Menu
Coordinate CorvidLabs agents through discovery, observation, execution, and verification.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Validate, route, inspect, or run agent.3md manifests and typed tool templates.
Inspect spec ownership, drift, coverage, review queues, and gaps with fledge atlas.
Verify or record reviewed commit provenance, policy gates, and audit exports with fledge attest.
Assess working-tree, staged, range, or CI change risk and gates with fledge augur.
Keep CI, pre-push verification, and release evidence tied to the current committed revision.
Build, modify, test, review, or release CorvidLabs Swift packages with Fledge-first and Swift 6 discipline.
Basé sur la classification professionnelle SOC
| name | agent-coordination |
| description | Coordinate CorvidLabs agents through discovery, observation, execution, and verification. |
Use this skill as the hub before asking an agent for status, intervening in its work, or
choosing a worktree. Product detail lives in sibling skills: let, rune,
fledge-workflows, and spec-sync / spec-sync-routing.
Do not skip ahead on session titles or branch names alone.
Stay inside the confirmed repository. Do not invoke Let:
fledge work status
fledge run --list
git status --short --branch
git ls-files
Read only tracked instructions and declared workflows. Mark Let discovery blocked and
continue with repository evidence. Let 0.2 project-scope finds still pull user catalogs for
skills/agents (and related kinds) by default, and where/worktrees can list siblings; do
not filter private results after they have already been read. Details and kind matrix: let
skill.
When the task explicitly permits local agent metadata, use the let skill with the
narrowest supported query (prefer find worktrees / find sessions with
--scope project). Treat every returned path and session id as private until public
provenance is proven independently. Do not run Let workbed write commands unless writes
are authorized.
Confirm repository, worktree, session, and active task first. Use the rune skill
(and fledge rune run|watch --help for the installed flag set):
fledge rune run --timeout=30 --json -- <agent-cli> … — bounded, non-interactivefledge rune watch --log=<path> -- <agent-cli> … — live PTY; keystrokes go herePublic Rune has no separate send subcommand. Send one scoped instruction only after
observation, and never while tests, commits, pushes, migrations, or independent review
are in progress—inspect Git and CI instead. Never drive the same session from two
operators, and never start a second agent in a worktree that already has an owner.
fledge run --list
fledge lanes list
fledge run <task>
fledge lanes run <lane>
Prefer declared tasks and lanes over ad hoc shell. Record task output with the exact
revision. Details: fledge-workflows skill.
spec-sync-routing to the generated local
skill; use shared spec-sync only for version-neutral principles.spec-sync skill.Highest trust first: committed tree and CI on that SHA → PR and review → Fledge/Spec Sync results on that SHA → worktree diff → Rune/session output → Let session cards.