- edit
- test
- release
developer tooling / 2026
Codex Orchestration
I evolved an open-source routing plugin into an event-driven delivery system that matches work to the right model, separates implementation from review, and keeps releases supervised.
view repositorythe problem
One model should not plan, build, review, and approve its own work.
Coding tasks vary from a quick factual check to a full product build, but a single undifferentiated workflow treats them as if they carry the same cost and risk.
Codex Orchestration classifies the intended outcome first, gives each role a narrow responsibility, and makes evidence—not momentum—the gate to release.
visual walkthrough
A task moves through distinct roles, not one long improvisation.
The workflow stays proportional for small work and adds checkpoints as risk grows.
-
01
Classify the outcome
A taxonomy-only orchestrator reads the current request and bounded continuity, chooses one of six work classes, returns the fixed route, and stops.
$ route current_request
CLASS BUILDIMPLEMENTER Sol / HighSUPERVISOR Sol / Extra High -
02
Separate making from judging
The implementer owns every edit and test. A different, read-only supervisor defines acceptance, reviews evidence at quiescent checkpoints, and can require precise corrections.
IMPLEMENTER owns change ≠SUPERVISOR owns review - acceptance
- evidence
- corrections
-
03
Advance on evidence
Builds pause at architecture, vertical-slice, and release-candidate checkpoints. Only supervisor readiness unlocks the already-prepared release plan.
01architectureboundaries + proof02vertical sliceend-to-end evidence03release candidateready, not released04releaseapproved + verified -
04
Keep the exact task intact
Versioned context bundles preserve the root-visible request and completion continuity. Amendments update the same roles and acceptance instead of silently losing earlier constraints.
revision 01initial requestrevision 02request + amendmentaccepted resultcompletion continuity
technical brief
README.md
A local Codex plugin built around strict role boundaries, exact continuity, and event-driven supervision.
01 / architecture
Root coordinates; narrow roles do one job each
A stable chat-scoped hook gives root a binary gate and the current task context. For routed work, a Terra / Extra High orchestrator returns only the relationship, work class, route, and reason. Root then starts the selected implementer and read-only supervisor against the same exact bundle.
The orchestrator never inspects the repository, performs task work, or judges acceptance. The supervisor never edits. Those constraints prevent responsibility from bleeding across roles.
02 / outcome taxonomy
Six classes keep cost proportional to the job
Fresh verification, audit, or substantial research.
Content-first documents, sheets, PDFs, and presentations.
Non-code deliverables where visual expression defines the outcome.
One existing behavior in one production component.
Existing behavior spanning components or material operational risk.
Any net-new code capability, with three supervised checkpoints.
Complexity remains telemetry rather than a hidden routing lever. The outcome class alone determines the fixed implementation lane, review lane, and checkpoint sequence.
03 / context continuity
Exact bundles replace lossy prompt retelling
Each task role independently loads a private, versioned JSON bundle containing the ordered root-visible conversation since the last accepted or cancelled objective. Later checkpoint relays carry only the new decision plus immutable acceptance.
current request
→ immutable context bundle
→ fixed route + acceptance
→ implementation checkpoint
→ continue | correct | ready to release
→ completion receipt
If the user amends active work, the hook writes a new bundle revision. The implementer pauses mutation while the same supervisor revises acceptance, preserving both the earlier request and the new constraint.
04 / release safety
The release plan exists before release authority
Editing, tests, builds, deploys, and migrations stop before each review.
Repository, deployment, decisive probe, and tunnel lifecycle are resolved at release candidate.
The same implementer performs the approved commit, push, deployment, and final probe.
Rendered or interactive outcomes use a bounded live check without turning the supervisor into an editor.
05 / stack
Small runtime, explicit protocol
- plugin surface
- Codex plugin manifest, hooks, custom agent profiles
- runtime
- Python 3, shell, JSON
- model lanes
- GPT-5.6 Sol, Terra, and Luna
- verification
- offline fixtures, protocol tests, install-state checks
- distribution
- Codex plugin marketplace, MIT license
Codex Orchestration is maintained at jessejaffe/codex-orchestration and builds on Danny McAteer's original sol-advisor project.