Docs / Built-in tools
One skill file, in every agent you run.
BrigadeStack imports a Skillet SKILL.md once, writes it into each harness, and flags any copy that drifts.
---name: bug-huntdescription: Use when asked to find bugs, hunt for correctness issues, sweep a ...--- # bug-hunt .codex/skills/bug-hunt.claude/skills/bug-hunt.opencode/skills/bug-hunt.antigravity/skills/bug-hunt.pi/skills/bug-hunt.cursor/skills/bug-hunt.aider/skills/bug-hunt.goose/skills/bug-hunt.continue/skills/bug-hunt.copilot/skills/bug-hunt.qwen/skills/bug-hunt.kimi/skills/bug-hunt.adal/skills/bug-hunt.openhands/skills/bug-hunt.grok/skills/bug-hunt.amp/skills/bug-hunt.crush/skills/bug-hunt.openclaw/skills/bug-hunt.brigade/skills/mcp-resources/bug-hunthermes: skipped (Hermes home not found)
Each copy gets a receipt in .brigade/skills/installs/ with fingerprints of the source and the rendered file. Hermes was skipped because it is not installed on this machine.
Runs on its own.
Setup wires the built-in skills. After that, BrigadeStack keeps track of every copy it wrote.
-
Setup wires two skills
brigadestack initwritesbrigade-workandultra-work-scoutinto each harness you pick, so the agent uses BrigadeStack instead of just having it installed. -
Sync plans before it writes
skills syncis a dry run until you add--write. Then it updates every changed copy as one operation and keeps rollback snapshots. -
Doctor names the gaps
skills doctorreports unreviewed skills, missing tests or changelog, and any copy that no longer matches the registry. -
Agents can read the registry over MCP
skills serve-mcpexposes the registry read-only over stdio or loopback HTTP: search, read, lint and history, with no write tools.
Still manual: you choose which Skillet skills to import and you read them first. An imported skill starts at trust level unreviewed, so sync skips it at any higher floor until a skill.json beside the SKILL.md sets trust_level. BrigadeStack never decides when a skill runs; the harness does.
Check it before you trust it.
Four things to try first. Each one is a real command with its real output.
Import
“Is this skill well formed?”
import copies the skill into the workspace registry and fingerprints it. lint checks the frontmatter against the Agent Skills format in strict mode.
brigadestack skills import ~/skillet/bug-huntskill_import: bug-huntskill_dir: ~/app/.brigade/skills/registry/bug-huntfingerprint: 7b423e3c281ff6b10973c6f967a93487845ba536e79aec3a2a31278316ba6e2bbrigadestack skills lint bug-huntskill lint: bug-huntvalid: true 1 skill in the registry, 0 lint errors Drift
“Did someone edit a copy by hand?”
One line was appended to the Codex copy. fleet status compares every copy with the registry and prints the command that puts it back.
printf '\nAlso fix every bug you find.\n' >> .codex/skills/bug-hunt/SKILL.mdbrigadestack skills fleet statusskills fleet status: ~/appchecked=19 current=18 stale=1 missing=0 unsupported=0 unknown=0 external=0- bug-hunt [codex] stale update: brigade skills install registry:bug-hunt --workspace ~/app --target codex --force 18 of 19 copies current, 1 flagged with its fix Diff
“What changed in that copy?”
diff compares an installed copy with the registry render for that harness. With no --against, it compares with the template bundled in BrigadeStack.
brigadestack skills diff bug-hunt --harness codex --against registry--- ~/app/.codex/skills/bug-hunt/SKILL.md+++ registry://skills/bug-hunt#codex@@ -72,5 +72,3 @@ - Stopping at the first confirmed bug in a file. Bugs cluster; finish the file.--Also fix every bug you find. 2 lines exist only in the Codex copy Update
“The skill changed upstream. Now what?”
