| name | operate-canonfig |
| description | Operate Canonfig 2 Source and Follower Machines safely. Use for discovery and proposal review, explicit profile publication, follower invitations, groups or revocation, profile selection, synchronization planning or apply, no-input automation, agent policy and harness bounds, schedules, status or doctor diagnostics, Follower Drift, Human Action Required, and interrupted-run recovery. |
Operate Canonfig
Use one-way, evidence-first operations. Followers consume Source Machine
authority; they never publish upstream.
Route the request
- For discovery, proposal review, publication, invitations, groups, and
revocation, read references/source-operations.md.
- For profile selection, plan/apply, status, diagnostics, drift, Human Action
Required, and recovery, read
references/follower-operations.md.
- Before changing schedules, credential handling, recipes, or harness paths,
read references/platform-boundaries.md.
Operating loop
- Observe with
canonfig status and bounded canonfig doctor.
- Identify the role and selected immutable Profile Revision.
- Plan with
canonfig sync --plan; review every target, dependency, agent
task, conflict, human action, and verification.
- Apply only an approved plan with
canonfig sync --apply.
- Verify the final outcome and evidence.
Converged requires every required
verification to pass.
- Preserve and report
HumanActionRequired, FollowerDrift, Failed, or
Interrupted exactly. Never reinterpret partial success as convergence.
Automation
Native schedules run:
canonfig sync --apply --no-input
For machine-readable output:
canonfig sync --apply --no-input --json
Scheduled runs never wait for approval. Human Action Required exits with code
3; Follower Drift exits with code 4. Keep failure output visible.
Agent policy
Inspect policy and harness separately:
canonfig agent policy
canonfig agent harness
deterministic-only invokes no agent. agent-propose records a bounded
proposal but does not execute it. agent-apply executes only the intersection
of the Agent Task bounds and local harness allowlists, followed by independent
verification.
Configure the smallest local allowlist:
canonfig agent harness codex --executable /opt/codex --allow-path /home/operator/.canonfig --allow-leaf-executable npm --allow-origin https://registry.npmjs.org --allow-capability restart --maximum-input-bytes 4096
Use exact HTTPS origins and platform-local paths. Elevation, login, restart,
and reboot remain denied unless both task and harness authorize them. A policy
change is not a repair for Human Action Required.
Safety boundary
- Publication is explicit. Discovery and agent output remain proposals.
- Invitations are short-lived and single-use; credentials are local references,
never copied from the Source Machine.
- Preserve modified follower skills. Canonfig does not merge or overwrite them
automatically.
- Preserve certificate pins, SQLite state, action journals, blobs, and
verification evidence during diagnosis.
- Run declared deterministic recipes before agent work. Keep recipes
platform-specific, version-aware, upstream-backed, and independently verified.
- Refuse requests to hide drift, erase evidence, bypass trust, broaden
allowlists without task need, or mark actions complete manually.
