Capture observations & summaries from Anthropic's Claude Code IDE extension (the surface that runs inside Cursor / VSCode) — distinct from the existing cursor platform

[github-actions[bot]]

Originally posted as issue #2160 by @zjl578254155
#2160

TL;DR

claude-mem currently treats two integration platforms as first-class:

• claude-code → the terminal CLI (npx @anthropic-ai/claude-code).
• cursor → Cursor's own native AI agent (the chat panel built into Cursor itself), wired up via bun run cursor:install, beforeSubmitPrompt, and .cursor/rules/claude-mem-context.mdc (see #838, #633).

There is a third surface that is currently uncovered: Anthropic's official Claude Code IDE extension (the VSCode-style extension shipped by Anthropic, which also runs inside Cursor since Cursor is a VSCode fork). This is neither the terminal CLI nor Cursor's own AI agent. In a session running on this surface:

• The MCP server registered via .mcp.json loads correctly → mem-search works.
• None of the lifecycle hooks fire, so no observations and no summaries are recorded for the entire session.

This issue requests treating the IDE extension as a third platform (e.g. claude-code-ide) alongside the two already supported.

Acknowledgement of existing work

I want to be explicit that I know about the existing Cursor support before filing this:

• package.json exposes cursor:install, cursor:uninstall, cursor:status, cursor:setup.
• #838 (closed, label integration:cursor) fixed a normalizer bug where Cursor's beforeSubmitPrompt event sends conversation_id / generation_id instead of session_id.
• Search snippet from the docs/repo: "Relevant past context is automatically injected into Cursor sessions via the .cursor/rules/claude-mem-context.mdc file."

All of that targets Cursor's own AI agent (its chat panel, governed by .cursor/rules/ and Cursor's hook events like beforeSubmitPrompt). My report is about a different surface, described below.

The three surfaces, side by side

Surface	What it is	claude-mem status
Terminal CLI	npx @anthropic-ai/claude-code in a terminal	✅ Fully supported via claude-code platform; PascalCase hooks (SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd) all fire.
Cursor's native AI agent	Cursor's built-in chat panel (Cursor's own product)	✅ Supported via cursor:install, beforeSubmitPrompt, .cursor/rules/claude-mem-context.mdc.
Anthropic's Claude Code IDE extension	The VSCode-style extension Anthropic ships, which runs in VSCode and also in Cursor (Cursor is a VSCode fork). Distinct UI from Cursor's chat panel.	❌ Not covered. MCP loads, but no hooks fire.

The third row is what this issue asks about.

Environment

• OS: macOS 24.3.0 (darwin, Apple Silicon)

• Anthropic Claude Code: npx @anthropic-ai/claude-code (latest at time of report)

• Host IDE: Cursor (latest), running Anthropic's official Claude Code IDE extension

• claude-mem versions present in ~/.claude/plugins/cache/thedotmack/claude-mem/: 12.3.8, 12.3.9, 12.4.7

• ~/.claude-mem/settings.json → worker port 37701, host 127.0.0.1, all defaults otherwise.

• ~/.claude/settings.json (relevant excerpt):

  "enabledPlugins": {
    "claude-mem@thedotmack": true,
    "superpowers@claude-plugins-official": true
  }

Reproduction

1. Verify the worker is healthy (in my case it was: curl http://127.0.0.1:37701/health → {"status":"ok","activeSessions":9,…}).
2. Terminal session — run claude in a terminal, do a few Read / Bash calls, exit. → Log shows the full chain: [HOOK] → PostToolUse: …, [QUEUE] ENQUEUED type=observation, [HOOK] → Stop: Requesting summary, [DB] STORED summaryId=…. ✅
3. Cursor session via the Claude Code IDE extension — open Cursor, open the Claude Code extension's chat panel (NOT Cursor's own chat panel), do a few Read / Bash calls, end the conversation. → Log shows only:

   [SYSTEM] Claude-mem search server started
   [SYSTEM] Worker available
   [SYSTEM] MCP server shutting down {reason=stdio-closed}
   

   No [HOOK], no [QUEUE], no [DB], no observation, no summary for this session.

