feat(prompt): 重新注册 bash，把"不要用 bash 搜索"写进工具描述

真实会话里模型仍习惯性先调 bash grep，被拒后才改用 grep 工具；系统提示词里已有 4 处
提醒，但模型下调用决定时看的是工具 schema 描述，而 Pi 内置 bash 描述对此只字未提。

- 新增 extensions/bash-override.ts：用 Pi 的 createBashToolDefinition 重新注册 bash，
  执行、输出结构与 codemode 说明完全沿用上游；shellPath/commandPrefix 按 ctx.cwd
  经 SettingsManager 惰性解析，与内置一致
- tools.jsonc 新增 descriptionAppend 字段，由 prompt-guard 注册时拼接到上游描述后
- bash 增加命令→工具对照规则（grep -n / grep -rl / find -name），并提示
  文件清单用 $(git ls-files) 传参、不要接 xargs（会触发权限审核）

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

AGENTS.md

@@ -20,7 +20,8 @@
 | `pi-notify/` | Kitty OSC 99 完成与权限确认通知，精确聚焦来源窗口 |
 | `pi-compaction/` | 压力触发的观察遮蔽（Pi context_edit）、阈值整段摘要、effective token、native compaction |
 | `pi-overflow/` | 新工具结果统一截断：超阈值原文落盘 + 有界视图 + `overflow_query` |
-| `extensions/fff-override.ts` | FFF `override` 模式接管 `find`/`grep`/`multi_grep` 与 `@` 补全 |
+| `extensions/fff-override.ts` | FFF `override` 模式接管 `find`/`grep` 与 `@` 补全，`multi_grep` 按上游保持关闭 |
+| `extensions/bash-override.ts` | 用 Pi 的 `createBashToolDefinition` 重新注册 `bash`，仅替换模型可见描述（追加 `tools.jsonc` `descriptionAppend`）；执行逻辑、输出结构、codemode 说明不变 |
 | `extensions/codegraph.ts` | 原生 `codegraph_explore`，自有 stdio JSON-RPC 直连 `codegraph serve`（`codegraph-client.ts`） |
 | `extensions/hashline.ts` | 部署 `config/pi-hashline-edit.json` 后加载 pi-hashline-edit 覆盖 `read`/`edit`（`hashline-config.ts`） |
 | `extensions/permission-system.ts` | 直接加载仓库内 `pi-permission-system/src/index.ts` |

README.md

@@ -2,7 +2,7 @@
 
 个人维护的 Pi 扩展组合包。安装一次即可加载并配置：
 
-- `@ff-labs/pi-fff@0.10.5`：使用官方 `override` 模式接管 `find`、`grep` 和 `multi_grep`。
+- `@ff-labs/pi-fff@0.10.5`：使用官方 `override` 模式接管 `find` 和 `grep`；上游默认关闭的 `multi_grep` 保持关闭。
 - 本仓库维护的 `pi-rtk-optimizer`：只处理 Bash/read 等输出压缩，默认不改写命令，也不处理搜索工具结果。
 - 本仓库自维护的 `pi-web-search`：对模型只暴露 `web_search` / `web_fetch` 两个工具；`web_search` 由当前主模型在 Tavily / Exa / Keenable 间路由，`web_fetch` 固定走 Exa，三家均以纯 `fetch` 直连（Exa 为 `mcp.exa.ai` 的 JSON-RPC，不经 MCP adapter）。
 - `extensions/codegraph.ts`：注册原生 `codegraph_explore` 工具，按索引根惰性启动常驻 `codegraph serve --mcp` 子进程，由组合包自有 stdio JSON-RPC 客户端直连（无 MCP adapter/SDK）；serve 的文件监听自动同步索引，无结果缓存；未初始化项目经确认后才 `codegraph init`，拒绝记入会话不再重复询问。
@@ -138,7 +138,7 @@ CodeGraph 的更新或遥测设置。
 
 ### 工具与搜索路由
 
-`config/prompt/routing.jsonc` 定义跨工具路由规则，由 `extensions/system-prompt.ts` 在每轮开始前按当前激活工具渲染为 `<tool_routing>` 分区。代码结构、调用关系和待修改 symbol 优先使用 CodeGraph，定义、引用、类型和修改后诊断优先使用 LSP；仓库字面搜索先用 FFF `find` 缩小文件或目录范围，再在已收敛的路径中用 `grep`/`multi_grep` 获取行号。宽泛搜索命中大量结果、发生截断或达到上限时，应继续缩小路径、glob 或 pattern，而不是提高 limit 或输出全部结果。
+`config/prompt/routing.jsonc` 定义跨工具路由规则，由 `extensions/system-prompt.ts` 在每轮开始前按当前激活工具渲染为 `<tool_routing>` 分区。代码结构、调用关系和待修改 symbol 优先使用 CodeGraph，定义、引用、类型和修改后诊断优先使用 LSP；本地字面搜索按已知信息直接选工具：确切标识符或文本用 `grep` 字面模式并限定 path/exclude，多个命名变体合并为一次 `grep` 正则 alternation，文件名、路径片段或功能概念用 FFF `find` 模糊全路径匹配，然后只 `read` 命中区域。对搜索结果做提取、去重、计数或聚合时，在 codemode 脚本中调用 `tools.grep` / `tools.find` 并用 JavaScript 后处理；bash 中不得运行 `grep`/`rg`/`find`/`fd`/`git grep`（策略 deny 的 reason 指明替代工具），程序需要文件清单时传 `$(git ls-files '<glob>')`。宽泛搜索命中大量结果、发生截断或达到上限时，应继续缩小路径、glob 或 pattern，而不是提高 limit 或输出全部结果。
 
 在线搜索默认按任务只选一个服务：关键词不确定、新闻/近期话题或需要较多候选来源时优先 Tavily；官方/API 文档、版本信息、论文和高确定性事实确认优先 Exa；中文政策、政府公告、中文网页、站点限定和日期筛选优先 Keenable。只有来源质量或不确定性确实需要时才跨服务复核，避免默认把三套结果全部塞进上下文。
 
@@ -211,9 +211,14 @@ Hashline 不负责压缩大文件；每行锚点本身还会增加少量 token
 
 `extensions/fff-override.ts` 强制设置 FFF 官方的 `PI_FFF_MODE=override`（显式传入
 `--fff-mode` 仍按 FFF 官方优先级覆盖它）。FFF 替换
-Pi 的 `find` 和 `grep`，并提供 `multi_grep`；RTK 对所有 `grep` 工具结果直接跳过，
+Pi 的 `find` 和 `grep`；`multi_grep` 上游标注为有害、默认关闭，wrapper 在加载前
+删除 `PI_FFF_MULTIGREP` 环境变量，使其永远无法注册。RTK 对所有 `grep` 工具结果直接跳过，
 也默认关闭 Bash 命令改写，因此不会把 `rg`、`grep`、`find` 或 `fd` 改写到 RTK。
 
+`extensions/bash-override.ts` 用 Pi 的 `createBashToolDefinition` 重新注册内置 `bash`，
+把「不要用 bash 搜索」的警告经 `tools.jsonc` 的 `descriptionAppend` 写进工具 schema，
+让模型在调用时就看到替代工具；执行逻辑、输出结构和 codemode 的 `tools.bash` 说明完全沿用上游定义。
+
 ### 输出压缩归 RTK
 
 RTK 保留 Bash、read、build、test、lint 和 Git 输出压缩。默认 `read` 的有损压缩

config/pi-permission-policy.jsonc

@@ -37,7 +37,6 @@
     "edit": "allow",
     "grep": "allow",
     "find": "allow",
-    "multi_grep": "allow",
     "ls": "allow",
     "codegraph_explore": "allow",
     "lsp_diagnostics": "allow",
@@ -103,10 +102,10 @@
     },
     "bash": {
       "*": "allow",
-      "rg*": "deny",
+      "rg*": { "action": "deny", "reason": "Bash content search is disabled. Locally use the grep tool (pattern + path/exclude; regex alternation for several names); to extract/dedupe/count matches, call tools.grep in codemode and process in JS; to filter command output, use overflow_query on the stored result. Over SSH use ssh_grep." },
-      "grep*": "deny",
+      "grep*": { "action": "deny", "reason": "Bash content search is disabled. Locally use the grep tool (pattern + path/exclude; regex alternation for several names); to extract/dedupe/count matches, call tools.grep in codemode and process in JS; to filter command output, use overflow_query on the stored result. Over SSH use ssh_grep." },
-      "find*": "deny",
+      "find*": { "action": "deny", "reason": "Bash file search is disabled. Locally use the find tool (fuzzy path, or a glob in path such as 'src/**/*.ts'); to pass a file list to a program use $(git ls-files '<glob>'). Over SSH use ssh_find." },
-      "fd*": "deny",
+      "fd*": { "action": "deny", "reason": "Bash file search is disabled. Locally use the find tool (fuzzy path, or a glob in path such as 'src/**/*.ts'); to pass a file list to a program use $(git ls-files '<glob>'). Over SSH use ssh_find." },
       "git *": "ask",
       "git status*": "allow",
       "git diff*": "allow",
@@ -114,7 +113,7 @@
       "git show*": "allow",
       "git rev-parse*": "allow",
       "git ls-files*": "allow",
-      "git grep*": "deny",
+      "git grep*": { "action": "deny", "reason": "Bash content search is disabled. Locally use the grep tool (pattern + path/exclude; regex alternation for several names); to extract/dedupe/count matches, call tools.grep in codemode and process in JS; to filter command output, use overflow_query on the stored result. Over SSH use ssh_grep." },
       "git branch": "allow",
       "git branch --show-current*": "allow",
       "git remote -v": "allow",

