zhihuiqz

Imported from jiangsanlin-crypto/zhihuiqz (AGENTS.md). Install upstream with npx skills add jiangsanlin-crypto/zhihuiqz. Copyright stays with the author.

Multi-Agent Workflow

The normal delivery pipeline has three logical AI work bodies plus one emergency account-backed escalation worker.

Canonical pipeline

Codex / product planning (OpenAI API Luna)
        ↓
OpenAI Validator / prototype validation (OpenAI API Luna)
        ↓
ChatGPT / formal implementation (owner account Chat 5.6 Sol High)
        ↓
Work GPT-6 / code review + targeted repair
        ↓
OpenAI Validator / QA + UI/UX + classification acceptance
        ↓
Codex / release review + release notes
        ↓
OpenAI Validator / deployment readiness gate
        ↓
GitHub Actions / merge + deploy + health verification / rollback
        ↓
status:done

For handoff compatibility, OpenAI Validator keeps agent ID workbuddy. This is not WorkBuddy Cloud and requires no WorkBuddy OAuth.

Emergency escalation

The owner's ChatGPT Work GPT-6 worker is the emergency/high-difficulty implementation path. It may take an implementation task when explicitly routed for urgency, repeated CI failure, a blocked normal implementation, or a complex recovery task. It must obey the same PR-branch, handoff, CI and production boundaries as the normal ChatGPT implementation role.

The emergency worker does not authorize direct main writes, merge or deployment.

Handoff rule

Every completed phase publishes <!-- agent-handoff:v1 -->.

A next phase may start only when:

• task ID matches;
• from/to logical agent matches;
• expected phase matches;
• previous handoff status is success;
• blockers are empty;
• handoff source SHA equals the current PR head SHA.

API-owned phases use GitHub Actions dispatch. Account Chat implementation is queued by PR labels and consumed automatically by the account scheduled worker. After Account Chat finishes, it queues agent:workreview + phase:code-review. Work GPT-6 independently reviews the implementation, performs only targeted repairs when necessary, verifies exact-final-SHA CI, then queues agent:workbuddy + phase:qa with status:todo last. That final label event starts OpenAI Validator QA.

Model and execution policy

• Codex API coordinator: gpt-6-luna / high.
• OpenAI Validator API: gpt-6-luna / high.
• Primary programmer: owner account ChatGPT 5.6 Sol / High.
• Mandatory senior reviewer/repair engineer: owner account ChatGPT Work GPT-6.
• Emergency programmer: owner account ChatGPT Work GPT-6, subject to the model actually available/configured in Work.
• API Sol implementation worker: retired; normal workload must be zero.

OpenAI now exposes GPT-6 Luna and GPT-6 Sol in the API as gpt-6-luna and gpt-6-sol. API coordination/validation uses gpt-6-luna / high. The API Sol programmer remains retired because normal implementation is account-backed.

Production policy

Automatic merge/deploy is allowed only when repository production controls, required CI and all release/deployment gates succeed. No agent may bypass failed gates or branch protection.

Real payment execution and real candidate production data require separate explicit authorization.