Arch / EndeavourOS now ships CUDA 13 (libcublas.so.13); ctranslate2
4.7.2 (which faster-whisper rides) wants CUDA 12 + cuDNN 9 and won't
load against the system libs. Travis got it working ad-hoc with a
shell export of LD_LIBRARY_PATH + manual pip install, but systemd
doesn't inherit either, so the next reboot would refire the
libcublas.so.12 load error.
Making it permanent + reproducible:
- pyproject `tower` extra now pins the CUDA-12 runtime as pip wheels
alongside faster-whisper:
nvidia-cublas-cu12; sys_platform == 'linux'
nvidia-cudnn-cu12>=9,<10; sys_platform == 'linux'
uv.lock resolves nvidia-cublas-cu12 12.9.2.10 + nvidia-cudnn-cu12
9.22.0.52. Dev side (no --extra tower) stays clean — verified by a
no-extra `uv sync` followed by `uv pip list | grep nvidia` returning
empty.
- deploy/tower/run-transcribe-worker.sh (new, +x): computes the venv's
CUDA-12 lib dirs at runtime via `uv run python` (resolving
nvidia.cublas / nvidia.cudnn through __path__ — they're PEP 420
namespace packages with no __file__), prepends them to
LD_LIBRARY_PATH, then execs `uv run second-brain transcribe-worker`.
No hard-coded python3.XX path so it survives Python upgrades. If the
wheels aren't installed it aborts with a clear "uv sync --extra
tower" hint instead of a silent libcublas load failure deep inside
ctranslate2.
- second-brain-transcribe.service: ExecStart now points at the
wrapper. Also moves StartLimitIntervalSec / StartLimitBurst from
[Service] into [Unit] where modern systemd expects them
(systemd-analyze verify previously flagged the misplaced keys as
silently ignored). Restart=always, EnvironmentFile, After=/Wants=
wg-quick@wg-lan.service, User=herbyadmin all unchanged.
- second-brain-transcribe.env.example: trimmed to just
SECOND_BRAIN_DATABASE_URL with the placeholder spelled out, plus a
clear pointer to the ready-to-scp env file generated on herbys-dev
at /opt/backups/postgres-consolidation/second-brain-transcribe.env
(mode 0600, regeneratable from credentials.env without ever echoing
the password). The committed example never carries a real secret.
- deploy/tower/README.md: documents the CUDA-13-vs-CUDA-12 gotcha
upfront ("don't `pacman -S cuda cudnn`"), the wrapper-based
ExecStart, the scp-from-dev EnvironmentFile recipe with the
password-regen one-liner, and the EnvironmentFile-vs-shell-export
note.
Verified locally on dev (no GPU):
- uv.lock resolves with the new tower deps.
- A throwaway venv installed with the same `nvidia-cublas-cu12
nvidia-cudnn-cu12>=9,<10` pins produces lib dirs containing
libcublas.so.12 and libcudnn.so.9 via the wrapper's path probe.
- systemd-analyze verify is clean except the expected
"/opt/projects/... not executable on this host" warning (the
wrapper exists only in the tower's checkout).
- 27 passed / 2 skipped in pytest; zero-check 5/5.
GPU large-v3 + live service start under systemd remain tower-only
validation steps.
Splits pull+transcribe (now tower-side, eager) from extract+embed
(stays on the dev scheduler). Three machine-coordination pieces land
together because they reference each other:
- v3 migration adds sources.claimed_by + claimed_at — observability +
stale-claim recovery columns. The actual race-safety primitive is
`SELECT ... FOR UPDATE SKIP LOCKED` in the new claim helper, so two
machines can poll the queue without doubling work.
- src/second_brain/claim.py owns the claim dance (claim_next_source,
release_claim, reap_stale_claims). Both stage gates filter by
source_type so the tower never grabs articles and the dev side never
grabs videos.
- src/second_brain/transcribe.py wraps faster-whisper (lazy-imported so
it stays out of the dev install). resolve_settings() reads env >
[whisper] block > defaults, falling back to int8 on cpu / float16 on
cuda when compute_type is unspecified. Default model large-v3.
- src/second_brain/scheduler/transcribe_worker.py is the long-running
poll loop. Reads pipeline_settings every iteration so the dashboard's
enable/window/max-items/max-video-length take effect within one
cycle. Reaps stale claims at startup. SIGTERM-clean. DB-unreachable
backs off with a log line; never crash-loops.
