mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-27 22:15:33 +00:00
f68eb9d905
Takeover of #2804 and #2805 (stacked), rebased onto current master. Rescan loop + dropped claims (#2804): - Zero-claim scans now write a status='empty' sentinel row so the (source_id, page_slug, content_hash, prompt_version) cache hits on every subsequent cycle instead of re-spending the extractor LLM call on the same unchanged page forever. - The idempotency index gains md5(claim_text): multi-claim pages keep every claim (the 4-column unique index silently dropped claims 2..N via ON CONFLICT DO NOTHING). proposals_inserted now counts rows that actually landed (RETURNING id), not insert attempts. - Migration renumbered 123 -> 125 (master shipped v123 configurable_fts_language and v124 after the PR was cut; the duplicate v123 would never have run under the version > current runner). Schema parity across schema.sql, pglite-schema.ts, schema-embedded.ts. Model-tier honoring (#2805), reworked for current master: - propose_takes, grade_takes, calibration_profile resolve their model through resolveModel(models.<phase> > models.default > env > getChatModel()) — per-phase config keys are now honored while the gateway chat model (already tier-resolved by reconfigureGatewayWithEngine) stays the default, preserving master's #2451 semantics (provider-prefixed stored ids, nested prefixes like openrouter:... intact) and the #2997 opts.model ?? getChatModel() idiom. The PR's original bare-tail-everywhere approach was dropped where it conflicted with #2451; grade_takes keeps the bare tail ONLY for its cache key / stored judge_model_id (continuity with historical rows), while the judge call itself gets the full resolved string — fixing the live label-vs-actual mismatch (call rode chat_model, telemetry recorded a hardcoded claude-sonnet-4-6). - brainstorm resolves via models.brainstorm > tier chain (was hardcoded anthropic:claude-sonnet-4-6). - gbrain models lists the four new per-task keys; output caps and context-window entries added for newer Claude models. Tests: test/propose-takes-rescan.test.ts (new, hermetic PGLite); config-key coverage in propose-takes/grade-takes tests; existing #2451 and gateway-chat-model tests pass unchanged. Co-authored-by: p3ob7o <p3ob7o@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>