config/prompt/routing.jsonc

@@ -16,16 +16,25 @@
       "text": "Use LSP for symbol definitions, references, hover/type information, and post-edit diagnostics when the language server supports the file."
     },
     {
-      "whenAny": ["find", "grep", "multi_grep"],
+      "whenAny": ["find", "grep"],
-      "text": "For literal search, narrow in stages instead of requesting a repository-wide dump:\n  1. Use find to identify likely files or directories. Constrain path/glob and keep the result limit small.\n  2. Use grep or multi_grep only inside the narrowed path or file to obtain matching line numbers. Start with a specific or literal pattern, a small context window, and a bounded limit.\n  3. Use read with offset/limit only for the exact matching region when full source is still needed."
+      "text": "Pick the local search tool by what you already know:\n  - Exact identifier or text: grep with a literal pattern, constrained by path/exclude. Do not run find first just to locate it.\n  - Several identifiers or naming variants: one grep with a regex alternation (e.g. user_id|userId|UserId) instead of several greps.\n  - A filename, path fragment, or feature concept: find (fuzzy whole-path match; use a glob in path for exact names).\n  - Then read only the matching region with offset/limit."
     },
     {
-      "whenAny": ["find", "grep", "multi_grep"],
+      "whenAny": ["find", "grep"],
       "text": "Do not run a broad search expected to return hundreds or thousands of files or matches. If results truncate or hit a limit, narrow the path, glob, or pattern instead of increasing the limit or dumping all results."
     },
+    {
+      "whenAll": ["codemode"],
+      "whenAny": ["find", "grep"],
+      "text": "To extract, deduplicate, count, or aggregate across search matches (what grep -o | sort -u | wc would do), call tools.grep / tools.find inside one codemode script and post-process the results in JavaScript."
+    },
     {
       "whenAny": ["bash"],
-      "text": "Use bash for builds, tests, git, package scripts, and other programs. Never search files by running grep, find, rg, or fd through bash; use the find/grep tools."
+      "text": "Use bash for builds, tests, git, package scripts, and other programs. Never run grep, rg, find, fd, or git grep in bash — standalone, piped, inside a loop, or in $(…). Search files with the grep/find tools; when a program needs a file list (wc, node, tsc), pass $(git ls-files '<glob>') as arguments rather than piping into xargs."
+    },
+    {
+      "whenAll": ["bash", "overflow_query"],
+      "text": "To filter a command's output, run it without | grep; if the result carries a [pi-overflow id=…] marker, search it with overflow_query instead of re-running."
     },
     {
       "whenAny": ["ssh_connect"],
@@ -59,12 +68,15 @@
     },
     {
       "whenAny": ["read"],
-      "text": "Use read only when exact source is needed or the file is small. Before editing, request the smallest useful offset/limit range; do not read a complete large file merely to copy, compare, hash, count, explore, or summarize it."
+      "text": "Use read only when exact source is needed or the file is small. Before editing, request the smallest useful offset/limit range; do not read a complete large source file merely to copy, compare, hash, count, or explore it. Reading project overview docs such as a README in full is fine."
     },
     {
       "whenAll": ["read", "edit"],
       "text": "For existing-file changes, obtain fresh LINE#HASH anchors with a small read (find/grep output has no anchors), then use anchored edit operations. Reuse fresh anchors returned by successful edits for chained changes; re-read only when anchors are stale or the next target region was not shown."
     },
+    {
+      "text": "When a tool call is denied by policy, use the alternative named in the denial and continue the subtask; do not drop it or present unverified facts as checked."
+    },
     {
       "text": "Stop searching once the current evidence is sufficient. Avoid repeating the same lookup through CodeGraph, find, grep, and read without a concrete information gap."
     }

config/prompt/subagent/agents/Explore.md

@@ -14,7 +14,7 @@ You locate code and answer "where is X / what references Y" questions for the de
 ## Hard limits
 - Do not create, modify, move, copy, or delete files, including temporary files anywhere (also /tmp).
 - Use bash only for read-only commands such as git status, git log, git diff, git show, and ls. Never use output redirection, heredocs, or any command that changes system state.
-- Never search through bash (grep, rg, find, fd); use the find and grep tools.
+- Never search through bash (grep, rg, find, fd); use the find and grep tools. For several names use one grep with a regex alternation (a|b); when a program needs a file list, pass $(git ls-files '<glob>').
 
 ## Approach
 - Match the requested breadth: "quick" is one targeted lookup, "medium" covers a few related locations, "very thorough" searches multiple locations and naming conventions.

config/prompt/subagent/agents/Plan.md

@@ -14,7 +14,7 @@ You explore the codebase and return an implementation plan to the delegating age
 ## Hard limits
 - Do not create, modify, move, copy, or delete files, including temporary files anywhere (also /tmp).
 - Use bash only for read-only commands such as git status, git log, git diff, git show, and ls. Never use output redirection, heredocs, or any command that changes system state.
-- Never search through bash (grep, rg, find, fd); use the find and grep tools.
+- Never search through bash (grep, rg, find, fd); use the find and grep tools. For several names use one grep with a regex alternation (a|b); when a program needs a file list, pass $(git ls-files '<glob>').
 
 ## Process
 1. Restate the goal and the constraints you were given. Record anything ambiguous as an open question instead of guessing.

config/prompt/tools.jsonc

@@ -2,23 +2,32 @@
   // Bundle-authoritative tool guidance, owned by extensions/system-prompt.ts.
   //
   // Every tool the bundle can expose must have an entry. Per field:
-  //   "snippet"     — replaces the tool's one-line <tools> entry (null hides it)
+  //   "snippet"           — replaces the tool's one-line <tools> entry (null hides it)
-  //   "guidelines"  — replaces the tool's <rules> lines
+  //   "guidelines"        — replaces the tool's <rules> lines
-  //   "description" — replaces the tool description at registration; only
+  //   "description"       — replaces the tool description at registration
-  //                   applied to third-party tools loaded through a bundle
+  //   "descriptionAppend" — appended to the description at registration (on top
-  //                   adapter (in-repo tools edit their own source instead)
+  //                         of "description" when both are set)
+  // "description"/"descriptionAppend" only apply to tools registered through a
+  // bundle adapter (prompt-guard), which includes the bash override and the
+  // third-party wrappers; other in-repo tools edit their own source instead.
   // An absent field keeps the upstream text. "inherit": true means the
   // upstream text was reviewed against this bundle and is kept as-is.
   "version": 1,
   "tools": {
     "bash": {
-      "snippet": "Run shell commands for builds, tests, git, package scripts, and other programs; search files with the find/grep tools, not bash"
+      "snippet": "Run shell commands for builds, tests, git, package scripts, and other programs; search files with the find/grep tools, not bash",
+      "descriptionAppend": "Not for searching files: use the grep tool for file contents and the find tool for paths (e.g. `grep -n X src/*.ts` → grep {pattern: \"X\", path: \"src/*.ts\"}); grep, rg, find, fd, and git grep are denied here. When a program needs many files, pass $(git ls-files '<glob>') as arguments instead of piping into xargs, which triggers a permission review.",
+      "guidelines": [
+        "You can inspect PI_* environment variables for current model and session details.",
+        "Search commands are denied in bash; call the tools instead: `grep -n X src/*.ts` → grep {pattern: \"X\", path: \"src/*.ts\"}; `grep -rl X src` → grep {pattern: \"X\", path: \"src/\"} and take the file names; `find . -name '*.ts'` → find {pattern: \"\", path: \"**/*.ts\"}.",
+        "When a program needs many files (wc, node, tsc), pass $(git ls-files '<glob>') as arguments; piping into xargs triggers a permission review."
+      ]
     },
     "read": { "inherit": true },
     "edit": {
       "guidelines": [
         "Use edit with LINE#HASH anchors from the latest read or edit result for all file changes; batch every change to one file into a single edit call.",
-        "find/grep/multi_grep results carry plain line numbers, not LINE#HASH anchors; read the exact target range before editing.",
+        "find/grep results carry plain line numbers, not LINE#HASH anchors; read the exact target range before editing.",
         "Never reuse anchors from an earlier read or edit response once a newer one exists for the same file; treat only the most recent response as carrying valid anchors.",
         "After a successful edit, the returned --- Anchors --- block replaces a re-read for nearby follow-up edits.",
         "On [E_STALE_ANCHOR], the error lists the stale refs and any content-matched candidate anchors; re-read the file to get current anchors before retrying."
@@ -28,19 +37,18 @@
     "write": { "inherit": true },
     "ls": { "inherit": true },
     "powershell": { "inherit": true },
-    "grep": { "inherit": true },
+    "grep": { "snippet": "Search file contents via the FFF index (git-aware, smart-case, frecency-ranked, regex or literal); replaces bash grep/rg" },
-    "find": { "inherit": true },
+    "find": { "snippet": "Find files by fuzzy path or glob via the FFF index; replaces bash find/fd" },
-    "multi_grep": { "inherit": true },
     "tool_search": { "inherit": true },
     "codemode": {
       "guidelines": [
         "Use codemode to batch two or more independent tool calls (Promise.allSettled), chain dependent lookups, or filter large output down to what you need; call a single tool directly instead of wrapping it in a script.",
         "Call edit, write, ssh_connect, ssh_cd, ask_user_question, and Agent directly, never from a codemode script.",
-        "Inside scripts, search with tools.grep, tools.find, or tools.multi_grep; never run grep, find, rg, or fd through tools.bash."
+        "Inside scripts, search with tools.grep or tools.find and do extraction, dedup, and counting in JavaScript; never run grep, find, rg, or fd through tools.bash."
       ]
     },
 
