Skip to content

scores and autonomy

Jump to bottom
mastalink edited this page Oct 9, 2026 · 2 revisions

Autonomous Scores & Conductor

Goal

Author, validate, and monitor Scores — versioned, deterministic JSON contracts that coordinate long-running autonomous development and operations runs with cryptographic evidence, resource locks, and policy gates.

Step-by-Step Instructions

1. What is a Score?

Unlike simple jobs or legacy workflows, a Score is an immutable, digest-bound contract that defines:

  • The complete goal hierarchy and dependency graph.
  • Required worker capabilities and role boundaries (independent authors and reviewers).
  • Resource locks (such as isolated repository worktrees or deployment lanes).
  • Mandatory test evidence and verification receipts.
  • Explicit human checkpoints and integer-cent budget limits.

2. Watching Score Runs in the Console

  1. Open Home. Each Score run appears under What's running as one card per run, with a state word, a sentence from the worker's progress and the step count from the real plan.
  2. Tap a Score card to watch its read-only live drawing.

See Console Home and Live Score for what the cards and the drawing show. The older Live look window on the Overview screen no longer exists.

3. Validating and Compiling Scores Offline

Before submitting any score, validate and compile it locally using the offline kernel:

# 1. Validate score structure against JSON Schema
python -m mco.orchestrator.scores validate examples\scores\via-cloud.score.json

# 2. Compile score into deterministic, digest-bound work packets
python -m mco.orchestrator.scores compile examples\scores\via-cloud.score.json

# 3. Preview planned tasks, dependencies, and capability requirements
python -m mco.orchestrator.scores preview examples\scores\via-cloud.score.json

The CLI Equivalent

Manage and advance Score Conductor runs from PowerShell:

# Check status of conductor and active score runs
mco score status --run-id <run-id>

# List the runs in the conductor database
mco score list

# Advance the conductor sweep engine
mco score tick --run-id <run-id>

# Grant local development policy authority
mco score grant-local examples/scores/via-cloud.score.json --run-id <run-id> --action <action> --resource <resource>

What You'll See

  • Independent Review Enforced: A worker cannot approve its own work. If codex-build-1 authors a change, the Conductor requires an independent reviewer (reviewer-beast or gemini-qa-1) before the task is accepted.
  • Worker Completion ≠ Task Acceptance: A worker returning completed does not advance the score. Only verified evidence artifacts and independent review receipts unlock downstream tasks.
  • Evidence Binding: Every event is tied to the score document digest (SHA-256) and run ID, guaranteeing that changing a score contract invalidates previously issued approvals.

If It Goes Wrong

1. "Score validation error: Cyclic dependency"

  • Cause: One goal packet depends on another in a circular loop.
  • Fix: Run scores validate to see the exact cycle path reported by the validator.

2. "Score run paused at human checkpoint"

  • Cause: The score reached a stage with checkpoint: true requiring explicit owner authorization.
  • Fix: Open the Governance or Approvals panel to sign off on the checkpoint decision.

3. "Budget reservation exceeded"

  • Cause: Planned tasks request more compute or token expenditure than the authorized integer-cent limit (authorized_budget_cents).
  • Fix: Issue an updated grant with higher budget limits using mco score grant-local.

Clone this wiki locally