Summary

Re-init no longer ignores the embedding settings users already wrote down in ~/.gbrain/config.json. gbrain init --pglite (no flags) on a brain whose persisted config declares a non-default embedding_model / embedding_dimensions / provider_base_urls now respects those values instead of silently falling back to OpenAI 1536. Closes #203.

Why

resolveAIOptions in src/commands/init.ts only consumed CLI flags. It never read loadConfig(), so aiOpts came out empty when no flags were passed. That tripped the if (embedding_model || chat_model) gate further down in initPGLite (line 386), so configureGateway was never called, _config stayed null, and initSchema()'s catch-block fell back to literal 1536 in pglite-engine.ts:212.

Net result on every re-init without flags: the user's persisted choice was silently overridden by the OpenAI default. v0.28.5's embedding-dim-check catches the resulting column mismatch on the next embed, but only AFTER the wrong-shape brain is already on disk and the user has spent another minute figuring out why their --embedding-model lmstudio:... setup keeps producing vector(1536) columns.

What changed

One function (resolveAIOptions) in src/commands/init.ts. Seeds out from loadConfig() at the top, before the existing flag-handling code. Flags still win on a per-field basis. New resolution chain:

1. ~/.gbrain/config.json via loadConfig() — NEW
2. --model SHORTHAND recipe lookup (sets model + dims)
3. --embedding-model VERBOSE (overrides shorthand)
4. --embedding-dimensions N (overrides recipe-derived dim)
5. --expansion-model / --chat-model (additive)

Gateway defaults still apply downstream when nothing in the chain populates a field — cold-start behavior unchanged. claw-test's fresh-install scenario keeps working.

Two edge cases caught during review and fixed in this PR:

• Model swap clears seeded dim. If config has lmstudio:nomic 768 and the user reruns init with --embedding-model openai:text-embedding-3-large (no --embedding-dimensions), the seeded 768 is cleared so the recipe-default derivation re-fires for the new provider. Without this, the user would silently get OpenAI configured at 768-dim.
• provider_base_urls rides along. Custom endpoint overrides (e.g. provider_base_urls.lmstudio pointing at http://localhost:1234/v1) are preserved through the seed + saveConfig cycle. Without this, re-init dropped the user's endpoint override and the gateway fell back to recipe defaults on next call.

Tests

test/init-config-first.test.ts — 7 cases, hermetic against the runner env via an isolateEnv() helper that unsets DATABASE_URL, GBRAIN_DATABASE_URL, and every GBRAIN_EMBEDDING_* / GBRAIN_*_MODEL / OPENAI_API_KEY / ANTHROPIC_API_KEY. Uses withEnv() from the canonical test helpers; no mock.module, no *.serial.test.ts quarantine needed. The 7 cases pin:

1. Config-only, no flags → returns config values
2. Flag-only, no config → existing flag behavior preserved
3. Both flags and config → flags override config per-field, config fills the gaps
4. Empty (cold-start, no config, no flags) → empty object (existing behavior preserved)
5. --embedding-model switches provider away from config → seeded dim cleared, recipe-default re-derives
6. provider_base_urls preserved through the seed cycle
7. Env-var-only (KNOWN LIMITATION, tracked at init: GBRAIN_EMBEDDING_MODEL/DIMENSIONS env vars ignored when no config.json AND no DATABASE_URL #1058)

Case 7 documents a separate bug class not closed by this PR: when there's no config.json AND no DATABASE_URL, loadConfig() early-returns null at config.ts:135, so the env-var merge below that early-return never fires. Setting GBRAIN_EMBEDDING_DIMENSIONS=768 env-only on a truly cold install still produces a 1536-dim schema. Filed at #1058 for separate fix.

Deliberate-failure run: DATABASE_URL=... GBRAIN_EMBEDDING_MODEL=... GBRAIN_EMBEDDING_DIMENSIONS=999 bun test test/init-config-first.test.ts → 7 pass, 0 fail. The isolateEnv() is adversarially proven, not just theoretical.

Verification

• bun run typecheck — clean
• bun run verify — all 13 pre-checks pass (check:privacy, check:test-isolation with 410 non-serial files scanned, etc.)
• bun test test/init-config-first.test.ts — 7 pass, 0 fail
• bun run test — full fast loop ran with 0 fails (two shards exceeded the 600s per-shard timeout cap; per gbrain CLAUDE.md that's the documented "infrastructure problem" classification, not test failures; all 4 shards' completed work reports 0 fail; new test is in shard 2 which completed in 1857 pass / 0 fail / 0 skip / rc=0)
• End-to-end smoke against a tempdir GBRAIN_HOME: column came up at vector(768) (verified via direct pg_attribute format_type probe) on a config-first init with no flags

Reviewer notes

Two adversarial reviews ran against this branch before opening the PR:

• A devil's-advocate critic loop (4 rounds, quieted with 4 findings raised + resolved — env-var-only path called out as KNOWN LIMITATION + filed as init: GBRAIN_EMBEDDING_MODEL/DIMENSIONS env vars ignored when no config.json AND no DATABASE_URL #1058, the bun -e smoke-test syntax error caught and re-run with ground truth, the rebase pre-check added before any force-push, and a test-isolation hole closed by isolateEnv()).
• codex review --base upstream/master 3 rounds: round 1 surfaced the model-swap dim-stale bug (fixed), round 2 surfaced provider_base_urls field-loss (fixed), round 3 clean.

Both reviewers were against the same branch tip after each fix; their findings did not overlap, which is the cross-model-review value-add CLAUDE.md Rule 5b describes.

🤖 Generated with Claude Code