-    "codegraph_explore": { "inherit": true },
+    "codegraph_explore": { "snippet": "Explore code structure from the CodeGraph index: symbols, relationships, call flows, and their source" },
     "lsp_diagnostics": { "inherit": true },
     "lsp_definition": { "inherit": true },
     "lsp_references": { "inherit": true },

extensions/AGENTS.md

@@ -24,7 +24,8 @@
 
 ## 各 wrapper 约定
 
-- `fff-override.ts`：强制 FFF 官方 `override` 模式，统一接管 `find`、`grep`、`multi_grep` 和 `@` 补全；FFF 独占本地字面搜索，远端服务器搜索由 `ssh_find` / `ssh_grep` 负责，RTK 不得接管搜索工具（见 `pi-rtk-optimizer/AGENTS.md`）。显式 CLI flag 仍遵循 FFF 官方优先级。
+- `fff-override.ts`：强制 FFF 官方 `override` 模式，统一接管 `find`、`grep` 和 `@` 补全；`multi_grep` 按上游保持关闭（加载前删除 `PI_FFF_MULTIGREP` 环境变量）。FFF 独占本地字面搜索，远端服务器搜索由 `ssh_find` / `ssh_grep` 负责，RTK 不得接管搜索工具（见 `pi-rtk-optimizer/AGENTS.md`）。显式 CLI flag 仍遵循 FFF 官方优先级。
+- `bash-override.ts`：用 `createBashToolDefinition` 重新注册内置 `bash`，目的只是让 `tools.jsonc` 的 `descriptionAppend` 落到模型可见描述里（prompt-guard 在注册时拼接）。ExtensionContext 不暴露 session settings，因此 `shellPath`/`commandPrefix` 在首次 execute 时按 `ctx.cwd` 经 `SettingsManager.create(ctx.cwd, getAgentDir())` 惰性解析并按 cwd 缓存，metadata 来自 `process.cwd()` 的定义；执行、输出结构与 codemode 说明完全委托上游定义，不得改动上游执行逻辑，也不得用 prepareLoadout 或 provider-payload 改写实现同样效果。
 - `codegraph.ts` + `codegraph-client.ts`：注册原生 `codegraph_explore` 工具，按索引根惰性常驻 `codegraph serve --mcp -p <root>` 子进程，经组合包自有 stdio 逐行 JSON-RPC 客户端直连（不经 shell、不带 pi-mcp-adapter 或 MCP SDK）；serve 进程内置文件监听自动同步索引，无结果缓存。子进程 10 分钟无调用自动关闭、会话结束统一关闭；调用中死亡换新客户端透明重试一次；其启动的上游 daemon（`.codegraph/daemon.sock`）退出后可多驻留 ≤5 分钟，属上游行为，组合包不管理。
 - CodeGraph 是机器级依赖，`install.sh`/`update.sh` 经官方安装器管理 `codegraph` CLI，组合包不内置其运行时。安装、升级和普通启动流程不得执行 `codegraph init`；只有 `codegraph_explore` 在未初始化项目上被调用、且用户在 `ctx.ui.confirm` 中明确确认后才可 `codegraph init` 当前目录，拒绝记入会话级 declined 不再重复询问；init 失败必须如实报告并让模型回退内置工具；不得改动 CodeGraph 更新或遥测设置。CLI 缺失、init 被拒或服务调用失败都不得抛出，必须以文本结果打回并引导模型改用 `find`/`grep`/`read`/LSP。已移除的 `plugin-init.ts`（原 `/plugin_init`）职责已并入本文件的惰性确认，不得恢复。
 - `hashline.ts` + `hashline-config.ts`：部署组合包权威 `config/pi-hashline-edit.json` 后加载 `pi-hashline-edit`；默认以 2 字符行哈希覆盖内置 `read` / `edit`，`grep` 与 `replaceText` 均关闭。它保证锚定编辑而不负责压缩，大文件上下文节省必须来自先定位/隔离分析、再做最小范围 `offset/limit` 读取。

extensions/bash-override.ts

@@ -0,0 +1,38 @@
+import {
+	createBashToolDefinition,
+	type ExtensionAPI,
+	getAgentDir,
+	SettingsManager,
+} from "@earendil-works/pi-coding-agent";
+import { guardExtensionApi } from "./prompt-guard.ts";
+
+type BashDefinition = ReturnType<typeof createBashToolDefinition>;
+
+/** Re-register Pi's bash so tools.jsonc can append to its model-facing description; execution stays upstream. */
+export default function bashOverrideExtension(pi: ExtensionAPI): void {
+	const sessionDefinitions = new Map<string, BashDefinition>();
+	const sessionDefinition = (cwd: string): BashDefinition => {
+		let definition = sessionDefinitions.get(cwd);
+		if (definition === undefined) {
+			const settings = SettingsManager.create(cwd, getAgentDir());
+			definition = createBashToolDefinition(cwd, {
+				shellPath: settings.getShellPath(),
+				commandPrefix: settings.getShellCommandPrefix(),
+			});
+			sessionDefinitions.set(cwd, definition);
+		}
+		return definition;
+	};
+
+	guardExtensionApi(pi).registerTool({
+		...sessionDefinition(process.cwd()),
+		execute: (toolCallId, params, signal, onUpdate, ctx) =>
+			sessionDefinition(ctx?.cwd ?? process.cwd()).execute(
+				toolCallId,
+				params,
+				signal,
+				onUpdate,
+				ctx,
+			),
+	});
+}

extensions/fff-override.ts

@@ -5,5 +5,6 @@ import { guardExtensionApi } from "./prompt-guard.ts";
 /** Load FFF as the bundle's canonical `find` and `grep` implementation. */
 export default function fffOverrideExtension(pi: ExtensionAPI): void {
 	process.env.PI_FFF_MODE = "override";
+	delete process.env.PI_FFF_MULTIGREP;
 	fffExtension(guardExtensionApi(pi));
 }

extensions/prompt-guard.ts

@@ -1,11 +1,28 @@
 import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
 
-import { loadToolCatalog, type ToolCatalog } from "./system-prompt-config.ts";
+import {
+	loadToolCatalog,
+	type ToolCatalog,
+	type ToolCatalogEntry,
+} from "./system-prompt-config.ts";
 
 type AnyMessage = { role?: string; content?: unknown } & Record<string, unknown>;
 
 type AnyHandler = (event: never, ctx: never) => unknown;
 
+/**
+ * The description a tool registers with under its catalog entry: a `description`
+ * override replaces the upstream text, a `descriptionAppend` is joined onto
+ * whichever base wins (override first, upstream otherwise) with a blank line.
+ */
+export function catalogDescription(description: string, entry: ToolCatalogEntry): string {
+	if ("inherit" in entry) return description;
+	const base = typeof entry.description === "string" ? entry.description : description;
+	return typeof entry.descriptionAppend === "string"
+		? `${base.trimEnd()}\n\n${entry.descriptionAppend}`
+		: base;
+}
+
 export interface PromptGuardOptions {
 	/**
 	 * Post-process the messages a wrapped `context` handler leaves behind.
@@ -19,9 +36,11 @@ export interface PromptGuardOptions {
  * Proxy an `ExtensionAPI` for third-party extensions so they cannot bypass the
  * bundle's single system-prompt owner (`extensions/system-prompt.ts`):
  *
- * - `registerTool` applies the tool's `description` override from
+ * - `registerTool` applies the tool's `description` override and
- *   `config/prompt/tools.jsonc` when the catalog carries one (lazily loaded;
+ *   `descriptionAppend` suffix from `config/prompt/tools.jsonc` when the
- *   a catalog error keeps the upstream description).
+ *   catalog carries them (lazily loaded; a catalog error keeps the upstream
+ *   description). An append composes onto the override when both are set, or
+ *   onto the upstream description when only the append is set.
  * - `before_agent_start` handlers keep every returned field except
  *   `systemPrompt`, which is dropped.
  * - `context` handlers are post-processed through `opts.contextFilter`.
@@ -42,10 +61,12 @@ export function guardExtensionApi(pi: ExtensionAPI, opts: PromptGuardOptions = {
 
 	const registerTool: ExtensionAPI["registerTool"] = (tool) => {
 		const entry = catalog()?.tools[tool.name];
-		if (entry !== undefined && !("inherit" in entry) && typeof entry.description === "string") {
+		const description =
-			pi.registerTool({ ...tool, description: entry.description });
+			entry === undefined ? tool.description : catalogDescription(tool.description, entry);
-		} else {
+		if (description === tool.description) {
 			pi.registerTool(tool);
+		} else {
+			pi.registerTool({ ...tool, description });
 		}
 	};
 

extensions/system-prompt-config.ts

@@ -10,7 +10,12 @@ import { fileURLToPath } from "node:url";
 /** Entry in `config/prompt/tools.jsonc`. */
 export type ToolCatalogEntry =
 	| { inherit: true }
-	| { snippet?: string | null; guidelines?: string[]; description?: string };
+	| {
+			snippet?: string | null;
+			guidelines?: string[];
+			description?: string;
+			descriptionAppend?: string;
+	  };
 
 export interface ToolCatalog {
 	version: 1;
@@ -152,7 +157,7 @@ function isStringArray(value: unknown): value is string[] {
 	return Array.isArray(value) && value.every((entry) => typeof entry === "string");
 }
 
