Astral Orchestrator · turn a request into a checked result

What is Astral Orchestrator?

A small flight crew for delegated work.

It is a local, open-source plugin for Codex. Delegated work travels with a crew: a captain who owns the plan, specialists who fly their leg, and an inspector who signs off before landing. Singularity deliberately keeps eligible work in one verified primary session instead. By default, no API key and no project-operated backend, no analytics. Just a steadier way of working.

Your current model leads

One lead turns your request into an outcome, a plan, and safety boundaries. The detected Sol, Luna, or Astra session stays responsible for the final call instead of drifting mid-conversation.

Specialists do the work

Astra takes selected deep-reasoning cards, Luna handles narrow repeatable pieces, and Sol handles context-heavy implementation. Each worker receives a bounded packet instead of your whole conversation history.

A fresh review signs off

Checks run against the actual changes first. Then a separate Sol reviewer, with no stake in producing the work, decides whether it is ready to hand back.

Visible while it runs

Live Astral status, without guessing.

The first user-facing progress update includes Astral status, followed by updates before and after every child launch, whenever state changes, and in the final handoff. Each row shows the lane's model, effort, state, and evidence. A planned route is never presented as proof that a worker actually ran.

Astra stays deliberate

Astra is a Medium-default worker for bounded cards whose reasoning depth justifies its cost. Ordinary work stays with Luna or Sol. Primary and child effort remain independent, and every accepted child needs matching runtime evidence.

Pick an altitude

Eight modes, one steady hand.

Tell Astral how careful to be. It raises safeguards when a request is riskier than the mode you picked, and it never broadens the work you asked for.

Comet

Tiny and reversible

The detected Sol, Luna, or Astra primary works directly and self-reviews. No delegation and no ceremony, just the explicit single-session option for small jobs.

Best for: obvious, easy-to-undo changes.

Orbit · default

Plan, delegate, verify

The primary plans, Astra, Luna, or Sol implements bounded work, and a fresh Sol review checks the result. The everyday route for real project work.

Best for: normal changes and projects.

Event Horizon

Fast, careful execution

Singularity-style YAGNI, necessary confirmation gates, parallel ready cards or a bounded hierarchy, targeted checks, and one concise Sol review.

Best for: credentials, payments, production, migrations.

Singularity · explicit

One focused session

One verified Sol, Luna, or Astra primary handles meaningful low- or medium-risk work larger than Comet. It uses a compact card, targeted context, and one proportional verification pass: no subagents or fresh reviewer.

Best for: disciplined multi-step work that does not need delegation.

Hypernova · explicit

Maximum native throughput

This explicit opt-in keeps the detected primary, uses Sol Max workers by default or Astra at its configured effort, and requires a fresh reviewer at the maximum safely available concurrency.

Best for: maximum throughput when speed matters more than token efficiency.

Pulsar · opt-in

Slower by design

A deliberately slower, evidence-oriented route for an explicit request. It is never auto-selected; Orbit remains the default for normal work.

Best for: a reproducible record of a route decision and its checks.

Morph · explicit

Choose the worker route

Morph uses an explicitly selected worker model and effort when the host can provide that capability; the route is capability-dependent. The provider may be external, and a fresh review is required before handoff.

Best for: a deliberate provider or capability choice.

Constellation · capacity-aware

Parallel first wave, when safe

Constellation starts a cost-aware parallel first wave only for independent, non-overlapping cards within host-advertised capacity. One primary consumes a slot; no extra Sol implementers run by default.

Best for: parallelizable work with a proven serial fallback.

Automatic primary checker

All modes first look for allowlisted local model/effort evidence. When evidence is absent, only modes other than Singularity and Hypernova can use a one-time user confirmation. Singularity and Hypernova require an observed, verified Sol, Luna, or Astra model and effort, must stop on unavailable evidence, and user confirmation cannot satisfy or override either requirement. Mismatch or invalid evidence blocks every route. Singularity uses one verified primary session with no subagents or fresh reviewer, and Event Horizon takes over for high-risk work. Morph does not claim every provider has native effort semantics, and Constellation does not claim every host supports multi-agent orchestration: when capacity or independence is uncertain, use the serial fallback.

How it works

One lead, the right lane, then a fresh look.

The usual Orbit or Event Horizon route, step by step. The detected primary keeps the outcome and safety boundaries; Astra, Luna, and Sol receive bounded work only after that plan is settled.

Tap a step to open it. Every lane is touch-friendly.

01 The detected model leads Your request becomes a clear outcome, a plan, and safety boundaries.

The current Sol, Luna, or Astra session settles requirements, architecture, and what done means. The primary checker detects its observed model and effort. If that route cannot be proven, Astral stops rather than quietly substituting another model.

The primary hands off bounded work

02 Astra takes difficult reasoning Bounded diagnosis or synthesis when the added cost is justified.

Astra runs at its configured effort, Medium by default, only for a bounded card whose reasoning depth, cross-domain synthesis, or difficult diagnosis justifies its added cost. Its effort is independent from the primary.

