Coordination core

Coordination Core

Coordinate delegated work through a portable contract. Keep product commands, model names, permission controls, and status strings in harness adapters.

Classify Before Starting

1. Collect only observable inputs defined below and in references/classification.md.
2. Run scripts/classify.py with those inputs when tool access exists. In an instruction-only surface, apply the same rules below verbatim.
3. Treat difficulty, risk, duration, coordination shape, requested capability tier, and reasoning effort as independent outputs.
4. Ask the selected adapter to resolve the requested tier against capabilities it can prove. Record unavailable or unknown; never infer availability from an unlabeled or shared usage percentage.
5. Reclassify only when a mechanical input changes. Do not relabel work merely because an agent sounds confident or uncertain.

Use the version 0.1 rules exactly:

• difficulty: hard if work crosses systems or a prior attempt failed; otherwise trivial for at most one file group; otherwise standard.
• risk: high for an irreversible or external action; otherwise elevated when a plan gate is required; otherwise low.
• duration: map small, medium, or large expected tool calls to short, medium, or long.
• coordination: team for at least two independent subtasks that cross systems; otherwise fan_out for at least two; otherwise single.
• requested_tier: max for hard and long; otherwise strong for hard or elevated/high risk; otherwise routine.
• reasoning_effort: high for hard difficulty or high risk; low for trivial and low risk; otherwise medium.

Emit exactly those six field names and declared values. Do not invent synonyms such as low difficulty, coordination_shape, or requested_capability_tier. If neither the rules nor the helper are available, report that classification is blocked instead of fabricating a result.

Coordinate Through Envelopes

Read references/protocol.md, then emit append-only JSON Lines records conforming to schemas/envelope.schema.json.

• Start with one assignment.
• Accept worker state, question, approval_request, evidence, and handoff records.
• Send one concrete next action per steer.
• Route approval requests to a human. Only a human may emit approval_decision; a timeout is a denial.
• Request independent exact-change review with review.
• Close a requested stop only after a terminal worker state; archive or idle is not stop evidence.
• Verify evidence against the referenced artifact, diff, test output, or live system. Do not use an agent summary as proof.

Run scripts/validate_envelope.py <events.jsonl> before trusting a stream.

Track Completion Separately

Maintain these ledger fields independently:

• implementation;
• local validation;
• live end-to-end validation;
• review;
• approval;
• release.

Use pending, in_progress, done, blocked, or not_applicable. Local tests never imply live validation, approval, or release.

Preserve Boundaries

• Scope reads and writes to the assignment.
• Keep events append-only and monotonically sequenced.
• Record the actual resolved worker identity or unknown, not merely the requested tier.
• Require a human decision for external actions.
• Mark unsupported adapter capabilities unavailable.
• Keep vendor policy in adapters so the core remains portable.