- adapters/youtube.py drops the torch-whisper transcribe path; pull
stays. Removes openai-whisper from the default deps and gates
faster-whisper behind a new `tower` extra (uv sync --extra tower).
- main.py: new `second-brain transcribe-worker` (--once for ad-hoc).
`process` now article-only on the pull side but still picks up
TRANSCRIBED of any source_type for the extract step.
Live-verified: migration applies clean, transcribe-worker --once
honors transcription_enabled=false gate.
Adds three new routes on the FastAPI web app:
- GET /dashboard — pipeline counts by status (per-status drill-down
via filtered /?status=… links), recent activity
(latest 10 sources), and a settings snapshot
card with an "edit →" shortcut.
- GET /settings — full editing form for pipeline_settings.
- POST /settings/save — HTMX endpoint that validates the partial form,
upserts the row, and returns the rerendered
card with a "Saved." or error flash. The card
is its own _settings_card.html partial so the
HTMX swap targets only the form region.
Form parsing lives in settings_store helpers (parse_time_or_none,
parse_int_or_none) — keeps the route thin and the empty-string-to-NULL
normalisation in one place. Validation rejects negative counts and
sub-1 GPU concurrency.
The base template grew a Dashboard + Settings nav so the new routes are
reachable without typing URLs. Added python-multipart to deps because
FastAPI's Form(...) raises at app import without it.
Ruff autofix removed pre-existing F401 unused imports across adapters,
context, compiler, extractor, scheduler, web, and main. Also reordered
models.py so its SQLAlchemy imports sit at the top of the file
(E402 was triggered by the inline `utcnow` helper definition).
Added a [dependency-groups] dev block (pytest, pytest-cov, ruff) so the
zero-check test gauntlet — which runs `pytest --cov --cov-report=term-missing`
— can resolve its plugins without a manual `uv pip install pytest-cov`.
Swap the SQLite backing store for petalbrain Postgres + pgvector, modeled
on vault-mcp. All second-brain relational tables now live in the
`second_brain` schema (owned by the lovebug role); embeddings are written
to the shared public.embeddings table.
Locked design decisions (per Travis):
- DB: existing petalbrain Postgres, second_brain schema, lovebug role.
- Connection: containerized homelab-postgres:5432, plain psycopg_pool
(min=1/max=10), no PgBouncer.
- ORM stays SQLAlchemy; int autoincrement PKs + naive UTC DateTime.
- Embeddings: reuse shared public.embeddings keyed by
(source_schema='second_brain', source_table='extractions', source_id,
model='nomic-embed-text'). Summaries only for this round.
- Pipeline: chunk_text → Ollama nomic-embed-text → delete-before-insert
upsert, with graceful degradation (no DB / no Ollama → log + skip).
- Alembic stands up second-brain's own schema; public.embeddings stays
out-of-band.
- File-based wiki compiler is unchanged.
No SQLite data import — starting clean.
This commit is the scaffolding only; `alembic upgrade head` and a smoke
test of the embedding path are the next checkpoint.
Run extraction under the Max OAuth subscription via `claude -p` instead
of the per-token Anthropic API. The new src/second_brain/llm/claude_cli.py
spawns the CLI in a hermetic tempdir so the host project's CLAUDE.md,
hooks, MCP config, and settings don't leak into the prompt. Uses
--json-schema with LLMExtraction.model_json_schema() so the CLI guarantees
valid structured output — replaces the brittle markdown-fence stripping
in the old engine. The Anthropic SDK is preserved as an optional "api"
backend selectable via config.
While in here, fix a handful of blockers that the smoke test surfaced:
- scheduler filtered ANALYZED instead of TRANSCRIBED, so it never
actually advanced any sources
- process command read sources in a closed session, raising
DetachedInstanceError before any work happened
- config.prompts_dir walked one parent too many, resolving outside
the project and forcing the fallback prompt for every domain
- compiler called git rev-parse against a vault that was never
git-init'd; now auto-inits with an empty seed commit and skips empty
commits cleanly
- datetime.utcnow() deprecated in 3.12+ — single utcnow() helper in
models.py keeps naive UTC semantics so no DB migration is needed
- sess.query(...).get() deprecated in SA 2.x → sess.get(...)
- dead `import anthropic` removed from compiler
Smoke test (article → process → accept → compile) succeeds end-to-end
with ANTHROPIC_API_KEY unset. a-review run saved under reviews/.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
