shipwright
CLAUDE.md16 starsChanged 24 days agoMITView on GitHub
Shipwright Harness is the open-source (MIT) autonomous delivery agent for Claude Code — a deployable agent plus the autonomous coding system that powers it. The system is a Claude Code plugin (shipwright) covering the full delivery loop — spec → plan → execute → review → deploy — alongside a metrics dashboard and a Shipwright agent. The brand is Shipwright Harness; the installable package is shipwright. The plugin is repo-agnostic: it runs its planning/execution/review/merge commands against any repository. This repo…
- Reads credentials
In this repository3 other files it gives its agents.
- marketplace-devSkill
- shipwright-brandSkill
- verifySkill
The file
We split this CLAUDE.md into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 57
Run the metrics service locally:
- Very specific: line 167
Universally-meaningful third-party vars: keep conventional names: GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET, PORT.
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
What Shipwright is
Shipwright Harness is the open-source (MIT) autonomous delivery agent for Claude Code — a deployable agent plus the autonomous coding system that powers it. The system is a Claude Code plugin (shipwright) covering the full delivery loop — spec → plan → execute → review → deploy — alongside a metrics dashboard and a Shipwright agent. The brand is Shipwright Harness; the installable package is shipwright.
The plugin is repo-agnostic: it runs its planning/execution/review/merge commands against any repository. This repo is both the source of the plugin and the codebase it ships against.
Architecture — four artifacts, sequenced A → B → C → D
| Phase | Artifact | Directory | What it is |
| A | Plugin (the system) | plugins/shipwright/ | Commands, skills, agents, scripts users /plugin install. Repo-agnostic. |
| B | Metrics dashboard | metrics/ | Stateless Hono service: task-store-backed JSON endpoints + a server-rendered dashboard. No database. |
| C | Shipwright agent | agent/ | Hono service + Prisma store; a thin autonomous runner: pick next ready task → build → ship PR → forward metrics. |
| D | Task store service | task-store/ | Postgres-backed task queue, PR tracking, and scoped tokens. Prisma schema defines Task, PullRequest, and TaskToken models; re-exported as @shipwright/task-store. Replaces the JSON file fallback. |
Supporting surfaces (not phased):
site/— Astro + Tailwind marketing site (shipwrightharness.com). Self-contained; not a Bun workspace; Playwright smoke tests.brand/— locked design system (BRAND.md,tokens.json) + CSS build + lint, consumed bysite/and brand artifacts. Editing brand artifacts triggers theshipwright-brandskill.state/— git-ignored local JSON task-store fallback + cached review state (only written when the GitHub backend isn't active).
Commands
go-task (Taskfile.yml) is the single local entrypoint; the root package.json mirrors a subset as bun run scripts.
task setup # bun install across all workspaces
task ci # lint → check-strings → typecheck → check-config-docs → check-version-sync → test:coverage → secret-scan (the merge-blocking gate; CI runs this exact chain)
task test # bun test (single file: bun test path/to/file.test.ts)
task test:coverage # bun test --coverage --coverage-reporter=lcov, then gate on aggregate 90/89 lines/functions (scripts/check-coverage.ts)
task lint # bunx biome lint .
task format # bunx biome format --write .
task typecheck # bun run --filter='*' --sequential typecheck
task check-strings # scan entire repo for banned/confidential identifiers (client names, internal infra IDs)The agent platform (runtime, admin, shared lib) lives in app-vitals/agent and is consumed as pinned @app-vitals/agent-* packages plus the ghcr.io/app-vitals/agent base image; to develop both together see docs/local-platform-development.md.
Database (admin service):
export DATABASE_URL_SHIPWRIGHT_ADMIN="postgresql://user:password@localhost:5432/shipwright_admin"
task db:provision # prisma migrate deploy (idempotent) — schema + migrations ship in @app-vitals/agent-adminMarketing site (run from site/, not part of bun test):
cd site && npm run dev # astro dev
cd site && npm run build # astro build
cd site && npm test # playwright test (*.spec.ts)
bunfig.tomlexcludessite/**from the rootbun testscan — Playwright's*.spec.tsfiles would otherwise crash Bun's runner. Keep site tests as*.spec.tsto stay isolated.
