Change Crew
Core Rule
Stay context-light by default. Do not preload references, templates, or every agent brief. Load the smallest set needed for the current risk.
The main Codex thread is the orchestrator. Use worker agents only for context isolation, independent judgment, or parallel non-overlapping work.
Pick A Profile
| Profile | Use when | Minimum behavior |
|---|---|---|
quick | Small, single-surface, low-risk change | No council ceremony. Identify scope, implement, review your diff, verify. |
standard | Normal feature/fix with some risk | Create task workspace, discover surfaces, freeze a short contract, implement, review, verify. |
full | Cross-repo, multi-surface, security/data/migration/release risk | Use deliberation, proposal critic, specialist agents, security, docs, bounded fix loops, memory capture. |
Default to quick unless the user asks for council workflow, multiple agents, auto mode, memory, security review, or the task clearly spans multiple surfaces.
Start
Use scripts instead of reading templates into context:
python3 skills/change-crew/scripts/init_task.py <repo> "<task title>"
python3 skills/change-crew/scripts/detect_surfaces.py <repo>
python3 skills/change-crew/scripts/discover_commands.py <repo>
Use scripts/init_memory.py <repo> only when persistent memory is needed or .change-crew/memory/ is missing and the task should record reusable constraints.
Gates
For standard and full, do not implement until these exist, even if short:
prd.md: user intent and acceptance criteria.surface-map.md: affected backend/web/mobile/data/infra/docs/tests surfaces.contract.md: API/schema/UI/client behavior and compatibility decisions.test-plan.mdorenvironment-report.md: commands and verification path.
For full, also require impact analysis, proposal critique, and explicit fix-loop policy. Read references/state-machine.md only when transitions or stop conditions are non-trivial.
Before standard or full freezes a contract, use references/grill-gate.md only when a key unknown could change scope, contract, risk, or acceptance criteria.
Agent Routing
Read references/agent-topology.md only before spawning multiple agents. Then load only the needed brief from references/agent-briefs/.
- Research: current external facts, APIs, vulnerabilities, or best practices.
- Proposal critic: score plans before approval and reduce sycophancy.
- Surface developer: one bounded backend/web/mobile/data/infra/docs/tests scope.
- Review: read-only quality/spec review.
- Security: read-only vulnerability/risk review.
- Verification: commands and evidence.
- Specialists: dependency, migration, external service, rollout/observability, memory.
Never run parallel developers on overlapping files. Prefer read-only specialists over broad developer prompts.
Reference Router
Load exactly one or two references at a time:
- Surfaces unclear:
references/surface-detection.md, thenreferences/surface-playbooks.md. - Unsure what to load or context is getting heavy:
references/context-loading.md. - Key unknown before contract or approval:
references/grill-gate.md. - UI/design task:
references/design-skill-integration.md. - Contracts or caller risk:
references/impact-analysis.md. - Proposal approval:
references/proposal-critique-policy.md. - Failed review/verification:
references/loop-policy.md. - Quality/security gate:
references/review-security-policy.md,references/risk-rubric.md. - Verification claims:
references/verification-policy.md. - Git/PR actions:
references/git-pr-policy.md. - Docs/API/migration notes:
references/documentation-policy.md. - Memory capture or conflict:
references/memory-policy.md. - Release of this skill:
references/publish-quality.md.
Do not read all references "just in case".
Memory
Memory is suggest-only by default. When the user explicitly says to remember a rule, workflow, command, or preference, write a memory suggestion immediately.
Before changing long-term memory, run:
python3 skills/change-crew/scripts/check_memory_conflicts.py <repo> <memory-file> --text "<proposed rule>"
If there is a conflict, show both rules and ask the user to keep, replace, or merge. Never silently overwrite memory.
Done
Do not claim done unless verification evidence exists or the blocker is explicit. For standard and full, record changed files, open findings, commands run, exit status, and remaining risks in final-report.md.