@andyqiu/codeforge 0.7.9 → 0.8.0

package/README.md

@@ -133,8 +133,12 @@ export KNOWLEDGE_API_KEY=你的-token
 
 ## 升级
 
+opencode 启动时会自动检查并升级，无需手动操作。
+
+如需手动升级：
+
 ```bash
-npm update -g @andyqiu/codeforge
+npm install -g @andyqiu/codeforge@latest
 ```
 
 ## 卸载

package/agents/codeforge.md

@@ -15,16 +15,15 @@ permissions:
   bash: deny
   webfetch: deny
 allowed_tools: [read, repo_map, task, plan_read, review_approval, session_merge]
-model: anthropic/claude-sonnet-4-6
-model_category: deep
-tier: deep
+model: google/gemini-3.1-pro
+model_category: balanced
+tier: balanced
 model_thinking:
   type: enabled
-  budget_tokens: 4000
+  thinking_level: low
 fallback_models:
-  - anthropic/claude-opus-4-8    # 路由判断失败 / 复杂场景升档
+  - anthropic/claude-sonnet-4-6
   - anthropic/claude-opus-4-8
-  - google/gemini-3-pro
 ---
 
 # CodeForge Agent
@@ -116,7 +115,7 @@ fallback_models:
 1. 派 planner 时，要求其 boomerang 摘要必须含 `plan_id: plan-YYYYMMDD-HHmmss-NNN` **独占一行**
 2. 派 coder 时，prompt 里塞 `plan_id` + `sessionId`（当前 session id），明示「**第一步调 `plan_read(plan_id=<id>)`** 拿完整方案」
 3. 派 reviewer 时按模式填写 prompt：
-   - **新模式（`[Session Merge Review]`）**：prompt 必须含 `sessionId` + `worktreePath` + `baseSha` + `plan_id`（四项均必填；reviewer 调 `review_approval` 时用 `["session:<sessionId>"]`）
+   - **新模式（`[Session Merge Review]`）**：prompt 必须含 `sessionId` + `worktreePath` + `baseSha` + `plan_id`（四项均必填；reviewer 调 `review_approval` 时用 `["session:<sessionId>"]` + `worktreePath`）。⚠️ **owner 归一**（ADR:worktree-approval-owner-reconciliation）：`sessionId` 应传 **worktree owner（root session id，即合入时 `session_merge` 用的那个 id）**，不确定时**不要**传子会话 id；即便误传子会话 id，`review_approval` 工具也会用 `worktreePath` 自动归一到 owner（child alias + owner key 双写），merge gate 兜底识别 —— 但显式传 owner 最稳
    - **审方案**：`review_target=plan_only` + `plan_id`
    - **审决策**：`review_target=decision_only` + 用户选项 + 候选派单
 4. **走 spec 路径时**，派 planner / coder 的 prompt 额外塞 `spec=<slug>`；派 coder 时再塞 `pre_ack=<PRE-id>,<PRE-id>`

package/agents/coder-deep.md

@@ -20,11 +20,10 @@ model_category: deep
 tier: deep
 model_thinking:
   type: enabled
-  budget_tokens: 8000
+  budget_tokens: 16000
 fallback_models:
-  - anthropic/claude-opus-4-8
-  - anthropic/claude-sonnet-4-6
-  - google/gemini-3-pro
+  - openai/gpt-5.5
+  - google/gemini-3.1-pro
 ---
 
 

package/agents/coder-quick.md

@@ -23,7 +23,7 @@ model_thinking:
   budget_tokens: 2000
 fallback_models:
   - anthropic/claude-opus-4-8
-  - anthropic/claude-sonnet-4-6
+  - google/gemini-3.1-pro
 ---
 
 

package/agents/coder.md

@@ -20,11 +20,10 @@ model_category: balanced
 tier: balanced
 model_thinking:
   type: enabled
-  budget_tokens: 4000
+  budget_tokens: 6000
 fallback_models:
   - anthropic/claude-opus-4-8
-  - anthropic/claude-opus-4-8
-  - google/gemini-3-pro
+  - google/gemini-3.1-pro
 ---
 
 

package/agents/discover-challenger.md

@@ -16,16 +16,15 @@ allowed_tools:
   # 故意不给 save_chat_insight：challenger 是过程角色，不沉淀
   # 故意不给 task：物理禁止派子 agent（避免嵌套对抗）
   # 故意不给 edit/write：零写权（worktree 也不允许写）
-model: anthropic/claude-opus-4-8
-model_category: deep
+model: openai/gpt-5.5
+model_category: ultrabrain
 tier: deep
 model_thinking:
   type: enabled
-  budget_tokens: 3000
+  reasoning_effort: high
 fallback_models:
   - anthropic/claude-opus-4-8
-  - anthropic/claude-sonnet-4-6
-  - google/gemini-3-pro
+  - google/gemini-3.1-pro
 ---
 