-const TOOL_ENTRY_KEYS = new Set(["inherit", "snippet", "guidelines", "description"]);
+const TOOL_ENTRY_KEYS = new Set(["inherit", "snippet", "guidelines", "description", "descriptionAppend"]);
 
 function validateToolEntry(name: string, raw: unknown): ToolCatalogEntry {
 	if (!isPlainObject(raw)) {
@@ -170,9 +175,9 @@ function validateToolEntry(name: string, raw: unknown): ToolCatalogEntry {
 		return { inherit: true };
 	}
 	const hasField =
-		"snippet" in raw || "guidelines" in raw || "description" in raw;
+		"snippet" in raw || "guidelines" in raw || "description" in raw || "descriptionAppend" in raw;
 	if (!hasField) {
-		fail("tools.jsonc", `tool "${name}" needs "inherit": true or at least one of snippet/guidelines/description`);
+		fail("tools.jsonc", `tool "${name}" needs "inherit": true or at least one of snippet/guidelines/description/descriptionAppend`);
 	}
 	if ("snippet" in raw && raw.snippet !== null && typeof raw.snippet !== "string") {
 		fail("tools.jsonc", `tool "${name}".snippet must be a string or null`);
@@ -183,10 +188,13 @@ function validateToolEntry(name: string, raw: unknown): ToolCatalogEntry {
 	if ("description" in raw && typeof raw.description !== "string") {
 		fail("tools.jsonc", `tool "${name}".description must be a string`);
 	}
+	if ("descriptionAppend" in raw && typeof raw.descriptionAppend !== "string") {
+		fail("tools.jsonc", `tool "${name}".descriptionAppend must be a string`);
+	}
 	return raw as ToolCatalogEntry;
 }
 
-function validateToolCatalog(raw: unknown): ToolCatalog {
+export function validateToolCatalog(raw: unknown): ToolCatalog {
 	if (!isPlainObject(raw) || raw.version !== 1 || !isPlainObject(raw.tools)) {
 		fail("tools.jsonc", 'expected { "version": 1, "tools": { ... } }');
 	}
@@ -321,8 +329,8 @@ export function isSubagentCustomPrompt(customPrompt: unknown): boolean {
  * Apply the tool catalog to the mutable prompt options: a `snippet` string
  * replaces the tool's `<tools>` line (`null` hides it), `guidelines` replaces
  * the tool's `<rules>` lines; `{inherit:true}` entries are left untouched.
- * `description` is not applied here — it is a registration-time concern handled
+ * `description` and `descriptionAppend` are not applied here — they are
- * by `extensions/prompt-guard.ts`.
+ * registration-time concerns handled by `extensions/prompt-guard.ts`.
  *
  * Returns the selected tool names that have no catalog entry at all.
  */

package.json

@@ -123,6 +123,7 @@
     "extensions": [
       "./pi-overflow/capture.ts",
       "./extensions/fff-override.ts",
+      "./extensions/bash-override.ts",
       "./pi-web-search/index.ts",
       "./extensions/codegraph.ts",
       "./pi-rtk-optimizer/index.ts",

pi-permission-system/CHANGELOG.md

@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Changed
 
+- **Descend into bash control-flow bodies when enumerating command units.** `for`/`while`/`until` loops, `if`/`elif`/`else` branches, `case` items, `{ … }` brace groups, function definitions, and `!`-negated commands were previously matched only as whole-statement text, so a denied command hidden in a loop body or behind `!` (e.g. `for d in */; do find "$d" …; done`, `! rm -rf x`) bypassed every bash pattern — as did a substitution inside a value (`x=$(grep foo f)`, `export X=$(find .)`). The enumerator still emits each enclosing statement as a unit and now also emits every condition, body, and negated command — plus substitutions hosted in value lists, case subjects, assignments, declarations, and other non-command positions — so the most restrictive decision wins; pure body containers (`do …; done`, `elif`/`else` clauses, `case` items) no longer produce their own units. Hard guards share the same enumeration and now reach those nested commands too.
 - **Replace the AutoReview text-JSON chat reviewer with a conversational tool-call reviewer.** The reviewer makes one independent model call per ask — a fixed review prompt, the rendered transcript, and a bounded JSON description of the planned action — and reaches its verdict only by calling one of two tools, `approve` or `reject`, each carrying a `reason`. A reply with no tool call counts as no verdict; multiple calls, unknown tools, and missing reasons count as invalid responses; neither is retried. An experimental TypeSafe Jev classifier rewrite explored in the working tree was dropped before release and never shipped.
 - **The review model follows the session by default.** With `autoReview.provider` and `autoReview.model` both unset, the reviewer calls the session's current model (`ctx.model`), resolved live per ask so a mid-session switch applies to the next request. Setting both pins a fixed reviewer; setting exactly one is a misconfiguration that warns once at `session_start` and defers every ask.
 - **A rejection defers instead of denying.** `reject` maps to `defer` and the model's `reason` is shown in the human permission prompt, labelled as model-generated. Every review failure — timeout, provider error, malformed or truncated reply, unresolved model — also defers with a fixed system-authored note. Pressing Esc during a review aborts it and denies outright, so no human dialog appears after the user already refused.

pi-permission-system/UPSTREAM.md

@@ -21,3 +21,7 @@ AutoReview 源码已整合进 `src/auto-review/`，不再保留或加载第二
 ## 本地维护
 
 本目录不是 submodule，不保留嵌套 `.git`、上游 `.pi` 本地状态、`dist/` 或构建产物。同步任一上游时必须从其源码仓库的明确 tag/commit 移植并与本地改动对照审查，不得整体覆盖，也不得使用 npm 预编译产物作为导入源。
+
+## 本地改动
+
+- Bash 命令枚举在保留整条控制流语句为独立 command unit 的基础上递归下降：`for`/`while`/`until`、`if`/`elif`/`else`、`case` 分支、`{ … }` 组合语句、函数定义与 `!` 取反命令体内的命令也按 bash pattern 逐条判定；赋值与声明值（`x=$(…)`、`export X=$(…)`）、值列表、case 词与 `[[ … ]]` 测试中的 `$(…)` 替换同样枚举，而 `do …; done`、`elif`/`else`、`case` 分支等纯容器不再单独产生 unit。藏在循环体、取反前缀或赋值载荷里的 `find`/`grep`/`rm` 等不再绕过 deny 或 hard guard。

pi-permission-system/docs/configuration.md

@@ -370,7 +370,10 @@ This holds wherever the substitution appears, not only in argument position.
 A substitution in a **redirect target** (`echo hi > $(rm *.txt)`, `cat < <(rm c)`, ``echo hi 2> `rm d` ``) and one in an **interpolating heredoc body** (`cat <<EOF` with `$(rm e)` in the body) are evaluated the same way.
 A quoted heredoc delimiter (`<<'EOF'` or `<<"EOF"`) does not interpolate, so its body is literal text and nothing in it is evaluated as a command.
 The enclosing command is still matched without its redirect, so a rule like `npm install` keeps matching `npm install > out.txt`.
-Control-flow bodies (`if`/`while`/`for`/`case`) and `{ … }` brace groups are not descended into; their contents are matched as part of the enclosing statement's text.
+Control-flow statements (`for`/`while`/`until` loops, `if`/`elif`/`else` branches, `case` items, `{ … }` brace groups, and function definitions) are evaluated as a whole unit *and* descended into: every command in their conditions and bodies is matched against the bash patterns in addition to the enclosing statement's text.
+So `for d in src docs; do find "$d" -name '*.ts'; done` evaluates both the whole loop and `find "$d" -name '*.ts'`; if `find *` is denied, the whole invocation is denied.
+Non-command parts of a control-flow statement — the `for` value list, the `case` subject and patterns, `[[ … ]]` tests — are not matched as commands themselves, but substitutions nested inside them still are, so `for f in $(find …)` and `if [[ "$(rm e)" = x ]]` gate the inner command.
+Negation and shell bookkeeping get the same treatment: `! grep -q x f` evaluates both `! grep -q x f` and the inner `grep -q x f`, and a substitution inside a value — `x=$(grep foo f)`, `export X=$(find .)`, `local y=$(rg z)` — is gated as a command substitution even though the assignment or declaration itself runs no command.
 
