stop hook prematurely calls /agentmemory/session/end, causing sessions to be marked as ended before user actually exits #745

Bug Description

The agentStop hook (plugin/scripts/stop.mjs / src/hooks/stop.ts) sends a request to /agentmemory/session/end every time the agent pauses after completing a response. In Claude Code CLI, this causes agentmemory to mark the session as completed while the user is still actively in the conversation — the session should only end when the user actually exits the CLI (which is the sessionEnd hook's responsibility).

Impact

• Context loss / session fragmentation: agentmemory believes the session is over and treats subsequent user messages as a new session.
• Downstream pipeline breakage: session-end.mjs is supposed to trigger slot-reflect, graph-extract, and consolidation once at true session exit. Because stop.mjs already ended the session, these lifecycle events either fire prematurely or against an already-completed session.
• Log noise: The old triggerVoid failed warning (pre-fix: replace non-existent sdk.triggerVoid with sdk.trigger + action: { type: 'void' } #731) was partly caused by stop.mjs repeatedly firing session/end and triggering event::session::stopped in a partially-teardown state.

Root Cause

In src/hooks/stop.ts, the agentStop event handler performs two requests:

fetch(`${REST_URL}/agentmemory/summarize`, ...)   // ✅ reasonable: summarize the turn
fetch(`${REST_URL}/agentmemory/session/end`, ...) // ❌ breaks Claude Code

The plugin's hook configuration (plugin/hooks/hooks.copilot.json) clearly separates the two events:

Calling session/end in agentStop conflates "agent paused" with "session ended" for Claude Code.

Historical Context

The /agentmemory/session/end call was intentionally added in commit 355124141625ccc0d740ae08ddaaf77fe2c165ae (#579, fixes #493) as a workaround for Codex, which does not fire a separate SessionEnd event. The commit message explicitly states:

"Codex does not fire a separate SessionEnd event, so its Stop hook is the only signal we get when a Codex session terminates."

"For Claude Code this is a harmless idempotent second call (session-end.mjs runs on the dedicated SessionEnd hook and sets the same endedAt + status fields). For Codex it's the only path that closes the lifecycle."

However, the assumption that this is "harmless idempotent" for Claude Code is incorrect in practice. Claude Code CLI fires agentStop after every assistant turn (not just at session exit), so the session gets repeatedly marked as completed mid-conversation.

Reproduction Steps

1. Start Claude Code CLI with the agentmemory plugin enabled.
2. Send any message to the agent and wait for it to complete its response.
3. Do not exit Claude Code — just let the agent finish and return to the prompt.
4. Observe the agentmemory service logs:

   [agentmemory] warn event::session::stopped trigger failed
   

   (or pre-fix: replace non-existent sdk.triggerVoid with sdk.trigger + action: { type: 'void' } #731: triggerVoid failed)
5. Check the session status in agentmemory state — it will show status: "completed" and endedAt set, even though the CLI is still open.

Expected Behavior

• Claude Code: stop.mjs (agentStop) should only call /agentmemory/summarize (to capture the turn's context) and must not call /agentmemory/session/end. session-end.mjs (sessionEnd) is the sole caller.
• Codex: Still needs a path to close the session lifecycle, since Codex lacks SessionEnd.

Suggested Fix

The fix needs to preserve Codex support while fixing Claude Code.

Option A: Agent-type guard (recommended)

Only call session/end in stop.mjs when running under Codex (or more generally, when no SessionEnd hook is available). The hook payload may contain an entrypoint or agent-type field that can be used to distinguish Codex from Claude Code.

// In stop.ts — only end session for agents that lack a dedicated SessionEnd hook
const isCodex = data.entrypoint === "codex" || /* other Codex identifiers */;
if (isCodex) {
  fetch(`${REST_URL}/agentmemory/session/end`, {
    method: "POST",
    headers: authHeaders(),
    body: JSON.stringify({ sessionId }),
    signal: AbortSignal.timeout(5000),
  }).catch(() => {});
}

Option B: Move Codex session-end to a Codex-specific hook/config

If the plugin system supports per-agent hook mappings, map Codex's agentStop to a separate script that includes session/end, while Claude Code continues to use the current stop.mjs without it.

Option C: Remove session/end from stop.mjs and accept Codex regression

Remove the call entirely. This fixes Claude Code but regresses Codex session lifecycle (sessions will show as active in the viewer until Codex adds a SessionEnd event).

Affected Files

• src/hooks/stop.ts (source)
• plugin/scripts/stop.mjs (build artifact, auto-generated)

Related

• Commit 355124141625ccc0d740ae08ddaaf77fe2c165ae — intentionally added session/end to stop.mjs as a Codex workaround (fix(hooks): stop also closes session for Codex (closes #493) #579, fixes Codex Stop hook does not mark sessions ended #493).
• v.0.9.24 sdk.triggerVoid is not a function #726 / fix: replace non-existent sdk.triggerVoid with sdk.trigger + action: { type: 'void' } #731 — the triggerVoid migration surfaced this because stop.mjs repeatedly triggered session/end, amplifying the triggerVoid is not a function noise.