fix: call sessions end endpoint from session-stop hook

[jmeiracorbal]

Summary

• Implement session-stop.sh to call POST /sessions/{id}/end when Claude Code stops, so ended_at is always set regardless of agent cooperation

Motivation

Closes #46

session-stop.sh was a placeholder that did nothing. The server already exposed POST /sessions/{id}/end, and the Stop hook already received session_id via stdin, but the connection was never made. As a result, every session in the database has ended_at = NULL and sessions grow unbounded with no duration data.

The fix mirrors the pattern already used in session-start.sh: read session_id from stdin, call the HTTP API.

Changes

• plugin/claude-code/scripts/session-stop.sh — replaced placeholder with a curl call to POST /sessions/{id}/end. Reads session_id from stdin (same as session-start.sh), exits cleanly if no session_id is present. Runs async (already configured in hooks.json) so it never blocks Claude's response.

Test plan

• go test ./... passes
• Manually tested end-to-end:
  1. Created a session via POST /sessions
  2. Confirmed ended_at = NULL in the database
  3. Simulated the hook: echo '{"session_id":"test","cwd":"/tmp"}' | bash session-stop.sh
  4. Confirmed ended_at was set correctly in the database

[Alan-TheGentleman]

LGTM

[Alan-TheGentleman]

Fix looks good — clean, mirrors session-start.sh, does exactly what it should.

One thing worth noting: if the server happens to be down when the session ends, this fails silently (which is fine for a hook). But the agent also can't call mem_session_end as a fallback because that tool isn't in the ToolSearch select list from #43. Not a blocker for this PR — just something to address in #43.

Also flagging that this only covers Claude Code. OpenCode, Gemini, and Codex don't have session-end hooks so their sessions still get ended_at = NULL. We should open a follow-up for that — the MCP tool exists, agents just need to be told to call it.

Approve from my side, good to merge.