Multi-agent analysis and planning toolkit for Claude Code. Each skill fans out parallel specialists, collects their reports, and synthesizes a unified result.
claude plugins add-marketplace KwanwooLee36/dispatch
claude plugins install dispatch| Skill | Purpose | Specialists | Key Output |
|---|---|---|---|
| skeptic | Adversarial review | 8 domain critics + synthesis | Scored report, findings by severity, improvement plan |
| recon | Competitive deep-dive | Per-competitor analysts | Gap matrix, differentiators, prioritized backlog |
| tribunal | Decision analysis | Per-option advocates + synthesis | Comparison matrix, recommendation with confidence |
| cartograph | Codebase mapping | 5 dimension specialists | Developer guide with architecture, data flow, conventions |
| landscape | Domain research | 4 research dimensions | Solutions catalog, opportunity ranking |
| migrate | Migration planning | 5 risk specialists + synthesis | Go/no-go, risk matrix, effort estimate |
Every specialist is biased against the project. They hunt for flaws within their domain, argue the project is broken, and provide no fixes. Synthesis merges their reports, scores the project 0-100 per category, and triages findings.
/dispatch:skeptic # Interactive menu — choose agents
/dispatch:skeptic full # All 8 agents — maximum coverage
/dispatch:skeptic quick # All 8 agents — capped exploration for speed
/dispatch:skeptic fix # Auto-fix Actionable Now items from latest report
/dispatch:skeptic arch # Architecture only
/dispatch:skeptic arch code security # Multiple specific agents
/dispatch:skeptic plan # Strategic improvement plan from findings
/dispatch:skeptic plan <type> # Category plan: arch|design|code|security|perf|dx|test|concept|debt
| Agent | Hunts For | Model |
|---|---|---|
| Architecture | Coupling, circular deps, wrong abstractions, scaling walls | Opus |
| Design | Leaky abstractions, pattern misuse, unclear intent, premature generalization | Opus |
| Code Quality | Deprecated deps, dead code, complexity, duplication, naming | Sonnet |
| Security | Injection, auth bypass, secrets, insecure defaults, OWASP top 10 | Opus |
| Performance | N+1 queries, allocations, missing caching, resource leaks | Sonnet |
| DX/Ergonomics | Confusing APIs, missing docs, bad error messages, onboarding friction, broken UX tooling | Sonnet |
| Concept & Strategy | Value prop, market fit, positioning, feasibility, differentiation | Opus |
| Testing | Test quality, coverage gaps, meaningless tests, flaky patterns, missing test types | Sonnet |
When the Concept & Strategy agent identifies competitors, skeptic prompts you: "Run recon?" If you say yes, competitive research agents dispatch immediately using the same research protocol as /dispatch:recon. Recon findings feed into synthesis, generating new findings from competitive gaps, missing moats, and positioning failures alongside the specialist critiques, and are written to .skeptic/recon-YYYY-MM-DD.md so a later /dispatch:skeptic plan can load them. Standalone /dispatch:recon additionally produces design notes in .skeptic/recon-designs/; the inline path does not.
/dispatch:skeptic plan reads existing recon (.skeptic/recon-*.md) and landscape (.landscape/survey-*.md) reports when they exist. Plan agents reference competitor approaches, flag table-stakes gaps, and note moat opportunities. If no reports exist, it proceeds without them.
Reports land in .skeptic/report-YYYY-MM-DD.md. Console prints the verdict, scores (0-100 per category), and finding counts. Severity levels: FATAL, MAJOR, MINOR, NITPICK. Future Work items route to docs/roadmap.md milestones when a roadmap exists, falling back to TODO.md backlog otherwise.
Reports accumulate in .skeptic/. Synthesis detects issues flagged in prior reviews, marks them as repeat offenders, escalates severity after 3+ appearances, and calculates score trends.
Reads competitor names from a prior /dispatch:skeptic run, researches each in depth, and synthesizes into a gap matrix, differentiator map, and backlog items with design suggestions.
/dispatch:recon # Full mode — one Opus agent per competitor, parallel
/dispatch:recon quick # Quick mode — single Sonnet agent, all competitors sequential
Requires a skeptic report with Concept & Strategy findings that name competitors. Produces per-competitor deep dives (overview, strengths, weaknesses, feature inventory), a gap matrix tagged STEAL/CONSIDER/IRRELEVANT, differentiator analysis (existing moat, potential moat, table stakes), and prioritized recommendations. Design notes go to .skeptic/recon-designs/. Gap items route to docs/roadmap.md milestones when a roadmap exists, falling back to TODO.md backlog otherwise.