 A leading environment-variable assignment prefix is stripped before matching, so the rule gates the underlying command rather than the prefix.
 So `AWS_PROFILE=prod aws ec2 …` is matched as `aws ec2 …` — a `aws *` rule applies even though the invocation begins with `AWS_PROFILE=`.

pi-permission-system/src/access-intent/bash/command-enumeration.ts

@@ -1,6 +1,7 @@
 import {
   EXECUTION_HOST_TYPES,
   forEachNestedExecution,
+  NESTED_EXECUTION_CONTEXTS,
 } from "#src/access-intent/bash/nested-execution";
 import type { TSNode } from "#src/access-intent/bash/parser";
 import {
@@ -72,6 +73,62 @@ const COMMAND_ENUM_DESCEND = new Set([
  */
 const COMMAND_ENUM_SKIP = new Set(["comment", "heredoc_end"]);
 
+/**
+ * Compound statement node types whose bodies really run commands:
+ * `for`/`while`/`until` loops (tree-sitter-bash folds `until` into
+ * `while_statement`), `if`/`case` statements, `{ … }` brace groups
+ * (`compound_statement`), function definitions, and `!`-negated commands.
+ *
+ * These were once emitted whole without descending (#306 deferred that), which
+ * let a gated command hide inside a loop body — a `for …; do find …; done`
+ * matched only patterns starting with `for`. Each is now emitted whole AND
+ * enumerated child-by-child; a `negated_command` emits `! grep …` whole and
+ * then `grep …` so the `!` prefix cannot defeat a command pattern either.
+ */
+const CONTROL_FLOW_TYPES = new Set([
+  "for_statement",
+  "c_style_for_statement",
+  "while_statement",
+  "if_statement",
+  "case_statement",
+  "compound_statement",
+  "function_definition",
+  "negated_command",
+]);
+
+/**
+ * Control-flow nodes that are pure body containers rather than statements in
+ * their own right: `do …; done`, `elif`/`else` clauses, and `case` items.
+ * They are descended exactly like {@link CONTROL_FLOW_TYPES} members but are
+ * not emitted as whole units — nobody writes a command pattern against
+ * `do …; done` or `a) find . ;;`, so emitting them only adds noise to prompts
+ * and approval suggestions.
+ */
+const CONTROL_FLOW_CONTAINERS = new Set([
+  "do_group",
+  "elif_clause",
+  "else_clause",
+  "case_item",
+]);
+
+/**
+ * Named children of a control-flow node that are statements in their own
+ * right and get full command enumeration (whole-emit rules, wrapper
+ * classification, substitution descent). Any other named child — words,
+ * `variable_name`, a for-loop's value list, a case subject or pattern, a
+ * `test_command` `[[ … ]]`, a comment — contributes only the nested
+ * `$(…)`/`<(…)` executions it hosts.
+ */
+const CONTROL_FLOW_STATEMENT_TYPES = new Set([
+  "command",
+  "list",
+  "pipeline",
+  "redirected_statement",
+  "subshell",
+  ...CONTROL_FLOW_TYPES,
+  ...CONTROL_FLOW_CONTAINERS,
+]);
+
 /**
  * Enumerate the command units of a bash program, in source order.
  *
@@ -82,8 +139,19 @@ const COMMAND_ENUM_SKIP = new Set(["comment", "heredoc_end"]);
  * subshells (`( … )`) — emitting each inner command as its own unit *in
  * addition to* the enclosing command, since those inner commands really execute
  * (#306).
- * Control-flow bodies and `{ … }` brace groups are emitted whole without
+ * Control-flow statements (loops, `if`/`elif`/`else`, `case`, `{ … }` groups,
- * descending (deferred).
+ * function definitions, `!` negation) are likewise emitted whole and then
+ * descended child-by-child: statement children (including nested control flow
+ * and conditions like `if grep -q …; then`) are enumerated as commands, while
+ * non-statement children (the for-loop value list, case subject/patterns,
+ * `[[ … ]]` tests, names) contribute only the nested executions they host.
+ * Pure body containers (`do …; done`, `elif`/`else` clauses, `case` items)
+ * are descended without a whole-unit emit of their own.
+ * Body commands keep the enclosing execution context — a loop inside `$(…)`
+ * runs there, a top-level loop's body runs in the current shell.
+ * Remaining named statements (`variable_assignment`, `declaration_command`,
+ * `test_command`, …) emit whole and still report the nested executions they
+ * host, so `x=$(grep foo f)` and `export X=$(find .)` gate the inner command.
  *
  * The enclosing command/subshell is always still emitted whole, so adding the
  * nested units can only ever produce a more-restrictive decision, never weaker.
@@ -135,9 +203,57 @@ function collectCommandsInto(
     return;
   }
 
-  // Any other named statement (compound_statement `{ … }`, if/while/for/case,
+  if (CONTROL_FLOW_TYPES.has(node.type)) {
-  // function_definition): emit whole, do not descend — deferred (#306).
+    out.push(makeUnit(node.text, context)); // never-weaker whole emit
+    descendControlFlowChildren(node, context, out);
+    return;
+  }
+
+  if (CONTROL_FLOW_CONTAINERS.has(node.type)) {
+    // Pure body container — enumerate its contents, emit no unit of its own.
+    descendControlFlowChildren(node, context, out);
+    return;
+  }
+
+  // Any other named statement (variable_assignment, declaration_command,
+  // test_command, …) keeps the whole-emit fallback, but a substitution in its
+  // payload still really runs — `x=$(grep foo f)` or `export X=$(find .)` —
+  // so hosted executions are enumerated in addition to the whole text.
   out.push(makeUnit(node.text, context));
+  collectHostedCommands(node, out);
+}
+
+/**
+ * Enumerate inside a control-flow statement, keeping the enclosing context.
+ *
+ * Statement children (`command`, `list`, `pipeline`, `redirected_statement`,
+ * `subshell`, or nested control flow such as a `do_group`)
+ * get full enumeration. Other named children are scanned only for nested
+ * executions: a `command_substitution` child directly (the `$(…)` in
+ * `for f in $(…)` or `case $(…) in`) is a context node in its own right — the
+ * strict-subtree {@link forEachNestedExecution} walk would not see it — so it
+ * is descended with its own context, while words, `variable_name`, test and
+ * expression subtrees go through the hosted-execution search.
+ */
+function descendControlFlowChildren(
+  node: TSNode,
+  context: BashCommandContext | undefined,
+  out: BashCommand[],
+): void {
+  for (let i = 0; i < node.childCount; i++) {
+    const child = node.child(i);
+    if (!child?.isNamed) continue;
+    if (CONTROL_FLOW_STATEMENT_TYPES.has(child.type)) {
+      collectCommandsInto(child, context, out);
+      continue;
+    }
+    const nestedContext = NESTED_EXECUTION_CONTEXTS.get(child.type);
+    if (nestedContext) {
+      descendCommandChildren(child, nestedContext, out);
+    } else {
+      collectHostedCommands(child, out);
+    }
+  }
 }
 
