repo-harness 0.9.1 → 0.9.2

package/README.es.md

@@ -85,7 +85,7 @@ artifacts.
 ## Novedades
 
 Las notas de versión viven en [`docs/CHANGELOG.md`](docs/CHANGELOG.md). La línea
-actual es `0.9.1`.
+actual es `0.9.2`.
 
 ## Cómo funciona
 
@@ -220,7 +220,7 @@ bun add -g repo-harness
 repo-harness install
 
 # Fallback con npx, con Bun ya en PATH porque el CLI corre sobre Bun
-npx -y repo-harness install
+npx -y repo-harness@latest install
 ```
 
 </details>
@@ -419,8 +419,8 @@ Guards habituales:
 
 ## Release actual
 
-- npm package: `repo-harness@0.9.1`
-- Generated workflow stamp: `repo-harness@0.9.1+template@0.9.1`
+- npm package: `repo-harness@0.9.2`
+- Generated workflow stamp: `repo-harness@0.9.2+template@0.9.2`
 - GitHub repository: `Ancienttwo/repo-harness`
 - Release history: [`docs/CHANGELOG.md`](docs/CHANGELOG.md)
 

package/README.fr.md

@@ -85,7 +85,7 @@ l'emportent.
 ## Nouveautés
 
 Les notes de version vivent dans [`docs/CHANGELOG.md`](docs/CHANGELOG.md). La
-ligne actuelle est `0.9.1`.
+ligne actuelle est `0.9.2`.
 
 ## Comment ça marche
 
@@ -224,7 +224,7 @@ bun add -g repo-harness
 repo-harness install
 
 # Fallback npx, avec Bun déjà sur PATH car le CLI s'exécute sur Bun
-npx -y repo-harness install
+npx -y repo-harness@latest install
 ```
 
 </details>
@@ -424,8 +424,8 @@ Guards courants :
 
 ## Release actuelle
 
-- npm package : `repo-harness@0.9.1`
-- Generated workflow stamp : `repo-harness@0.9.1+template@0.9.1`
+- npm package : `repo-harness@0.9.2`
+- Generated workflow stamp : `repo-harness@0.9.2+template@0.9.2`
 - GitHub repository : `Ancienttwo/repo-harness`
 - Release history : [`docs/CHANGELOG.md`](docs/CHANGELOG.md)
 

package/README.ja.md

@@ -75,7 +75,7 @@ review、checks、handoff と食い違う場合は、source artifacts を優先
 ## What's New
 
 リリースノートは [`docs/CHANGELOG.md`](docs/CHANGELOG.md) にあります。現在の
-ラインは `0.9.1` です。
+ラインは `0.9.2` です。
 
 ## 仕組み
 
@@ -201,7 +201,7 @@ bun add -g repo-harness
 repo-harness install
 
 # npx fallback。CLI runtime は Bun なので、Bun が PATH 上に必要です。
-npx -y repo-harness install
+npx -y repo-harness@latest install
 ```
 
 </details>
@@ -398,8 +398,8 @@ hook がブロックしたときは、まず terminal の構造化された出
 
 ## 現在の Release
 
-- npm package：`repo-harness@0.9.1`
-- Generated workflow stamp：`repo-harness@0.9.1+template@0.9.1`
+- npm package：`repo-harness@0.9.2`
+- Generated workflow stamp：`repo-harness@0.9.2+template@0.9.2`
 - GitHub repository：`Ancienttwo/repo-harness`
 - Release history：[`docs/CHANGELOG.md`](docs/CHANGELOG.md)
 

package/README.md

@@ -85,7 +85,7 @@ active plan, contract, review, checks, or handoff, the source artifacts win.
 ## What's New
 
 Release notes live in [`docs/CHANGELOG.md`](docs/CHANGELOG.md). The current line
-is `0.9.1`.
+is `0.9.2`.
 
 ## How It Works
 
@@ -234,6 +234,10 @@ concrete sprint instead of reinterpreting the original chat.
 
 ## First 5 Minutes
 
+<p align="center">
+  <img src="docs/images/repo-harness-install-donkey-carrot.png" alt="Pixel art donkey following a carrot for repo-harness installation" width="900">
+</p>
+
 This is the fastest path for an AI tooling owner evaluating whether the workflow is
 safe to adopt in a real repo. It separates the machine-level runtime bootstrap
 from the repo-local contract install, so a dry run can show exactly what will
@@ -256,20 +260,20 @@ curl -fsSL https://raw.githubusercontent.com/Ancienttwo/repo-harness/main/instal
 irm https://raw.githubusercontent.com/Ancienttwo/repo-harness/main/install.ps1 | iex
 ```
 
