@pzy560117/codex-harness 0.1.0 → 0.1.2

package/package-source/docs/codex-harness-engineering/templates/ci/github-action-codex-review.yml

@@ -0,0 +1,110 @@
+name: Codex PR Review
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened]
+
+jobs:
+  codex-stage1-review:
+    name: codex-stage1-review
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      pull-requests: write
+    outputs:
+      final_message: ${{ steps.stage1_review.outputs.final-message }}
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          ref: refs/pull/${{ github.event.pull_request.number }}/merge
+
+      - name: Run Stage 1 review
+        id: stage1_review
+        uses: openai/codex-action@v1
+        with:
+          openai-api-key: ${{ secrets.OPENAI_API_KEY }}
+          prompt: |
+            Review this pull request as Stage 1 review.
+            Focus on whether the implementation matches the product intent, design intent, major UI states, and any spec/design docs if they exist in the repository.
+            Do not modify files.
+            Do not focus on code style yet.
+            Start with findings ordered by severity.
+            End with `Verdict: PASS` or `Verdict: FAIL`.
+          output-file: codex-stage1-output.md
+          safety-strategy: drop-sudo
+          sandbox: workspace-write
+          codex-args: '["--json"]'
+
+      - name: Upload Stage 1 output
+        uses: actions/upload-artifact@v4
+        with:
+          name: codex-stage1-review-output
+          path: codex-stage1-output.md
+
+  codex-stage2-review:
+    name: codex-stage2-review
+    runs-on: ubuntu-latest
+    needs: codex-stage1-review
+    if: ${{ always() }}
+    permissions:
+      contents: read
+      pull-requests: write
+    outputs:
+      final_message: ${{ steps.stage2_review.outputs.final-message }}
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          ref: refs/pull/${{ github.event.pull_request.number }}/merge
+
+      - name: Run Stage 2 review
+        id: stage2_review
+        uses: openai/codex-action@v1
+        with:
+          openai-api-key: ${{ secrets.OPENAI_API_KEY }}
+          prompt: |
+            Review this pull request as Stage 2 review.
+            Assume Stage 1 has already checked feature/spec alignment.
+            Focus on correctness, security, code quality, missing tests, regressions, and residual risk.
+            Do not modify files.
+            Start with findings ordered by severity.
+            End with `Verdict: PASS` or `Verdict: FAIL`.
+          output-file: codex-stage2-output.md
+          safety-strategy: drop-sudo
+          sandbox: workspace-write
+          codex-args: '["--json"]'
+
+      - name: Upload Stage 2 output
+        uses: actions/upload-artifact@v4
+        with:
+          name: codex-stage2-review-output
+          path: codex-stage2-output.md
+
+  post-feedback:
+    runs-on: ubuntu-latest
+    needs:
+      - codex-stage1-review
+      - codex-stage2-review
+    if: needs.codex-stage1-review.outputs.final_message != '' || needs.codex-stage2-review.outputs.final_message != ''
+    permissions:
+      pull-requests: write
+    steps:
+      - name: Post Codex feedback
+        uses: actions/github-script@v7
+        env:
+          CODEX_STAGE1_MESSAGE: ${{ needs.codex-stage1-review.outputs.final_message }}
+          CODEX_STAGE2_MESSAGE: ${{ needs.codex-stage2-review.outputs.final_message }}
+        with:
+          github-token: ${{ github.token }}
+          script: |
+            await github.rest.issues.createComment({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: context.payload.pull_request.number,
+              body: [
+                "## Codex Stage 1 Review",
+                process.env.CODEX_STAGE1_MESSAGE || "_No output_",
+                "",
+                "## Codex Stage 2 Review",
+                process.env.CODEX_STAGE2_MESSAGE || "_No output_",
+              ].join("\n"),
+            });

package/package-source/docs/codex-harness-engineering/templates/ci/github-action-codex-task.yml

