Runtime: MCP servers not fully loaded on session start and session unrecoverable after interrupt

[icoe44]

Summary

When using the Codepilot Runtime (codepilot_runtime), two distinct issues are observed:

1. MCP server loading is incomplete: Only a subset of configured MCP servers are synced to the agent loop. The third-party MCP server (codepilot-image-gen) is consistently missing from the tool list, while only playwright and pencil are loaded.

2. Session becomes unrecoverable after interrupt: Interrupting the agent mid-response causes the session to enter a permanently broken state, with Tool result is missing errors persisting across all subsequent turns.

---

Environment

• Codepilot version: 0.64.0 (upgraded from 0.63.0 on this session start)
• OS: Windows 11 Pro 10.0.22621
• Runtime mode: codepilot_runtime (pinned in session)
• Project path: C:\Users\Administrator\codepilot

---

Issue 1: MCP servers not fully loaded

Observed behavior

Every agent loop turn logs the same line:

[agent-loop] Syncing 2 MCP servers: playwright, pencil

Only 2 MCP servers are synced. The tool list for the agent contains 58 tools total, consisting of built-in tools + Playwright MCP tools + Pencil MCP tools. No tools from the codepilot-image-gen MCP server are present.

Evidence

Log excerpt (repeated across all 7 turns in the session):

[agent-loop] Syncing 2 MCP servers: playwright, pencil
[agent-loop] Session 5857d9d0e7910ef2d8eebd4286c78124: model=deepseek-v4-flash, tools=[Read, Write, Edit, Bash, Glob, Grep, Skill, Agent, codepilot_dashboard_pin, ..., mcp__playwright__*, mcp__pencil__*] (58 total)

The codepilot-image-gen MCP tools are consistently absent from all 58-tool listings across all turns (lines 96, 120, 137, 181, 203, 222, 247, 272, 299, 317 of the log).

Related error

Every turn also logs:

[harness-home] Canonical projection failed { runtimeId: 'codepilot_runtime', reason: 'r is not a function' }

This error occurs on every single request (7 occurrences in the session log). It precedes the tool listing and may be related to the MCP loading failure.

Comparison: Claude Code Runtime works correctly

Under the same working directory, same MCP configuration, and same machine, using the Claude Code CLI Runtime:

• codepilot-image-gen MCP server is loaded automatically on session start
• Agent can directly invoke codepilot_generate_image without any searching or manual configuration
• MCP log confirms successful tool execution: Tool 'codepilot_generate_image' completed successfully in 34s

---

Issue 2: Session unrecoverable after interrupt

Observed behavior

When the user interrupts the agent mid-response, the session enters a state where all subsequent turns fail with:

[agent-loop] streamText error: Tool result is missing for tool call call_7718bc899ed247aea103f200.
[agent-loop] Error: No output generated. Check the stream for errors.

Evidence

Interrupt log:

[codex.runtime] interrupt (route) — no active turn for 5857d9d0e7910ef2d8eebd4286c78124
[interrupt] fan-out { sessionId: '5857d9d0...', attempted: { native: true, codex_runtime: true, sdk: false } }

After the interrupt, the user sent 6 more messages. All of them failed with the same Tool result is missing error. Even switching the model (from deepseek-v4-flash to doubao-seed-2.1-turbo to claude-sonnet-5) and provider did not recover the session. The error persisted across model changes, and the sdk_session_id was repeatedly cleared and reset.

The session eventually became so broken that the user had to open a new session (478f123a...), but even that new session showed the same MCP loading issue (only 2 MCP servers synced).

---

Root cause analysis (speculative)

The recurring [harness-home] Canonical projection failed { runtimeId: 'codepilot_runtime', reason: 'r is not a function' } error suggests:

1. The harness-home module (likely responsible for resolving the home directory / config path for the agent harness) is failing during canonical projection — a path normalization operation
2. The error 'r is not a function' indicates a JavaScript runtime error, possibly a function reference that was expected but not found (e.g., a callback, a method on a config object, or a module export)
3. This failure may prevent the MCP configuration from being fully discovered and loaded, since MCP server definitions typically reside in the home directory config

---

Suggested investigation areas

1. harness-home module: Examine the canonical projection function for the runtimeId: 'codepilot_runtime' code path. The 'r is not a function' error suggests a variable r is expected to be a function but is undefined or of a different type.

2. MCP server discovery: Compare how the Codepilot Runtime discovers MCP servers vs. the Claude Code Runtime. The Codepilot Runtime may be using a different config resolution strategy that doesn't pick up all MCP servers.

3. Interrupt recovery: The interrupt handler attempts recovery via native and codex_runtime but neither succeeds. The Tool result is missing error suggests the interrupt left the tool call state machine in an inconsistent state.

4. MCP server definition location: The codepilot-image-gen MCP server is only referenced in the project's settings.local.json permissions list. Its mcpServers definition may be in a non-standard location that the Codepilot Runtime doesn't check.

---

Configuration files

Project settings.local.json:

{
  "permissions": {
    "allow": [
      "mcp__codepilot-image-gen__codepilot_generate_image"
    ]
  }
}

Global ~/.claude/settings.json: Contains env vars, plugin config, theme, and permissions — but no mcpServers section relevant to codepilot-image-gen.

