Incident Commander

Use this skill for one command decision inside an incident that already belongs to a Runx agency case. It checks the folded case and fixed roster, asks the canonical ops-desk skill for one bounded decision, then enforces the incident rules before returning incident_turn.

The result is one bounded, receipt-backed command turn: an exact dispatch, approval escalation, verified resolution from a supplied member receipt, or actionable blocker. This skill does not append an agency event, acquire a case lease, send a message, call a provider, or close external state. The agency driver remains responsible for the expected-version append and for binding this run's sealed receipt to the case stream.

Composes

• ops-desk#advance

Required context

Supply case_id, driver_id, incident_objective, folded case_state, and a fixed roster. The objective must be one of begin, assign, send, resolve, or postmortem.

The roster has exactly three entries:

• commander, which normally names ops-desk and its incident command scope;
• responder_lead, which names the response skill and its allowed scope;
• comms_lead, which names send-as or slack-notify and its allowed scope.

Each entry includes a principal, skill, and nonempty scope ceiling. Roles and principals must be unique. An assign turn also needs case_state.named_owner to match one roster role or principal.

case_state is a bounded projection, not the incident event history. Durable history stays behind the agency data source. Include declaration status, severity, scope, current turn, and only the evidence or pending decision needed for this command turn.

Communications

A send starts from case_state.pending_escalation.proposed_handoff. That object binds the exact comms_lead principal, send-as or slack-notify, runner plan, channel, audience, and SHA-256 content digest.

The optional approval must contain that same principal and a nonempty reason. Without approval, a valid handoff returns awaiting_approval. An approval from another principal is refused. A matched approval can return a downstream planning handoff with state ready_for_planning; delivery_status remains not_sent and provider delivery remains not_executed.

Run the named skill separately. send-as plans a provider-neutral send. slack-notify owns its governed Slack delivery lane. Return the later sealed receipt through member_result and advance the agency case again. A delivery claim without runx:receipt:sha256:<digest> is refused.

Resolution and durable state

Only resolve may return resolved. It needs either case_state.resolution_evidence_ref or a successful member_result with a valid Runx receipt reference. The result still says agency_state: not_persisted; the agency driver must append the turn under compare-and-swap.

The normal driver sequence is:

1. Open the case with canonical agency open if it does not exist.
2. Advance agency to obtain the folded state and bounded dispatch context.
3. Run Incident Commander with that context and keep its sealed receipt.
4. Pass the result and receipt back to agency for the expected-version append.
5. Run a named responder or communications skill only when the accepted turn calls for it, then feed that sealed member receipt into the next agency turn.

The agency receipt proves persistence. The Incident Commander receipt proves the local command decision. A downstream provider receipt and its readback prove delivery. Do not substitute one receipt for another owner's effect.

Outcomes

incident_turn.decision is one of:

• advanced: a bounded dispatch or verified member result can continue;
• awaiting_approval: a valid communications proposal needs its exact roster principal's approval;
• resolved: linked resolution evidence permits the agency driver to persist closure;
• needs_input: required operator context, such as an assignment owner, is absent;
• refused: supplied state, roster, approval, scope, or evidence contradicts the incident rules.

Every result includes case and turn identity, a reason, validation findings, and effect state. Missing information is not reported as a negative operational fact. A refusal does not mutate the case.

Recovery

For needs_input, correct the folded projection or name the roster owner and rerun. For awaiting_approval, keep the proposed audience and digest unchanged, obtain approval from the exact comms_lead, and rerun. For refused, resolve the listed validation finding rather than weakening the roster or scope. For a provider or responder failure, preserve its receipt and let agency decide the next bounded turn.

Inspect and test the package with the repository-built CLI:

./crates/target/debug/runx skill inspect ./skills/incident-commander --json
./crates/target/debug/runx harness ./skills/incident-commander --json

Registry publication is separate from authoring. Preserve the contributor identity mossony/incident-commander; claim registry availability only after a publish readback for the exact package digest.