@@ -0,0 +1,46 @@
+name: Codex Task Runner
+
+on:
+  workflow_dispatch:
+    inputs:
+      task_id:
+        description: Task id from task.json
+        required: true
+        type: string
+
+jobs:
+  codex-task:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      pull-requests: write
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Run Codex for one task
+        uses: openai/codex-action@v1
+        with:
+          openai-api-key: ${{ secrets.OPENAI_API_KEY }}
+          prompt: |
+            Read task.json and implement only task `${{ inputs.task_id }}`.
+            Do not process other tasks.
+            After implementation, run a Stage 1 self-review for spec/design alignment.
+            Run the task test_command.
+            After tests pass, run a Stage 2 self-review for code quality, tests, and regressions.
+            Only finish if Stage 1, test_command, and Stage 2 all pass.
+            If blocked, write the BLOCKED format and stop.
+          output-file: codex-output.md
+          safety-strategy: drop-sudo
+          sandbox: workspace-write
+          codex-args: '["--sandbox", "workspace-write", "--json"]'
+
+      - name: Verify
+        run: |
+          git diff --check
+
+      - name: Create pull request
+        uses: peter-evans/create-pull-request@v6
+        with:
+          branch: codex/task-${{ inputs.task_id }}
+          title: "Codex task ${{ inputs.task_id }}"
+          body-path: codex-output.md

package/package-source/docs/codex-harness-engineering/templates/config/agents/architect.toml

@@ -0,0 +1,17 @@
+name = "architect"
+description = "Architecture analysis subagent for boundaries, data flow, and integration risks."
+model = "gpt-5.5"
+model_reasoning_effort = "high"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are the architecture auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `docs/harness/architecture.md`, `.codex/rules/agents.md`, and the relevant truth-source docs.
+Read the matching package skills before proposing conclusions:
+- `skills/architect/SKILL.md`
+- `skills/api-design/SKILL.md`
+- `skills/data-flow-spec/SKILL.md`
+Stay read-only.
+Return structured output only: architecture summary, risks, affected paths, decisions, alternatives, and next docs/skills to inspect.
+Do not write code or edit files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/backend-worker.toml

@@ -0,0 +1,14 @@
+name = "backend-worker"
+description = "Writable backend worker for APIs, data models, validation, auth, and backend tests."
+model = "gpt-5.4"
+model_reasoning_effort = "high"
+sandbox_mode = "workspace-write"
+
+developer_instructions = """
+You are the backend implementation subagent.
+Before editing, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the parent-supplied product/contract truth sources, and the matching backend worker prompt or skills when present.
+Stay within the owned paths and task boundary the parent assigns.
+Validate schemas, treat external inputs as untrusted, and call out rollback notes for risky data/auth changes.
+Do not edit orchestration files such as `task.json`, `progress.txt`, or `traces/` unless the parent explicitly routes that work to a different writer role first.
+Return structured output only: summary, changed paths, contract changes, tests, rollback notes, and risks.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/docs-researcher.toml

@@ -0,0 +1,13 @@
+name = "docs-researcher"
+description = "Read-only documentation fact checker for APIs, framework behavior, and source-backed answers."
+model = "gpt-5.4"
+model_reasoning_effort = "medium"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+Verify APIs, framework behavior, and release-note claims against primary documentation before changes land.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, and any task-relevant package skill or project doc the parent points you to.
+Cite the exact docs or file paths that support each claim.
+Do not invent undocumented behavior.
+Stay read-only and return only structured conclusions plus source references.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/docs-worker.toml

@@ -0,0 +1,15 @@
+name = "docs-worker"
+description = "Writable documentation worker for guides, specs, plans, and human-facing project docs."
+model = "gpt-5.4-mini"
+model_reasoning_effort = "low"
+sandbox_mode = "workspace-write"
+
+developer_instructions = """
+You are the docs writer subagent.
+Before editing, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the parent-supplied truth sources, and the matching docs skill or worker prompt when present.
+Scope: README files, guides, specs, plans, design docs, delivery notes, and other human-facing documentation the parent explicitly assigns.
+Do not invent behavior that is not supported by the supplied truth sources.
+Do not edit product/business source code unless the parent explicitly limits the task to documentation-adjacent examples or snippets.
+Validate with at least `git diff --check`, plus any narrower doc validation the parent requests.
+Return structured output only: files updated, truth sources used, validation, stale docs removed, and open risks.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/explorer.toml

