feat(recall): link an OMP subagent's session to its parent for adoption

docs/usage-guide.md

@@ -1361,7 +1361,7 @@ Recall counts every doc it returns (`recalled_count`). A returned doc is **adopt
 | Copilot CLI | Yes | No: the subagent has a session of its own, and no hook links it to its parent |
 | Cursor | Yes | No: as for Copilot CLI |
 | OpenCode | Yes | Yes: the `task` call links the subagent's session to its parent |
-| OMP | Yes, settled only by the claim of its `bash` call | No: the subagent's session is not linked to its parent |
+| OMP | Yes, settled only by the claim of its `bash` call | Yes: the subagent's session file sits under its parent's, whose session header links the two sessions (verified against OMP 18.4.8) |
 | Pi | Yes | None: TeamAI deploys no subagent to Pi |
 | ZCode | Yes | No: ZCode runs no hooks inside a subagent |
 | OpenClaw, Hermes, Kiro, JoyCode | No: no PostToolUse hook | No |
@@ -1371,7 +1371,7 @@ Recall counts every doc it returns (`recalled_count`). A returned doc is **adopt
 **Known limits.**
 
 - **Cursor, Copilot CLI and ZCode subagents.** A recall run in a subagent never credits the main agent's reads: Cursor and Copilot CLI give the subagent its own session with no link to the parent, and ZCode runs no hooks in it. A recall the main agent runs itself does vote.
-- **OMP.** The subagent's session is not linked to its parent, so the subagent path gives no adoption. OMP sets no session variable in its shell, so a run settles only through the claim of the `bash` call that ran it: when OMP moves a large output into an artifact, the `run=` line and the vote are lost, and an OMP started from a Claude Code shell records its run under the Claude session until that claim corrects it.
+- **OMP.** The subagent path needs the main session's file on disk: when it has none (`--no-session`), the subagent is not linked to its parent and gives no adoption. OMP sets no session variable in its shell, so a run settles only through the claim of the `bash` call that ran it: when OMP moves a large output into an artifact, the `run=` line and the vote are lost, and an OMP started from a Claude Code shell records its run under the Claude session until that claim corrects it.
 - **Searches not counted.** OMP's `grep` (a markdown tree) and Cursor's `Grep` add no evidence; opening the doc still counts. ZCode prints `Grep` lines relative to its working directory, so a ZCode search run from a directory inside the team repo is not counted.
 - **No PostToolUse hook.** OpenClaw, Hermes, Kiro and JoyCode record their recalls but no reads, so these recalls never vote.
 - **Older CLIs.** A member on an older TeamAI still votes from the session transcript, and that path keys a skill, a doc in a subdirectory or a wiki page by its file's basename (`SKILL`, `setup`) rather than the id recall prints (`retry`, `common/setup`), so those votes miss the doc. Top-level learnings and docs are unaffected, and upgrading ends it.
@@ -2158,7 +2158,7 @@ These paths are verified against the ZCode desktop app: profiles created in its
 