Each advocate argues FOR one option with evidence. Synthesis identifies table stakes, conflicts, and recommends the strongest path forward.
/dispatch:tribunal "Postgres vs DynamoDB"
/dispatch:tribunal "monolith vs microservices vs modular monolith"
/dispatch:tribunal # Interactive prompt for decision and options
Requires 2-4 options. Project context (codebase) is mandatory so decisions stay grounded. Produces per-option advocate reports, a comparison matrix, and a recommendation with confidence assessment.
Five dimension specialists analyze structure, data flow, conventions, infrastructure, and domain logic in parallel. Produces a single unified developer guide.
/dispatch:cartograph # Full analysis, all dimensions
/dispatch:cartograph quick # Quick scan (all Sonnet, 15-file cap)
/dispatch:cartograph focus:auth # Focus all dimensions on auth subsystem
/dispatch:cartograph focus:api quick # Focus + quick mode
| Dimension | Lens | Model |
|---|---|---|
| Structure | Module boundaries, entry points, dependency graph | Opus |
| Data Flow | APIs, state, storage, transformations, middleware chains | Opus |
| Conventions | Naming patterns, style, idioms, error handling, testing | Sonnet |
| Infrastructure | Build, deploy, CI/CD, containers, external services | Sonnet |
| Domain | Business logic, key entities, workflows, domain vocabulary | Sonnet |
Output: docs/cartograph.md.
Four parallel research agents map solutions, approaches, community insights, and business dynamics for a given domain. No codebase dependency.
/dispatch:landscape "real-time collaboration tools" # Full mode — 4 agents, broad research
/dispatch:landscape quick "state management" # Quick mode — focused research
/dispatch:landscape # Interactive — prompts for domain
Domain string is 2-10 words; whole-category inputs ("technology", "database") are rejected with narrower suggestions, and a longer-than-10-word domain warns rather than fails. Produces a solution catalog with positioning and pricing, approaches ranked by adoption, and an opportunity ranking of gaps and emerging patterns. Opportunity items route to docs/roadmap.md milestones when a roadmap exists, falling back to TODO.md backlog otherwise.
Five parallel specialists investigate API compatibility, dependencies, tests, infrastructure, and data layer. Cross-cutting landmine detection with a go/no-go recommendation.
/dispatch:migrate "Express to Fastify" # Full analysis
/dispatch:migrate quick "webpack to vite" # Quick mode (Sonnet agents)
/dispatch:migrate # Interactive prompt for migration intent
| Dimension | Model | Investigates |
|---|---|---|
| API Surface | Opus | Compatibility, transformation rules, breaking changes |
| Dependencies | Sonnet | Package ecosystem, replacement research, version alignment |
| Tests | Sonnet | Framework migration, coverage gaps, effort estimate |
| Infrastructure | Sonnet | CI/CD, deployment, monitoring changes |
| Data | Opus | Schema, ORM, data access patterns, rollback strategy |
Console prints go/no-go verdict, top 5 risks, and effort estimate. Full plan goes to docs/migration-YYYY-MM-DD-{slug}.md. Migration steps route to docs/roadmap.md milestones when a roadmap exists, falling back to TODO.md backlog otherwise — always after you confirm.
Five skills — skeptic (via /dispatch:skeptic plan), cartograph, landscape, migrate, and tribunal — can persist their primary output as a design doc in docs/designs/. On the first time that directory is created in a project, they append a ## Design Docs orientation block to the project's CLAUDE.md, so future sessions (switch-in, dian, or any session start) read docs/designs/INDEX.md and open individual docs on demand, inheriting the context without paying for the whole directory.
Recon is the exception — it does not use docs/designs/ and does not auto-inject into CLAUDE.md. Recon's design notes are fragmentary: one note per gap tagged STEAL in its Inbound Gap Matrix. It writes each to .skeptic/recon-designs/[gap-name].md and links them from the main recon report (.skeptic/recon-YYYY-MM-DD.md). This keeps recon's many small, competitive-intelligence fragments co-located with the skeptic/recon reports they derive from, rather than scattering them across docs/designs/ and bloating the always-read CLAUDE.md auto-inject. Recon surfaces these notes through its report and its routed backlog/roadmap items instead.
Works standalone. Integrates with Kerd when available: reviews noted in session logs (dian), findings persist to vault (kivna), and skeptic complements slainte's factual audits with adversarial interpretation.
MIT