Orchestrate it out

Orchestrate It Out

Use this topology for substantial implementation work:

GPT-5.6 Sol (read-only planning)
        |
        v
sealed comprehensive plan with exactly five bounded workstreams
        |
        v
Fable 1   Fable 2   Fable 3   Fable 4   Fable 5
        |
        v
receipts + mechanically enforced scope/tests/final gate
        |
        v
DONE or BLOCKED — no semantic reviewer

Run it

From the target repository:

orchestrate-it-out "implement the exact substantial task" --project "$PWD" --trusted-task

Inspect the full Sol prompt, pinned topology, and command sequence without model calls or NTM mutation:

orchestrate-it-out "implement the exact substantial task" --project "$PWD" --dry-run

Stay with the command until it reports DONE or BLOCKED. Do not declare the task finished merely because the NTM panes became idle. Completion requires five atomic COMPLETE receipts and a passing final evidence artifact.

Trigger exclusions take precedence over the magic phrase: do not run on a dirty repository, for a task too small to support five genuine bounded workstreams, or when the user asks for independent review. A security-sensitive task may still use this topology when the user explicitly chooses it after the absence of independent review is clear; do not infer that choice. An explicit request to execute orchestrate-it-out, “orchestrate it out,” or the exact Sol/five-Fable workflow authorizes adding --trusted-task. A request to inspect or discuss the workflow does not.

What the controller enforces

• Sol is pinned to gpt-5.6-sol and plans before any Fable session exists.
• Sol receives a JSON output schema and runs in Codex's read-only sandbox by default. If that sandbox is unavailable, stop and explain the issue; use --uncontained-orchestrator only when the operator explicitly accepts it.
• The plan must contain IDs fable-1 through fable-5, specific context, decisions, steps, exact write scopes, dependencies, invariants, non-goals, deliverables, acceptance criteria, and direct-argv verification commands.
• A structural validator rejects cycles, unsafe paths/commands, glob scopes, and unordered scope overlap. It seals the task, baseline, plan, and validation.
• Only after sealing does the controller create exactly five claude-fable-5 panes. It cross-checks NTM's manifest against tmux and sends each mechanically rendered brief to one exact pane—never a provider broadcast.
• Dependency waves may run concurrently only where Sol declared disjoint work. Workers may not spawn agents, nest this workflow, commit, push, install, deploy, publish, or contact external services.
• The final gate verifies unchanged HEAD, protected ignored-file fingerprints, union write scope, git diff --check, every workstream command, and every final integration command. Verification runs without a shell, with bounded output, closed stdin, scrubbed sensitive environment variables, timeouts, and repository-mutation detection.

Deliberate omission

Do not add a critic, reviewer pane, worker cross-review, review schema, or semantic repair round. This workflow intentionally trusts Sol's plan and the five Fable implementations, then requires mechanical evidence. “No reviewer” means no independent post-build semantic review; Sol and the workers still use semantic reasoning while planning and implementing. The handoff must state: Mechanically green; independent semantic review intentionally omitted.

Use $build-it-out instead when independent semantic review is part of the request, or when risk warrants review and this no-review topology was not explicitly chosen.

Safety boundary

Live execution requires --trusted-task and a clean Git repository. Fable workers run unattended with the caller's Claude permissions. The Sol sandbox, prompts, seals, and Git gates reduce mistakes and stale-state errors; they are not a hostile same-account security boundary. Repository content, task context, and bounded evidence are sent through the configured Codex and Claude clients, subject to those providers' accounts, retention, and cost policies.

If either exact pinned model is unavailable or the default Sol read-only sandbox cannot start, stop as BLOCKED; never silently substitute a model. The operator may explicitly choose --uncontained-orchestrator for the documented sandbox fallback.

Verify the package

node --test tests/orchestrate-it-out.test.mjs
python3 /path/to/skill-creator/scripts/quick_validate.py .