ai-contract-platform/AGENTS.md

AI-assisted work

This repo no longer uses gstack.

• Do not require, install, or verify ~/.Codex/skills/gstack before working.
• Do not block AI-assisted work when gstack is missing.
• Use the current AI coding tool's built-in capabilities and the repository's scripts directly.
• For web browsing, use the available browser or web tools directly; do not require /browse.

Mandatory release-closeout efficiency rules

These rules apply to every AI-assisted development, QA, release, deployment, and UAT task in this repository. Before such work, read:

• README.md, section Mandatory Delivery Policy / 强制交付规则
• project-docs/AI合同平台-生产上线加速与开发效率整改方案-2026-09-21.md
• project-docs/AI合同平台-多场景完整交付与开发收口方案-2026-09-10.md

For the current RC-05 through RC-08 closeout, the delivery baseline is 60 assets, 14 Prompt modules, and 144 formal evaluation scenarios. Do not silently reduce it to a procurement-only pilot, and do not expand it with new non-blocking features before this closeout finishes. This temporary scope freeze ends only when the release owner records completion and updates the baseline; it is not a permanent restriction on later product development.

1. State one deliverable, exit condition, and time box before starting a task. Do not automatically invent a follow-on phase when the requested result is complete.
2. During ordinary implementation, default to the direct failing example and affected tests. Expand validation for security, migration, shared-contract, other high-risk changes, or an explicit user requirement. Run the complete deterministic gate once after integration and candidate freeze; run real-model certification and UAT only against that same immutable candidate.
3. Do not create a formal tag, source manifest, release image, SBOM, full model matrix, or formal UAT evidence while relevant source files are still changing.
4. Parallel writers must use separate Git worktrees/branches with declared file ownership. One integrator merges them before freeze. Never generate immutable evidence from a shared worktree while another agent is editing it.
5. After freeze, an evidence-backed P0 may directly invalidate the candidate; a P1 requires release-owner classification. Batch approved fixes and target at most one refreeze. If another proven P0 or approved P1 appears, remain NO_GO, invalidate the candidate, and escalate the decision; never ship merely to satisfy the refreeze target. Record P2/P3 issues in the post-launch backlog.
6. Keep raw JSON, JUnit, per-case outputs, archives, and diagnostics outside the source worktree in persistent build-artifact storage. Track only required summaries, manifests, indexes, and hashes. Never bulk-delete existing untracked files without ownership review.
7. Treat real-model runs as bounded background release jobs: canary first, no silent retry or fallback, stop dispatching a failing class, and preserve every failure in the denominator. Do not repeatedly poll or rerun to manufacture a pass rate.
8. Track external dependencies such as policy approvals, domains, TLS certificates, Provider credentials, production hosts, and named approvers separately from engineering work. Report the owner and impact; do not replace missing external material with more product code.
9. Preserve user-owned changes, especially docs/architecture/local-development.md and start-dev.ps1, unless the user explicitly asks to modify them.
10. Speed never authorizes bypassing tenant isolation, contract-fact integrity, version protection, migration safety, backup/restore, secrets/TLS, failure recovery, formal policy approval, or final public cutover authorization. These efficiency rules are not standing authorization for production deployment, migration, rollback, cutover, real credentials/data use, risk acceptance, or legal signing; each still requires the appropriate named authority.

The only sequence for the current closeout is FREEZE → BUILD → RC05-RUN → RC06-UAT → RC07-RECOVERY → RC08-PRODUCTION. Until corresponding real evidence exists, keep internal controlled pilot and public release at NO_GO, and never describe local tests or AI review as legal approval, third-party notarization, or production authorization.