 ### Oh My Pi
 
-Oh My Pi (OMP) is available as a built-in target. TeamAI deploys skills, rules, and subagents to OMP's native directories — `.omp/skills/`, `.omp/rules/`, and `.omp/agents/` at project scope, and `~/.omp/agent/skills/`, `~/.omp/agent/rules/`, and `~/.omp/agent/agents/` at user scope (user-scope resources live under the agent directory `~/.omp/agent/`, a different prefix from the project one, so TeamAI switches prefixes with the scope). Instructions (`claudemd`) deploy to the matching `AGENTS.md`, and MCP servers merge into `~/.omp/agent/mcp.json` / `<project>/.omp/mcp.json` (Claude `mcpServers` shape — see the MCP section above). Skills are one-level `<name>/SKILL.md` bundles and TeamAI fills in a `description` on sync, which OMP's native skill provider requires to discover a skill. These paths follow OMP's documented discovery layout (verified against OMP 18.2.5). Hooks ride OMP's extension runner: `teamai pull` writes a single generated extension to `~/.omp/agent/extensions/teamai-hooks.ts` (never a project copy — OMP auto-loads both roots and would double-dispatch every event), which forwards OMP's `session_start` / `session_stop` / `before_agent_start` / `tool_result` events to the same `teamai hook-dispatch` entry point every other agent uses, gated on the session `cwd`. Every event carries the OMP session id (`ctx.sessionManager.getSessionId()`; a subagent has its own), and `tool_result` also the tool's text output and a status from `isError`, so upvote **adoption** runs for OMP's main agent: OMP sets no session variable in its shell, so a recall joins the session of the `bash` call that ran it, and a `read` with a line selector (`x.md:50-200`, `x.md:raw`) counts as a read of the file. From OMP 18.3.2 a subagent's events also carry its `ctx.agent` id and name, so the `teamai-recall` subagent's own reads never count. A doc the main agent opens after a subagent's recall is not upvoted yet: the subagent's session is not linked to its parent. The `session_stop` handler returns nothing, so a dispatch can never force a session continuation, and there is no matcher-scoped post-tool-use pass because OMP's tool ids are lowercase (`bash`, `read`, …) and it has no `Skill` / `TodoWrite` tool. `teamai uninstall` removes the extension. OMP profiles (`OMP_PROFILE` / `PI_CODING_AGENT_DIR` / `PI_CONFIG_DIR`), which relocate the agent directory, are not supported; the default `~/.omp/agent/` layout is used.
+Oh My Pi (OMP) is available as a built-in target. TeamAI deploys skills, rules, and subagents to OMP's native directories — `.omp/skills/`, `.omp/rules/`, and `.omp/agents/` at project scope, and `~/.omp/agent/skills/`, `~/.omp/agent/rules/`, and `~/.omp/agent/agents/` at user scope (user-scope resources live under the agent directory `~/.omp/agent/`, a different prefix from the project one, so TeamAI switches prefixes with the scope). Instructions (`claudemd`) deploy to the matching `AGENTS.md`, and MCP servers merge into `~/.omp/agent/mcp.json` / `<project>/.omp/mcp.json` (Claude `mcpServers` shape — see the MCP section above). Skills are one-level `<name>/SKILL.md` bundles and TeamAI fills in a `description` on sync, which OMP's native skill provider requires to discover a skill. These paths follow OMP's documented discovery layout (verified against OMP 18.2.5). Hooks ride OMP's extension runner: `teamai pull` writes a single generated extension to `~/.omp/agent/extensions/teamai-hooks.ts` (never a project copy — OMP auto-loads both roots and would double-dispatch every event), which forwards OMP's `session_start` / `session_stop` / `before_agent_start` / `tool_result` events to the same `teamai hook-dispatch` entry point every other agent uses, gated on the session `cwd`. Every event carries the OMP session id (`ctx.sessionManager.getSessionId()`; a subagent has its own), and `tool_result` also the tool's text output and a status from `isError`, so upvote **adoption** runs for OMP's main agent: OMP sets no session variable in its shell, so a recall joins the session of the `bash` call that ran it, and a `read` with a line selector (`x.md:50-200`, `x.md:raw`) counts as a read of the file. From OMP 18.3.2 a subagent's events also carry its `ctx.agent` id and name, so the `teamai-recall` subagent's own reads never count. A subagent's session file sits under its parent's, whose header names the parent session, so the extension links the two on the subagent's tool calls, and a doc the main agent opens after a subagent's recall is upvoted (verified against OMP 18.4.8). The `session_stop` handler returns nothing, so a dispatch can never force a session continuation, and there is no matcher-scoped post-tool-use pass because OMP's tool ids are lowercase (`bash`, `read`, …) and it has no `Skill` / `TodoWrite` tool. `teamai uninstall` removes the extension. OMP profiles (`OMP_PROFILE` / `PI_CODING_AGENT_DIR` / `PI_CONFIG_DIR`), which relocate the agent directory, are not supported; the default `~/.omp/agent/` layout is used.
 
