[houlianpi]

Summary

Streamline the shared agent workflow without weakening write authorization, independent review, verification coverage, or release permissions.

This consolidates the six previously uncommitted workflow changes with the agreed general improvements. It changes agent instructions only, not product code, project SPECs, CI configuration, dependencies, or release artifacts.

Related issue

No linked issue. This is maintainer-requested workflow-control maintenance, explicitly authorized and discussed directly; it does not change a public product contract.

Design document

No separate design document: documentation/workflow-control maintenance. The maintainer confirmed the general design before implementation:

• Local verification follows diff impact; CI owns the full suite and platform matrix. Expanding local verification requires a concrete evidence gap.
• Default team: Tech Lead coordinates, Developer implements, and one independent combined Reviewer reviews. PM and Designer join only when needed for actual decisions.
• Reuse matching raw verification evidence; independently inspect actual changes and their authority.
• Continue authorized work autonomously, escalate genuine decisions or external permissions, and deliver useful milestones promptly.

These are general rules, not a special fast path for version bumps.

SPEC updates

None. This is workflow-control-only maintenance under the existing root authorization exception, not project development or a project no-SPEC-delta bug fix. No project SPEC was changed or used to invent workflow requirements; product behavior and architecture remain untouched.

The project-SDD checklist below is not applicable to this maintenance task and remains unchecked for that reason.

Changes

• AGENTS.md: clarify default role selection, proportional work, autonomy, and pause/safety boundaries.
• Developer and Reviewer agents: clarify maintenance versus project-development gates, frozen-input discipline, evidence reuse, and independent judgment.
• Tech Lead agent: make PM/Designer participation conditional, keep ordinary technical decisions within the team, and report milestones and stage-specific blockers promptly.
• spec-driven: assign local focused checks and full CI coverage, reuse evidence with affected-only invalidation, and prefer one compact complete review package over repeated repository copies.
• spec-implementation-audit: retain full bidirectional change-scope coverage, domain/risk checks, and repair-introduced behavior review without re-auditing unrelated history.
• Preserve the previously uncommitted parallel-verification, combined-review, and consolidated-repair improvements.
• Preserve explicit project write authorization, SPEC confirmation, independent review, complete findings, the automatic repair limit, and separate external-operation permissions.

Verification

• git --no-pager diff --check — passed before commit.
• git diff --cached --check — passed when committing.
• git --no-pager diff --check origin/main...HEAD — passed before publication.
• python3 inline validation using existing PyYAML 6.0.3 — all five YAML frontmatters parsed successfully and remained byte-identical to their pre-edit versions; 21 relative links, anchors, and explicit workflow references passed.
• Independent full six-file workflow-document review — passed with no findings, covering both the prior uncommitted changes and the new optimization.
• Static scenario review — passed for a clear bug, ambiguous cross-module work, a high-risk one-line change, a CI-pending PR, reused versus invalidated evidence, external approval, and user pause.
• Commit identity checks — the six committed file contents matched the reviewed snapshot; no product or unrelated files were included.

No product pytest/npm/build suite was run locally: this is documentation-only maintenance with no runtime or build changes. The documented area-focused testing policy is retained; any CI checks run by GitHub remain separate evidence.

Evidence

• Reviewed commit: 9acdbb4.
• Complete reviewed diff: six files, 201 insertions / 92 deletions; SHA-256 1b91e8aa7acdfd0d4cf61b6f4a6bd9b9edeb162acb70462c15225fec1c489b40.
• No screenshots or raw session logs are uploaded; there is no product UI change.
• Runtime efficacy is not established. A single paired replay used identical business code and model, but both runs stopped before implementation because of CLI execution permissions and skill-registration failures. The observed 125.8 s versus 100.6 s blocked-session exit times are not evidence of a 20% development productivity gain.

Release or documentation impact

Agent workflow documentation changes only. No README, product setup, packaging, release workflow, public media, or version changes. The separate release task remains paused; this PR does not authorize or trigger PyPI publication.

Checklist

Spec-driven development

• I reviewed root SPEC.md and every relevant module SPEC.md.
• Any required design and SPEC confirmation gates occurred before implementation.
• The implementation, tests, and documentation agree with the current confirmed SPEC files.
• I used the current confirmed SPEC files, not the design document, as the implementation source of truth.

Quality and safety

• I ran focused tests and required formatting or lint checks, or explained why a check was not applicable.
• I updated user-facing documentation when behavior or setup changed, or it is not applicable.
• I did not commit credentials, tokens, account data, personal data, local .env files, generated logs, reports, screenshots, or workspace output.
• I reviewed any included screenshots, videos, logs, reports, and Run artifacts for secrets, private paths, device identifiers, and account data.
• I did not add large binary media to Git history unless the repository already owns that exact artifact and the size is justified.