Run the metrics service locally:
task api # start metrics dashboard in offline mode → http://localhost:3460/dashboard
task ui # same as task api (API and UI are one process)
task dev # dev supervisor: starts metrics + Ctrl-C kills all children
task stack # full dev stack in a tmux session (6 panes) — requires tmux + Docker + state/dev-agent.env
task hitl # human-in-the-loop runner: boots task-store (:3002) + admin (:3001), then loops — pulls the next ready task and launches Claude Code with the right command
task agent-workspace-pull -- <id-or-name> # mirror a real, already-deployed agent's workspace locally (clones its repos, installs its plugins) — no dispatch loop. Requires SHIPWRIGHT_API_URL + SHIPWRIGHT_ADMIN_API_KEY; optional SHIPWRIGHT_WORKSPACE_PULL_ROOT. The task defaults SHIPWRIGHT_SPECIALTIES_DIR (symlinking agent/workspace templates) and SHIPWRIGHT_DEFAULT_AGENT_TYPE=coding when unset
task store:local # boot a local task store (Postgres + published image via Docker) and print the SHIPWRIGHT_TASK_STORE_URL/TOKEN exports; `-- --down` removes it (wraps scripts/local-store.sh)task stack needs state/dev-agent.env — if it's missing, the first task stack run auto-creates it from state/dev-agent.env.example and exits; fill in CLAUDE_CODE_OAUTH_TOKEN (or ANTHROPIC_API_KEY) in the newly-created file, then re-run task stack. See docs/quickstart.md for the full flow.
task stack (scripts/dev-tmux.ts) launches one tmux session named shipwright with a 6-pane dashboard: metrics (SQLite, :3460), admin (CRUD API + UI, :3001), task-store (:3002), chat-svc (the platform-hosted chat service from app-vitals/agent, :3003 — run from a local platform checkout at $AGENT_PLATFORM_DIR, default ../agent, or from the ghcr.io/app-vitals/agent-chat image when SHIPWRIGHT_CHAT_IMAGE is set; see docs/local-platform-development.md), agent in Docker with the chat poll loop enabled, and a scratch logs shell. task stack creates zero agents automatically: the agent pane prints a pointer to http://localhost:3001/admin/agents/new and polls the admin DB (scripts/wait-for-agent.ts) until a developer creates one via the UI, then seeds that agent's chat token and starts its container with SHIPWRIGHT_AGENT_ID set to the resolved id — no stack restart required, and relaunching with an agent already created proceeds immediately. Chatting with the agent happens in the browser via the admin console's Chat tab (/admin/chat). It runs a Prisma migrate deploy preflight before the admin pane so the admin service's Postgres schema is up to date; the preflight first checks Postgres is reachable and, on macOS, prints the exact brew/createdb commands and offers to run them ([y/N]) before launching. The agent pane builds the Docker image from agent/Dockerfile and runs it with --env-file state/dev-agent.env so secrets are injected at runtime without appearing in the command or the image. Closing the session (tmux kill-session -t shipwright) stops every pane. task stack is additive — it does not touch task dev, which stays the no-tmux fallback the quickstart depends on; if tmux isn't installed, task stack fails fast and points you at task dev. The command/pane-env sequence is built by a pure, injected-exec builder (mirrors scripts/dev.ts) and unit-tested in scripts/dev-tmux.unit.test.ts. scripts/seed-dev-agent.ts still exists as a standalone, explicitly-invoked convenience script — task stack no longer calls it.
Shipwright Learned Facts
Auto-maintained by Shipwright's toolchain detection — edits here are overwritten on the next run.
- Scoped-command variants:
lintScopeddetected —git diff --name-only --diff-filter=ACMR {base}...{head} -- '*.ts' '*.tsx' '*.js' '*.jsx' | xargs -r bunx biome lint, which scopes lint to changed JS/TS files instead of the whole repo. NotypecheckScoped/testScopedvariant is available:task typechecktype-checks the whole project graph by design, and there's no monorepo affected-graph tool wired in, so both run full-repo. - Enforced per-check verification budget:
600s(fallback-10m— CI-derived numbers for this repo'slint / typecheck / testjob were observed non-representative of a fresh sandbox checkout, and an earlier 150s estimate padded from CI's 91s proved undersized against a full localtask testrun at ~236s, so the flat 10-minute constant is used instead). Skip-locally classifications: none yet.