 ### DeepSeek Harness
 

docs/usage-guide.zh-CN.md

@@ -1241,7 +1241,7 @@ recall 会为返回的每篇文档计数（`recalled_count`）。运行 recall 
 | Copilot CLI | 支持 | 不支持：subagent 有自己的会话，且没有 hook 将其关联到父会话 |
 | Cursor | 支持 | 不支持：同 Copilot CLI |
 | OpenCode | 支持 | 支持：`task` 调用将 subagent 的会话关联到父会话 |
-| OMP | 支持，仅通过其 `bash` 调用的认领确定归属 | 不支持：subagent 的会话未关联到父会话 |
+| OMP | 支持，仅通过其 `bash` 调用的认领确定归属 | 支持：subagent 的会话文件位于父会话文件之下，父会话文件的会话头把两个会话关联起来（对照 OMP 18.4.8 验证） |
 | Pi | 支持 | 不适用：TeamAI 不向 Pi 部署 subagent |
 | ZCode | 支持 | 不支持：ZCode 在 subagent 内不运行 hook |
 | OpenClaw、Hermes、Kiro、JoyCode | 不支持：没有 PostToolUse hook | 不支持 |
@@ -1251,7 +1251,7 @@ recall 会为返回的每篇文档计数（`recalled_count`）。运行 recall 
 **已知限制。**
 
 - **Cursor、Copilot CLI 和 ZCode 的 subagent。** 在 subagent 中运行的 recall 从不为主 agent 的读取计分：Cursor 和 Copilot CLI 给 subagent 分配独立会话且不关联父会话，ZCode 在 subagent 内不运行 hook。主 agent 自己运行的 recall 可以正常投票。
-- **OMP。** subagent 的会话未关联到父会话，因此 subagent 路径不产生采纳。OMP 不在 shell 中设置会话变量，因此 run 只能通过运行它的 `bash` 调用的认领确定归属：OMP 把大段输出转存为 artifact 时，`run=` 行和投票都会丢失；从 Claude Code shell 启动的 OMP 会先把 run 记在 Claude 会话下，直到该认领将其纠正。
+- **OMP。** subagent 路径要求主会话文件已写入磁盘：主会话没有会话文件时（`--no-session`），subagent 无法关联到父会话，也不产生采纳。OMP 不在 shell 中设置会话变量，因此 run 只能通过运行它的 `bash` 调用的认领确定归属：OMP 把大段输出转存为 artifact 时，`run=` 行和投票都会丢失；从 Claude Code shell 启动的 OMP 会先把 run 记在 Claude 会话下，直到该认领将其纠正。
 - **不计入的搜索。** OMP 的 `grep`（markdown 树形输出）和 Cursor 的 `Grep` 不产生证据；打开文档仍然计入。ZCode 打印的 `Grep` 行是相对其工作目录的路径，因此从团队仓库内的目录发起的 ZCode 搜索不计入。
 - **没有 PostToolUse hook。** OpenClaw、Hermes、Kiro 和 JoyCode 会记录其 recall，但不记录读取，因此这些 recall 从不投票。
 - **旧版 CLI。** 使用旧版 TeamAI 的成员仍从会话 transcript 投票，该路径以文件的 basename（`SKILL`、`setup`）而非 recall 打印的 id（`retry`、`common/setup`）作为 skill、子目录中的文档或 wiki 页面的键，因此这些投票落不到该文档上。顶层的 learnings 和文档不受影响，升级后即可解决。
@@ -2023,7 +2023,7 @@ ZCode 已作为内置目标支持。Skills 下发到 `.zcode/skills/`（ZCode 
 