 function makeUnit(

pi-permission-system/test/access-intent/bash/command-enumeration.test.ts

@@ -0,0 +1,179 @@
+import { afterEach, beforeEach, describe, expect, it } from "vitest";
+import {
+  resetWarmBashParser,
+  warmBashParser,
+} from "#src/access-intent/bash/parser";
+import { parseBashCommandsSync } from "#src/access-intent/bash/sync-commands";
+
+describe("collectCommands — control-flow descent", () => {
+  beforeEach(async () => {
+    resetWarmBashParser();
+    await warmBashParser();
+  });
+  afterEach(() => {
+    resetWarmBashParser();
+  });
+
+  it("emits a for-loop's body commands in addition to the whole statement", () => {
+    expect(
+      parseBashCommandsSync('for d in a b; do find "$d" -name x; done'),
+    ).toEqual([
+      { text: 'for d in a b; do find "$d" -name x; done' },
+      { text: 'find "$d" -name x' },
+    ]);
+  });
+
+  it("emits a command substitution in the for value list as a nested context", () => {
+    expect(
+      parseBashCommandsSync('for f in $(find . -name x); do echo "$f"; done'),
+    ).toEqual([
+      { text: 'for f in $(find . -name x); do echo "$f"; done' },
+      { text: "find . -name x", context: "command_substitution" },
+      { text: 'echo "$f"' },
+    ]);
+  });
+
+  it("emits the condition and body commands of a redirected while loop", () => {
+    expect(
+      parseBashCommandsSync('while read l; do grep x "$l"; done < list'),
+    ).toEqual([
+      { text: 'while read l; do grep x "$l"; done' },
+      { text: "read l" },
+      { text: 'grep x "$l"' },
+    ]);
+  });
+
+  it("emits an if's condition, branch, and else commands", () => {
+    expect(
+      parseBashCommandsSync("if grep -q x f; then echo y; else rm z; fi"),
+    ).toEqual([
+      { text: "if grep -q x f; then echo y; else rm z; fi" },
+      { text: "grep -q x f" },
+      { text: "echo y" },
+      { text: "rm z" },
+    ]);
+  });
+
+  it("emits a case item's commands", () => {
+    expect(parseBashCommandsSync("case $x in a) find . ;; esac")).toEqual([
+      { text: "case $x in a) find . ;; esac" },
+      { text: "find ." },
+    ]);
+  });
+
+  it("emits a command substitution in the case subject as a nested context", () => {
+    expect(
+      parseBashCommandsSync("case $(rm x) in a) find . ;; esac"),
+    ).toEqual([
+      { text: "case $(rm x) in a) find . ;; esac" },
+      { text: "rm x", context: "command_substitution" },
+      { text: "find ." },
+    ]);
+  });
+
+  it("descends a brace group", () => {
+    expect(parseBashCommandsSync("{ grep x f; }")).toEqual([
+      { text: "{ grep x f; }" },
+      { text: "grep x f" },
+    ]);
+  });
+
+  it("descends a function definition's body", () => {
+    expect(parseBashCommandsSync("f() { grep x f; }")).toEqual([
+      { text: "f() { grep x f; }" },
+      { text: "{ grep x f; }" },
+      { text: "grep x f" },
+    ]);
+  });
+
+  it("descends nested loops", () => {
+    expect(
+      parseBashCommandsSync("for a in 1; do for b in 2; do find .; done; done"),
+    ).toEqual([
+      { text: "for a in 1; do for b in 2; do find .; done; done" },
+      { text: "for b in 2; do find .; done" },
+      { text: "find ." },
+    ]);
+  });
+
+  it("descends a c-style for loop", () => {
+    expect(
+      parseBashCommandsSync("for ((i=0;i<3;i++)); do echo i; done"),
+    ).toEqual([
+      { text: "for ((i=0;i<3;i++)); do echo i; done" },
+      { text: "echo i" },
+    ]);
+  });
+
+  it("emits no container units for a clean loop body", () => {
+    expect(parseBashCommandsSync('for d in a b; do echo "$d"; done')).toEqual([
+      { text: 'for d in a b; do echo "$d"; done' },
+      { text: 'echo "$d"' },
+    ]);
+  });
+
+  it("does not emit a [[ ]] test as a separate command unit", () => {
+    expect(parseBashCommandsSync("if [[ -f x ]]; then echo y; fi")).toEqual([
+      { text: "if [[ -f x ]]; then echo y; fi" },
+      { text: "echo y" },
+    ]);
+  });
+
+  it("catches a substitution hosted inside a [[ ]] test", () => {
+    expect(
+      parseBashCommandsSync('if [[ "$(rm x)" = y ]]; then echo y; fi'),
+    ).toEqual([
+      { text: 'if [[ "$(rm x)" = y ]]; then echo y; fi' },
+      { text: "rm x", context: "command_substitution" },
+      { text: "echo y" },
+    ]);
+  });
+});
+
+describe("collectCommands — assignment, declaration, and negation", () => {
+  beforeEach(async () => {
+    resetWarmBashParser();
+    await warmBashParser();
+  });
+  afterEach(() => {
+    resetWarmBashParser();
+  });
+
+  it("emits a substitution in a plain assignment value as a nested context", () => {
+    expect(parseBashCommandsSync("x=$(grep foo f)")).toEqual([
+      { text: "x=$(grep foo f)" },
+      { text: "grep foo f", context: "command_substitution" },
+    ]);
+  });
+
+  it("emits a substitution in an export declaration value", () => {
+    expect(parseBashCommandsSync("export X=$(find .)")).toEqual([
+      { text: "export X=$(find .)" },
+      { text: "find .", context: "command_substitution" },
+    ]);
+  });
+
+  it("emits a substitution in a local declaration inside a function body", () => {
+    expect(parseBashCommandsSync("f() { local y=$(rg z); }")).toEqual([
+      { text: "f() { local y=$(rg z); }" },
+      { text: "{ local y=$(rg z); }" },
+      { text: "rg z", context: "command_substitution" },
+    ]);
+  });
+
+  it("emits a negated command's inner command as a normal unit", () => {
+    expect(parseBashCommandsSync("! grep -q x f")).toEqual([
+      { text: "! grep -q x f" },
+      { text: "grep -q x f" },
+    ]);
+  });
+
+  it("emits a negated condition's inner command inside an if", () => {
+    expect(parseBashCommandsSync("if ! grep -q x f; then :; fi")).toEqual([
+      { text: "if ! grep -q x f; then :; fi" },
+      { text: "! grep -q x f" },
+      { text: "grep -q x f" },
+      { text: ":" },
+    ]);
+  });
+});

pi-permission-system/test/handlers/gates/bash-command-metamorphic.test.ts

@@ -14,10 +14,13 @@ import { BashProgram } from "#src/access-intent/bash/program";
 import { resolveBashCommandCheck } from "#src/handlers/gates/bash-command";
 import { pathFlavorForPlatform } from "#src/path/path-flavor";
 import { PathNormalizer } from "#src/path-normalizer";
+import { PermissionResolver } from "#src/permission-resolver";
 import type { ScopedPermissionResolver } from "#src/permission-resolver";
-import type { PermissionState } from "#src/types";
+import { SessionRules } from "#src/session-rules";
+import type { DenyWithReason, PermissionState } from "#src/types";
 
 import { makeCheckResult } from "#test/helpers/handler-fixtures";
+import { createInMemoryManager } from "#test/helpers/manager-harness";
 
 /** Decision strength ordering: deny (2) > ask (1) > allow (0). */
 const STRENGTH: Record<PermissionState, number> = {
@@ -146,3 +149,103 @@ describe("bash command gate — nested execution hosts do not weaken", () => {
     expect(await decide("cat <<'EOF'\n$(rm x)\nEOF", resolver)).toBe("allow");
   });
 });
+
+/**
+ * The same totality property for control-flow bodies.
+ *
+ * A `for`/`while`/`if`/`case`/function was once emitted whole without
+ * descending, so `for d in …; do find …; done` only ever matched patterns
+ * starting with `for` — a `find*: deny` rule was bypassed. The enumerator now
+ * emits the whole statement AND its body commands, so hosting a gated command
+ * inside control flow cannot weaken its decision.
+ */
+describe("bash command gate — control-flow bodies do not weaken", () => {
+  /** Real pattern matching over a bash map: `*` allow, the rest deny-with-reason. */
+  function makePolicyResolver(
+    bash: Record<string, PermissionState | DenyWithReason>,
+  ): ScopedPermissionResolver {
+    const manager = createInMemoryManager({
+      global: { permission: { "*": "allow", bash } },
+    });
+    return new PermissionResolver(manager, new SessionRules());
+  }
+
+  async function checkWith(
+    command: string,
+    resolver: ScopedPermissionResolver,
+  ) {
+    const program = await BashProgram.parse(
+      command,
+      new PathNormalizer(pathFlavorForPlatform(process.platform), "/cwd"),
+    );
+    return resolveBashCommandCheck(command, program.commands(), undefined, resolver);
+  }
+
+  it("denies the reported repro: find inside a for-loop command substitution", async () => {
+    const resolver = makePolicyResolver({
+      "*": "allow",
+      "find*": { action: "deny", reason: "use the find tool" },
+      "grep*": { action: "deny", reason: "use the grep tool" },
+    });
+
+    const result = await checkWith(
+      'cd /Users/sia/code/my-pi && ls -1 extensions/ && echo "=== LOC by ext ===" && for d in pi-*/; do n=$(find "$d" -name \'*.ts\' -not -path \'*/node_modules/*\' -not -path \'*/dist/*\' -not -path \'*/test*\' | xargs wc -l 2>/dev/null | tail -1 | awk \'{print $1}\'); echo "$d $n"; done && echo "=== extensions ts ===" && wc -l extensions/*.ts | tail -3',
+      resolver,
+    );
+
+    expect(result.state).toBe("deny");
+    expect(result.command).toBe(
+      'find "$d" -name \'*.ts\' -not -path \'*/node_modules/*\' -not -path \'*/dist/*\' -not -path \'*/test*\'',
+    );
+    expect(result.matchedPattern).toBe("find*");
+    expect(result.reason).toBe("use the find tool");
+    expect(result.commandContext).toBe("command_substitution");
+  });
+
+  it.each([
+    'for d in a b; do find "$d" -name x; done',
+    'while read l; do grep x "$l"; done < list',
+    "if grep -q x f; then echo y; else echo n; fi",
+    "case $x in a) find . ;; esac",
+    "{ grep x f; }",
+    "f() { grep x f; }",
+    "for a in 1; do for b in 2; do find .; done; done",
+    "x=$(grep foo f)",
+    "export X=$(find .)",
+    "f() { local y=$(rg z); }",
+    "! grep -q x f",
+    "if ! grep -q x f; then :; fi",
+  ])("denies a gated command inside `%s`", async (command) => {
+    const resolver = makePolicyResolver({
+      "*": "allow",
+      "find*": "deny",
+      "grep*": "deny",
+      "rg*": "deny",
+    });
+
+    expect((await checkWith(command, resolver)).state).toBe("deny");
+  });
+
+  it("denies a hard-guard payload hidden in a loop body", async () => {
+    const resolver = makePolicyResolver({
+      "*": "allow",
+      "rm -rf*": "deny",
+    });
+
+    const result = await checkWith("for f in a; do rm -rf /; done", resolver);
+    expect(result.state).toBe("deny");
+    expect(result.matchedPattern).toBe("rm -rf*");
+  });
+
+  it("allows a control-flow statement whose body commands are allowed", async () => {
+    const resolver = makePolicyResolver({
+      "*": "allow",
+      "find*": "deny",
+      "grep*": "deny",
+    });
+
+    expect(
+      (await checkWith('for d in a b; do echo "$d"; done', resolver)).state,
+    ).toBe("allow");
+  });
+});

pi-permission-system/test/seed-config.test.ts

@@ -7,6 +7,7 @@ import {
   autoReviewConfigFileSchema,
   normalizeAutoReviewConfig,
 } from "#src/auto-review/config";
+import { stripJsonComments } from "#src/config-loader";
 import { unifiedConfigSchema } from "#src/config-schema";
 
