b-open-io/orchestrator

Orchestrator

Keep one main seat in control while using other agents for the work they do best. This skill supports both Claude Code and Codex. The current main model is whatever the user selected; never infer, rename, or pin it.

Default topology

Current Claude or Codex main
├── native specialist agents: exploration, review, testing, domain expertise
├── Grok worker lane: bounded implementation volume
└── Fable advisor lane: read-only second opinions at commitment boundaries

On a Claude Code main, a fourth lane exists: the native Workflow tool for deterministic staged fan-outs (opt-in-gated, Claude-only). Treat it as a dispatch lane under the same ownership rules — see ../coordinator/references/native-workflows.md.

The main owns:

• task interpretation and the controlling plan
• decomposition, file ownership, and shared interfaces
• acceptance criteria and verification commands
• review of every worker diff
• reconciliation of specialist and advisor findings
• final tests, user-facing synthesis, commits, pushes, and pull requests

Workers and advisors do not inherit that ownership. The main can accept, reject, or reconcile their results using direct evidence.

Compose the detailed protocols

Use the existing skills instead of duplicating their full manuals:

• Load Coordinator before dispatching implementation. It owns specs, disjoint file ownership, worker preflight, structured final reports, adversarial diff review, and main-seat verification.
• Load Advisor before consulting an independent model. It owns consult packaging, read-only boundaries, commitment timing, the verdict contract, and reconciliation.
• Load Wave Coordinator when the requested fan- out exceeds currently free host slots or needs staged diversity.
• Before dispatching to a generic specialist, match the unit against the roster in ../deploy-agent-team/references/agent-roster.md (or ask Front Desk) and pass the specific subagent_type. Use a generic agent only when no roster agent fits, and say so explicitly in the dispatch note.

Apply Coordinator and Advisor together when the workflow has both workers and an advisor. Coordinator governs execution; Advisor governs judgment consults.

Host adapter

Codex main

• Prefer installed bopen_* custom agents for named specialists. Use built-in explorer or worker agents as an explicit fallback.
• If a named adapter is absent, say so and offer the explicit bopen-tools:codex-agent-setup skill. Do not claim the persona was spawned.
• Keep wave control in the main thread. With Codex's default max_depth = 1, direct children cannot recursively fan out; that is usually desirable.
• Use Grok for bounded implementation when authorized. Do not launch a second Codex CLI merely to reproduce what a native Codex subagent can do.
• Use the Advisor skill's Fable CLI channel for an independent Claude opinion.

Claude Code main

• Prefer plugin-qualified Claude agents for specialists.
• Use Grok or Codex as implementation lanes when authorized and economical.
• Prefer Claude's native advisor or a read-only premium Claude subagent when available. Use an external advisor only when it adds independence.

External data boundaries

External lanes are optional and must be transparent:

• A Grok dispatch can send its prompt, specification, code excerpts, and other repository content to xAI.
• A Fable consult can send its consult package and repository files inspected by read tools to Anthropic.

Before the first use of each external lane, state what content will be shared and obtain approval unless the user already explicitly authorized that lane for the task. Never send secrets, credentials, unrelated proprietary content, or a broader repository snapshot than the assignment needs.

Orchestration sequence

1. Orient in the main. Inspect the request, applicable instructions, repo state, available agent roster, and relevant evidence. Do not delegate a premise that has not been checked.
2. Select the topology. Decide which work stays in the main, which native specialists are useful, which implementation units fit Grok, and whether a Fable consult reaches a real commitment boundary.
3. Preflight lanes. Verify native agent availability. For Grok, inspect the complete grok models output, require BOPEN_WORKER_MODEL, and confirm its exact value exists before dispatch. For Fable, verify Claude CLI authentication and use ${BOPEN_ADVISOR_MODEL:-fable} without claiming that alias is permanently the latest model.
4. Disclose external sharing. Obtain any required approval before sending repository content to xAI or Anthropic.
5. Gather specialist evidence. Use native agents for independent research, architecture, security, testing, documentation, or domain analysis. Give each a bounded, self-contained assignment and require a complete report — for background-dispatched agents, instruct them explicitly to deliver that report via the host's messaging mechanism before going idle; an idle notification is not a deliverable (see Coordinator's Background Subagent Etiquette for the full pattern).
6. Consult at a commitment boundary. If warranted, package the goal, constraints, state, evidence, and one decision for the read-only advisor. Advice returns to the main; it never becomes an edit task.
7. Write worker specs in the main. Partition implementation by disjoint file ownership or isolate it with worktrees. Pin shared interfaces verbatim, include exact acceptance commands, name forbidden files, and require the Coordinator final-report contract.
8. Dispatch and keep working. Run independent workers in parallel while the main prepares later specs, reviews completed evidence, or plans verification. Stop at a barrier before cross-unit synthesis or git operations.
9. Review adversarially. Inspect every diff, especially tooling, dependency, security, and sandbox-workaround surfaces. Reconcile disagreements with evidence or one focused follow-up.
10. Verify and ship from the main. Re-run acceptance in the main environment. Only the main commits, pushes, opens a PR, or reports completion.

Failure behavior

• A missing or unauthenticated lane is unavailable, not permission to silently absorb or reroute the work. Report it and choose another lane explicitly.
• Infrastructure failure is not a worker-quality strike. Preserve the spec and retry or reroute it without accepting environment-driven workarounds.
• Two corrected quality misses trigger the Coordinator escape hatch.
• Advisor disagreement is surfaced and reconciled; never silently ignore it.
• Do not increase Codex agent depth, install adapters, install CLIs, or change global configuration without user authorization.

Final report

Report the topology actually used, external providers consulted, work returned by each lane, files changed, verification results, unresolved disagreements, and any lane that was unavailable. Never imply a worker or advisor ran when it did not.