-<details>
-<summary>Already have Bun? Use Bun directly, or npx as a fallback</summary>
+If Bun is already on PATH, you can skip the shell installer:
 
 ```bash
-# Bun (recommended)
+# Bun one-shot bootstrap
+bunx repo-harness@latest install
+
+# Or install the persistent CLI first
 bun add -g repo-harness
 repo-harness install
 
 # npx fallback, with Bun already on PATH because the CLI runs on Bun
-npx -y repo-harness install
+npx -y repo-harness@latest install
 ```
 
-</details>
-
 ### 2. Bootstrap the host runtime once
 
 ```bash
@@ -279,15 +283,17 @@ repo-harness install
 `install` is the first-run global bootstrap path. It installs the current npm
 package as the global CLI, refreshes repo-harness skill aliases, installs
 user-level hook adapters, configures Waza runtime skills, persists a brain root
-under `~/.repo-harness/config.json`, and configures CodeGraph MCP. In an
-interactive terminal it asks Y/n before installing the external skills and
-CodeGraph pieces (Enter keeps today's default of installing both); non-TTY
+under `~/.repo-harness/config.json`, and configures CodeGraph MCP. The command
+is idempotent: when the CLI is already installed from Bun's global package
+source, it skips the CLI reinstall and still refreshes the host runtime pieces.
+In an interactive terminal it asks Y/n before installing the external skills
+and CodeGraph pieces (Enter keeps today's default of installing both); non-TTY
 runs and `--json` stay unprompted with the same default-on behavior. Passing
 `--no-external-skills` or `--no-codegraph` explicitly also skips that item's
-prompt unprompted, which is the escape hatch for PTY-allocating CI (for
-example `docker run -t`). It does not apply repo-local workflow files to the
-current directory. `repo-harness init` remains a compatibility alias for
-existing scripts.
+prompt unprompted, which is the escape hatch for PTY-allocating CI (for example
+`docker run -t`). It does not apply repo-local workflow files to the current
+directory. `repo-harness init` remains a compatibility alias for existing
+scripts.
 
 For an Agent-owned, read-only bootstrap audit, run `repo-harness setup check
 --json` or add `--check-updates` for version and adopted-repo refresh
@@ -627,8 +633,8 @@ Most common guards:
 
 ## Current Release
 
-- npm package: `repo-harness@0.9.1`
-- Generated workflow stamp: `repo-harness@0.9.1+template@0.9.1`
+- npm package: `repo-harness@0.9.2`
+- Generated workflow stamp: `repo-harness@0.9.2+template@0.9.2`
 - GitHub repository: `Ancienttwo/repo-harness`
 - Release history: [`docs/CHANGELOG.md`](docs/CHANGELOG.md)
 

package/README.zh-CN.md

@@ -75,7 +75,7 @@ review、checks 或 handoff 冲突，以 source artifacts 为准。
 
 ## What's New
 
-Release notes 见 [`docs/CHANGELOG.md`](docs/CHANGELOG.md)，当前版本线是 `0.9.1`。
+Release notes 见 [`docs/CHANGELOG.md`](docs/CHANGELOG.md)，当前版本线是 `0.9.2`。
 
 ## 工作原理
 
@@ -198,6 +198,10 @@ source of truth，Codex Goal mode 只围绕具体 sprint 恢复和推进，而
 
 ## 前 5 分钟
 
+<p align="center">
+  <img src="docs/images/repo-harness-install-donkey-carrot.png" alt="repo-harness 安装引导的 pixel art 驴和萝卜 banner" width="900">
+</p>
+
 这是评估一个真实仓库是否适合接入该 workflow 的最快路径。它把机器级 runtime
 bootstrap 和 repo-local contract install 分开，所以 dry-run 能先展示会改什么，
 再决定是否应用。
@@ -218,20 +222,20 @@ curl -fsSL https://raw.githubusercontent.com/Ancienttwo/repo-harness/main/instal
 irm https://raw.githubusercontent.com/Ancienttwo/repo-harness/main/install.ps1 | iex
 ```
 
-<details>
-<summary>已经有 Bun？优先直接用 Bun，也可以把 npx 作为备选</summary>
+如果 Bun 已经在 `PATH` 上，可以跳过 shell installer：
 
 ```bash
-# Bun（推荐）
+# Bun 一步 bootstrap
+bunx repo-harness@latest install
+
+# 或者先安装持久化 CLI
 bun add -g repo-harness
 repo-harness install
 
 # npx 备选；仍要求 Bun 已在 PATH 上，因为 CLI runtime 是 Bun
-npx -y repo-harness install
+npx -y repo-harness@latest install
 ```
 
