millrace-ai 0.15.3

Millrace runtime rewrite baseline package

Millrace

Other agents win sprints. Millrace wins marathons.

Raw agent harnesses - Codex, Claude Code, Gemini - are built for sprints. Give them a tight spec in a greenfield repo, let them rip, ship the result. They are very good at this.

Millrace starts where those runs end. When the work spans sessions, accumulates state, needs to survive a crash without losing context, or has to gate on real acceptance criteria rather than "the agent said it was done", that is the problem Millrace was built for.

Millrace is a governed runtime for long-running agent work. The harness still does the local stage work. Millrace owns the queue, the compiled plan, runtime state, recovery paths, and closure behavior around that work.

If you are an agent and your harness supports repo-local skills, for example Codex or Claude Code, install docs/skills/millrace-ops-agent-manual/SKILL.md to become a Millrace Operator before deciding whether work should enter Millrace. It teaches when to use Millrace, when not to, and how to run, monitor, and intervene safely.

Agent Install Examples

• Codex: if your environment supports repo-local skills, load docs/skills/millrace-ops-agent-manual/ as the local skill package. The entry file is docs/skills/millrace-ops-agent-manual/SKILL.md.
• Claude Code: use the same package path or entry file as the local project skill, then follow the autonomy handshake inside the skill before deciding whether work should enter Millrace.
• Other harnesses: if the harness accepts a SKILL.md, point it at docs/skills/millrace-ops-agent-manual/SKILL.md; if it supports directory packages, prefer docs/skills/millrace-ops-agent-manual/.

Runtime Lifecycle

Need the full implementation-accurate flow instead of the landing-page view? See the standalone lifecycle chart.

flowchart TD
    A["Initialize workspace, then compile the frozen plan"] --> B{"Deterministic tick loop"}
    B --> C["Process control inputs:<br/>mailbox commands, watcher intake, reconciliation"]
    C --> D{"Scheduler claim decision"}
    D -- planning incident or spec --> E["Planning loop:<br/>interpret specs and incidents,<br/>govern remediation, emit executable work"]
    D -- execution task --> F["Execution loop:<br/>build, verify, repair, recover, update"]
    D -- nothing claimable --> G{"Completion behavior eligible?"}
    G -- yes --> H["Arbiter closure pass"]
    G -- no --> I["Idle until the next tick"]
    E --> J["Runtime applies results,<br/>persists state, and routes the next action"]
    F --> J
    H --> J
    J --> B
    I --> B

Millrace does not try to replace raw harness reasoning with a thicker prompt. It wraps long-horizon work in a real runtime:

• workspace bootstrap is explicit: run millrace init before operator commands
• managed baseline refresh is explicit: run millrace upgrade to preview or apply packaged workspace asset updates
• compile happens at startup and again only on explicit config reload
• compile tracks input fingerprints so operators can see whether the persisted compiled plan is current or stale
• planning and execution are claim domains inside one deterministic scheduler, not two concurrent lanes
• stage results are routed by the runtime, not by direct stage-to-stage handoffs
• Arbiter activates only when the scheduler finds no lineage work left and closure behavior is actually ready
• runtime startup and config reload refuse to keep running on a stale last-known-good plan when current compile inputs no longer match

The shipped core already includes separate planning and execution loops, typed terminal results, compiler-governed completion behavior, and persisted run artifacts for post-run inspection.

Early Proof

Millrace already has a useful public benchmark, and the right read is not "Millrace already beats raw Codex on absolute final quality." The useful read is that framework-driven orchestration is already competitive on hard, long-horizon work while being much more efficient.

On the first substantive public A/B benchmark, both systems were aimed at the same target: a parity-first modern Fabric port of Aura Cascade, a ten-year-old Minecraft mod. The stronger direct-agent condition, raw Codex on gpt-5.4 xhigh, finished at 95 / 100. Millrace, running as a staged daemon workflow on routed gpt-5.3-codex high / xhigh, finished at 94 / 100.

Metric	Raw Codex	Millrace
Final score	95 / 100	94 / 100
Total tokens	1,071,700,018	241,046,303
Wall-clock span	72h 23m 20.320s	28h 02m 36.972s
Active runtime	18h 04m 07.914s	12h 36m 15.515s

That means raw Codex used about 4.45x Millrace's total tokens, took about 2.58x the wall-clock span, and still used about 1.43x Millrace's active runtime.

