[yw0nam]

Summary

uv run python -m memory_base.eval.sandbox serves the real REST API on a throwaway Postgres and prints its URL, a key, the author, and an example curl. A PR's runtime evidence then comes from real HTTP requests without touching the deployment and without a temporary test file.

main(wait)
  install _stop for SIGINT and SIGTERM          (first signal: ignore both, raise KeyboardInterrupt)
  with throwaway_postgres() as database:        (eval/longmemeval.py, db.Dockerfile on tmpfs)
      key = asyncio.run(_prepare(url))          (schema; keys.new_key("sandbox"); close_pool in finally)
      with EvalApi() as api:                    (eval/mcp_writer.py, uvicorn on a free loopback port)
          print URL / KEY / AUTHOR / curl … /notes
          wait()
  KeyboardInterrupt → "sandbox stopped", exit 0; previous handlers restored

The API subprocess inherits the environment: DB_URL points at the throwaway DB, and the embedder, reranker and chat settings come from .env. EvalApi.__enter__ now cleans up (terminates the subprocess, removes its temp dir) on any exception during boot, including a failed Popen and Ctrl-C; its timeout RuntimeError is unchanged.

Owner decisions: reuse the eval parts (throwaway_postgres, EvalApi, keys.new_key) rather than a new bash/compose path; the API uses the real .env services.

Review notes (GPT spec review ×3, two-axis review, a fix round, final two-axis review):

• Applied: both signals installed explicitly (a background start inherits SIGINT as ignored), a second stop signal cannot abort cleanup, argv dropped, the README line moved into the Development code block.
• Not changed: the boot sequence repeats mcp_writer.run's three lines rather than gaining a shared helper; _prepare_schema stays private.
• Known limit: a stop during throwaway_postgres's own docker run can leave a container in the Created state (--rm never applies to a container that did not start). Remove it with docker rm -f $(docker ps -aq --filter label=memory-base-longmemeval).

Related issues

None.

Runtime evidence (required for ingest / retrieval / serving behavior)

Background starts from the worktree (implementer and spec reviewer, independently):

run	/notes with key	without key	exit	stderr	labelled containers left
kill -INT	200	401	0	sandbox stopped	none
kill -TERM	200	401	0	sandbox stopped	none
INT twice, 1 s apart	200	401	0	sandbox stopped	none
TERM twice	—	—	0	sandbox stopped	none
INT during docker build	—	—	0	sandbox stopped	none
INT during docker run	—	—	0	sandbox stopped	one Created container (known limit above)

Printed by a run (the key lives only in that throwaway DB):

URL http://127.0.0.1:33799
KEY xuNyDjwvDj2CUm-8Gxv40zUv1G_D7J0al0W1sSWRUuU
AUTHOR claude-code
curl -s -H 'X-API-Key: xuNy…' -H 'X-Author: claude-code' http://127.0.0.1:33799/notes   → []

RED before each change: ImportError: cannot import name 'sandbox', assert [] == [True] (EvalApi boot interrupt); then AttributeError: … no attribute '_stop', assert not True (temp dir left after a failed Popen).

Verification

• uv run pytest passes — tests/eval unit: 124 passed
• uv run ruff format --check . and uv run ruff check . pass
• docker compose build mcp passes (Dockerfile / dependency changes) — N/A
• Runtime evidence attached above, or N/A justified

Checklist (core principles)

• Selective storage — notes stay distilled; verbatim text lives only in doc_rows and the turns lane
• Source-specific logic stays out of the source-agnostic core (ingest pipeline, retrieval, serving)
• No hardcoded endpoints/models — configuration lives in .env
• New/changed behavior ships a test in this PR (or skip-tests label justified)