Context-light delivery workflow for Codex and SKILL.md-compatible coding agents.
Change Crew starts light, then adds structure only when the change deserves it:
quick scope -> implement -> review -> verify
standard/full risk -> contract -> agents -> critique -> review -> secure -> verify -> fix -> ship
It is built for the part of AI coding that usually gets messy: cross-surface changes, hidden risks, vague handoffs, skipped verification, and agents that agree too quickly. It deliberately avoids loading every prompt, template, and agent brief up front.
Most agentic coding failures happen at boundaries:
- frontend and backend implement different contracts;
- mobile, web, backend, data, and infra agents work from different assumptions;
- proposal authors drift into optimistic plans without an independent critic;
- review finds issues but does not turn them into actionable fix work;
- security checks are skipped or mixed into normal review;
- hidden callers, migrations, rollback, and docs are forgotten;
- verification is claimed without command output;
- agents debate forever instead of shipping.
Change Crew makes those boundaries explicit with artifacts, gates, independent critique, and bounded fix loops.
- Multi-surface orchestration for backend, web, mobile, desktop, data, infra, docs, and tests
- Quick, standard, and full profiles so small changes do not pay the full workflow cost
- PRD, impact analysis, research, deliberation, contract, test plan, implementation plan, verification report, and final report templates
- Independent proposal critic to reduce sycophancy and weak plans
- Grill Gate questioning for the one unknown that can change the plan
- Read-only review and security gates
- High-density core and specialist agent briefs for research, critique, surface development, review, security, verification, dependency, migration, external service, rollout, observability, and memory work
- Fix packets that turn findings into bounded implementation work
- Surface detection and command discovery scripts
- Long-term memory templates for project constraints, commands, decisions, and lessons learned
- Memory conflict checks before writing new long-term rules
- Design-skill routing for web and mobile UI tasks
- Context-loading policy so unused references and agent briefs stay unloaded
codex plugin marketplace add HuangJingwang/change-crewThen open the Codex plugin directory, choose the Change Crew marketplace, and install change-crew.
After installing, restart Codex or start a new thread so bundled skills are reloaded.
/plugin marketplace add HuangJingwang/change-crew
/plugin install change-crew@change-crew
After installing, start a new Claude Code session before testing the skill.
The rename to Change Crew changes the repository, marketplace, plugin, skill, and generated workspace identifiers. GitHub redirects the old repository URL, but existing installations should remove the old plugin or skill and reinstall change-crew.
Existing task data is not moved automatically. Rename .ship-council/ to .change-crew/ and rename each ship-council.json task manifest to change-crew.json before continuing a previous task.
mkdir -p ~/.codex/skills
git clone https://github.com/HuangJingwang/change-crew /tmp/change-crew
cp -R /tmp/change-crew/skills/change-crew ~/.codex/skills/Restart Codex after installing.
mkdir -p ~/.claude/skills
git clone https://github.com/HuangJingwang/change-crew /tmp/change-crew
cp -R /tmp/change-crew/skills/change-crew ~/.claude/skills/For other agents, copy skills/change-crew into that agent's skills directory.
Ask your agent to use the skill. For small tasks, prefer the quick profile:
Use change-crew in quick profile to fix this backend validation bug.
For normal feature work:
Use change-crew to implement organization filtering in the work item list.
For larger tasks:
Use change-crew in auto mode, up to 3 fix loops, to add password reset across web and backend.
For design-heavy UI work:
Use change-crew to redesign the onboarding flow. Route visual decisions through the design skill integration and require screenshot verification.
For plan stress-testing before implementation:
Use change-crew standard profile and run Grill Gate before freezing the contract.
Change Crew has two separate controls:
- Profile controls prompt weight and workflow depth:
quick,standard, orfull. - Mode controls autonomy:
semi-autoorauto.
| Profile | Use When | Behavior |
|---|---|---|
| Quick | Small, single-surface, low-risk change | No full council ceremony; implement, self-review, verify |
| Standard | Normal feature/fix with some risk | Minimal task workspace, surface discovery, short contract, review, verification |
| Full | Multi-surface, data/security/migration/release risk | Deliberation, proposal critic, specialist agents, security, docs, bounded fix loops, memory capture |
Default behavior is intentionally light. The agent should not read every reference or spawn every agent unless the task risk justifies it.
Change Crew is intentionally split into many files, but they are not meant to load together.
