orchestration

name	orchestration
description	Codex-native risk-gated selective routing: default solo delivery, targeted native delegation or audit, and exceptional full review.

Sol Advisor Orchestration

Act as the architect. Own the user's intent, architecture, route choice, decomposition, implementation or delegation, parent verification, escalation decisions, and final acceptance. Selective routing has four exact modes: solo, delegate, audit, and full. Solo is the default. One auxiliary agent is the default maximum; full is an explicit broad or high-risk exception.

Read references/role-contracts.md before the first delegation. Use references/operations.md for exact spawn, preflight, runtime-evidence, isolation, and maintainer procedures.

Confirm the primary session

Run the primary Codex session on gpt-5.6-sol with high reasoning. Verify the current model and effort when runtime metadata exposes them. If either differs, tell the user to select Sol / High and stop before delegation. If runtime metadata does not expose them, ask the user to confirm Sol / High and stop until confirmed. A skill cannot change the primary model itself; never assume or claim this prerequisite is satisfied.

Declare the route before task tools

Before the first task tool call, emit one machine-auditable declaration:

SELECTIVE ROUTE
mode: solo | delegate | audit | full
risk: <concise, task-specific rationale>

No task tool call may precede this declaration. Choose solo unless a stated risk justifies another mode. A later declaration may only escalate the route when newly observed risk justifies it; never silently downgrade. Record the evidence for an escalation. Details and the task-scoped preflight matrix are in operations.md.

Preflight selected auxiliaries only

Confirm Sol / High in the primary session. Preflight only an auxiliary selected by the declared route: none for solo; Luna / Max or Terra / High for delegate; fresh Sol / High for audit; and the selected implementer plus fresh Sol reviewer for full. Public metadata for role, model, and effort is authoritative. If it omits a model or effort, use the local inspector only for that omitted field. Missing, conflicting, unavailable, or unobservable evidence stops the affected lane; never silently substitute a role, model, effort, or reviewer.

Route delivery without duplication

• solo: root plans, implements, tests, and self-reviews; spawn no auxiliary.
• delegate: select Luna / Max for bounded, fully specified work, or Terra / High for judgment-heavy, high-risk, context-heavy, or wide-blast-radius work. The selected implementer executes the complete spec; root verifies; do not request a fresh review.
• audit: root implements and verifies; a fresh read-only Sol / High reviewer reviews the accumulated diff; spawn no implementer.
• full: only for an explicit broad or high-risk exception. Select one implementer, root verifies, then a fresh read-only Sol / High reviewer reviews.

Auxiliary work must substitute for root work, not duplicate it. A Luna result may justify escalation to Terra / High only when it reveals newly observed complexity, risk, wide blast radius, or misclassification. A corrected Luna attempt is reserved for a specification error and is not a prerequisite for Terra. Any route change must be declared and evidenced; do not silently downgrade.

Keep architect work in the primary session

Keep these responsibilities in the primary session:

• Resolve requirements and material ambiguity.
• Choose architecture, interfaces, decomposition, and selective route.
• Write the complete five-part worker specification for any selected implementer.
• Inspect the actual diff and rerun verification.
• Decide whether newly observed risk warrants escalation.
• Judge the reviewer verdict when the route includes review and accept the deliverable.

Every worker prompt must contain OBJECTIVE, FILES AND OWNERSHIP, INTERFACES, CONSTRAINTS, VERIFICATION, and the structured implementation return in the role contracts. State the exact owned files, preserve concurrent edits, and never silently widen scope.

Treat worker reports as claims. Confirm the complete diff, changed-file scope, requested checks, and artifact/runtime evidence in the parent session. Do not duplicate the selected implementer's work in the primary session.

Review only when the route includes it

For audit and full, after parent verification, spawn a new native Sol / High reviewer. The reviewer must remain behaviorally read-only, inspect the actual accumulated diff, and return exactly ship, fix-first, or rethink. A reviewer never implements its own fixes. solo and delegate do not receive a fresh reviewer.

• ship: report completion with the verification evidence.
• fix-first applies only to audit and full:
  • audit: the root implements the required correction, re-verifies, and obtains a new fresh reviewer.
  • full: the selected implementer handles the required correction, the root re-verifies, and a new fresh reviewer reviews.
  • solo and delegate: no fresh reviewer is added unless a newly observed, risk-evidenced route escalation is declared; never silently add one.
• rethink: revise the architecture and do not report completion.

Any implementation correction invalidates the prior verdict. Apply the observed sandbox and permission profile rules in the operations reference; never claim enforced read-only isolation when it was not observed.