-</details>
-
 ### 2. 先做一次 host runtime bootstrap
 
 ```bash
@@ -240,8 +244,10 @@ repo-harness install
 
 `install` 是首次全局引导入口。它把当前 npm 包安装成全局 CLI，刷新 repo-harness
 skill aliases，安装 user-level hook adapters，配置 Waza runtime skills，把 brain
-root 持久化到 `~/.repo-harness/config.json`，并配置 CodeGraph MCP。它不会把当前目录
-默认迁移成 repo-local workflow。`repo-harness init` 保留为兼容 alias，给已有脚本用。
+root 持久化到 `~/.repo-harness/config.json`，并配置 CodeGraph MCP。这个命令是
+幂等的：如果 CLI 已经来自 Bun global package source，它会跳过 CLI 重装，但仍继续
+刷新 host runtime pieces。它不会把当前目录默认迁移成 repo-local workflow。
+`repo-harness init` 保留为兼容 alias，给已有脚本用。
 
 如果要让 Agent 做只读 bootstrap audit，运行 `repo-harness setup check
 --json`；需要版本和已接入仓库刷新提示时加 `--check-updates`。`setup check`
@@ -452,8 +458,8 @@ hook block 工作时，先看 terminal 里的结构化输出。核心字段是
 
 ## 当前 Release
 
-- npm package：`repo-harness@0.9.1`
-- Generated workflow stamp：`repo-harness@0.9.1+template@0.9.1`
+- npm package：`repo-harness@0.9.2`
+- Generated workflow stamp：`repo-harness@0.9.2+template@0.9.2`
 - GitHub repository：`Ancienttwo/repo-harness`
 - Release history：[`docs/CHANGELOG.md`](docs/CHANGELOG.md)
 

package/assets/hooks/AGENTS.md

@@ -49,8 +49,8 @@ Owns the runtime-harness-hook-adapters capability boundary declared in .ai/conte
 - Architecture domain: `runtime-harness`
 - Architecture capability: `hook-adapters`
 - Architecture module: `docs/architecture/modules/runtime-harness/hook-adapters.md`
-- Last architecture event: 2026-07-06T13:54:06+0800
-- Last changed path: `assets/hooks/codex-delegation-advisor.sh`
+- Last architecture event: 2026-07-09T11:39:06+0800
+- Last changed path: `assets/hooks/prompt-guard.sh`
 - Severity: high
 - Change type: workflow-surface
 - Module responsibility: Keep this block aligned with the local boundary described by surrounding human-owned context.

package/assets/hooks/CLAUDE.md

@@ -49,8 +49,8 @@ Owns the runtime-harness-hook-adapters capability boundary declared in .ai/conte
 - Architecture domain: `runtime-harness`
 - Architecture capability: `hook-adapters`
 - Architecture module: `docs/architecture/modules/runtime-harness/hook-adapters.md`
-- Last architecture event: 2026-07-06T13:54:06+0800
-- Last changed path: `assets/hooks/codex-delegation-advisor.sh`
+- Last architecture event: 2026-07-09T11:39:06+0800
+- Last changed path: `assets/hooks/prompt-guard.sh`
 - Severity: high
 - Change type: workflow-surface
 - Module responsibility: Keep this block aligned with the local boundary described by surrounding human-owned context.

package/assets/hooks/prompt-guard.sh

@@ -1054,7 +1054,7 @@ prompt_guard_engine_call
 if [[ "$PG_ENGINE_STATE" != "ok" ]]; then
   if [[ "$PG_ENGINE_STATE" == "legacy" ]]; then
     echo "[PromptGuard] Advisory: the installed repo-harness CLI predates the prompt-verdict protocol; prompt intent gates are degraded to advisory for this prompt."
-    echo "[PromptGuard] Refresh the CLI with: bun add -g repo-harness@latest && repo-harness install. Fallback: npx -y repo-harness install."
+    echo "[PromptGuard] Refresh the CLI with: bun add -g repo-harness@latest && repo-harness install. Fallback: npx -y repo-harness@latest install."
   else
     echo "[PromptGuard] Advisory: prompt-guard decision engine is unavailable (repo-harness CLI or bun not found); prompt intent gates are degraded to advisory for this prompt."
     echo "[PromptGuard] Edit-layer guards still enforce plan and contract scope. Install the repo-harness CLI to restore prompt decisions."

package/assets/reference-configs/external-tooling.md

@@ -495,10 +495,15 @@ guessed mapping.
 
 | Upstream frontmatter | Codex TOML |
 |---|---|
-| `model: opus`, `effort: max` | `model = "gpt-5.5"`, `model_reasoning_effort = "xhigh"` |
-| `model: sonnet`, `effort: max` | `model = "gpt-5.5"`, `model_reasoning_effort = "medium"` |
+| `model: opus`, `effort: max` | `model = "gpt-5.6-sol"`, `model_reasoning_effort = "xhigh"` |
+| `model: sonnet`, `effort: max` | `model = "gpt-5.6-terra"`, `model_reasoning_effort = "medium"` |
 | `tools: [...]` present | `sandbox_mode = "read-only"` |
 
+The Codex generator also rewrites the exact upstream provider label in the
+description (`Opus 4.8 at max effort` or `Sonnet 5 at max effort`) to the
+mapped GPT-5.6 model and reasoning level. A missing label fails closed so the
+installed metadata cannot claim a different model from the TOML settings.
+
 `developer_instructions` is the upstream `.md` body plus the canonical
 EXECUTION_BOUNDARY anti-extras clause, kept byte-identical to the
 `EXECUTION_BOUNDARY` constant in `scripts/contract-run.ts` so every generated

package/assets/reference-configs/global-working-rules.md

@@ -5,6 +5,12 @@ Use this content for user-level `~/.codex/AGENTS.md` and `~/.claude/CLAUDE.md` w
 ```md
 # Global Working Rules
 