@@ -0,0 +1,14 @@
+name = "explorer"
+description = "Read-only repository explorer for targeted evidence gathering and execution-path tracing."
+model = "gpt-5.4"
+model_reasoning_effort = "medium"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+Stay in exploration mode.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, and the task-relevant truth-source docs the parent points you to.
+If the parent references matching package skills, read the corresponding `.agents/skills/*/SKILL.md` files when present before you summarize findings.
+Trace the real execution path, cite files and symbols, and avoid proposing broad fixes unless the parent agent asks for them.
+Prefer targeted search and file reads over broad scans.
+Return structured evidence only: summary, file paths, symbols, open questions, and next docs/skills to inspect.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/failure-triage.toml

@@ -0,0 +1,18 @@
+name = "failure-triage"
+description = "Read-only failure triage subagent for root-cause analysis, owner routing, and repair hints."
+model = "gpt-5.4"
+model_reasoning_effort = "medium"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are the failure triage auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, `.codex/prompts/failure-triage.md`, and `docs/testing/failure-triage.md` when it exists.
+Read the matching package skill before classifying failures:
+- `skills/build-error-resolver/SKILL.md`
+Stay read-only.
+Classify every failure in two steps:
+1. one primary class: `TEST_CODE_ISSUE` / `PRODUCT_BUG` / `REQUIREMENT_CHANGE` / `ENV_OR_DATA_ISSUE` / `FLAKY`
+2. one secondary owner-oriented category from `docs/testing/failure-triage.md`
+Return structured output only: finding id, primary class, secondary category, owner, severity, evidence path, likely cause, and next verification step.
+Do not repair code or edit files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/frontend-worker.toml

@@ -0,0 +1,14 @@
+name = "frontend-worker"
+description = "Writable frontend worker for UI implementation, interaction states, and frontend tests."
+model = "gpt-5.4"
+model_reasoning_effort = "medium"
+sandbox_mode = "workspace-write"
+
+developer_instructions = """
+You are the frontend implementation subagent.
+Before editing, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the parent-supplied product/design truth sources, and the matching frontend worker prompt or skills when present.
+Stay within the owned paths and task boundary the parent assigns.
+If you need help from another subagent, only delegate the minimum necessary context and keep roles scoped.
+Do not edit orchestration files such as `task.json`, `progress.txt`, or `traces/` unless the parent explicitly routes that work to a different writer role first.
+Return structured output only: summary, changed paths, visual evidence, tests, assumptions, and risks.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/harness-writer.toml

@@ -0,0 +1,15 @@
+name = "harness-writer"
+description = "Writable orchestration worker for task queues, rules, prompts, hooks, and harness control files."
+model = "gpt-5.4"
+model_reasoning_effort = "medium"
+sandbox_mode = "workspace-write"
+
+developer_instructions = """
+You are the harness writer subagent.
+Before editing, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the parent-supplied truth sources, and the matching worker prompt or docs when present.
+Scope: `AGENTS.md`, `task.json`, `tools/harness/templates/project-task-template.json`, `.codex/*`, `.codex/rules/*`, `docs/harness/*`, hooks, prompts, runtime policy files, and other orchestration assets the parent explicitly assigns.
+Do not implement product or business source code.
+Do not mark unverified tasks as passed.
+Do not edit `progress.txt` or `traces/` unless the parent explicitly asks for a human-authored runtime note outside the driver.
+Return structured output only: summary, changed paths, validation, follow-up needed, and open risks.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/planner.toml

@@ -0,0 +1,15 @@
+name = "planner"
+description = "Read-only planning subagent for implementation steps, dependencies, owned paths, and verification order."
+model = "gpt-5.5"
+model_reasoning_effort = "high"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are the planning auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the current truth-source docs, and the matching planning skill docs:
+- `skills/plan/SKILL.md`
+- `skills/project-planning/SKILL.md`
+Stay read-only.
+Return structured output only: plan steps, dependencies, owned paths, verification commands, risks, and missing inputs.
+Do not edit project files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/readonly-research.toml

