interactive-story/plan/STATUS.md

Where things stand

Read this first when picking the project back up. Updated at the end of a working session; the per-phase plan files hold the detail, this holds the thread.

Last updated: 2026-08-17.

---

The live URL is not the one in render.yaml

Production is https://ai-dnd-1gmp.onrender.com. Render appended a suffix to the ai-dnd service name in render.yaml, and plain ai-dnd.onrender.com belongs to a different, suspended service that answers 503 with "suspended by its owner" — which is easy to mistake for this deploy being down. The authoritative link is the one the project page points at (docs/index.html), not the service name in the blueprint. GET /api/health on the real host returns {"ok":true}.

---

Needs a human

The free tier's storage ceiling is closer than the egress work suggested. The Neon database is 99.6 MB of a 512 MB allowance and actions.context_snapshot is essentially all of it. See "Storage, which this plan did not cost" in plan/13. This is now the most likely thing to break the deploy, ahead of anything on the read path.

---

Pick up here

plan/13-memory-embedding-cost.md, step 6 — infinite scroll upward in Play.jsx.

Opening a finished adventure is comfortably the largest single read in the app — a turn is now 122 kB, Insights 118 kB, the Memories drawer 24 kB. Measured on production (2026-08-17), the largest real adventure is 607 actions and 589.5 kB in one response; the 426.7 kB the harness reports is a 200-action fixture whose actions are about twice as heavy as real ones (994 B/action in production). So the fixture overstates width and understates length — real stories get longer than it models, which is the direction that hurts. The backend already has the windowing primitives (context/history.py: tail_range, slice_, count), and GET /adventures/{id}/actions exists. What is missing is a paged shape for it and a Play.jsx that loads the newest turns and fetches older ones as the reader scrolls up.

Watch for: the story is a flat list today, and the story tree replaces it (plan/14-phase-story-tree.md). Paging that reads by position from the end survives that change; paging that assumes Action.index is a dense 0..n sequence does not.

After that: the tree itself. Its design is settled in plan/14; nothing about it has been built.

---

What happened on 2026-08-16

Four commits, all on the egress work that has to land before the tree.

1. A byte meter, in the repo this time — 7ee5cee

backend/tools/dbmeter.py + backend/tools/stress_session.py.

cd backend
.venv/Scripts/python.exe -m tools.stress_session
.venv/Scripts/python.exe -m tools.stress_session --shapes turn --repeat 2
.venv/Scripts/python.exe -m tools.stress_session --no-embeddings   # the old blind spot

It counts bytes at the DBAPI cursor, not queries — both egress blowouts this project has had were one query fetching a column nobody read, and a statement count showed nothing wrong in either. It drives a production-shaped synthetic adventure through the real routes with only the LLM and the embedding endpoint faked.

The memory bank is ON by default and that is the whole point. The previous harness ran without an embedding model configured; embedding providers are BYOK-only by construction, so retrieve_memories returned early every time and the heaviest read in a turn never happened. --no-embeddings reproduces that deliberately — the gap is 29x.

It calibrates against the two figures measured directly on production: 426.7 kB for a 200-action page load against 423 KB, and 3,258.7 kB for one turn against 3,153 kB. It runs on SQLite, so treat absolutes as production-shaped and compare before/after.

2. Packed float32 embeddings — c568648

Migration 38 + app/vectors.py. A 1536-dimension vector as a JSON list is ~31 KB; the same numbers as float32 are 6,144 bytes. Not a precision trade — the endpoints compute in float32 and render that into JSON, so converting back is bit-exact. Nothing re-embeds, no API calls.

The backfill is the one in migrations.py that cannot be portable SQL, so it comes through Python, batched. Migration SQL can now be a {dialect: sql} map (BLOB vs BYTEA have no common spelling).

3. Ranking the bank without reading the bank — b7e53ae

retrieve_memories walked adventure.memories, loading every row with its vector. It now asks SQL which memories are in play (an id and a flag per row), ranks against vectors held in process, and fetches text only for the top-K it picks.

Two more callers were doing the same thing, and the production SQL could not see either: _evict_over_capacity walked the bank to count it, _embed_pending walked it to find rows with no vector. A played turn cost 6.4 MB, not the 3.2 the plan assumed.

shape	before	cold	warm
one turn	3,258.7 kB	723.4 kB	122.3 kB
run_post_turn	3,139.1 kB	0.7 kB	0.7 kB
Insights	3,223.7 kB	117.9 kB	117.9 kB
Memories drawer	~3.1 MB	23.7 kB	23.7 kB

A played turn is turn + post-turn: 6.4 MB → 123 kB, 52x.