+Rule 0: You may spend as much time as needed thinking. Do not send optional commentary progress messages. Use tools only when they are required. For tasks that do not require tools, complete the reasoning first, then answer in final.
+
+Reasoning: Prefer first principles over pattern matching. Before solving, first identify the observable and controllable conditions. For quantitative logic problems, before the final answer, you must prove the strategy is sufficient in the worst case. Numeric answers must have their arithmetic rechecked.
+
+Generality: These are general working rules. Do not tailor behavior to any specific evaluation or expected answer.
+
 - Use Chinese by default for this user; keep technical terms in English. If the user writes in another language, mirror that language.
 - Act as an engineering collaborator: finish the concrete task, verify it, then report conclusion, actual change, reason, verification, and residual risk.
 - Prefer direct execution over repeated confirmation. Stop to ask only when continuing would likely produce output contrary to the user's intent.

package/assets/skill-version.json

@@ -1,6 +1,6 @@
 {
-  "version": "0.9.1",
-  "templateVersion": "0.9.1",
+  "version": "0.9.2",
+  "templateVersion": "0.9.2",
   "skillName": "repo-harness",
   "contractId": "tasks-first-harness-v1",
   "compatibility": {
@@ -195,6 +195,10 @@
     {
       "version": "0.9.1",
       "description": "Hardens the Bun-first installer and setup-check path with bunx skills probes, PATH readback hints, and explicit Codex delegation-mode selection"
+    },
+    {
+      "version": "0.9.2",
+      "description": "Hardens CLI environment diagnostics and hook execution intent while aligning the generated GPT-5.6 agent fleet with current Codex runtime guidance"
     }
   ],
   "generatedProjectStamp": {

package/assets/templates/helpers/install-agent-fleet.sh

@@ -61,11 +61,26 @@ const EXECUTION_BOUNDARY = [
   "If the requested outcome cannot be completed without expanding scope, fail closed: stop, name the missing decision, and cite the exact file/section that blocks execution.",
 ].join("\n");
 
-// Only (opus, max) and (sonnet, max) are recognized. Any other model or a
-// non-"max" effort is a fail-closed error -- the generator never guesses a mapping.
+// Only (opus, max) and (sonnet, max) are recognized. Each mapping also owns the
+// exact provider label rewritten in Codex metadata so the generated role never
+// claims to run a different model. Any mismatch is a fail-closed error.
 const MODEL_EFFORT_MAP = {
-  opus: { max: { model: "gpt-5.5", effort: "xhigh" } },
-  sonnet: { max: { model: "gpt-5.5", effort: "medium" } },
+  opus: {
+    max: {
+      model: "gpt-5.6-sol",
+      effort: "xhigh",
+      sourceDescription: "Opus 4.8 at max effort",
+      targetDescription: "GPT-5.6 Sol at extra high reasoning",
+    },
+  },
+  sonnet: {
+    max: {
+      model: "gpt-5.6-terra",
+      effort: "medium",
+      sourceDescription: "Sonnet 5 at max effort",
+      targetDescription: "GPT-5.6 Terra at medium reasoning",
+    },
+  },
 };
 
 function fetchSource(agent) {
@@ -133,6 +148,9 @@ function validateFrontmatter(parsed) {
   if (!mapped) {
     return { ok: false, reason: `unmapped model/effort combination: ${parsed.model}/${parsed.effort}` };
   }
+  if (!parsed.description.includes(mapped.sourceDescription)) {
+    return { ok: false, reason: `description missing expected model label: ${mapped.sourceDescription}` };
+  }
   return { ok: true, mapped };
 }
 
