orchestrate-cost-aware-agents

Orchestrate cost-aware agents

Treat this as the Codex-native product entry point. Keep Python and CLI commands as internal implementation and diagnostic details unless the user explicitly asks for them.

Before taking task actions, read ../../SKILL.md completely and follow its scheduler, safety, evidence, recovery, budget, and ArchMarshal compatibility requirements. Resolve every relative script or reference path from the plugin root, two directories above this file.

Interact with the user in Codex: translate natural-language intent into the bounded CostMarshal workflow, report durable outcomes, and expose a manual command only when recovery or diagnosis genuinely requires it.

Natural-language control plane

Classify the user's request before running the internal engine:

• Set up CostMarshal: discover an existing CostMarshal project first, including compatible v2 state. If none exists, obtain only the missing objective, writable workspace, reviewed low/medium/high provider catalog, budget, and governance choice. Preview provider/profile changes before applying them, never request secret values in chat, then run init with a stable project name.
• Plan or explain: use providers, budget, and read-only route. Explain the safety floor, complete admitted chain, per-step reservation, historical acceptance evidence, and why a cheaper route was rejected. Do not create a task or start a provider.
• Do new work: inspect status, work-graph, and relevant model-memory scopes. Decompose the request into bounded work packages with roles, dependencies, deliverables, gates, and teaching policy. Create each package with new-task, run a read-only route explanation, then dispatch --start only when its dependencies are accepted and it remains within the user's workspace, budget, paths, capabilities, and acceptance criteria. Start the scheduler in bounded cycles and stop monitoring only at gated leader acceptance, explicit failure, budget exhaustion, a recoverable pause, or user stop. For legitimately low-floor, text-only analysis, debugging proposals, and scout/reviewer packages routed to LongCat, prefer report-only proposal-api: send committed allowlisted context, grant no writes or tools, and require Codex Leader review. Do not relabel a medium/high-risk task merely to force a cheaper tier. A severity-0/1 defect may receive one precise feedback retry; otherwise continue to the admitted Codex Worker or let the Codex Leader implement. Treat low/medium/high as cost/safety routing authority, never as a fixed intelligence ranking.
• Coordinate a large project: inspect repositories and workstreams before task creation. Register only explicit committed Git roots, bind each task to one repository and optional Workstream, and respect Workstream dependencies, concurrency, and CNY allocations. At a milestone, freeze a complete staged integration plan and request a Leader Gate; never claim cross-repository atomicity or mutate source repositories while planning.
• Audit or monitor: use JSON status, dashboard, work-graph, workstreams, repositories, production-status, model-memory, providers, budget, validate, and read-only governance-status. Summarize dependency readiness, integration Gates, evaluations, errors, cost, and any unpromoted policy candidate; never infer success only from a live process.
• Check Codex compatibility: use codex-native-status before diagnosing native child-agent or App Server availability. This probe makes no provider call. A failed handshake disables only the native attempt team and preserves the original single-agent execution contract; it never authorizes a provider, isolation, credential, or safety fallback.
• Resume or recover: run recover read-only first. Show the exact restart plan before --restart-missing; preserve sealed routes, generations, attempts, reservations, runtime receipts, and leader ownership. Never silently respawn an uncertain actor.
• Stop: use the actor's durable identity with stop-actor --stop-runtime, then verify terminal state and cleanup receipts. Do not kill by an unverified PID or process name.