 # Discover Challenger

package/agents/discover.md

@@ -40,11 +40,10 @@ model_category: deep
 tier: deep
 model_thinking:
   type: enabled
-  budget_tokens: 6000
+  budget_tokens: 16000
 fallback_models:
-  - anthropic/claude-opus-4-8
-  - anthropic/claude-sonnet-4-6
-  - google/gemini-3-pro
+  - openai/gpt-5.5
+  - google/gemini-3.1-pro
 ---
 
 # Discover Agent

package/agents/planner.md

@@ -20,11 +20,11 @@ model_category: deep
 tier: deep
 model_thinking:
   type: enabled
-  budget_tokens: 8000
+  budget_tokens: 16000
 fallback_models:
   - anthropic/claude-opus-4-8
   - anthropic/claude-sonnet-4-6
-  - google/gemini-3-pro
+  - google/gemini-3.1-pro
 ---
 
 <!--

package/agents/reviewer-lite.md

@@ -15,12 +15,15 @@ permissions:
   bash: allow
   webfetch: deny
 allowed_tools: [plan_read, bash, read, task, review_approval]
-model: anthropic/claude-sonnet-4-6
+model: google/gemini-3.1-pro
 model_category: balanced
 tier: balanced
+model_thinking:
+  type: enabled
+  thinking_level: medium
 fallback_models:
-  - anthropic/claude-sonnet-4-5
-  - google/gemini-2.5-flash
+  - anthropic/claude-sonnet-4-6
+  - openai/gpt-5.5
 ---
 
 <!-- ADR:reviewer-multi-profile -->
@@ -165,10 +168,13 @@ reviewer 的 `read` 工具**仅允许**读以下路径（白名单）：
                                                            // 模式 2：传 ["decision:<hash>"]
     notes: "审阅摘要（建议 ≤ 500 字）",
     decisionLine: "APPROVE",                               // 可选
+    worktreePath: "<worktree_path>",                       // 模式 3 必传：prompt 的 worktree_path（owner 归一反查）
   })
   ```
   否则 codeforge / `/merge` workflow 后续会被 `plugins/tool-policy.ts` / merge-loop 硬拦截。`REQUEST_CHANGES` / `BLOCK` **不**调此工具。
 
+  - **模式 3 owner 归一**（ADR:worktree-approval-owner-reconciliation）：子会话场景下 prompt 里的 `session_id` 可能是子会话 id，而 worktree 实际归属父/root（canonical owner）。`review_approval` 工具会用 `session:<sid>` + 你传的 `worktreePath` 把审批**自动归一**到 owner key（并写一条 child alias），使父 session 合入时无需重复 review。**你只需照传 `session:<sessionId>` + `worktreePath`，归一由工具内部完成。**
+
 - **两层 token 语义独立**（ADR:decision-token-vs-approval-verdict-layering）：
 
   | 层 | 出现位置 | 合法值 | 用途 |

package/agents/reviewer.md

@@ -15,16 +15,15 @@ permissions:
   bash: allow
   webfetch: deny
 allowed_tools: [plan_read, bash, read, task, review_approval]
-model: anthropic/claude-opus-4-8
+model: openai/gpt-5.5
 model_category: ultrabrain
 tier: deep
 model_thinking:
   type: enabled
-  budget_tokens: 2000
+  reasoning_effort: xhigh
 fallback_models:
   - anthropic/claude-opus-4-8
-  - anthropic/claude-sonnet-4-6
-  - google/gemini-3-pro
+  - google/gemini-3.1-pro
 ---
 
 <!-- ADR:reviewer-multi-profile -->
@@ -164,10 +163,13 @@ reviewer 的 `read` 工具**仅允许**读以下路径（白名单）：
                                                            // 模式 2：传 ["decision:<hash>"]
     notes: "审阅摘要（建议 ≤ 500 字）",
     decisionLine: "APPROVE",                               // 可选
+    worktreePath: "<worktree_path>",                       // 模式 3 必传：prompt 的 worktree_path（owner 归一反查）
   })
   ```
   否则 codeforge / `/merge` workflow 后续会被 `plugins/tool-policy.ts` / merge-loop 硬拦截。`REQUEST_CHANGES` / `BLOCK` **不**调此工具。
 