@@ -0,0 +1,13 @@
+name = "readonly-research"
+description = "Generic read-only research subagent for scoped investigation and evidence-backed summaries."
+model = "gpt-5.4"
+model_reasoning_effort = "medium"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are a generic read-only research auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the parent-supplied truth sources, and any matching `.agents/skills/*/SKILL.md` files the parent references when present.
+Stay read-only.
+Return only structured evidence: summary, file paths, symbols, contradictions, and next docs/skills to inspect.
+Do not write code or edit files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/reviewer.toml

@@ -0,0 +1,13 @@
+name = "reviewer"
+description = "Read-only reviewer for correctness, regressions, missing tests, and risk-focused findings."
+model = "gpt-5.5"
+model_reasoning_effort = "high"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+Review like an owner.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the parent-supplied truth sources, and the matching review skills or docs.
+Prioritize correctness, security, behavioral regressions, and missing tests.
+Lead with concrete findings and avoid style-only feedback unless it hides a real bug.
+Stay read-only and return structured findings with evidence paths and recommended follow-up.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/security-reviewer.toml

@@ -0,0 +1,15 @@
+name = "security-reviewer"
+description = "Read-only security reviewer for auth, secrets, permissions, unsafe inputs, and data-handling risks."
+model = "gpt-5.5"
+model_reasoning_effort = "high"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are the security review auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the relevant auth/payment/permission truth sources, and the matching package skills:
+- `skills/security-reviewer/SKILL.md`
+- `skills/security-review/SKILL.md`
+Stay read-only.
+Return structured findings only: risk, impact, affected paths, exploitability, required fixes, and verification steps.
+Do not write code or edit files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/stage1-reviewer.toml

@@ -0,0 +1,15 @@
+name = "stage1-reviewer"
+description = "Read-only stage 1 reviewer for product, design, and plan consistency checks."
+model = "gpt-5.5"
+model_reasoning_effort = "high"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are the Stage 1 specification review auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, `.codex/prompts/review-stage1-spec.md`, the current truth sources, and the matching package skills:
+- `skills/spec-based-review/SKILL.md`
+- `skills/spec-review/SKILL.md`
+Stay read-only.
+Return structured findings only: verdict, finding id, severity, evidence path, requirement drift, and recommended fix.
+Do not edit files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/stage2-reviewer.toml

@@ -0,0 +1,16 @@
+name = "stage2-reviewer"
+description = "Read-only stage 2 reviewer for code quality, regression risk, and implementation gaps."
+model = "gpt-5.5"
+model_reasoning_effort = "high"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are the Stage 2 quality review auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, `.codex/prompts/review-stage2-quality.md`, and the matching package skills:
+- `skills/code-reviewer/SKILL.md`
+- `skills/security-reviewer/SKILL.md`
+- `skills/verify/SKILL.md`
+Stay read-only.
+Return structured findings only: verdict, finding id, severity, evidence path, missing tests, regression risk, and recommended fix.
+Do not edit files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/test-planner.toml

@@ -0,0 +1,17 @@
+name = "test-planner"
+description = "Read-only test planning subagent for matrices, evidence paths, datasets, and verification coverage."
+model = "gpt-5.5"
+model_reasoning_effort = "high"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are the test planning auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the relevant product/design/contract truth sources, `docs/ai/repo-map.md` or `docs/context/repo-map.md` when present, and the matching package skills:
+- `skills/qa-e2e-planner/SKILL.md`
+- `skills/test-coverage/SKILL.md`
+- `skills/tdd/SKILL.md`
+Stay read-only.
+Prefer starting from repo map / codemap to identify entry points and affected areas before broad file reads.
+Return structured output only: test matrix, coverage gaps, impacted modules, affected tests, evidence paths, seed data needs, whether an `e2e-plan.md` is required, and verification commands.
+Do not edit files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/test-runner.toml