Check Reason Classified At (none)
Before you commit — this repository is going public
This repo is private today but destined to be a public, MIT open-source project. Git history is permanent. Scrub before the commit, not after the push.
The rule: review every change before staging it. Stage specific files — never git add -A/-u blindly. When unsure whether something is proprietary, ask before committing.
Scrub for: secrets & credentials (ANTHROPIC_API_KEY, GH_TOKEN, SESSION_SECRET, .env contents, private keys) · client/customer/partner names · internal infra identifiers (cloud project names, analytics project IDs, internal hostnames) · internal PR/issue/Slack/Jira links · local filesystem paths revealing usernames (/Users/<name>/...) · financials, compensation, PII · the identity of whoever operates/deploys this shipwright instance (e.g. an employer/company name), even when it arrives verbatim in an inbound task-store task description (such as an internal org/repo citation like org-name/repo#123).
task check-strings (banned-strings scan) is the CI backstop — not a substitute for this discipline. Internal, build-time-only notes live in the git-ignored CLAUDE.local.md; read it for operational context before working in this repo.
How work is tracked
Two surfaces — do not conflate them:
- Automated task store (what
/shipwright:dev-taskand the agents read/write): the HTTP Shipwright task-store service (artifact D). Individual tasks likePV-1.2live here. Query it directly —gh issue listdoes not see these tasks.curl -sf -H "Authorization: Bearer $SHIPWRIGHT_TASK_STORE_TOKEN" \ "$SHIPWRIGHT_TASK_STORE_URL/tasks?ready=true" | jq '.tasks' - Manual human planning (historical): GitHub Issues under the
shipwright-ossmilestone, each with a machine-readable```shipwrightYAML block and astatus:*label. A record/planning surface for humans; the automated loop does not read it.gh issue list --milestone shipwright-oss --state open --label status:pending
Status lifecycle:
pending → in_progress → pr_open → merged → deployed | releasedplus approved, done, blocked, cancelled.
Tasks sharing a non-blank session field (set only by /shipwright:plan-session) roll up into a
Session — browsable in the admin UI's Sessions tab (/admin/sessions), which shows per-session
task status and follow/notification controls. A background alert sweeper pushes "waiting" / reminder /
"completed" notifications to session followers; see docs/task-store.md
and docs/agent-web-ui.md for the full model.
Execution loop
- Pick a
pendingtask whose everydependenciesentry is done. - Branch from the task's
branchfield (feat/sw-x-y-slug) — never work onmain. - Build + land tests in the same PR, at the correct layer (no "tests later").
- Open a PR; move the status through its lifecycle.
Driven by Shipwright's own commands: /shipwright:dev-task → /shipwright:review / /shipwright:patch → /shipwright:merge. The shipwright-loop cron is the sole supported autonomous driver — it drains the queue by dispatching enabled pipeline phases (dev-task, plan, review, patch, merge) in a single multi-step run until work is exhausted. The plan phase (PDR-4.1, /shipwright:plan-session --autonomous) ships disabled and is double-gated: it additionally requires the SHIPWRIGHT_AGENT_AUTONOMOUS_PLAN_SESSION_ENABLED env var, so with that unset the loop behaves exactly as the four-phase loop did.
Task-store connection is env-var-only: both SHIPWRIGHT_TASK_STORE_URL and SHIPWRIGHT_TASK_STORE_TOKEN must be set for task operations to function. There is no GitHub fallback and no file-based config — the provisioner injects these two vars into managed GKE agents, and local installs must set them explicitly. If task operations seem to no-op, check SHIPWRIGHT_TASK_STORE_URL first.
Deploy model
direct — ships via Helm chart + GHCR-pinned image tags, with no staging/production GitHub Environments and no deploy-staging→canary→promote-to-production pipeline.