@@ -142,8 +160,9 @@ function tomlBasicString(value) {
 
 function generateToml(parsed, mapped) {
   const lines = [];
+  const description = parsed.description.replace(mapped.sourceDescription, mapped.targetDescription);
   lines.push(`name = ${tomlBasicString(parsed.name)}`);
-  lines.push(`description = ${tomlBasicString(parsed.description)}`);
+  lines.push(`description = ${tomlBasicString(description)}`);
   lines.push(`model = ${tomlBasicString(mapped.model)}`);
   lines.push(`model_reasoning_effort = ${tomlBasicString(mapped.effort)}`);
   if (parsed.hasTools) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "repo-harness",
-  "version": "0.9.1",
+  "version": "0.9.2",
   "description": "Installs, migrates, audits, and repairs repo-local agentic development harnesses",
   "license": "MIT",
   "repository": {

package/scripts/install-agent-fleet.sh

@@ -61,11 +61,26 @@ const EXECUTION_BOUNDARY = [
   "If the requested outcome cannot be completed without expanding scope, fail closed: stop, name the missing decision, and cite the exact file/section that blocks execution.",
 ].join("\n");
 
-// Only (opus, max) and (sonnet, max) are recognized. Any other model or a
-// non-"max" effort is a fail-closed error -- the generator never guesses a mapping.
+// Only (opus, max) and (sonnet, max) are recognized. Each mapping also owns the
+// exact provider label rewritten in Codex metadata so the generated role never
+// claims to run a different model. Any mismatch is a fail-closed error.
 const MODEL_EFFORT_MAP = {
-  opus: { max: { model: "gpt-5.5", effort: "xhigh" } },
-  sonnet: { max: { model: "gpt-5.5", effort: "medium" } },
+  opus: {
+    max: {
+      model: "gpt-5.6-sol",
+      effort: "xhigh",
+      sourceDescription: "Opus 4.8 at max effort",
+      targetDescription: "GPT-5.6 Sol at extra high reasoning",
+    },
+  },
+  sonnet: {
+    max: {
+      model: "gpt-5.6-terra",
+      effort: "medium",
+      sourceDescription: "Sonnet 5 at max effort",
+      targetDescription: "GPT-5.6 Terra at medium reasoning",
+    },
+  },
 };
 
 function fetchSource(agent) {
@@ -133,6 +148,9 @@ function validateFrontmatter(parsed) {
   if (!mapped) {
     return { ok: false, reason: `unmapped model/effort combination: ${parsed.model}/${parsed.effort}` };
   }
+  if (!parsed.description.includes(mapped.sourceDescription)) {
+    return { ok: false, reason: `description missing expected model label: ${mapped.sourceDescription}` };
+  }
   return { ok: true, mapped };
 }
 
@@ -142,8 +160,9 @@ function tomlBasicString(value) {
 
 function generateToml(parsed, mapped) {
   const lines = [];
+  const description = parsed.description.replace(mapped.sourceDescription, mapped.targetDescription);
   lines.push(`name = ${tomlBasicString(parsed.name)}`);
-  lines.push(`description = ${tomlBasicString(parsed.description)}`);
+  lines.push(`description = ${tomlBasicString(description)}`);
   lines.push(`model = ${tomlBasicString(mapped.model)}`);
   lines.push(`model_reasoning_effort = ${tomlBasicString(mapped.effort)}`);
   if (parsed.hasTools) {

package/scripts/lib/project-init-lib.sh

@@ -2278,6 +2278,14 @@ pi_root_context_content() {
 This is the root routing contract for Claude Code and Codex. Keep it short:
 load this first, then follow the repo-owned artifacts it names.
 