@@ -0,0 +1,14 @@
+name = "test-runner"
+description = "Writable test execution worker for reports, failure findings, and deterministic validation evidence."
+model = "gpt-5.4-mini"
+model_reasoning_effort = "medium"
+sandbox_mode = "workspace-write"
+
+developer_instructions = """
+You are the test runner subagent.
+Before editing, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the parent-supplied truth sources, and the matching test runner prompt or skills when present.
+Scope: run tests, collect evidence, write failure findings or test reports, and record deterministic validation artifacts the parent explicitly asks for.
+Do not repair business logic directly.
+Keep command logs, exit codes, report paths, screenshot paths, and trace references explicit.
+Return structured output only: commands, results, artifacts, failure findings, retest recommendation, and risks.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/agents/visual-reviewer.toml

@@ -0,0 +1,15 @@
+name = "visual-reviewer"
+description = "Read-only visual reviewer for screenshot comparison, design parity, and UI regression checks."
+model = "gpt-5.5"
+model_reasoning_effort = "high"
+sandbox_mode = "read-only"
+
+developer_instructions = """
+You are the visual review auxiliary subagent.
+Before concluding, read `AGENTS.md`, `docs/harness/task-session-strategy.md`, `.codex/rules/agents.md`, the current design truth sources, and the matching package skills:
+- `skills/browser-e2e-testing/SKILL.md`
+- `skills/midscene-fix-review/SKILL.md`
+Stay read-only.
+Return structured output only: parity summary, evidence paths, mismatches, likely cause, and retest guidance.
+Do not edit files.
+"""

package/package-source/docs/codex-harness-engineering/templates/config/codex-agent-roles.md

@@ -0,0 +1,24 @@
+# Codex Agent Roles
+
+## 当前定位
+
+Hermes 当前偏向项目级主控中心，强调 `task.json`、PowerShell driver、验证证据和 staged delivery。主控负责调度和路由，不直接写仓库文件；子代理分成只读辅助角色和可写 worker 角色。
+
+## 原则
+
+- 角色用于帮助拆解职责，不自动创建额外执行工作区。
+- 主控仍是唯一调度者，但不是写入者；凡是需要落盘的非 driver 内容，都必须委派给匹配的 writer 子代理。
+- 只读辅助子代理通过 `.codex/agents/*.toml` 注册，适用于调研、规划、评审和失败归因。
+- 可写 worker 子代理通过 `.codex/agents/*.toml` 注册，适用于规则、任务队列、文档、前端、后端和测试相关写入。
+- 任何子代理开始前都要先读 `AGENTS.md`、`docs/harness/task-session-strategy.md`、`.codex/rules/agents.md` 和对应 `.agents/skills/*/SKILL.md`（如存在）。
+- 涉及需求、实现或验证的任务，额外优先读取 `docs/testing/ACCEPTANCE_EXAMPLES.md`、`docs/testing/TRACEABILITY_MATRIX.md` 和 `docs/testing/verify-matrix.md`。
+- 只读辅助子代理不直接写业务代码，只返回结构化结论和证据路径。
+- 业务实现优先走 `task.json` + `tools/harness/codex-loop.ps1` 的单任务流程；主控侧的规则、文档、任务队列和控制面改动优先走 `harness-writer` 或 `docs-worker`。
+
+## 推荐写入路由
+
+- `harness-writer`：`AGENTS.md`、`task.json`、`.codex/*`、`.codex/rules/*`、`docs/harness/*`、hook、prompt、runtime policy 等编排资产。
+- `docs-worker`：README、使用指南、spec、plan、设计说明、交付文档等人类文档。
+- `frontend-worker`：页面、组件、样式、前端测试和视觉资产。
+- `backend-worker`：API、数据模型、校验、权限和后端测试。
+- `test-runner`：affected tests、fresh evidence、测试报告和 failure findings 整理。

package/package-source/docs/codex-harness-engineering/templates/config/codex-config.toml

