…, migrate command

Schema templating:
- pglite-schema.ts: getPGLiteSchema(dims, model) generates schema with correct
  vector dimensions for any provider (additive export, PGLITE_SCHEMA_SQL preserved)
- pglite-engine.ts: initSchema() uses getPGLiteSchema() from active provider
- postgres-engine.ts: initSchema() string-replaces vector dim/model in SCHEMA_SQL

Config persistence:
- config.ts: GBrainConfig gains embedding_provider + embedding_dimensions fields;
  loadConfig() propagates them to env vars so getActiveProvider() stays consistent
  across sessions without re-setting env manually

Init-time provider selection:
- gbrain init --provider gemini [--dimensions N]
  Creates brain with vector(768) schema and persists choice to config.json

Provider migration:
- gbrain migrate --provider gemini [--dimensions N] [--dry-run]
  Full migration: ALTER vector column + re-embed all chunks + update config table
  Handles both PGLite and Postgres engines. Safe to resume if interrupted.

Tests:
- 3 live Gemini API tests added (auto-skip when GOOGLE_API_KEY absent)
- Set GOOGLE_API_KEY in .env.local to run them (Bun auto-loads .env.local)
- .env.local.example added to repo as setup guide

Total: 19 embedding-provider tests, 2616 unit tests — 0 fail

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

- Model: text-embedding-004 → gemini-embedding-001 (the available stable model)
  gemini-embedding-001 supports 1–3072 dims via outputDimensionality (Matryoshka)
- Default dimensions remain 768; max raised from 768 to 3072
- Added 1536-dim mode (OpenAI-compat: switch providers without ALTER TABLE)
- Live API tests: 3 calls verified in 751ms total — batch + single + dim truncation
- .env.test.local: API keys only (no provider override) so schema stays at 1536
  for all existing tests; live Gemini tests instantiate GeminiEmbedder directly
- .env.local.example: scrubbed real keys, added NEVER put real keys warning

All 2609 unit tests pass. Live Gemini API verified.

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

- pglite-schema-provider.test.ts: 6 tests for getPGLiteSchema() pure function
  (all 3 substitution targets, custom dims, OpenAI-compat mode, default args)
- config-embedding-provider.test.ts: 4 tests for env-var propagation logic
  (propagates when unset, does not override when already set, for both
  embedding_provider and embedding_dimensions)
- migrate-provider-args.test.ts: 8 tests for provider instantiation and
  dimsChange logic (same/different dims, API key guard, no-retry on config error)
- embedding-provider.test.ts: +1 test for unknown provider fallback to OpenAI

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

- migrate-provider.ts: validate --dimensions is integer 1-3072, reject NaN
  (was: parseInt of non-numeric string produces NaN which silently passes
  GeminiEmbedder constructor and gets interpolated into ALTER TABLE SQL)
- migrate-provider.ts: remove duplicate ?? 768 default (GeminiEmbedder
  constructor already defaults to DEFAULT_DIMS=768)
- embedding-provider.test.ts: use it.skipIf pattern for live API tests
  instead of silent early-return (tests were green when not running)
- embedding-provider.test.ts: add resetActiveProvider() to isEmbeddingAvailable
  beforeEach/afterEach for proper test isolation
- embedding-provider.test.ts: add boundary dimension tests (1 and 3072)

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

1. embedding.ts: change EMBEDDING_MODEL/EMBEDDING_DIMENSIONS from eager
   module-level consts to lazy functions getEmbeddingModel/getEmbeddingDimensions.
   Consts were evaluated at import time, before loadConfig() could propagate
   the persisted provider choice to env — caused silent ordering bugs.

2. migrate-provider.ts: add remote=false guard parameter; throws if called
   with remote=true. Command does destructive DDL and mutates process.env,
   both unsafe in MCP/remote context.

3. providers/retry-utils.ts: extract exponentialDelay() and sleep() shared
   helpers from openai-embedder.ts and gemini-embedder.ts into a single file.
   Both had identical implementations — one change point now instead of two.

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

Critical:
- operations.ts: replace hardcoded OPENAI_API_KEY check with isEmbeddingAvailable()
  in put_page handler. Gemini users had all embeddings silently skipped on every
  put_page call (noEmbed was always true when OPENAI_API_KEY is unset).

High:
- gemini-embedder.ts: add Number.isInteger() to constructor guard to catch NaN.
  NaN fails both < 1 and > 3072 comparisons, silently creating an invalid provider.

Low:
- init.ts: use !== undefined instead of truthiness check for embeddingDimensions.
  Prevents dim=0 or dim=1 from being silently skipped when setting the env var.

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

Mirrors the GBRAIN_EMBEDDING_PROVIDER abstraction for query expansion.
Extracted the hardcoded Anthropic/Haiku call from expansion.ts into a
provider interface with two implementations: AnthropicExpander (default,
same behavior) and GeminiExpander (gemini-1.5-flash via function calling).

New files:
- src/core/expansion-provider.ts — ExpansionProvider interface + factory
- src/core/providers/anthropic-expander.ts — Claude Haiku impl (extracted)
- src/core/providers/gemini-expander.ts — Gemini Flash impl
- test/expansion-provider.test.ts — 10 unit tests

Config:
  GBRAIN_EXPANSION_PROVIDER=anthropic|gemini (default: anthropic)
  GOOGLE_API_KEY or GEMINI_API_KEY required when using gemini

Updated docs: INSTALL_FOR_AGENTS.md, README.md, .env.testing.example
All 2648 unit tests pass.

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

GeminiEmbedder.embed() and embedBatch() now run via injected mock client —
no GOOGLE_API_KEY required. AnthropicExpander missing-key test isolates
env var deletion to prevent parallel test file race conditions.

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

…AGENTS

Step 2 previously buried Gemini as a paragraph footnote after the OpenAI
block. Agents would parse "Ask the user for these" and present only
OPENAI_API_KEY + ANTHROPIC_API_KEY. Restructured as Option A / Option B
so the agent surfaces both upfront and asks the user to choose.

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>

…ander

Pre-upstream-contribution cleanup:
- Remove // FORK: comments from embedding.ts, hybrid.ts
- Remove FORK: comment from anthropic-expander.ts docblock
- Add explicit ANTHROPIC_API_KEY check before SDK call so tests
  get a clear error message instead of SDK's generic auth error

Co-authored-by: Al's bot <aloysiusmartis@users.noreply.github.com>