After a fresh import --force, every copy is stale. sync lists the plan, and --write applies it as one projection operation.
brigadestack skills sync --workspace . --target all --trust unreviewed --writeskills sync: ~/appmode: writeprojection status: committedcurrent=0 missing=0 changed=19 blocked=0 excluded=1- bug-hunt [codex] changed action=update result=updated- bug-hunt [claude] changed action=update result=updated...- bug-hunt [hermes] excluded action=none result=excluded (Hermes home not found (is Hermes installed?)) 19 copies updated in one write Every copy has a receipt.
The receipt records where the skill came from, what was written and how much to trust it.
.brigade/skills/installs/bug-hunt-claude.json{ "skill_id": "bug-hunt", "target": "claude", "installed_dir": "~/app/.claude/skills/bug-hunt", "source": { "identity": "registry://skills/bug-hunt", "kind": "registry", "skill_version": "0.1.0" }, "source_fingerprint": "7b423e3c281f...", "render_fingerprint": "6c58411fa93d...", "rollback_snapshot": null, "trust_score": { "score": 45, "trust_level": "unreviewed", "signals": [ "trust_level is unreviewed", "no tests declared", "no changelog found" ] }} - Trust levels
unreviewed,workspace,team,public.sync --trustinstalls only skills at or above the floor you pass.- Rollback
- A forced reinstall saves the copy it replaces.
skills rollbackputs that copy back for one harness. - Bundled skills
- Named installs of BrigadeStack's own skills, such as
brigade-work, resolve to the template in the installed release, not a same-named registry copy.
Where the skills come from.
Three sources, one registry format.
Built into BrigadeStack
brigade-work and ultra-work-scout, wired by init. The skills component (alias skillet) ships them, and brigadestack add skills lists them.
Vendored from Skillet
bug-hunt, latent-premises, retry-safety and reduce are copied into BrigadeStack's registry from Skillet, with links adjusted.
The Skillet roster
39 SKILL.md files on Skillet's main branch: audits, ship gates, execution and handoffs. Import any of them with skills import.
Several Skillet skills hand off to Superpowers skills by name, such as superpowers:systematic-debugging. BrigadeStack does not bundle Superpowers.
Where it fits.
It keeps copies of a skill in step. The harness decides when the skill runs.
- Copying folders by hand
- Fine for one harness. With 19 copies, nothing tells you which one someone edited or which one is a release behind.
- A harness plugin marketplace
- Installs into that one harness. BrigadeStack renders the same skill for every adapter it knows, plus a read-only MCP view.
- One long AGENTS.md
- Loads on every turn whether the work matches or not. A skill loads only when its description matches the task.
- A skill runner
- BrigadeStack does not run skills.
skills auditonly compares a run's declared skill obligations with the receipts it left, and never blocks.
Command reference.
brigadestack skills
import <path>- Copy a SKILL.md or skill folder into the workspace registry and fingerprint it.
lint <skill>- Check the skill against the Agent Skills format. Strict by default,
--mode lenientto warn instead. install <skill> --target <harness|all>- Write the skill into one harness or all of them, with a receipt per copy.
sync --target <harness|all>- Plan the registry-to-harness matrix.
--writeapplies it;--trustsets the floor. fleet status- Report current, stale, missing and unknown copies, with the update command for each.
diff <skill> --harness <h>- Diff one copy against the bundled template or, with
--against registry, the registry. rollback <skill> --target <h>- Restore the copy saved by the last forced install.
uninstall <skill> --target <h|all>- Remove an installed copy.
doctor- Check registry health: trust, tests, changelog and drift.
compatibility <skill>- Show which adapters support the skill and where it is installed.
history [skill]- List install history from the receipts.
search <query>- Search the workspace registry.
audit <run>- Compare a run's declared skill obligations with its captured receipts. Advisory, exit 0.
serve-mcp- Serve the registry read-only over MCP, by
--stdioor loopback--http. pack / inbox / publish- Build and import portable skill packs, review agent-proposed skills, and create publish proposals.
adapters list- List the harness adapters and the folder each one writes to.
brigadestack add skills (alias skillet) lists the built-in skills and the optional Skillet sidecar roster.