+## General Working Rules
+
+Rule 0: You may spend as much time as needed thinking. Do not send optional commentary progress messages. Use tools only when they are required. For tasks that do not require tools, complete the reasoning first, then answer in final.
+
+Reasoning: Prefer first principles over pattern matching. Before solving, first identify the observable and controllable conditions. For quantitative logic problems, before the final answer, you must prove the strategy is sufficient in the worst case. Numeric answers must have their arithmetic rechecked.
+
+Generality: These are general working rules. Do not tailor behavior to any specific evaluation or expected answer.
+
 ## Root Workflow Contract
 
 - Keep sibling `CLAUDE.md` and `AGENTS.md` files aligned. Claude Code consumes `CLAUDE.md`; Codex consumes `AGENTS.md`.

package/src/cli/commands/doctor.ts

@@ -10,6 +10,7 @@ import * as fs from 'fs';
 import * as os from 'os';
 import * as path from 'path';
 import { spawnSync } from 'child_process';
+import { fileURLToPath } from 'url';
 import { ALL_TARGETS } from '../installer/targets/registry';
 import { checkCodegraph, type CodegraphCheckResult } from '../tools/codegraph';
 import { CLI_VERSION } from './status';
@@ -19,6 +20,8 @@ import { ROUTES } from '../hook/route-registry';
 
 const TRUST_STATE_LINE = /^\[hooks\.state\."[^"]+\/\.codex\/hooks\.json:/;
 const PACKAGE_NAME = 'repo-harness';
+const PACKAGE_ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..', '..', '..');
+const MIN_CODEX_CLI_VERSION = '0.144.0';
 const UPDATE_CHECK_ENV = 'REPO_HARNESS_CHECK_UPDATES';
 const LATEST_VERSION_ENV = 'REPO_HARNESS_LATEST_VERSION';
 
@@ -105,7 +108,7 @@ function parseVersion(value: string): number[] | null {
   return match.slice(1).map((part) => Number(part));
 }
 
-function compareVersions(a: string, b: string): number | null {
+export function compareVersions(a: string, b: string): number | null {
   const left = parseVersion(a);
   const right = parseVersion(b);
   if (!left || !right) return null;
@@ -116,14 +119,70 @@ function compareVersions(a: string, b: string): number | null {
   return 0;
 }
 
-function readLatestPackageVersion(): { version?: string; error?: string } {
-  if (process.env[LATEST_VERSION_ENV]) {
-    return { version: process.env[LATEST_VERSION_ENV] };
+function checkCodexCliVersion(): DoctorCheckResult {
+  const id = 'codex-cli-version';
+  const describe = `Codex CLI supports generated GPT-5.6 agent profiles (>= ${MIN_CODEX_CLI_VERSION})`;
+  const resolved = findCommandOnPath('codex');
+  if (!resolved) {
+    return { id, describe, status: 'na', detail: 'codex not found on PATH' };
   }
 
-  const result = spawnSync('npm', ['view', PACKAGE_NAME, 'version', '--json'], {
+  const result = spawnSync(resolved, ['--version'], {
     encoding: 'utf-8',
     timeout: 5000,
+    env: process.env,
+  });
+  if (result.status !== 0 || result.error) {
+    const error = result.stderr || result.stdout || String(result.error?.message ?? result.error ?? 'codex --version failed');
+    return { id, describe, status: 'warn', detail: `path=${resolved}; ${error.trim()}` };
+  }
+
+  const output = result.stdout.trim();
+  const match = output.match(/^codex-cli (\d+\.\d+\.\d+)$/);
+  const version = match?.[1] ?? '';
+  const comparison = compareVersions(version, MIN_CODEX_CLI_VERSION);
+  if (comparison === null) {
+    return {
+      id,
+      describe,
+      status: 'warn',
+      detail: `path=${resolved}; unable to parse version from ${JSON.stringify(output)}`,
+    };
+  }
+  if (comparison < 0) {
+    return {
+      id,
+      describe,
+      status: 'warn',
+      detail: `path=${resolved}; current=${version}; minimum=${MIN_CODEX_CLI_VERSION}`,
+    };
+  }
+  return {
+    id,
+    describe,
+    status: 'ok',
+    detail: `path=${resolved}; current=${version}; minimum=${MIN_CODEX_CLI_VERSION}`,
+  };
+}
+
+export function readLatestPackageVersion(env?: NodeJS.ProcessEnv): { version?: string; error?: string } {
+  const activeEnv = env ?? process.env;
+  if (activeEnv[LATEST_VERSION_ENV]) {
+    return { version: activeEnv[LATEST_VERSION_ENV] };
+  }
+
+  const result = spawnSync(process.execPath, [
+    'pm',
+    'view',
+    PACKAGE_NAME,
+    'version',
+    '--json',
+    '--registry=https://registry.npmjs.org',
+  ], {
+    encoding: 'utf-8',
+    timeout: 5000,
+    cwd: PACKAGE_ROOT,
+    env: activeEnv,
   });
   if (result.status !== 0 || result.error) {
     return { error: result.stderr || result.stdout || String(result.error?.message ?? result.error ?? 'npm view failed') };
@@ -439,6 +498,7 @@ export function runDoctor(cwd: string = process.cwd()): DoctorReport {
   const securityReport = runSecurityScan({ cwd });
   checks.push(checkPath());
   checks.push(checkVersion());
+  checks.push(checkCodexCliVersion());
   checks.push(checkCliUpdate());
   for (const target of ALL_TARGETS) {
     if (target.supportsLocation('global')) {