 /** The bundle's authoritative policy seed, as shipped. */
@@ -16,10 +17,10 @@ const SEED_PATH = fileURLToPath(
 
 function readSeed(): Record<string, unknown> {
   const raw = readFileSync(SEED_PATH, "utf8");
-  const stripped = raw
+  // Parse with the same string-aware stripper the loader uses: a naive regex
-    .replace(/^\s*\/\/.*$/gm, "")
+  // strips `/*`-looking runs inside string values (e.g. a `'src/**/*.ts'`
-    .replace(/\/\*[\s\S]*?\*\//g, "");
+  // glob in a deny reason) and corrupts the JSON.
-  return JSON.parse(stripped) as Record<string, unknown>;
+  return JSON.parse(stripJsonComments(raw)) as Record<string, unknown>;
 }
 
 describe("bundled policy seed", () => {

pi-tool-search/CHANGELOG.md

@@ -3,6 +3,7 @@
 ## Unreleased — my-pi maintained source
 
 ### Changed
+- Removed `multi_grep` from the `filesystem-navigation` group: the bundle wrapper keeps FFF's upstream `PI_FFF_MULTIGREP` opt-in disabled, so the tool is never registered and the group now covers only `ls`.
 - Pinned Pi's built-in `codemode` (exposure `model-only`) in `alwaysEnabled` so batching several tool calls through one script never depends on a `tool_search` activation; `platformToolPolicy` now also drops tools with `hidden` or `codemode` (script-only) exposure so they can never enter the catalog or the unknown-tools path.
 - Removed the `PI_TOOL_SEARCH_CONFIG_PATH` environment override; the extension reads the bundle `config/tool-search.json`, and tests inject a path through the `createToolSearchExtension({ configPath })` factory.
 - Removed the retired context-mode integration: the `ctx_execute` / `ctx_execute_file` / `ctx_batch_execute` pins left `alwaysEnabled` and the four `context-*` bundle groups were dropped. Large tool results are now bounded by pi-overflow and recoverable through the pinned `overflow_query`.

pi-tool-search/README.md

@@ -26,7 +26,7 @@ Run `/tool-search-status` to inspect the current catalog source, pinned count, d
 
 | Group | Bundle tools |
 |---|---|
-| `filesystem-navigation` | `ls`, `multi_grep` |
+| `filesystem-navigation` | `ls` |
 | `ssh-connection` | `ssh_connect` |
 | `ssh-remote-files` | `ssh_read`, `ssh_write`, `ssh_edit` |
 | `ssh-remote-search` | `ssh_find`, `ssh_grep` |

pi-tool-search/extensions/bundle-groups.ts

@@ -25,11 +25,11 @@ export const BUNDLE_GROUP_DEFINITIONS: BundleGroupDefinition[] = [
   {
     id: "filesystem-navigation",
     title: "Filesystem navigation",
-    summary: "Inspect directory layouts and run bounded multi-pattern literal searches.",
+    summary: "Inspect directory layouts.",
-    useWhen: ["You need directory entries or several literal searches after narrowing a path"],
+    useWhen: ["You need the alphabetical directory entries of an unknown path"],
     avoidWhen: ["Code relationships or types are the real question"],
-    keywords: ["files", "directory", "list", "multi grep", "目录", "文件", "多模式搜索"],
+    keywords: ["files", "directory", "list", "目录", "文件"],
-    tools: ["ls", "multi_grep"],
+    tools: ["ls"],
   },
   {
     id: "ssh-connection",

tests/bundle-config.test.ts

@@ -16,7 +16,10 @@ import {
 import { CONFIG_PATH as RTK_CONFIG_PATH } from "../pi-rtk-optimizer/src/constants.ts";
 import { CONFIG_PATH as COMPACTION_CONFIG_PATH } from "../pi-compaction/src/config.ts";
 import { WORKTREE_CONFIG_PATH } from "../pi-worktree/src/settings.ts";
-import { resolvePreamblePath } from "../extensions/system-prompt-config.ts";
+import {
+  resolvePreamblePath,
+  stripJsoncComments,
+} from "../extensions/system-prompt-config.ts";
 
 const repositoryRoot = resolve(dirname(fileURLToPath(import.meta.url)), "..");
 const bundleConfigDir = join(repositoryRoot, "config");
@@ -48,9 +51,9 @@ const BUNDLE_CONFIG_FILES = [
   "p10k.zsh",
 ];
 
-/** Strip JSONC comments (line and block) — same approach as the seed test. */
+/** Strip JSONC comments (line and block) without touching string literals. */
 function stripJsonc(raw: string): string {
-  return raw.replace(/^\s*\/\/.*$/gm, "").replace(/\/\*[\s\S]*?\*\//g, "");
+  return stripJsoncComments(raw);
 }
 
 test("every bundle config file exists and is non-empty", async () => {

tests/extension-load.test.ts

@@ -58,6 +58,11 @@ test("all package extensions load together without global registration conflicts
 		packageJson.pi.extensions.includes("./extensions/codegraph.ts"),
 		"the bundle must expose the native codegraph_explore tool",
 	);
+	assert.equal(
+		packageJson.pi.extensions.indexOf("./extensions/bash-override.ts"),
+		packageJson.pi.extensions.indexOf("./extensions/fff-override.ts") + 1,
+		"the bash description override must register right after fff-override",
+	);
 	assert.ok(
 		packageJson.pi.extensions.includes("./pi-web-search/index.ts"),
 		"the bundle must load the unified pi-web-search tools",
@@ -292,21 +297,51 @@ test("all package extensions load together without global registration conflicts
 	const home = await mkdtemp(join(tmpdir(), "my-pi-extension-load-"));
 	const mockBin = join(home, "bin");
 	await mkdir(mockBin, { recursive: true });
-	// Probe loaded LAST: records every registered tool (name + exposure) and the
+	// Probe loaded LAST: records every registered tool (name, exposure,
-	// final active set so the test can assert the pinned built-in codemode.
+	// description, source) and the final active set so the test can assert the
+	// pinned built-in codemode. Its async setup also runs the bash override's
+	// registered definition end to end — executing `echo hi` through the same
+	// delegation path a model call would use.
 	const probeOutput = join(home, "probe-tools.json");
 	const probePath = join(home, "probe.ts");
 	await writeFile(
 		probePath,
 		[
 			'import { writeFileSync } from "node:fs";',
-			"export default function (pi: any): void {",
+			`import bashOverrideExtension from ${JSON.stringify(join(repositoryRoot, "extensions", "bash-override.ts"))};`,
+			"export default async function (pi: any): Promise<void> {",
+			"\tlet executed: unknown;",
+			"\ttry {",
+			"\t\tlet registered: any;",
+			"\t\tbashOverrideExtension({ registerTool: (tool: any) => { registered = tool; } } as any);",
+			"\t\tconst result = await registered.execute(",
+			'\t\t\t"probe-1",',
+			'\t\t\t{ command: "echo hi" },',
+			"\t\t\tundefined,",
+			"\t\t\tundefined,",
+			`\t\t\t{ cwd: ${JSON.stringify(home)}, sessionManager: { getSessionId: () => "probe-session", getSessionFile: () => undefined } } as any,`,
+			"\t\t);",
+			"\t\texecuted = {",
+			"\t\t\ttext: result?.content?.[0]?.text,",
+			"\t\t\texitCode: result?.structuredContent?.exit_code,",
+			"\t\t\thasOutputSchema: result?.structuredContent !== undefined,",
+			"\t\t\tregisteredDescription: registered?.description,",
+			"\t\t};",
+			"\t} catch (error) {",
+			"\t\texecuted = { error: String(error) };",
+			"\t}",
 			'\tpi.on("session_start", () => {',
 			"\t\twriteFileSync(",
 			`\t\t\t${JSON.stringify(probeOutput)},`,
 			"\t\t\tJSON.stringify({",
-			"\t\t\t\tall: pi.getAllTools().map((tool: any) => ({ name: tool.name, exposure: tool.exposure })),",
+			"\t\t\t\tall: pi.getAllTools().map((tool: any) => ({",
+			"\t\t\t\t\tname: tool.name,",
+			"\t\t\t\t\texposure: tool.exposure,",
+			"\t\t\t\t\tdescription: tool.description,",
+			"\t\t\t\t\tsourcePath: tool.sourceInfo?.path,",
+			"\t\t\t\t})),",
 			"\t\t\t\tactive: pi.getActiveTools(),",
+			"\t\t\t\texecuted,",
 			"\t\t\t}),",
 			"\t\t);",
 			"\t});",
@@ -337,10 +372,52 @@ test("all package extensions load together without global registration conflicts
 	assert.doesNotMatch(`${result.stdout}\n${result.stderr}`, /Failed to load extension|conflicts with/u);
 