+  - **模式 3 owner 归一**（ADR:worktree-approval-owner-reconciliation）：子会话场景下 prompt 里的 `session_id` 可能是子会话 id，而 worktree 实际归属父/root（canonical owner）。`review_approval` 工具会用 `pendingIds` 里的 `session:<sid>` + 你传的 `worktreePath` 把审批**自动归一**到 owner key（并写一条 child alias），使父 session 合入时无需重复 review。**你只需照传 `session:<sessionId>` + `worktreePath`，归一由工具内部完成，无需手算 owner。**
+
 - **两层 token 语义独立**（ADR:decision-token-vs-approval-verdict-layering）：
 
   | 层 | 出现位置 | 合法值 | 用途 |

package/bin/codeforge.mjs

@@ -139,7 +139,8 @@ function uninstallOpencode({ scope }) {
 }
 
 function detectOpencode() {
-  const r = spawnSync("opencode", ["--version"], { stdio: "pipe", encoding: "utf8", shell: true })
+  const opencodeCmd = process.platform === "win32" ? "opencode.cmd" : "opencode"
+  const r = spawnSync(opencodeCmd, ["--version"], { stdio: "pipe", encoding: "utf8" })
   if (r.status === 0) return { installed: true, version: (r.stdout ?? "").trim() }
   return { installed: false }
 }
@@ -249,7 +250,7 @@ async function cmdUpgrade(args) {
     log(`当前版本：${currentVersion} → 最新版本：${latestVersion}`)
   }
 
-  const npmCmd = "npm"
+  const npmCmd = process.platform === "win32" ? "npm.cmd" : "npm"
   // 有明确版本号时用精确版本，绕过 @latest dist-tag 缓存
   const installTarget = latestVersion
     ? `@andyqiu/codeforge@${latestVersion}`
@@ -263,7 +264,7 @@ async function cmdUpgrade(args) {
     return 0
   }
 
-  const r = spawnSync(npmCmd, npmArgs, { stdio: "pipe", shell: true })
+  const r = spawnSync(npmCmd, npmArgs, { stdio: "pipe" })
   if (r.status !== 0) {
     if (r.stderr) process.stderr.write(r.stderr)
     if (r.stdout) process.stderr.write(r.stdout)

package/codeforge.json

@@ -3,7 +3,7 @@
   "_doc": "CodeForge 总配置文件。所有 agent 的模型映射 + 全局运行时偏好。改完跑 `npm run sync:models` 把 models 同步到 agents/*.md。",
 
   "models": {
-    "_doc": "agent → 主模型 + fallback 链。效果优先：主模型用各家最新旗舰，fallback 跨家保高可用。Anthropic Opus 4.8 / Sonnet 4.6 / Haiku 4.5；OpenAI GPT-5.5 (2026-04-23) / GPT-5.4；Google Gemini 3 Pro / 2.5 Pro。",
+    "_doc": "agent → 主模型 + fallback 链。三家分工：Anthropic 负责生产（规划+编码），OpenAI 负责审查（跨家偏见隔离），Gemini 负责调度和轻量快速任务。",
     "default": "anthropic/claude-opus-4-8",
 