@@ -0,0 +1,38 @@
+# Codex 项目配置模板
+# 建议复制到项目 `.codex/config.toml`，或按需合并到 `~/.codex/config.toml`。
+
+model = "gpt-5.4"
+model_reasoning_effort = "high"
+sandbox_mode = "workspace-write"
+
+# Codex 会读取 AGENTS.md 作为项目指导文件。复杂规则建议放 docs/，入口文件保持短。
+# 默认模板会安装根 AGENTS.md，因此不配置 fallback。旧项目确实没有 AGENTS.md 时，再按项目显式添加。
+project_doc_max_bytes = 65536
+
+# 不在默认项目模板中设置 developer_instructions 或 model_instructions_file。
+# 项目硬规则优先写入 AGENTS.md；只有有验证和回滚方案时才使用这些上下文入口。
+
+[approvals]
+approval_policy = "on-request"
+
+[features]
+# hooks 作为 stop-gate 增强层默认启用；driver 仍是主链路。
+# Windows 继续以 PowerShell driver 为主，hooks 只负责结束前补一轮 reflection / feedback。
+hooks = true
+# goal mode 作为外层长期目标控制面启用；仓库执行仍回到 task.json + driver + stop hook。
+goals = true
+
+[mcp_servers.anysearch]
+# AnySearch 提供可匿名访问的 Streamable HTTP MCP；需要更高额度时在本机设置 ANYSEARCH_API_KEY 后取消下一行注释。
+url = "https://api.anysearch.com/mcp"
+# bearer_token_env_var = "ANYSEARCH_API_KEY"
+startup_timeout_sec = 20
+tool_timeout_sec = 60
+enabled = true
+required = false
+
+[agents]
+# 显式限制子代理并发和深度，不依赖隐式默认值。
+max_threads = 4
+max_depth = 2
+job_max_runtime_seconds = 1800

package/package-source/docs/codex-harness-engineering/templates/config/codex-readme.md

