claude-multi-agent-pipeline

Generate interactive HTML diagrams (flowcharts, sequence diagrams, class diagrams, state diagrams, ER diagrams, Gantt charts, etc.) using Mermaid.js. Opens the result in the browser. Use when the user says "/diagram", "draw a flowchart", "make a diagram", "show me a flow", "sequence diagram", "class diagram", "state diagram", or any request for a visual diagram.

Graph-node skill: reads .ship-manifest.json (or story/epic IDs), launches one background agent per story to write plans/*.md files, applies critique checklist, updates DB with plan_file paths. Artifact-chain-aware version of /draft-plan. Consumes manifest from /plan-stories, produces plans/*.md for /env-preflight.

Spec-driven feature decomposition: accepts a FeatureSpec JSON (product, pattern, entity, fields, permissions, integrations, UI) and decomposes it into dependency-ordered stages using deterministic pattern DAGs. Creates stories via PM tools and writes .ship-manifest.json for downstream /draft-plans and /run-stories consumption. Use when the user says "/factory spec.json", "/factory --pattern crud-ui <inline JSON>", or "/factory --dry-run spec.json".

Merge a story worktree branch into its dev branch, push, clean up the worktree and branches, and mark the story done. Use when the user says "/merge-worktree", "/merge-worktree story-NNN", or "merge worktree".

Execute stories in parallel (where safe) using background agents, each working in an isolated git worktree branched from a shared dev branch. Handles dependency ordering, write-file conflict detection, and reports results. Use when the user says "/run-stories", "/run-stories story-NNN", "/run-stories epic-NNN", or any combination of story and epic IDs.

Bridge between presearch briefings (or inline args) and the /draft-plans skill. Parses inputs, dispatches the planner agent to create an epic and stories in the DB, and writes .ship-manifest.json for downstream consumption. Three invocation modes: briefing path, inline title+items, or resume an existing epic. Also handles execute mode for existing plan files (bypasses planner, creates DB entries directly). Use when the user says "/plan-stories presearch/topic.md", "/plan-stories \"Title\" 1. Feature 2. Feature", or "/plan-stories epic-NNN".

Universal entry point for all code changes — from 1-file fixes to multi-story epics. Auto-classifies scope and routes internally (quickfix for <=5 files, full pipeline otherwise). Never skip /ship — it handles everything.

Generate a UI component library via Gemini. Creates reusable components with clean interfaces that ui-coder agents import instead of generating per-story. Use /gen-library before /ship for projects with UI work.

Project introspection for implementation planning. Reads the codebase, queries recorded decisions and OpenMemory, identifies existing patterns, constraints, and testable assertions. NO web research — that's /research's job. Scout answers "what does THIS project require?" not "what exists in the world?" Reads .clarify-<slug>.json and/or .research-<slug>.json if provided. Writes presearch/.scout-<slug>.json with findings, constraints, patterns, and testable assertions. Bootstrap mode (--bootstrap <path>) scans an unfamiliar repo and generates CLAUDE.md, decisions.sql, and .claude/refs/ convention artifacts. Use when the user says "/scout <topic>", "/scout --clarify presearch/.clarify-foo.json", "/scout --research presearch/.research-foo.json", "/scout --deep <topic>", or "/scout --bootstrap /path/to/repo".

Standalone quickfix pipeline: validates criteria, writes plan, launches coder in worktree, merges via /merge-worktree. Accepts optional --context flag for artifact chain ingestion (upstream clarify/research findings enrich the plan).

Pipeline output quality validation. Generates a real app via /ship, builds it, and grades wiring quality with deterministic checks. Default mode ("quality") generates a recipe app and validates build + wiring. "ship" mode runs the legacy structural pipeline validation.

Self-critique loop: iteratively improve whatever Claude just produced — plans, code, architecture, design proposals. Loops until NMIP (No Material Improvements Possible) on all 5 lenses, then escalates to Gemini for fresh eyes. Stores blind spots for future sessions. Use when the user says "/critique", "/critique path/to/file.md", "/critique 'topic'", "/critique --deep", or "/critique --plans <paths>".

Integrated review + build + test + acceptance criteria verification for a dev branch. Launches a reviewer agent on the full diff, runs build and tests, and walks acceptance criteria from all plan files. Use after stories merge into dev — independently of /ship. Invoked as "/verify .ship-manifest.json", "/verify epic-NNN", or "/verify" (auto-detects manifest in cwd).

Unified dual-engine audit: runs Gemini large-context analysis followed by Claude critique, synthesizing both into a scored report with per-finding acceptance criteria. Use when the user says "/audit", "audit the codebase", "audit story-NNN", or "audit <path>". Replaces both old /audit and /gemini-audit. Supports --claude-only and --gemini-only for single-engine backward compat. Supports scoping to files, directories, story diffs, or time ranges. Section filters, scoring, and all existing flags preserved.

Gemini synthesis + Claude critique + scope check + decision recording. Reads scout artifacts (presearch/.scout-<slug>.json), walks prev pointers to load clarify, research, and scope data, writes human-readable presearch briefing (presearch/<slug>.md). Standalone invocable. Use when the user says "/briefing <slug>", "/briefing --scout path/to/.scout-foo.json", or "/briefing --refine presearch/existing.md".

Detect project constraints, analyze ambiguities via Gemini, batch Q&A to the user, and write a typed JSON artifact (.clarify-<slug>.json) for downstream consumption. Use when the user says "/clarify <topic>", "/clarify path/to/requirements.md", "/clarify --quick <topic>", or "/clarify --research <path> <topic>".