 ### Oh My Pi
 
-Oh My Pi（OMP）已作为内置目标支持。TeamAI 将 Skills、Rules 和 Subagents 下发到 OMP 的原生目录——项目级为 `.omp/skills/`、`.omp/rules/` 和 `.omp/agents/`，用户级为 `~/.omp/agent/skills/`、`~/.omp/agent/rules/` 和 `~/.omp/agent/agents/`（用户级资源位于 agent 目录 `~/.omp/agent/` 下，与项目级前缀不同，TeamAI 会随作用域自动切换）。指令（`claudemd`）下发到对应的 `AGENTS.md`；MCP Server 合并进 `~/.omp/agent/mcp.json` / `<project>/.omp/mcp.json`（Claude `mcpServers` 结构，见上文 MCP 章节）。Skills 采用一层 `<name>/SKILL.md` 目录结构，TeamAI 在同步时补全 `description`——OMP 原生 skill 发现要求该字段。以上路径遵循 OMP 官方文档的发现布局（对照 OMP 18.2.5 验证）。Hooks 走 OMP 的 extension runner：`teamai pull` 会生成唯一的 extension 写入 `~/.omp/agent/extensions/teamai-hooks.ts`（绝不写项目副本——OMP 会同时加载两个根并导致每个事件双派发），它把 OMP 的 `session_start` / `session_stop` / `before_agent_start` / `tool_result` 事件转发给所有 agent 共用的 `teamai hook-dispatch` 入口，并按会话 `cwd` 做项目门控。每个事件都携带 OMP 会话 id（`ctx.sessionManager.getSessionId()`；subagent 有自己的会话），`tool_result` 还带上工具的文本输出和根据 `isError` 得出的状态，因此 upvote **采纳（adoption）**在 OMP 主 agent 上生效：OMP 不在其 shell 中设置会话变量，所以 recall 归入运行它的那次 `bash` 调用所在的会话；带行选择器的 `read`（`x.md:50-200`、`x.md:raw`）计为对该文件的读取。从 OMP 18.3.2 起，subagent 的事件还会携带其 `ctx.agent` 的 id 和名称，因此 `teamai-recall` subagent 自身的读取从不计入。主 agent 在 subagent recall 之后打开的文档暂不会被 upvote：subagent 的会话尚未关联到父会话。`session_stop` 处理器不返回任何值，分发绝不会强制会话继续；由于 OMP 的工具名是小写（`bash`、`read` 等）且没有 `Skill` / `TodoWrite` 工具，post-tool-use 不做 matcher 定向分发。`teamai uninstall` 会移除该 extension。OMP 的 profile（`OMP_PROFILE` / `PI_CODING_AGENT_DIR` / `PI_CONFIG_DIR`，会迁移 agent 目录）暂不支持，使用默认的 `~/.omp/agent/` 布局。
+Oh My Pi（OMP）已作为内置目标支持。TeamAI 将 Skills、Rules 和 Subagents 下发到 OMP 的原生目录——项目级为 `.omp/skills/`、`.omp/rules/` 和 `.omp/agents/`，用户级为 `~/.omp/agent/skills/`、`~/.omp/agent/rules/` 和 `~/.omp/agent/agents/`（用户级资源位于 agent 目录 `~/.omp/agent/` 下，与项目级前缀不同，TeamAI 会随作用域自动切换）。指令（`claudemd`）下发到对应的 `AGENTS.md`；MCP Server 合并进 `~/.omp/agent/mcp.json` / `<project>/.omp/mcp.json`（Claude `mcpServers` 结构，见上文 MCP 章节）。Skills 采用一层 `<name>/SKILL.md` 目录结构，TeamAI 在同步时补全 `description`——OMP 原生 skill 发现要求该字段。以上路径遵循 OMP 官方文档的发现布局（对照 OMP 18.2.5 验证）。Hooks 走 OMP 的 extension runner：`teamai pull` 会生成唯一的 extension 写入 `~/.omp/agent/extensions/teamai-hooks.ts`（绝不写项目副本——OMP 会同时加载两个根并导致每个事件双派发），它把 OMP 的 `session_start` / `session_stop` / `before_agent_start` / `tool_result` 事件转发给所有 agent 共用的 `teamai hook-dispatch` 入口，并按会话 `cwd` 做项目门控。每个事件都携带 OMP 会话 id（`ctx.sessionManager.getSessionId()`；subagent 有自己的会话），`tool_result` 还带上工具的文本输出和根据 `isError` 得出的状态，因此 upvote **采纳（adoption）**在 OMP 主 agent 上生效：OMP 不在其 shell 中设置会话变量，所以 recall 归入运行它的那次 `bash` 调用所在的会话；带行选择器的 `read`（`x.md:50-200`、`x.md:raw`）计为对该文件的读取。从 OMP 18.3.2 起，subagent 的事件还会携带其 `ctx.agent` 的 id 和名称，因此 `teamai-recall` subagent 自身的读取从不计入。subagent 的会话文件位于父会话文件之下，父会话文件的头部写明父会话 id，因此 extension 会在 subagent 的工具调用中关联这两个会话，主 agent 在 subagent recall 之后打开的文档会被 upvote（对照 OMP 18.4.8 验证）。`session_stop` 处理器不返回任何值，分发绝不会强制会话继续；由于 OMP 的工具名是小写（`bash`、`read` 等）且没有 `Skill` / `TodoWrite` 工具，post-tool-use 不做 matcher 定向分发。`teamai uninstall` 会移除该 extension。OMP 的 profile（`OMP_PROFILE` / `PI_CODING_AGENT_DIR` / `PI_CONFIG_DIR`，会迁移 agent 目录）暂不支持，使用默认的 `~/.omp/agent/` 布局。
 