 	const probe = JSON.parse(await readFile(probeOutput, "utf8")) as {
-		all: { name: string; exposure?: string }[];
+		all: { name: string; exposure?: string; description?: string; sourcePath?: string }[];
 		active: string[];
+		executed?: {
+			text?: string;
+			exitCode?: number;
+			hasOutputSchema?: boolean;
+			registeredDescription?: string;
+			error?: string;
+		};
 	};
 	const codemode = probe.all.find((tool) => tool.name === "codemode");
 	assert.equal(codemode?.exposure, "model-only", "Pi's built-in codemode must load with model-only exposure");
 	assert.ok(probe.active.includes("codemode"), "Tool Search must pin codemode into the active set");
+	assert.ok(
+		probe.all.every((tool) => tool.name !== "multi_grep"),
+		"FFF's upstream-flagged multi_grep must stay unregistered (fff-override deletes PI_FFF_MULTIGREP)",
+	);
+	assert.ok(
+		!probe.active.includes("multi_grep"),
+		"multi_grep must not appear in the active tool set",
+	);
+	const bash = probe.all.find((tool) => tool.name === "bash");
+	assert.ok(bash, "the bundle must expose the bash tool");
+	assert.ok(
+		bash.description?.startsWith("Execute a bash command in the current working directory."),
+		"the bash override must keep Pi's upstream description prefix",
+	);
+	assert.ok(
+		bash.description?.endsWith("which triggers a permission review."),
+		"the registered bash description must end with the tools.jsonc descriptionAppend",
+	);
+	assert.ok(
+		(bash.sourcePath ?? "").endsWith("extensions/bash-override.ts"),
+		`bash must be registered by extensions/bash-override.ts, got ${bash.sourcePath}`,
+	);
+	const executed = probe.executed ?? {};
+	assert.equal(executed.error, undefined, `bash override execute must not throw: ${executed.error}`);
+	assert.equal(executed.exitCode, 0, "echo hi must exit 0 through the delegated definition");
+	assert.match(executed.text ?? "", /hi/);
+	assert.equal(
+		executed.hasOutputSchema,
+		true,
+		"the bash override must keep the upstream outputSchema (structuredContent for codemode)",
+	);
+	assert.ok(
+		executed.registeredDescription?.includes("Not for searching files"),
+		"the catalog descriptionAppend must reach the registered bash description",
+	);
 });

tests/system-prompt.test.ts

@@ -21,6 +21,7 @@ import {
 	stripJsoncComments,
 	validateRoutingConfig,
 	validateSectionsConfig,
+	validateToolCatalog,
 	type MutablePromptOptions,
 	type RoutingRule,
 } from "../extensions/system-prompt-config.ts";
@@ -28,7 +29,7 @@ import systemPromptExtension, {
 	createSystemPromptComposer,
 	getSystemPromptDumpPath,
 } from "../extensions/system-prompt.ts";
-import { guardExtensionApi } from "../extensions/prompt-guard.ts";
+import { catalogDescription, guardExtensionApi } from "../extensions/prompt-guard.ts";
 import { BUNDLE_GROUP_DEFINITIONS } from "../pi-tool-search/extensions/bundle-groups.ts";
 import { BUNDLE_TOOL_SEARCH_DEFAULTS } from "../pi-tool-search/extensions/config.ts";
 
@@ -93,6 +94,20 @@ test("routing scope validation: main/subagent accepted, other values rejected",
 	);
 });
 
+test("tools.jsonc validation: descriptionAppend accepted as a string, rejected otherwise", () => {
+	assert.doesNotThrow(() =>
+		validateToolCatalog({ version: 1, tools: { bash: { descriptionAppend: "extra" } } }),
+	);
+	assert.throws(
+		() => validateToolCatalog({ version: 1, tools: { bash: { descriptionAppend: 5 } } }),
+		/descriptionAppend must be a string/,
+	);
+	assert.throws(
+		() => validateToolCatalog({ version: 1, tools: { bash: { descriptionAppend: true } } }),
+		/descriptionAppend must be a string/,
+	);
+});
+
 test("sections.jsonc validation: subagentExclude defaults to [] and rejects bad shapes", () => {
 	assert.deepEqual(validateSectionsConfig({ version: 1, order: ["tool_routing"] }), {
 		version: 1,
@@ -255,7 +270,7 @@ test("bundled routing rules select the same guidance the retired tool-routing ex
 	const routing = loadRoutingRules();
 
 	const literal = renderRouting(routing.rules, ["find", "grep", "read"]);
-	assert.match(literal, /Use find to identify likely files or directories/);
+	assert.match(literal, /Pick the local search tool by what you already know/);
 	assert.match(literal, /narrow the path, glob, or pattern instead of increasing the limit/);
 
 	const ssh = renderRouting(routing.rules, ["ssh_connect", "ssh_cd", "ssh_bash"]);
@@ -267,7 +282,7 @@ test("bundled routing rules select the same guidance the retired tool-routing ex
 
 	const minimal = renderRouting(routing.rules, ["read"]);
 	assert.doesNotMatch(minimal, /codegraph_explore first/);
-	assert.doesNotMatch(minimal, /For literal search/);
+	assert.doesNotMatch(minimal, /Pick the local search tool/);
 	assert.doesNotMatch(minimal, /ssh_connect only when/);
 	assert.match(minimal, /Stop searching once the current evidence is sufficient/);
 });
@@ -397,7 +412,6 @@ test("composer splices the preamble, applies the catalog, and orders sections",
 			read: "reads files",
 			bash: "runs commands",
 			edit: "edits",
-			codegraph_explore: "explores",
 			tool_search: "searches",
 		},
 		toolGuidelines: { edit: ["upstream edit rule"] },
@@ -414,6 +428,11 @@ test("composer splices the preamble, applies the catalog, and orders sections",
 		"the bundle preamble replaces Pi's default preamble",
 	);
 	assert.match(output, /<tool_routing>\n- Use codegraph_explore first/);
+	assert.match(
+		output,
+		/<tools>[\s\S]*?- codegraph_explore: Explore code structure from the CodeGraph index/,
+		"tools.jsonc must add a snippet for codegraph_explore, which has no upstream <tools> entry",
+	);
 	assert.match(output, /search files with the find\/grep tools, not bash/);
 	assert.match(output, /LINE#HASH anchors/);
 	assert.ok(
@@ -604,7 +623,38 @@ test("prompt guard applies catalog description overrides at registration", () =>
 	const override = loadToolCatalog().tools.edit;
 	assert.ok(override && !("inherit" in override) && typeof override.description === "string");
 	assert.equal(tools[0].description, override.description);
-	assert.equal(tools[1].description, "upstream grep", "inherit entries keep the upstream description");
+	assert.equal(tools[1].description, "upstream grep", "entries without description fields keep the upstream description");
+});
+
+test("prompt guard appends the catalog descriptionAppend to the upstream description", () => {
+	const { pi, tools } = makeRecordingPi();
+	const guarded = guardExtensionApi(pi as unknown as ExtensionAPI);
+
+	guarded.registerTool({ name: "bash", description: "upstream bash" } as never);
+	guarded.registerTool({ name: "read", description: "upstream read" } as never);
+
+	const bashEntry = loadToolCatalog().tools.bash;
+	assert.ok(bashEntry && !("inherit" in bashEntry) && typeof bashEntry.descriptionAppend === "string");
+	assert.equal(tools[0].description, `upstream bash\n\n${bashEntry.descriptionAppend}`);
+	assert.equal(tools[1].description, "upstream read", "inherit entries keep the upstream description");
+});
+
+test("catalogDescription composes override, append, and inherit", () => {
+	assert.equal(
+		catalogDescription("upstream", { description: "override", descriptionAppend: "append" }),
+		"override\n\nappend",
+		"an append sits on top of the description override when both are set",
+	);
+	assert.equal(
+		catalogDescription("upstream  \n", { descriptionAppend: "append" }),
+		"upstream\n\nappend",
+		"trailing whitespace on the base is trimmed before the blank line",
+	);
+	assert.equal(
+		catalogDescription("upstream", { inherit: true }),
+		"upstream",
+		"inherit entries are untouched",
+	);
 });
 
 test("prompt guard strips a returned systemPrompt but keeps other fields", async () => {