Feature request – automatic orchestration from a plan file (Boss+ Workers)

Which codex product are you using: CLI, app, extension, web?

Have you tried building the proposed solution already? If so, what pieces are missing in codex that prevent you from doing so?

Which codex product are you using: CLI, app, extension, web? I’m using the Codex VS Code extension (local project, multi-chat setup). Have you tried building the proposed solution already? If so, what pieces are missing in codex that prevent you from doing so? Yes — I already built a manual version

SEF = BOSS

## My requested capabilities from Codex

• I need plan-driven automation from tasks.md, not manual chat orchestration.
• SEF should read tasks automatically and execute incrementally (step-by-step), not all at once.
• SEF should auto-assign W1/W2/W3 with strict scope boundaries.
• Each task must have a checkpoint gate (tests/checks/diff) before moving forward.
• Persistent task states are required: TODO / IN_PROGRESS / BLOCKED / DONE.
• I need automatic decision log + change log per step.
• Small fixes should be auto-applied only if they stay strictly within requirements.
• If a fix would change requirements, execution must stop and request explicit confirmation.
• No deviation from original specification without approval.
• Goal: reduce manual overhead (no repeated manual START/VALIDATE per micro-step).

-----------------------------------------------------------------------------------------------------

Current workflow (BOS+ Workers, local on my PC)

• I run 4 chats in parallel:
• 1 SEF chat (extra-high reasoning, strongest model) for planning, orchestration, and final validation.
• 2 Worker chats (high reasoning) for complex technical implementation.
• 1 Worker chat (medium reasoning) for fast execution tasks (typically QA, quick fixes, operational checks).
• All chats work on the same local repository with full access.

Role initialization (manual, explicit)

• At the beginning, each chat receives a role identity prompt:
• SEF = orchestrator + validator.
• W1/W2/W3 = scoped implementers.
• After initialization, each chat keeps its role across runs.

Coordination artifacts (7 .md files)

• AI/RULES.md — global mandatory rules for all chats (scope, allowed/disallowed changes, delivery format).
• AI/tasks.md — technical backlog ordered by runs, with dependencies and goals.
• AI/SEF.md — active run plan: worker tasks, allowed scope, status, final validation.
• AI/start.md — standard start prompts for each worker.
• AI/worker1-backend.md — Worker 1 execution report.
• AI/worker2-frontend.md — Worker 2 execution report.
• AI/worker3-dbqa.md — Worker 3 execution report.

How a run works in practice

• SEF takes an item from tasks.md and turns it into a concrete run in SEF.md.
• Each worker receives a scoped task (file boundaries + definition of done).
• Workers implement in parallel and report in their own .md file:
• files changed,
• impacted APIs/endpoints,
• test steps,
• risks/limitations.
• SEF validates both reports and actual code/diffs, then decides:
• approved (continue),
• needs_fix (targeted fix run).

Benefits

• True parallelization: backend/frontend/qa move at the same time.
• Strong traceability: every run has history, scope, decisions, and verdict.
• Regression control: separated responsibilities + centralized SEF validation.
• Scalable process: can increase/decrease worker count without redesigning the workflow.
• Auditability: standardized report format across workers.

Current limitations

• Orchestration is not fully automated: I manually trigger worker starts.
• Scope/lock enforcement is procedural, not hard-enforced by the platform.
• Risk of inconsistency if a worker report does not perfectly match final code.
• Coordination overhead is high: SEF validation is manual and frequent.
• If a run is ambiguous, rework and fix-runs increase.

Conclusion

• The process is effective for iterative development at speed.
• Quality depends on discipline: RULES.md + SEF.md + real validation.
• The next major improvement is orchestration automation.

Did you achieve good results?

This feature request hasn't received enough upvotes, so closing.