, and the PR stats. On the teams backend each wave is verifier-gated: the lead waits on the verifier's actual task result (never self-verifies) and a coverage gate before aggregation backfills any missing verdict, so a PR cannot open while a verifier is still outstanding. git is optional: when absent (no binary or not a repo) all git operations are skipped. Before opening the PR, syncs a code-atlas architecture index (runs /code-atlas:update when .code-atlas/ is present) so the index reflects what was implemented. Verification coverage is configurable via a .plan-runner.yml verify_mode (per-agent | per-wave | last-wave-only) or a --verify flag; lower modes verify fewer waves, record the rest as SKIPPED (distinct from UNVERIFIABLE), and force a draft PR with a banner. Final step opens or updates a proper PR via the plan-runner:pr skill (conventional title, rich body, draft when bugs remain).

Generate a navigable, version-controlled multi-page Markdown wiki from a codebase, engineered for on-demand retrieval as Claude's context while staying human-readable. Two-pass outline-then-fill generation across a parallel agent swarm of planner, diagram author, page writers, and index synthesizer; consumes a code-atlas index as ground truth when present, else self-scans. Pure static Markdown (no embeddings), Mermaid diagrams derived from the dependency graph, per-page source provenance, and git-blob hash-diff staleness detection that regenerates only stale pages. A SessionStart hook loads the wiki index at session start

Vulnerability-aware dependency upgrade orchestrator across 7 ecosystems (npm, Python, Go, Rust, Java, Kotlin, C#); detects outdated packages, queries OSV.dev for CVEs, recommends the safest-yet-most-recent version per package, then executes wave-by-wave with build/typecheck/test verification and clean git rollback on failure