Maestro Setup Agent Skill

Maestro Setup

Tune a Maestro-enabled repository harness from current repository evidence.

Activate with a known session id:

maestro hook record --event skill_activation --skill maestro-setup --session <session_id>

Recipe checkpoint: Maestro's main workflow is the loop. Use maestro status

for current state and maestro loop next as the read-only router when the next

lifecycle is not obvious. Setup writes still use the existing Maestro verbs

named by this skill (init, install, sync, doctor, upgrade,

uninstall, and shell-init), not hidden setup state. Rule: loop next

recommends; outcome/proof/memory verbs write. maestro loop next --chain

explains current chain position without writing; maestro loop outcome appends

structured outcomes and transition receipts after native work; `maestro loop

trace <card> audits card-scoped receipts; maestro loop improve` is a read-only

proposer. Do not use hidden stores, hidden schedulers, silent recipe mutation,

or proof/QA bypass.

Use

• After maestro init.
• After maestro install.
• When an already-initialized repo needs its Maestro setup updated to the

running binary with minimal repeated token cost.

• When maestro doctor reports setup or local agent integration problems.
• When a repo's build/test/harness instructions are missing or stale.

Do

• Run pwd, then maestro status.
• If the repo is already initialized and the request is only to update Maestro

setup, take the cheap rerun path before any broader read-in:

• Run maestro sync --dry-run.
• If it reports only current/skipped resources, do not run maestro sync.
• If it reports drift, run maestro sync, then report only changed resources,

backup paths, and blockers.

• Run maestro sync --global-skills only when the user asks for global

skills or maestro doctor reports global skill drift.

• Finish with maestro doctor, then maestro status.
• If the repo is not initialized, run maestro init --dry-run.
• If dry-run prints operating on <path> and that path is not the intended

current project root, stop and ask the user to pick or create the root before

any write.

• Only after the root is correct, run maestro init --yes.
• Run maestro doctor.
• If no agent integration is installed, run maestro install --agent codex

unless the user asked for another agent.

• Inspect repo structure, build/test commands, and workflow constraints.
• Read in the existing agent and doc instructions as a BOUNDED set: at the

repo root, and under each folder matched by the projects: globs in

.maestro/harness/harness.yml, read AGENTS.md, CLAUDE.md, README.md,

and top-level docs/*.md. Stay shallow (one level per location, no deep

crawl) and skip outsized files (roughly 64 KB and up) so a vendored doc dump

cannot flood context. With no projects: declared, this is the repo root

alone.

10. Synthesize what you read into the SINGLE root harness guidance, one section

per project (a single section when nothing is declared). This is read-in

only: never write maestro-managed guidance into a sub-project's own

AGENTS.md/CLAUDE.md -- maestro install/sync write managed blocks at

the repo root alone. Cite the inspected files; do not tune from guesses.

11. Run maestro doctor, then maestro status.

Stop

• maestro init --dry-run writes nothing; use it before init writes.
• Native harness layer setup route: use maestro install --dry-run before

agent mirror writes, maestro sync --dry-run before setup refresh writes,

maestro capability to explain optional provider state, and

maestro maturity to read context, proof gaps, friction, maturity level, and

next owner before claiming setup readiness. Generated CLI references prove

command shape; Harness and targeted skills teach the workflow.

• maestro init --yes keeps existing files and creates missing files.
• maestro sync --dry-run is the cheap already-initialized update check; do not

run broader setup read-in when it proves the setup is current and no doctor

blocker remains.

• Use maestro init --force only for deliberate refresh; it backs up managed

files first.

• Do not tune the harness from guesses, package-manager defaults, or stale chat

memory.

• Per-project docs are read-in only. Maestro owns one root scope; never write

maestro-managed guidance into a sub-project's AGENTS.md/CLAUDE.md.

Done

• Setup is healthy, or the remaining setup blocker is explicit.
• The next handoff is visible from maestro status.
• Harness guidance changes cite inspected files or commands.
• Completion criterion: every write was preceded by the matching dry-run or

current-state read, and every changed managed resource is named with its

backup path or blocker.