Skill v1.0.0

version: "1.0.0" name: team-coordination description: Coordinate an enabled leader-and-execution-team workflow with explicit roles, ownership, versioned context acknowledgements, evidence and recovery.

The leader owns scope, decisions and acceptance. Execution owners investigate and implement. An independent reviewer checks current artifacts; do not approve your own runtime. Reuse a small team and bounded advisor help. This workflow grants no new permissions.

For managed projects, use one .codex-team/state.json through scripts/harness.py; legacy ledger CLI refuses managed state. Preserve critical decisions and rationale in role packets. Register real supported IDs with honest provenance, claim canonical write paths, and require current revision/epoch/digest ACKs. Identity registration remains a claim, not OS authentication.

Read the relevant file in roles/ and policy.json. Exact model availability is client/account dependent. The managed adapter initializes, queries model/list, verifies thread model, then rechecks current instructions before dispatch. It is a read-only pilot; direct clients that bypass it retain advisory guidance.

Use trusted Python with site/environment isolation and no bytecode writes. Runtime source is tools/codex-team-harness/ in the matching reviewed oozoofrog/oozoofrog-plugin checkout. Obtain and verify that source before runtime use; a skills-only Plugin does not install the runtime or global guidance. From that verified directory use python3 -I -S -B run.py harness ... or run.py managed-run .... For installed scripts, an approved isolated loader must add only this script directory before importing sibling modules. Never use an untrusted environment.

Instruction/contract changes invalidate ACKs. Guarded policy_refresh refuses active/uncertain runs. Hold uncertain ownership until supported terminal evidence. Reservations are conservative accounting; byte/4 context estimates cover the explicit packet only. Two review passes by default; unresolved material findings require a scoped decision. Handoff artifact hashes and fresh file-backed independent review before acceptance.

For session model/effort inspection, explicit change requests, or an unmanaged session's requested transition to new managed sessions, read session control. Use run.py session-control or its long-lived Controller API. Separate requested/accepted/pending/live configured settings from unavailable runtime identity. Keep one authoritative project record, real packet receipts and atomic ownership transfer; a registration/ACK alone is not migration. Stop the source at an owner checkpoint, preserve old chat/code, and never silently create a standalone replacement for an unavailable native target.

When managed packets exceed the unchanged context guard or a reset needs portable evidence, read checkpoints. Preserve canonical history; use an owner- and independently reviewed working view with exact live obligations, hash-pinned snapshots, allowlisted evidence and fresh packet ACKs. Installation, administrative policy migration, archive verification and actual work resumption are separate steps.

For Astra leader/advisor defaults, read profiles. Model, reasoning and requested service mode are separate. Core Sol and support Luna remain unchanged. Desired defaults and configured observations do not attest current provider execution.

For a failed-review correction under the same owner while retaining claims, read same-owner rework. Use the explicit CAS/idempotent transition and fresh packet receipt; do not release/reassign merely to reset task state or review counts.

For a human-authorized project budget increase, read budget administration. Preserve defaults, reviews and accounting; use exact project/CAS approvals and actual fresh packet ACKs.

For human-authorized source instrumentation after a retained failed review, read bounded diagnostics. Keep the original task, owner, failed reviews and budget; use one exact approved diagnostic cycle and fresh packet ACK. Stop after its separate evidence result.