package/src/cli/commands/global-runtime.ts

@@ -5,6 +5,7 @@ import { fileURLToPath } from "url";
 import { configureBrainRoot, defaultBrainRootChoice, expandHomePath } from "./brain-root";
 import { syncCrossReviewSkills } from "./init";
 import { runInstall, type InstallTargetSpec } from "./install";
+import { compareVersions, readLatestPackageVersion } from "./doctor";
 import { configureCodegraph } from "../tools/codegraph";
 import { runProcess as runBoundedProcess } from "../../effects/process-runner";
 
@@ -175,16 +176,31 @@ function isBunGlobalPackageSource(sourceRoot: string, env?: NodeJS.ProcessEnv):
   }
 }
 
+// Best-effort: readLatestPackageVersion() already swallows offline/npm-missing/
+// timeout failures into `.error`, so any lookup failure just yields no hint —
+// this must never turn the "skipped" step into a "failed" one.
+function updateAvailableHint(version: string | null, env?: NodeJS.ProcessEnv): string {
+  if (!version) return "";
+  const activeEnv = env ?? process.env;
+  if (activeEnv.REPO_HARNESS_CHECK_UPDATES !== "1") return "";
+  const latest = readLatestPackageVersion(env);
+  if (!latest.version) return "";
+  const comparison = compareVersions(version, latest.version);
+  if (comparison === null || comparison >= 0) return "";
+  return `; latest=${latest.version} available — run: repo-harness update`;
+}
+
 function installCli(sourceRoot: string, cwd: string, env?: NodeJS.ProcessEnv, installSpec?: string): GlobalRuntimeStep {
   const version = packageVersion(sourceRoot);
   const name = packageName(sourceRoot);
   if (installSpec === undefined && isBunGlobalPackageSource(sourceRoot, env)) {
+    const base = version
+      ? `already installed from Bun global package source; version=${version}`
+      : "already installed from Bun global package source";
     return {
       step: "install repo-harness CLI",
       status: "skipped",
-      detail: version
-        ? `already installed from Bun global package source; version=${version}`
-        : "already installed from Bun global package source",
+      detail: `${base}${updateAvailableHint(version, env)}`,
     };
   }
   const spec = installSpec ?? (existsSync(join(sourceRoot, "package.json")) ? sourceRoot : "repo-harness");

package/src/cli/commands/init-hook.ts

@@ -311,6 +311,7 @@ function doctorChecks(
 ): InitHookCheck[] {
   const checks: InitHookCheck[] = [];
   for (const entry of report.checks) {
+    if (target === 'claude' && entry.id === 'codex-cli-version') continue;
     const source: InitHookCheckSource = entry.id === 'security-config' ? 'security' : 'doctor';
     let checkStatus: InitHookCheckStatus = entry.status;
 

package/src/cli/hook/prompt-intents.ts

@@ -82,12 +82,40 @@ const EXPLICIT_EXECUTION_LINE = re(
   String.raw`^${SP}*(please\s+)?(implement\s+(this|the)|execute\s+(this|the)|start\s+(implementation|executing|coding)|go ahead|proceed|ship it|开始(实现|执行|落实|写)|执行计划|落实计划|批准执行|批准|直接(改|做|实现|执行|落地)|动手|开干|可以(开始|执行|干)|可以干|干吧|做吧)(${SP}|$)`,
 );
 
