sourabhJainR

agency-runtime-v9

Conflict-aware multi-specialist execution planning for Agency Runtime, with deterministic scheduling, explicit resource ownership, serialized mutations, and parallel read-only work.

sourabhJainR 0 Updated 1mo ago
GitHub

Install

npx skillscat add sourabhjainr/auren/agency-runtime-v9

Install via the SkillsCat registry.

SKILL.md

Agency Runtime v9

Contract

The planner decides who may work when; it does not execute commands or mutate files.

Every specialist work unit must declare:
ROLE | MUTATION_MODE | READ_PATHS | WRITE_PATHS | DEPENDENCIES | PRIORITY.

Scheduling rules

  1. Read-only specialists with no dependency or resource conflict may share a wave.
  2. Any write/write overlap is a conflict and must be serialized.
  3. A write/read overlap is a conflict and must be serialized.
  4. Independent mutations are still serialized so mutation order remains explicit and auditable.
  5. Dependencies create ordering edges; a dependent specialist cannot enter an earlier wave.
  6. Unknown dependencies block the plan instead of being silently ignored.
  7. Read-only work must never declare write paths.
  8. Mutating work must explicitly declare write paths.

Determinism

Plans are stable for the same work specification. Ordering uses priority, role, specialist name, dependencies, and resource conflicts. Every plan exposes a digest suitable for provenance and regression history.

Provenance

Record the execution-plan digest alongside specialist selection, evidence, verification, regression, and release events. Do not put secrets or raw sensitive payloads into scheduling metadata.

Host boundary

The AI coding orchestrator remains responsible for sandboxing, credentials, permissions, command execution, external side effects, and the final mutation mechanism. The planner only supplies an auditable execution schedule.

Failure handling

A missing dependency or invalid resource declaration is a planning failure. A conflict is not itself a failure; it is a scheduling fact that must result in serialization. Never report parallel execution where the plan serialized the work.

Categories