Interactive spec builder: takes a product name + natural language description and produces a full feature spec document (markdown) with machine-readable FeatureSpec JSON. Covers objective, user stories, requirements, acceptance criteria, constraints, integration points, out of scope, AI boundaries, and the JSON extract for /factory. Use when the user says "/spec <product> <description>", "/spec --edit path/to/spec.md", or "/spec --validate path/to/spec.md".

View, add, edit, remove, export, and import behavioral preferences stored in correction_groups. Use when the user says "/prefs", "/prefs list", "/prefs add", "/prefs edit", "/prefs remove", "/prefs export", or "/prefs import".

Switch the active hook profile (minimal, standard, strict) or show the current one. Use when the user says "/hook-profile", "/hook-profile strict", "/hook-profile minimal", "/hook-profile standard", "change hook profile", or "show hook profile".

Bird's-eye roadmap view with milestone progress, at-risk detection, and quick actions (check-off, archive, add, target, reorder, health). Use when the user says "/roadmap", "/roadmap check story-NNN", "/roadmap archive story-NNN", "/roadmap add 'Title'", "/roadmap target epic-NNN YYYY-MM-DD", "/roadmap reorder", or "/roadmap health".

Friction and skill health dashboard: friction events by category and trend, skill invocation counts, outcome correlation, and pattern detection. Use when the user says "/skill-health", "skill health", "show friction", or "show skill metrics".

Adversarial Claude-vs-Gemini debate tool. Claude and Gemini form independent positions, then debate back-and-forth until convergence or the round cap. Claude participates as a genuine debater — not a summarizer. Produces a battle-tested synthesis written to ~/.claude/arguments/finals/<slug>.md. Use when the user says "/argue <topic>", "/argue plan <topic>", "/argue bug <symptom>", or "/argue tech <question>".

Facilitated BrainLift creation: domain research via Gemini + WebSearch first, then orientation and Socratic prompts that draw out the human's genuine DOK 3 insights and DOK 4 SPOVs. Assembles a complete BrainLift document. Use when the user says "/brainlift <topic>", "/brainlift refine <path>", or "/brainlift --deep <topic>".

Core project research pipeline: routes /scope -> /clarify+/scout (parallel) -> /briefing. No web search — that's /research's job. Produces a structured briefing (presearch/<slug>.md) for /ship. Use when the user says "/presearch <topic>", "/presearch path/to/requirements.md", "/presearch presearch/existing-briefing.md", or "/presearch --deep <topic>".

Deep web research + full presearch pipeline. Runs /scope, then fan-out/fan-in web research (Claude + Gemini in parallel), then calls /presearch with the research artifact to complete the pipeline (clarify+scout -> briefing). Use when web research is needed before building. Use when the user says "/research <topic>", "/research --scope .scope-foo.json <topic>", or "/research --deep <topic>".

Unified detect-diagnose-fix-revalidate loop. Iteratively validates code in a worktree via a strictly gated validation pyramid (compile -> lint -> tests), launches fix agents to resolve failures, tracks error hashes for oscillation detection, and escalates models (Sonnet -> Opus) when fixes stall. Atomic git commits per passing layer enable granular rollback. Returns DONE/NEED_DECISION/BLOCKED. Callable by /run-stories, /quickfix, and standalone via "/fix-loop --worktree-path <path>".

Run Gemini planning on a story/epic, then immediately convert the result into Claude-style plan files. Use when the user says "/draft-plan", "plan this story", "ask gemini to plan story-NNN", "plan epic-NNN", or any natural-language request to plan a story or epic.

Capture lightweight project definition, detect existing stack, and classify domain complexity to determine whether deep research is needed before clarify. Use when the user says "/scope <topic>" or "/scope --skip-qa <topic>".

Seed OpenMemory with existing knowledge from key-prompts, decisions, tool learnings, and behavioral prefs. Safe to re-run — uses per-entry dedup.

Scan plan files for external service dependencies and present an env var checklist. Accepts a .ship-manifest.json path, space-separated plan file paths, or is called by /ship after Step 3c. Skips silently when no external services are detected. Use when the user says "/env-preflight .ship-manifest.json", "/env-preflight plans/foo.md plans/bar.md", or called from /ship pipeline.

Symlink the current npm project's local source as the global install, or restore the published version. Use when the user says "/link-local", "/link-local unlink", or "test local package".

Two-phase todo queue: append raw items to .claude/todos.md, then bulk-plan them via Gemini's pm_plan_items. Use when the user says "/todo <items>", "/todo plan", "/todo list", or "/todo clear".

Audit a Flutter codebase and write a structured AUDIT.md report via a foreground Sonnet agent. Use when the user says "/flutter-audit", "flutter audit the codebase", "flutter audit story-NNN", or "flutter audit <path>". Supports all flags from /audit plus --state-mgmt for explicit state management pattern selection and --opus to force Opus model. Section filters include: security, bugs, completeness, quality, widgets, state, performance, platform, pubspec.

Archive specific stories or epics by ID, delete merged git branches, and prune stale worktrees. Use when the user says "/cleanup", "/cleanup story-NNN", "/cleanup epic-NNN", "/cleanup git", or "/cleanup all".

Run the project linter against a story's worktree. Quick pre-merge sanity check without running the full unit-tester pipeline. Use when the user says "/lint", "/lint story-X", or "lint story X".

Cross-reference git worktree state against the DB to find and fix mismatches. Detects orphaned worktrees, stale DB records, and stories stuck in wrong states. Use when the user says "/recover", "recover", or "fix worktree state".

Find the root cause of a bug using Gemini's large context window. Returns a structured diagnosis with most likely location, contributing factors, reproduction steps, and fix direction. Use when the user says "/find-bug", "find bug", or "find-bug <symptom>". Delegates to the Gemini MCP find_bug tool for large-context root-cause analysis. Supports --model flag.