Test conventions
Tests land with the code, at the correct layer — same PR, no "add tests later" tasks. Layer is encoded in the filename:
| Suffix | Layer | What it covers |
*.unit.test.ts | unit | pure logic, no I/O |
*.integration.test.ts | integration | real dependency behavior via recorded fixtures / injected doubles |
*.smoke.test.ts | smoke | HTTP route contracts — Hono endpoints via in-process app.request() (no real socket); a real Bun.serve() boot is permitted only when there's no in-process request seam (e.g. a bare-Bun.serve() server with no Hono app factory) |
*.spec.ts (in site/) | e2e | the site in a real browser via Playwright |
*.content.test.ts | content | markdown/prompt-content-assertion tests — e.g. checking that a command's or skill's Markdown body contains expected sections/instructions/wording. Distinct from unit/integration/smoke: no real I/O boundary, just asserting on static content. |
Test isolation (hard rule): inject time via a Clock; test external clients (task store, GitHub) with recorded fixtures. No mock.module(), no global.fetch/global.* overrides — Bun shares the test process, so leaked globals break sibling suites.
Hidden-directory gotcha: bun test's recursive walker never descends into a dot-prefixed directory (e.g. .claude/) — no bunfig.toml setting or CLI flag changes this (confirmed against bun test --help: --path-ignore-patterns only removes matches from an existing walk, it can't add dot-dirs back in). A test file placed there (e.g. .claude/commands/*.content.test.ts) silently never runs under plain bun test or task ci. scripts/find-hidden-dir-tests.ts finds such files via Bun.Glob's dot: true option and runs each one as an explicit ./-prefixed path (bun resolves an explicit path directly regardless of the walker) — wired into task test and task ci (via test:coverage) so this runs automatically; no per-file setup needed for a new hidden-directory test.
Conventions
- No new coupling: the plugin stays repo-agnostic; the metrics service and the agent depend on no external platform service.
- Local-first: everything runs offline by default (fixtures / injected doubles / scratch queue); live external calls only when env explicitly enables them.
- Conventional Commits — required (a
pr-title-lintworkflow enforces PR titles); releases are automated via semantic-release. - Lint/format: Biome (2-space indent, organize-imports). Run
task lintbefore committing. - License: MIT across all artifacts.
Env var namespacing convention
Env vars read by Shipwright services are namespaced so each service is portable across any infra:
- Suite-wide:
SHIPWRIGHT_<THING>— e.g.SHIPWRIGHT_SESSION_SECRET,SHIPWRIGHT_ENCRYPTION_KEY,SHIPWRIGHT_AGENT_API_KEY. - Per-subservice:
SHIPWRIGHT_<SUBSERVICE>_<THING>— e.g.SHIPWRIGHT_ADMIN_ALLOWED_EMAILS,SHIPWRIGHT_ADMIN_APP_BASE_URL. - DB connection strings:
DATABASE_URL_SHIPWRIGHT_<SUBSERVICE>— never bareDATABASE_URL(collides with the host's own DB var). - Universally-meaningful third-party vars keep conventional names:
GOOGLE_CLIENT_ID,GOOGLE_CLIENT_SECRET,PORT. - Secret manager IDs map 1:1 to env var names: lowercase-kebab ↔ uppercase-snake (e.g.
shipwright-admin-allowed-emails↔SHIPWRIGHT_ADMIN_ALLOWED_EMAILS).
Database env vars
Each Prisma service reads its own DATABASE_URL_* — never a shared connection.
| Variable | Service | Schema |
DATABASE_URL_SHIPWRIGHT_ADMIN | @shipwright/admin | prisma/schema/ in @app-vitals/agent-admin (via admin/prisma.config.ts) |
DATABASE_URL_SHIPWRIGHT_TASK_STORE | @shipwright/task-store | task-store/prisma/schema.prisma |
DATABASE_URL_SHIPWRIGHT_CHAT | platform-hosted chat service (agent-chat image) | schema owned by app-vitals/agent |
The schema uses provider = "postgresql". All database connection strings must be Postgres connection strings: postgresql://user:password@host:5432/database.
For the full configuration reference (all env vars, agent config, policy config), see docs/configuration.md (agent runtime env vars split into docs/configuration-agent.md).