---

Log reference

• Session ID: 5857d9d0e7910ef2d8eebd4286c78124 (broken session)
• Log file: %APPDATA%\codepilot\logs\codepilot-main.log
• MCP log (working): %LOCALAPPDATA%\claude-cli-nodejs\Cache\C--Users-Administrator-codepilot\mcp-logs-codepilot-image-gen\2026-08-03T05-34-03-041Z.jsonl

[github-actions]

感谢反馈！这个 issue 似乎缺少必要的复现信息，已标记 needs-repro。

请用 issue 模板、在最新版 v0.56.2+ 上补齐：版本号、操作系统、Runtime、Provider / 模型、复现步骤、日志 / session id（请脱敏，不要包含 API key）。

直接在本 issue 下回复补充即可 —— 补全后会自动取消 needs-repro 标记、进入正常排队；长期无回应会按 stale 流程提醒、最终关闭（关闭后随时可重新打开）。

[icoe44]

补充具体日志证据

从日志中确认了两个问题：

问题 1：新会话 MCP 未加载

Codepilot Runtime 日志 显示每次只同步 2 个 MCP 服务器：

[agent-loop] Syncing 2 MCP servers: playwright, pencil

codepilot-image-gen 这个 MCP 服务器从未被列入同步列表，导致 agent 工具列表中完全缺失 codepilot_generate_image 工具（工具列表共 58 个，不含 image-gen 相关工具）。

对比：Claude Code 官方 Runtime 中，codepilot-image-gen MCP 正常加载，codepilot_generate_image 工具可用（MCP 日志显示调用成功，耗时 34s）。

问题 2：打断会话必炸

每次会话启动时都有报错：

[harness-home] Canonical projection failed { runtimeId: 'codepilot_runtime', reason: 'r is not a function' }

这个错误在所有请求轮次中持续出现（第 1、2、3 次对话都有此错误）。打断时日志显示：

[codex.runtime] interrupt (route) — no active turn for 5857d9d0...
[interrupt] fan-out { sessionId: '5857d9d0...', attempted: { native: true, codex_runtime: true, sdk: false } }

中断后会话无法恢复，只能开新会话。

总结

• [harness-home] Canonical projection failed 可能是核心根因 —— 这个错误导致 MCP 服务器列表加载不完整，也可能是会话状态损坏的元凶
• 'r is not a function' 指向某个 JS 运行时错误，可能是 harness-home 模块的 bug
• 版本：Codepilot v0.64.0（从日志中 Version changed from 0.63.0 to 0.64.0 确认）

[icoe44]

问题 3：对比验证 — Claude Code 官方 Runtime 完全正常

在相同的机器、相同的 MCP 配置下，使用 Claude Code 官方 Runtime 做对比测试：

1. MCP 自动加载：新会话启动后，直接说"帮我生成一张图"，agent 立即调用 codepilot_generate_image 工具，不需要任何搜索或翻配置
2. 打断恢复：多次打断 agent 响应，会话始终稳定，打断后能正常继续对话
3. 生图成功：首次调用即成功，耗时约 34s

这说明：

• MCP 配置本身没有问题（settings.local.json 中的权限配置正确）
• 问题出在 Codepilot Runtime 的 MCP 加载逻辑上，可能和 harness-home 模块的 'r is not a function' 错误有关
• 打断会话崩溃也是 Codepilot Runtime 特有的问题，官方 Runtime 无此问题

[github-actions]

感谢反馈！这个 issue 似乎缺少必要的复现信息，已标记 needs-repro。

请用 issue 模板、在最新版 v0.56.2+ 上补齐：版本号、操作系统、Runtime、Provider / 模型、复现步骤、日志 / session id（请脱敏，不要包含 API key）。

直接在本 issue 下回复补充即可 —— 补全后会自动取消 needs-repro 标记、进入正常排队；长期无回应会按 stale 流程提醒、最终关闭（关闭后随时可重新打开）。

[icoe44]

补充复现信息（按 issue template 要求）：

版本号: v0.64.0（从日志 Version changed from 0.63.0 to 0.64.0 确认）
操作系统: Windows 11 Pro 10.0.22621
Runtime: codepilot_runtime（session pin）
Provider / 模型: 先后测试了 deepseek-v4-flash、doubao-seed-2.1-turbo、claude-sonnet-5，均复现
Session ID: 5857d9d0e7910ef2d8eebd4286c78124（已脱敏，不含 API key）

复现步骤（MCP 未加载）:

1. 打开 Codepilot 桌面客户端，使用 codepilot_runtime
2. 开始新会话
3. 输入"帮我生成一张图"
4. 观察 Agent 行为：无法调用生图工具，开始搜索本地配置

复现步骤（打断崩溃）:

1. 在 Agent 响应过程中，按 ESC 或点击停止
2. 尝试继续对话
3. 会话返回错误 Tool result is missing for tool call
4. 后续所有消息均失败，即使切换模型/Provider 也无法恢复

日志文件路径: %APPDATA%\codepilot\logs\codepilot-main.log