     "agents": {
@@ -12,11 +12,11 @@
         "model": "anthropic/claude-opus-4-8",
         "category": "deep",
         "tier": "deep",
-        "thinking": { "type": "enabled", "budget_tokens": 8000 },
+        "thinking": { "type": "enabled", "budget_tokens": 16000 },
         "fallback_models": [
           "anthropic/claude-opus-4-8",
           "anthropic/claude-sonnet-4-6",
-          "google/gemini-3-pro"
+          "google/gemini-3.1-pro"
         ]
       },
       "coder": {
@@ -24,11 +24,10 @@
         "model": "anthropic/claude-sonnet-4-6",
         "category": "balanced",
         "tier": "balanced",
-        "thinking": { "type": "enabled", "budget_tokens": 4000 },
+        "thinking": { "type": "enabled", "budget_tokens": 6000 },
         "fallback_models": [
           "anthropic/claude-opus-4-8",
-          "anthropic/claude-opus-4-8",
-          "google/gemini-3-pro"
+          "google/gemini-3.1-pro"
         ]
       },
       "coder-quick": {
@@ -39,7 +38,7 @@
         "thinking": { "type": "enabled", "budget_tokens": 2000 },
         "fallback_models": [
           "anthropic/claude-opus-4-8",
-          "anthropic/claude-sonnet-4-6"
+          "google/gemini-3.1-pro"
         ]
       },
       "coder-deep": {
@@ -47,23 +46,21 @@
         "model": "anthropic/claude-opus-4-8",
         "category": "deep",
         "tier": "deep",
-        "thinking": { "type": "enabled", "budget_tokens": 8000 },
+        "thinking": { "type": "enabled", "budget_tokens": 16000 },
         "fallback_models": [
-          "anthropic/claude-opus-4-8",
-          "anthropic/claude-sonnet-4-6",
-          "google/gemini-3-pro"
+          "openai/gpt-5.5",
+          "google/gemini-3.1-pro"
         ]
       },
       "reviewer": {
-        "_doc": "审查者：Opus 4.8 主（稳定性优先），GPT-5.5 降 fallback（跨家族偏见隔离）。thinking 2000：审查 prompt 一般不大，长思考链浪费。",
-        "model": "anthropic/claude-opus-4-8",
+        "_doc": "审查者：GPT-5.5 主（跨家族偏见隔离），Opus 4.8 降 fallback。thinking 2000：审查 prompt 一般不大，长思考链浪费。",
+        "model": "openai/gpt-5.5",
         "category": "ultrabrain",
         "tier": "deep",
-        "thinking": { "type": "enabled", "budget_tokens": 2000 },
+        "thinking": { "type": "enabled", "reasoning_effort": "xhigh" },
         "fallback_models": [
           "anthropic/claude-opus-4-8",
-          "anthropic/claude-sonnet-4-6",
-          "google/gemini-3-pro"
+          "google/gemini-3.1-pro"
         ]
       },
       "discover": {
@@ -71,49 +68,58 @@
         "model": "anthropic/claude-opus-4-8",
         "category": "deep",
         "tier": "deep",
-        "thinking": { "type": "enabled", "budget_tokens": 6000 },
+        "thinking": { "type": "enabled", "budget_tokens": 16000 },
         "fallback_models": [
-          "anthropic/claude-opus-4-8",
-          "anthropic/claude-sonnet-4-6",
-          "google/gemini-3-pro"
+          "openai/gpt-5.5",
+          "google/gemini-3.1-pro"
         ]
       },
       "discover-challenger": {
-        "_doc": "强对抗子 agent：Opus 4.8 主（稳定），GPT-5.5 降 fallback（跨家族偏见隔离），结论由离散触发组合锁死。",
-        "model": "anthropic/claude-opus-4-8",
+        "_doc": "强对抗子 agent：GPT-5.5 主（跨家族偏见隔离），Opus 4.8 降 fallback，结论由离散触发组合锁死。",
+        "model": "openai/gpt-5.5",
         "category": "ultrabrain",
         "tier": "deep",
-        "thinking": { "type": "enabled", "budget_tokens": 3000 },
+        "thinking": { "type": "enabled", "reasoning_effort": "high" },
         "fallback_models": [
           "anthropic/claude-opus-4-8",
+          "google/gemini-3.1-pro"
+        ]
+      },
+      "codeforge": {
+        "_doc": "编排者（调度专用）：Gemini 3.1 Pro 主（轻量快速路由），Sonnet 4.6 / Opus 4.8 备用。codeforge 无写权限，只路由不生产，用 Gemini 降成本。category=balanced 对应模型档次，tier=balanced 对应任务难度（路由判断本身是轻量任务）。",
+        "model": "google/gemini-3.1-pro",
+        "category": "balanced",
+        "tier": "balanced",
+        "thinking": { "type": "enabled", "thinking_level": "low" },
+        "fallback_models": [
           "anthropic/claude-sonnet-4-6",
-          "google/gemini-3-pro"
+          "anthropic/claude-opus-4-8"
         ]
       }
     },
 
-    "categories": {
+      "categories": {
       "deep": {
-        "_doc": "用于规划 / 架构 / 复杂推理（Opus 4.8 + 8K thinking）",
+        "_doc": "用于规划 / 架构 / 复杂推理（Opus 4.8 + 16K thinking）",
         "model": "anthropic/claude-opus-4-8",
-        "thinking": { "type": "enabled", "budget_tokens": 8000 },
+        "thinking": { "type": "enabled", "budget_tokens": 16000 },
         "fallback_models": ["anthropic/claude-opus-4-8", "anthropic/claude-sonnet-4-6"]
       },
       "ultrabrain": {
-        "_doc": "用于代码审查 / 独立评估 / 跨视角（Opus 4.8 旗舰，GPT-5.5 备用）",
-        "model": "anthropic/claude-opus-4-8",
+        "_doc": "用于代码审查 / 独立评估 / 跨视角（GPT-5.5 主，Opus 4.8 备用；跨家偏见隔离）",
+        "model": "openai/gpt-5.5",
         "fallback_models": ["anthropic/claude-opus-4-8", "anthropic/claude-sonnet-4-6"]
       },
       "balanced": {
         "_doc": "速度+智能平衡（Sonnet 4.6 + 4K thinking）",
         "model": "anthropic/claude-sonnet-4-6",
         "thinking": { "type": "enabled", "budget_tokens": 4000 },
-        "fallback_models": ["openai/gpt-5.4", "google/gemini-3-pro"]
+        "fallback_models": ["openai/gpt-5.5", "google/gemini-3.1-pro"]
       },
       "quick": {
         "_doc": "短上下文 / 快速校验（Haiku 4.5）",
         "model": "anthropic/claude-haiku-4-5",
-        "fallback_models": ["anthropic/claude-sonnet-4-6", "openai/gpt-5.4"]
+        "fallback_models": ["anthropic/claude-sonnet-4-6", "openai/gpt-5.5"]
       }
     },