Debugging
- A 401 with no
WWW-Authenticateheader is not the auth middleware.createAdminAuthMiddlewarealways setsWWW-Authenticateon its bearer-path 401s. A{"error":"Unauthorized"}401 without that header is surfaced via the app'sonErrorfrom a downstream call inside a handler — typically aKubernetesClientrequest the K8s API rejected. Tell-tale: read-only routes (DB only) return 200 while a route that calls the K8s API (agentreconcile/provision) 401s with the same token — that's a calls-the-dependency correlation, not an auth or routing bug. Check the response header (curl -sD - -o /dev/null …) before suspecting the token. Common cause of the K8s 401: the pod's ServiceAccount was deleted+recreated (new UID), invalidating the bound token in the already-running pod even though itsexpis far in the future — restart the pod, and don't render the SA as a Helm hook (hooks recreate it everyhelm upgrade).
Reference
To load additional context into a session, add @docs/filename.md entries here — don't create separate CLAUDE-REFERENCE.md or similar files.
- docs/architecture.md — the four-artifact A→B→C→D design (plugin / metrics / agent / task-store), supporting surfaces, and workspace layout
- docs/testing.md — the five-layer test model (unit / integration / smoke / e2e / content), run commands, speed budgets, and the isolation contract
- docs/metrics.md — metrics service (B): JSON endpoints, server-rendered dashboard, dual auth (Bearer / session), and environment
- docs/agent.md — Shipwright agent (C): run modes, HTTP surfaces (runtime + admin CRUD APIs), the sixteen-model Prisma store, and encryption notes
- docs/agent-web-ui.md — Shipwright agent (C) browser-facing surfaces: admin chat UI, sessions list UI, session alert sweeper, public read-only task board, dev auto-login, and PWA shell — split out of docs/agent.md to stay under the docs line-count threshold
- docs/configuration-agent.md — Shipwright agent (C) runtime env var reference (Claude/Anthropic, Slack, GitHub, Shipwright platform, server, database, provisioning, workspace/tooling, voice, dev-only) — split out of docs/configuration.md to stay under the docs line-count threshold
- docs/agent-ops.md — Shipwright agent (C) operations: tool management/narrowing (floor vs. allowed tools), default system crons, environment variables, and baked marketplaces
- docs/local-platform-development.md — working on app-vitals/agent and Shipwright together:
bun linkof@app-vitals/agent-{runtime,admin,lib}, which dev flows (task dev,task hitl, seed-dev-agent,task stack, minikube) pick the link up, running a local platform image, and how a platform release reaches Shipwright - docs/agent-key-files.md — Shipwright agent (C) key files: per-file reference table for
admin/srcandagent/src - docs/platform/agent-api.md — the admin CRUD API (D): agent CRUD, auth paths, env vars, and runtime config
- docs/platform/agent-api-ops.md — the admin CRUD API (D), operational surfaces: cron jobs and cron run endpoints (create/update/delete, system-cron reconciliation, run records and token/cost stats)
- docs/agent-api-resources.md — the admin CRUD API (D), per-agent resources: allowed-tools list, API tokens, plugins, chat-token-usage endpoints, and the work-queue snapshot
- docs/platform/github-multi-org.md — multi-org GitHub App: one App installed on several orgs, discovery and per-owner token routing, operator and org-admin setup, the admin installations card, and troubleshooting
- docs/platform/github-enterprise-apps.md — enterprise GitHub apps (optional, per deployment): the installer and lookup Apps, exact permissions, setup, why there is no shared published App, the admin isolation exception, key storage, and the admin settings page
- docs/extending.md — extending Shipwright without forking it: installing a companion plugin, command namespacing, custom crons for repo-specific automation, and the lightweight
.claude/shipwright/override pattern - docs/task-store.md — task store service (D): the sole HTTP-service backend — tasks, PR tracking, tokens — and troubleshooting
- docs/task-store-pr-tracking.md — task store PR origin metrics, verification checks, skip tracking, same-branch exclusivity guard, and session archive sweep
- docs/mcp-tools.md — generated MCP server tool reference (name, description, HTTP method/path, parameters, body) derived from task-store/openapi.json + the tool allowlist; regenerate with
bun run generate:mcp-docs - docs/chat.md — pointer: the chat service now lives in the platform repo (