+// Unlike the legacy execution verbs above, this newly supported phrase must
+// remain an actual command at the start of a line. A whitespace-only prefix
+// keeps quoted examples, questions, and negations out of the execution path.
+const DIRECT_MODIFICATION_LINE = re(
+  String.raw`^\s*(请\s*)?直接修改`,
+);
+const DIRECT_MODIFICATION_INLINE_PAYLOAD =
+  /“[^”\n]*”|「[^」\n]*」|『[^』\n]*』|"[^"\n]*"|`[^`\n]*`/gu;
+const DIRECT_MODIFICATION_QUESTION = re(
+  String.raw`([?？]|[吗么呢]${SP}*$|是不是|能不能|可不可以|要不要|应不应该|该不该|会不会|行不行|好不好|对不对|合不合适)`,
+);
+const DIRECT_MODIFICATION_NON_COMMAND = re(
+  String.raw`(不合适(?:吧|${SP}*($|[,，;；。]))|(?:是)?不(?:对|应该|行)(?:吧|的?${SP}*($|[,，;；。]))|不要这么做|不是(我|我们)?的?要求|只是(一个|个)?示例|仅作示例|作为示例)`,
+);
+
+function isDirectModificationCommandLine(line: string): boolean {
+  const topLevelText = line.replace(DIRECT_MODIFICATION_INLINE_PAYLOAD, '');
+  return (
+    DIRECT_MODIFICATION_LINE.test(line) &&
+    !DIRECT_MODIFICATION_QUESTION.test(topLevelText) &&
+    !DIRECT_MODIFICATION_NON_COMMAND.test(topLevelText)
+  );
+}
+
+function hasDirectModificationCommandLine(ctx: PromptIntentContext): boolean {
+  return isDirectModificationCommandLine(ctx.firstLine);
+}
+
 export function promptHasExplicitExecutionCommandLine(ctx: PromptIntentContext): boolean {
-  return EXPLICIT_EXECUTION_LINE.test(ctx.text);
+  return EXPLICIT_EXECUTION_LINE.test(ctx.text) || hasDirectModificationCommandLine(ctx);
 }
 
 export function isExplicitExecutionStartLine(ctx: PromptIntentContext): boolean {
-  return EXPLICIT_EXECUTION_LINE.test(ctx.firstLine);
+  return EXPLICIT_EXECUTION_LINE.test(ctx.firstLine) || hasDirectModificationCommandLine(ctx);
 }
 
 const PLAN_EXECUTION_PROJECTION_LINE = re(
@@ -104,6 +132,7 @@ const TRIGGER_QUESTION = re(
 );
 
 export function isTriggerQuestionPrompt(ctx: PromptIntentContext): boolean {
+  if (hasDirectModificationCommandLine(ctx)) return false;
   return TRIGGER_QUESTION.test(ctx.firstLine);
 }
 
@@ -130,6 +159,7 @@ const PLAN_REFINEMENT = re(
 );
 
 export function isPlanRefinementIntent(ctx: PromptIntentContext): boolean {
+  if (hasDirectModificationCommandLine(ctx)) return false;
   if (PLAN_REFINEMENT_EXEC.test(ctx.firstLine)) return false;
   return PLAN_REFINEMENT.test(ctx.firstLine);
 }
@@ -179,6 +209,7 @@ export function isDiagnosticQuestionIntent(ctx: PromptIntentContext): boolean {
   if (isExecutionApprovalIntent(ctx)) return false;
   if (isEmbeddedApprovedPlanIntent(ctx)) return false;
   if (isPlanShapedMarkdownIntent(ctx)) return false;
+  if (hasDirectModificationCommandLine(ctx)) return false;
   if (DIAGNOSTIC_DIRECT.test(ctx.text)) return true;
   return DIAGNOSTIC_TOPIC.test(ctx.text) && DIAGNOSTIC_QUESTION.test(ctx.text);
 }
@@ -198,6 +229,7 @@ export function isReviewReleaseAdvisoryIntent(ctx: PromptIntentContext): boolean
   if (isEmbeddedApprovedPlanIntent(ctx)) return false;
   if (isPlanShapedMarkdownIntent(ctx)) return false;
   if (isExecutionApprovalIntent(ctx)) return false;
+  if (hasDirectModificationCommandLine(ctx)) return false;
   // Review/check prompts often say "execute /check" or "执行 checklist". Those
   // route to evaluator evidence, not implementation.
   if (REVIEW_RELEASE_CODING_VERB.test(ctx.text)) return false;
@@ -266,7 +298,9 @@ export function isNextSliceOrStatusAdvisoryIntent(ctx: PromptIntentContext): boo
   return false;
 }
 
-const IMPLEMENT_VERB = re('(implement|execute|build it|do it|go ahead|proceed|ship it|实现|执行|开始写|动手|开干)');
+const IMPLEMENT_VERB = re(
+  '(implement|execute|build it|do it|go ahead|proceed|ship it|实现|执行|开始写|动手|开干)',
+);
 
 export function isImplementIntent(ctx: PromptIntentContext): boolean {
   if (isTriggerQuestionPrompt(ctx)) return false;
@@ -280,6 +314,7 @@ export function isImplementIntent(ctx: PromptIntentContext): boolean {
   if (isPassiveWorktreeStatusIntent(ctx)) return false;
   return (
     IMPLEMENT_VERB.test(ctx.text) ||
+    hasDirectModificationCommandLine(ctx) ||
     isExecutionApprovalIntent(ctx) ||
     isEmbeddedApprovedPlanIntent(ctx) ||
     isPlanShapedMarkdownIntent(ctx)