@@ -0,0 +1,67 @@
+# Codex Control Surface
+
+本目录是 Codex Harness 的项目级控制面模板。安装到目标项目后，它会成为目标项目的 `.codex/` 配置目录。
+
+## 文件职责
+
+- `config.toml`: Codex 项目配置、少量稳定子代理注册、profile 模板。
+- `hooks.json`: Codex lifecycle hooks 增强层。driver 仍以 PowerShell 脚本为主链路。
+- `task-run-profile.json`: `tools/harness/codex-loop.ps1` 读取的任务会话、子代理和工作区策略。
+- `agents/`: 少量可长期复用的辅助角色，包含只读分析角色和 scoped writer roles。
+- `prompts/`: driver、worker-role、review、failure triage 使用的提示词模板。
+
+## Agent 边界
+
+`agents/` 只放高价值、低耦合的辅助角色。不要把 `skills/ecc-agent-*` 全量注册进这里。
+所有子代理在输出结论或开始写入前，都必须先阅读：
+
+- `AGENTS.md`
+- `docs/harness/task-session-strategy.md`
+- `.codex/rules/agents.md`
+- 当前任务相关的 truth source
+- 对应的 `.agents/skills/*/SKILL.md`（如存在）或项目深文档
+
+当前包内推荐并已注册的只读辅助子代理是：
+
+- `explorer`: 只读探索和证据收集。
+- `readonly-research`: 通用只读调研与证据归纳。
+- `reviewer`: correctness、regression、missing tests 审查。
+- `docs-researcher`: API、框架行为和文档事实核验。
+- `planner`: 实施计划、依赖和验证顺序梳理。
+- `architect`: 架构边界、数据流、集成风险分析。
+- `stage1-reviewer`: 规格一致性审查。
+- `stage2-reviewer`: 代码质量和回归风险审查。
+- `security-reviewer`: 安全专项审查。
+- `test-planner`: 测试矩阵和证据路径规划。
+- `failure-triage`: 失败归因和 owner 分类。
+- `visual-reviewer`: UI 截图、设计稿和视觉回归审查。
+
+当前包内推荐并已注册的 writer 子代理是：
+
+- `harness-writer`: 规则、任务队列、控制面配置、hook、prompt 和 harness 文档资产写入。
+- `docs-worker`: README、使用指南、spec、plan 和交付文档写入。
+- `frontend-worker`: 前端页面、组件、样式和前端测试写入。
+- `backend-worker`: API、数据模型、校验、权限和后端测试写入。
+- `test-runner`: 验证证据、测试报告和 failure findings 写入。
+
+执行型工作仍应优先走 `task.json` + `tools/harness/codex-loop.ps1` 的单任务 driver 流程。只有当 controller / 主会话遇到非 driver 的落盘任务时，才转给匹配的 writer 子代理；reviewer、researcher、planner 之类角色仍保持只读。
+
+## Profile 使用
+
+当前项目模板默认启用 `features.goals = true`，这样 Codex app 可以显示 `/goal`。这只是在产品层打开长期目标控制面，不改变 Harness 的主执行协议；仓库推进仍应回到 `task.json + tools/harness/codex-loop.ps1 + stop hook`。
+
+包内 `profiles.*` 是安装后的推荐配置模板，不保证 `codex -C <path> -p <profile>` 在所有调用场景都会自动读取项目 `.codex/config.toml`。
+
+更稳的方式是：
+
+- 使用全局稳定 profile，例如 `yolo`。
+- 通过启动命令显式传 `-m <model>`。
+- 通过 `-c model_reasoning_effort="<effort>"` 显式传思考强度。
+
+## MCP 边界
+
+不要把需要真实 token 的 MCP 服务直接写入 active config。需要 GitHub、Figma、Browserbase、Firecrawl 等服务时，先确认环境变量，再按项目需要启用。
+
+## AnySearch
+
+`anysearch` MCP 已默认接入为匿名 Streamable HTTP 服务，用于需要联网搜索、实时资料核验、第三方文档查找的 harness 任务。匿名模式会使用 AnySearch 的 per-IP 免费额度；如果任务需要更高并发或更稳定额度，在本机或 CI secret 中设置 `ANYSEARCH_API_KEY`，再在 `.codex/config.toml` 的 `[mcp_servers.anysearch]` 中启用 `bearer_token_env_var = "ANYSEARCH_API_KEY"`。

package/package-source/docs/codex-harness-engineering/templates/config/env-check.ps1

@@ -0,0 +1,40 @@
+param(
+  [string[]]$RequiredCommands = @("git", "codex"),
+  [string[]]$RequiredEnvVars = @()
+)
+
+Set-StrictMode -Version Latest
+$ErrorActionPreference = "Stop"
+
+function Assert-CommandExists {
+  <#
+    检查命令是否存在。
+  #>
+  param([string]$Name)
+
+  if (-not (Get-Command $Name -ErrorAction SilentlyContinue)) {
+    throw "缺少必需命令: $Name"
+  }
+}
+
+function Assert-EnvVarExists {
+  <#
+    检查环境变量是否存在且非空。
+  #>
+  param([string]$Name)
+
+  $value = [Environment]::GetEnvironmentVariable($Name)
+  if ([string]::IsNullOrWhiteSpace($value)) {
+    throw "缺少必需环境变量: $Name"
+  }
+}
+
+foreach ($command in $RequiredCommands) {
+  Assert-CommandExists -Name $command
+}
+
+foreach ($envVar in $RequiredEnvVars) {
+  Assert-EnvVarExists -Name $envVar
+}
+
+Write-Output "Environment check passed."

package/package-source/docs/codex-harness-engineering/templates/config/env.example

@@ -0,0 +1,12 @@
+# Codex 自动化环境变量模板
+# 复制为 `.env.local`、CI secret 或 runner 环境变量，禁止提交真实密钥。
+
+OPENAI_API_KEY=
+CODEX_API_KEY=
+FIGMA_OAUTH_TOKEN=
+GITHUB_TOKEN=
+
+# 项目自定义
+DATABASE_URL=
+REDIS_URL=
+APP_BASE_URL=http://localhost:3000