agent-chatimage); details indocs/platform/chat-service.md - docs/session-notifications.md — session-follow subsystem (C): the follow/mute + notification-prefs data model, the session-notification detail-level policy (operator ceiling × per-subscription opt-in × per-call ceiling), the
/admin/settings/notificationspage and follow/unfollow routes, and thesession-scope.tsvisibility model (not yet wired into most routes) - docs/deploy-kubernetes.md — Kubernetes deployment guide: auth modes, plus pointers to the networking model, cloud provider guides, agent provisioning model, and optional add-ons
- docs/deploy-kubernetes-networking.md — Kubernetes networking model:
networking.typemodes, exposing the task-store externally, and the opt-in chat and MCP server services — split out of docs/deploy-kubernetes.md to stay under the docs line-count threshold - docs/deploy-kubernetes-providers.md — per-provider Kubernetes deployment walkthroughs: Minikube (local), GKE (Gateway API + cert-manager), EKS (ALB), Traefik, and cloud-native (any cluster)
- docs/deploy-kubernetes-provisioning.md — Kubernetes agent runtime provisioning model: RBAC, ServiceAccounts, the provisioner env contract, provisioning values, chat-service token provisioning, and the task-store claim TTL constraint — split out of docs/deploy-kubernetes.md to stay under the docs line-count threshold
- docs/platform/anthropic-usage-by-agent.md — IT runbook: split the Anthropic usage report by per-agent service account and map each
svac_…ID to its agent name and owner - docs/deploy-kubernetes-addons.md — optional Kubernetes deploy add-ons: agent voice (STT/TTS), Web Push notifications, bringing your own PostgreSQL, and bundled ingress-controllers/cert-manager
- docs/helm-repo.md — installing the published
shipwrightHelm chart, how chart publishing and version bumps are automated - docs/quickstart.md — local onboarding: metrics-only quickstart and the full dev stack (
task stack) - docs/test-readiness/test-system.md — the authoritative test blueprint: layer matrix, boundary rules, per-component budgets, CI pipeline shape, and the full isolation contract
- docs/test-readiness/naming.md — file-suffix → layer naming convention and the
bunfig.tomlrunner-exclusion config (e2e + reserved canary suffix), gating M1 rename tasks - docs/migration.md — breaking changes and migration steps across versions (e.g.
AgentProvisioner.reconcile()interface change) - docs/observability.md — Sentry error/log reporting: what's collected, what's scrubbed, how to disable, and self-hosted Sentry support
- docs/consolidation.md — consolidation-patrol system: the ledger format, the decisions registry, and the cron (disabled by default, opt-in per agent)
- docs/prompt-audit.md — prompt-audit patrol:
prompt-scanfinding classes and report-only contract, the propose-onlyprompt-fixtask filer and its guardrails, the ledger and decisions registry, thecheck-prompt-audit.tsprecheck, the disabled-by-defaultprompt-audit-maintenancecron, and cost/auth/version gates - docs/prompt-audit-calibration.md — the signed calibration record that gates the prompt-audit eval tier's quality claims: arms, paired McNemar/MDE results, known-bad detection, A/A noise, replayability of dev-task/review/patch under
claude plugin eval, and the current FAIL verdict (tier off, cost-only findings) - docs/suppressions.md — finding suppressions:
.shipwright/suppressions.ymlmatch keys,false_positivevsaccepted_riskexpiry, which scan/fix skills honor it, how expired entries resurface, and the hitl proposal flow - docs/site-docs-freshness.md — site-docs-freshness (SDR): the
site/docs-source-map.jsonformat and its structural test enforcement, the page-scopedcheck-site-docs-freshness.tsprecheck and its exit-code contract, and current cron-wiring status - docs/test-readiness/ — the test-readiness pipeline: Phase 1 inventory (current test surface), Phase 2 ideal system design (blueprint), Phase 3 migration plan (from status quo to blueprint), Phase 4 executable roadmap (concrete tasks). The suite includes the test-readiness decisions registry (
.claude/shipwright/test-readiness-decisions.md) — same tier asconsolidation-decisions.md— for recording human-reviewed coverage decisions.test-inventoryandtest-fixboth consume the registry to avoid re-flagging already-resolved ambiguous items — see Step 0 intest-inventory/SKILL.mdand Step 4.5 intest-fix/SKILL.mdfor implementation details. - docs/platform/specialties.md — authoring guide for Specialty manifests: field reference, blank template, worked examples, the FLOOR_TOOLS/no-secrets/crons-disabled rules, and a "Specified, not implemented" spec for GitHub-hosted custom Specialty source references
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