02 Luna takes narrow work Focused, repeatable tasks with a settled instruction.

Think of a rename, a small transform, or a tight template: fully specified and easy to verify. Luna receives a bounded packet, not your whole conversation, at its configured effort.

02 Sol takes context-heavy work Broader implementation, integration, and debugging.

When the work needs project context, like wiring a feature, chasing a bug across files, or careful refactoring, Sol steps in after the primary has settled the plan, at its configured effort.

03 Checks run, then a fresh Sol review The actual changes are inspected before anything comes back to you.

The smallest relevant checks run against the real diff. Then a separate Sol reviewer performs one concise workspace-write review-and-repair pass. Event Horizon adds necessary confirmation gates without isolation ceremony or repeated reviewer cycles.

This normal multi-agent route is a documented heuristic based on the work at hand. Every child runs its selected model at its configured or Adaptive-selected effort, and a mismatched lane is blocked and explained instead of hidden. Singularity is the deliberate one-session exception.

Pulsar mode

Freeze the plan. Keep the evidence.

Pulsar is an explicit opt-in route, not a claim that every task needs more process. It creates a deliberate, inspectable record while the detected primary keeps the requirements, architecture, and final decisions.

One fixed graph, parallel ready items

The primary freezes one canonical dependency graph and named acceptance checks, then uses deterministic Astra/Luna/Sol routing at the configured efforts. One selected parent owns integration while independent frozen items may run in parallel or through a bounded child hierarchy.

Ambiguity gets the same question

Only a genuine Luna/Sol routing ambiguity can trigger optional, identical read-only probes. They receive the same frozen card and checks; the primary records the decisive facts and keeps the final route decision.

Resumable proof, then a fresh look

Pulsar keeps private, resumable evidence locally without secrets, raw prompts, or a project-operated service. Frozen checks run on each attempt, and a fresh Sol review supplies the final verdict.

Pulsar evidence

Fewer tokens in context. More eyes on the work.

Astral loads instructions progressively, so each run reads only what its route needs. These numbers are not estimates: they come from the committed, reproducible benchmark in the repository. They are static instruction-context measurements, not a score for task quality.

Instruction-context tokens, by loading path

Static instruction-context measurements · tiktoken 0.13.0 · o200k_base · v3.6.0 · 2026-08-21

Pulsar explicit evidence route 12,401

Orbit / full normal project route 10,223

Comet path tiny tasks 4,722

Core skill always loaded 3,403

Scope: Comet loads the core skill and mode/risk reference only; static instruction-context measurements are not task quality, latency, price, or total-run tokens. Regenerate the published v3.6.0 measurement with one command from benchmarks/ in the repository.

53.8% instruction-context tokens Comet avoids versus Orbit / full loading

5,501 instruction-context tokens Comet avoids versus Orbit / full loading

3,403 tokens for the entire core skill in the published v3.6.0 measurement

Token-efficient by design

Progressive disclosure is measured, not hoped for: the Comet route avoids 5,501 instruction-context tokens (53.8%) compared with Orbit / full loading.

A distinct check before handoff

Splitting planning, doing, and checking means the reviewer had no hand in the work it judges. Route proof blocks silent model or effort swaps, so the workflow you configured is the one that actually runs.

Evidence you can reproduce

The chart regenerates from one command in benchmarks/. Read the benchmark guide and inspect the committed context-footprint JSON. The local scorecard runs locally on your supplied trials; it can assess them, but it is not a standing product-performance claim.

Optional Adaptive routing chooses an eligible GPT-6 worker model and effort once per delegated task, including tasks transported through OpenCodex. It never changes the running primary or authorizes execution. The separate session plugin starts off and uses an explicitly selected TypeSafe or OpenRouter Jev key; see routing setup.

Honest scope: these static instruction-context measurements do not measure task quality, latency, price, or total-run tokens. They do not by themselves prove that every orchestrated run uses fewer total tokens.

The current repository verification passed 100+ automated tests and package verification. This validates behavior and contracts: progressive context loading, exact pinned routes, objective checks plus fresh review, and a local privacy/no-analytics runtime posture. It does not prove Astral beats single-Sol.

The first end-to-end pilot is retained as invalid exploratory evidence after fresh review found protocol defects. Read the invalid-pilot disclosure. No valid outcome comparison exists. This page does not publish outcome, token, time, or quality numbers from that pilot.

Ready when you are

Two minutes to a steadier route.

Two terminal commands add the GitHub marketplace and install the plugin in your local Codex and project environment. The optional native-profile setup remains available when you want it.

Open the quick install

Need a hand?

Start with the canonical docs, then use the public issue tracker for bugs, confusing behavior, or improvement ideas.

Open documentation Read support guidance

Want the facts first?

Read the concise privacy policy and terms before installing or using the project.

Read privacy Read terms