That wall-clock gap is not pure model speed. The raw Codex run needed repeated manual continuation prompts whenever the operator was away from the keyboard, while Millrace kept progressing through a staged runtime. Even after accounting for that, the active-runtime gap still favors Millrace.

Read the full public evidence pack here:

• codex-vs-millrace-mc-mod

How Millrace Fits With Raw Harnesses

Millrace is not a replacement for Codex, Claude Code, Aider, or similar raw agent harnesses. It is the runtime layer you put around them when the work is too long-running, stateful, or recovery-sensitive to trust to a single session.

Think of the split this way:

• the raw harness reasons locally, edits code, and emits a stage result
• Millrace decides which stage runs next and what contract that stage receives
• Millrace persists queue state, runtime snapshots, artifacts, and recovery context after each handoff
• the operator or ops agent decides when work enters the runtime and how the workspace is configured

If a direct Codex or Claude Code session is enough, use the direct session. Millrace matters when the work has crossed out of sprint territory.

When To Use Millrace

Use Millrace when:

• the work will outlast a single agent session
• you want explicit stage gates instead of "done enough" chat conclusions
• recovery and resumability matter
• you need durable state, queue artifacts, and run history under <workspace>/millrace-agents/
• completion has to clear a real closure pass rather than informal optimism
• an operator or ops agent is intentionally managing intake and runtime control

Do not use Millrace when:

• the task is small, bounded, and cleanly handled in one direct session
• the work is exploratory and governance would add more overhead than value
• single-session throughput matters more than persistence and recovery
• nobody is available to manage runtime configuration, intake, and workspace hygiene

60-Second Proof

Install:

pip install millrace-ai

Then point Millrace at a workspace:

export WORKSPACE=/absolute/path/to/your/workspace

millrace init --workspace "$WORKSPACE"
millrace compile validate --workspace "$WORKSPACE"
millrace run once --workspace "$WORKSPACE"
millrace status --workspace "$WORKSPACE"

That flow proves seven things quickly:

• workspace bootstrap is explicit and creates the managed baseline under millrace-agents/
• the selected mode compiles into one persisted compiled_plan.json before execution
• compile output fingerprints the selected mode, runtime config, and packaged assets so compile show / status can report whether the plan is current or stale
• that compiled plan carries node bindings, intake entries, recovery policies, closure-target activation, and post-stage routing
• the shipped default_codex mode freezes closure behavior directly into that single compiled artifact
• status and run inspection carry compiled-plan identity so operators can tie runtime activity back to the frozen plan that produced it
• the runtime can execute a deterministic tick and report persisted status

When the packaged workspace baseline changes, use millrace upgrade first to preview the managed-file classifications, then millrace upgrade --apply to apply safe baseline updates. If compile inputs drift and the persisted plan is stale, runtime startup and config reload refuse to keep running on the stale plan.

Canonical shipped modes today:

• default_codex
• default_pi

Compatibility alias:

• standard_plain -> default_codex

Read By Journey

Need the single dense system explainer first? Start with docs/millrace-technical-overview.md.

Start Here

• docs/runtime/README.md
• docs/skills/millrace-ops-agent-manual/SKILL.md if you are operating Millrace as an agent

Run It

• docs/runtime/millrace-cli-reference.md
• docs/runtime/millrace-runtime-architecture.md

Understand It

• docs/runtime/millrace-compiler-and-frozen-plans.md
• docs/runtime/millrace-modes-and-loops.md
• docs/runtime/millrace-arbiter-and-completion-behavior.md
• docs/runtime/millrace-runner-architecture.md

Extend It

• docs/runtime/millrace-entrypoint-mapping.md
• docs/runtime/millrace-loop-authoring.md
• docs/skills/millrace-loop-authoring/SKILL.md
• docs/source-package-map.md

Status

Millrace ships as a maintained pre-1.0 runtime line. If you depend on exact behavior, pin to a patch version and verify against the current CLI and docs rather than assuming every newer build is identical.

License

See LICENSE.