 ### DeepSeek Harness
 

skill-data/core/references/troubleshooting.md

@@ -203,9 +203,11 @@ judge (`TEAMAI_UPVOTE_JUDGE=1`) credits that. `teamai stats` shows each recent
 session's runs, recalled docs and adopted docs. Per agent:
 
 - **Claude Code, Codex (0.134+), CodeBuddy (2.103.1+), WorkBuddy, Qoder,
-  OpenCode**: both a recall the main agent runs and one the `teamai-recall`
-  subagent runs are credited when the main agent opens the doc.
-- **Cursor, Copilot CLI, ZCode, OMP, Pi**: only a recall the main agent runs
+  OpenCode, OMP**: both a recall the main agent runs and one the
+  `teamai-recall` subagent runs are credited when the main agent opens the doc.
+  On OMP the subagent's recall needs the main session's file on disk, so a
+  `--no-session` run credits only the main agent's own recalls.
+- **Cursor, Copilot CLI, ZCode, Pi**: only a recall the main agent runs
   itself. A subagent's recall is not linked to the main session, and Pi has no
   TeamAI subagent.
 - **OpenClaw, Hermes, Kiro, JoyCode**: no PostToolUse hook, so recalls never

src/__tests__/helpers/pi-extensions.ts

@@ -1,3 +1,5 @@
+import fs from 'node:fs';
+import path from 'node:path';
 import vm from 'node:vm';
 import { EventEmitter } from 'node:events';
 import { buildPiExtensionSource } from '../../pi-hooks.js';
@@ -12,7 +14,7 @@ export interface ExtensionDispatch {
 /** The ExtensionContext fields the extensions read: Pi and OMP share `cwd` and `sessionManager`; OMP adds `agent`. */
 export interface ExtensionContext {
   cwd: string;
-  sessionManager?: { getSessionId: () => string };
+  sessionManager?: { getSessionId: () => string; getSessionFile?: () => string | undefined };
   agent?: { kind: 'main' | 'sub'; id: string; name: string; depth?: number; parentId?: string };
 }
 
@@ -65,6 +67,6 @@ export function loadOmpExtension(): LoadedExtension {
     })();
     return { quiet: () => ({ nothrow: () => run }) };
   };
-  const on = register(buildOmpExtensionSource(), { $, Response });
+  const on = register(buildOmpExtensionSource(), { $, Response, fs, path, Buffer });
   return { on, dispatches };
 }