Empirical evidence from my machine

Excerpt from ~/.claude-mem/logs/claude-mem-2026-04-27.log (local time; current wall-clock is 20:00):

[2026-04-27 17:29:57.690] [INFO ] [HOOK  ] → Stop: Requesting summary {hasLastAssistantMessage=true}
[2026-04-27 17:30:25.164] [INFO ] [SDK   ] [session-121] ← Response received  <summary>…</summary>
[2026-04-27 17:32:58.290] [INFO ] [WORKER] Broadcasting processing status …
                                  ← end of last terminal session
[2026-04-27 19:56:34.958] [INFO ] [SYSTEM] MCP server shutting down {reason=stdio-closed}
[2026-04-27 19:56:37.718] [INFO ] [SYSTEM] Claude-mem search server started
[2026-04-27 19:56:37.733] [INFO ] [SYSTEM] Worker available {}
[2026-04-27 19:56:55.809] [INFO ] [SYSTEM] Claude-mem search server started
[2026-04-27 19:56:57.737] [INFO ] [SYSTEM] Claude-mem search server started
                                  ← Cursor (IDE extension) session(s) from here on
                                  ← Many tool calls happened in this window. Zero [HOOK] entries.

Between 19:56 and 20:00 I executed many Read / Bash / WebFetch calls inside the Claude Code IDE extension running in Cursor. Not a single [HOOK], [QUEUE], or [DB] entry was produced for them. The Claude-mem search server started lines confirm .mcp.json is being honored — it's specifically the hook layer that's silent on this surface.

Why I think this is happening (best-effort root cause)

Anthropic's Claude Code IDE extension appears to use a different hook event vocabulary than the terminal CLI — specifically camelCase event names and a smaller event set. Strong precedent that this is the case:

superpowers@claude-plugins-official/5.0.7 ships two hook files side by side:

File	Event keys
hooks/hooks.json	SessionStart (PascalCase — terminal CLI form)
hooks/hooks-cursor.json	sessionStart (camelCase — IDE-extension form)

So superpowers already encodes the assumption that the IDE-side surface needs camelCase event names, and it ships an explicit hooks-cursor.json for that case. (The naming "cursor" there is because Cursor is the most common IDE host that surfaced the gap; the same file applies whenever the Claude Code IDE extension runs, including in plain VSCode.)

claude-mem's hooks/hooks.json registers only PascalCase events:
SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SessionEnd. Nothing matches what the IDE extension fires, so nothing fires.

Structural parallel to #1270

This is the same shape as #1270 ("Codex CLI: read now via MCP, write when hooks land"):

• The IDE extension surface can already read from claude-mem today via the MCP server (mem-search works in Cursor's Claude Code panel — I've verified).
• The IDE extension surface cannot currently write observations/summaries because its hook events don't match what hooks.json registers.

The key difference vs Codex: the Claude Code IDE extension does fire hooks — they just have different event names (camelCase) than the terminal CLI does, evidenced by superpowers/hooks-cursor.json.

Suggested fix

Add a third platform alongside the existing claude-code and cursor integrations — provisional name claude-code-ide. Concretely:

1. Ship a sibling hook file (e.g. hooks/hooks-claude-code-ide.json, mirroring superpowers/hooks-cursor.json) using the camelCase event names the IDE extension fires.
2. Re-use the existing worker-service.cjs hook claude-code <event> worker entrypoints — only the trigger names change, not the worker contract.
3. Optionally add claude-code-ide:install / :status scripts in package.json parallel to the cursor:* ones, so users can install just this surface explicitly.

I deliberately do not suggest adding camelCase keys directly to the existing terminal hooks.json, because that conflates the IDE-extension surface with the terminal CLI surface. The cleaner architectural move is a new platform entry, matching how cursor is its own thing today.

If only a subset of camelCase hook events is supported by the IDE extension at the time of writing (e.g. sessionStart and userPromptSubmit for sure, postToolUse / stop / sessionEnd uncertain), even partial support — say sessionStart + userPromptSubmit + postToolUse — would unlock observation capture for IDE-extension sessions and would already be a huge win.

Open questions for the maintainer

1. Is there an authoritative reference (Anthropic docs, your own notes) for the full set of hook events Anthropic's Claude Code IDE extension fires, and the exact event names? superpowers only exercises sessionStart, so I can't infer the full set from there alone.
2. Do you agree the IDE-extension surface should be a third named platform (claude-code-ide or similar), or would you rather express it as a sub-mode of the existing claude-code platform with a --surface=ide style switch?
3. Would a PR adding a hooks-claude-code-ide.json (initially with whatever events I can confirm work) plus a claude-code-ide:install script be welcome?

Workarounds I'm using in the meantime

• Sessions I want recorded → claude in a terminal.
• Cursor's Claude Code IDE extension → only used for casual / throwaway prompts, on the assumption nothing will be remembered.

Happy to attach the full claude-mem-2026-04-27.log, run experiments to enumerate which camelCase event names the IDE extension actually fires, or test draft hook files. Thanks for the great plugin and for the work you've already done on the Cursor-native side.