The loading model is:
metadata -> SKILL.md -> 0-2 triggered references -> exact agent brief only if a worker is dispatched
Practical rules:
quickshould usually read onlySKILL.md.standardreads only the references needed for the current phase.fullcan use more references, but still one phase at a time.- agent briefs load only for the exact agent being used.
- templates and assets are copied by scripts, not read into context by default.
references/context-loading.mdexists for routing audits and context bloat debugging.
Grill Gate borrows the useful part of grill-me: one sharp question at a time.
It runs only when one unresolved answer could change scope, contract, risk, test plan, rollout, or acceptance criteria. It must search the repo first, ask one question, explain why it matters, and stop after the riskiest branch is resolved.
| Mode | Use When | Behavior |
|---|---|---|
| Semi-auto | You want approval checkpoints | Stops after PRD, contract, high-risk findings, repeated loop failure, and final report |
| Auto | You want the agent to keep moving | Runs normal gates automatically for up to 3 fix loops, but still stops for destructive operations, missing credentials, or policy changes |
flowchart LR
A["Profile Selection"] --> B["Surface Discovery"]
B --> C["Impact Analysis"]
C --> D["Research"]
D --> E["Deliberation"]
E --> F["Proposal Critique"]
F --> G["Contract"]
G --> H["Test Strategy"]
H --> I["Implementation"]
I --> J["Review"]
J --> K["Security"]
K --> L["Verification"]
L --> M{"Pass?"}
M -- "no" --> N["Fix Packets"]
N --> I
M -- "yes" --> O["Memory Capture"]
O --> P["Done"]
Change Crew writes inspectable files into the target repository:
.change-crew/tasks/2026-07-09-org-filter/
change-crew.json
prd.md
surface-map.md
impact-analysis.md
research.md
deliberation.md
proposal-critique.md
contract.md
test-plan.md
environment-report.md
implementation-plan.md
findings/
fix-packets/
git-pr-plan.md
verification-report.md
retrospective.md
memory-suggestions.md
final-report.md
Long-term project knowledge can live beside task history:
.change-crew/memory/
project-profile.md
coding-constraints.md
surface-map.md
verification-recipes.md
security-constraints.md
decisions.md
lessons-learned.md
By default, Change Crew suggests memory updates instead of silently writing long-term memory. If a new rule conflicts with existing memory, Change Crew must show both rules and ask which one to keep, replace, or merge.
- Start light.
- Debate before coding.
- Independent critique before major approval.
- One sharp question beats a broad questionnaire.
- Contract before parallel work.
- Impact analysis before contract.
- Test strategy before implementation.
- Files over paraphrased handoffs.
- Review and security are read-only gates.
- Findings become fix packets.
- Verification must run, not pretend.
- Documentation and rollout are checked before done.
- Design-heavy web/mobile tasks route to specialist design guidance instead of inventing generic UI.
- Reusable constraints become explicit memory suggestions.
- Loops have a time-to-live.
- References are loaded one or two at a time, never all at once.
- Agent briefs are execution cards, not essays.
.codex-plugin/plugin.json
.agents/plugins/marketplace.json
.claude-plugin/plugin.json
.claude-plugin/marketplace.json
skills/change-crew/
SKILL.md
agents/openai.yaml
references/
agent-briefs/
assets/templates/
scripts/
python3 skills/change-crew/scripts/init_task.py <repo> "task title"
python3 skills/change-crew/scripts/init_memory.py <repo>
python3 skills/change-crew/scripts/check_memory_conflicts.py <repo> verification-recipes.md --text "rule"
python3 skills/change-crew/scripts/detect_surfaces.py <repo>
python3 skills/change-crew/scripts/discover_commands.py <repo>
python3 skills/change-crew/scripts/validate_artifacts.py <task-dir>
python3 skills/change-crew/scripts/merge_findings.py <task-dir>| Problem | Fix |
|---|---|
| Plugin does not appear in Codex | Run codex plugin marketplace list, then restart Codex. |
| Skill does not trigger automatically | Invoke it explicitly with change-crew or choose the bundled skill from the plugin. |
| Updated plugin still shows old behavior | Upgrade the marketplace, reinstall the plugin, and start a new thread. |
| Manual skill install not detected | Confirm skills/change-crew/SKILL.md is under your agent's skills directory and restart the agent. |
- GitHub Action for validating skill packaging
- More realistic demo traces
- Optional visual report output
Change Crew includes executable helper scripts. Review them before installing from any fork. Review and security agents are intentionally read-only, and generated findings must include concrete evidence before they can block completion.
Issues and pull requests are welcome. Good contributions make the workflow more reliable, more general across project types, or easier to verify.
MIT