Migrations 39/40 add memories.embedded, migration 41 drops the capacity default 200 → 80 for rows still on the old default.

---

What happened on 2026-08-17

No new behaviour — a verification pass on what shipped the day before, because every number in the section above had been measured on SQLite against a synthetic fixture. Full write-up in plan/13 under "Verified on production".

It holds. schema_version is 41 on the live Postgres with embedding_blob bytea and embedded boolean present, so migration 38's dialect map is correct against a real server. The backfill is complete (134/134). The packed vectors are 5.04x smaller than the JSON on real data — 30,971 → 6,144 bytes a memory, as predicted.

SQLite was not lying. tools.stress_session can now target Postgres via AIDND_STRESS_DATABASE_URL, and every shape agrees within 0.5% — the warm turn is 121.1 kB on Postgres against 122.3 kB on SQLite, with memories down to 1.7 kB of it. Run it against a throwaway database only; the harness writes, so it refuses any target whose name does not contain stress or scratch.

Two corrections came out of it, both above: the page-load model has the wrong shape (too heavy per action, far too short), and the storage ceiling was never costed.

Things worth remembering

The vector cache needs no invalidation callbacks, and that is why it is safe. A stored vector can only change through memorybank.set_vector, which drops that one entry. Anything that removes a memory from play — eviction, deletion, pruning, an edit clearing the vector — falls out of the catalogue query, and entries missing from the catalogue are dropped on the next read. So there is no hook anyone can forget to call. It is in-process and assumes one worker, which is what the deploy runs.

Weigh new columns in bytes, not rows. The comment on Memory.embedding said "fine at bank sizes of a few hundred" and was wrong by the only measure that mattered: a few hundred JSON vectors is ten megabytes, fetched fresh every turn.

A deferred column needs a cheap flag beside it. memories.embedded exists because once the vector is deferred, every "is this embedded?" check becomes a 6 KB lazy load, once per row down the Memories drawer. Exactly the same shape as actions.variant_count beside actions.variants. Expect to need this for any future heavy column.

Any egress measurement must run with an embedding model set. This is the second time that omission has hidden the biggest number in the room.

Production has real users on it now. Measure it without reading it. Counts, sum(octet_length(...)) and pg_total_relation_size answer every sizing question asked so far, and none of them return anyone's story, memory text or email. When a real Postgres is needed for a write path, create a throwaway database beside the real one and drop it after — never point a harness at the production database.

octet_length is the egress number, not the on-disk number. Postgres TOAST compresses big JSON — context_snapshot is 150.8 MB uncompressed but ~89 MB stored — and decompresses before sending. Size reads with octet_length, size the storage bill with pg_total_relation_size, and do not mix them up.

---

Still open from plan/13

• Step 6, infinite scroll upward — the pick-up item above.
• Query-count / byte assertions per endpoint, extending tests/test_egress.py against production-sized fixtures. dbmeter is importable from tests (from tools import dbmeter) and was built with this in mind; nothing uses it there yet.
• Explicit column projections on read paths, so the next heavy column is opt-in. Done for the memory paths, not as a general rule.
• Drop memories.embedding (the JSON column) in a follow-up migration. It is still written by set_vector and read by nothing, kept so a rollback finds the vectors. tests/test_memory_retrieval.py has a guard asserting nothing selects it. Measured on production: dropping it reclaims 4.05 MB, 4% of the database.
• context_snapshot and the 512 MB ceiling — new, and now the biggest open item. See the two sections named above. The egress case for leaving it in the database still stands; the storage case does not.

Deliberately not taken: moving context_snapshot out of the database (~$0.02/mo, costs nothing on reads now that it is deferred), and pgvector (breaks the SQLite dev parity this codebase protects on purpose).

---

Running things

cd backend
.venv/Scripts/python.exe -m pytest tests/          # 225 tests
.venv/Scripts/python.exe -m tools.stress_session   # egress report (SQLite)

# Same harness against a real Postgres. The target must be a THROWAWAY database
# — this writes a synthetic adventure, and it refuses any name without
# 'stress'/'scratch' in it.
AIDND_STRESS_DATABASE_URL=postgresql://…/stress_scratch \
    .venv/Scripts/python.exe -m tools.stress_session

On Windows the report's box-drawing characters crash the default cp1252 console; prefix with PYTHONIOENCODING=utf-8.

Port 8000 is shared with the job-pipeline app, which will squat it and silently shadow the AI-DnD API — free it before running the backend, or move the vite proxy.