@3030-labs/wotw 0.8.4 → 0.9.0

package/CHANGELOG.md

@@ -11,6 +11,60 @@ always non-breaking.
 
 ---
 
+## [0.9.0] — 2026-05-31 — Multi-LLM + verify-and-harness + Group C + redaction security
+
+This release ships the multi-provider LLM abstraction, the standing gold-fact
+regression harness and its CI, the Group C residuals (token-estimation accuracy
++ per-provider extraction quality), and two credential-redaction security fixes.
+All live provider baselines (Anthropic / OpenAI / Gemini) are recorded and
+committed; Ollama ships untested (no local server at release time).
+
+### Security
+- **`AQ.`-format Google API keys are now redacted.** The `gemini-api-key`
+  redaction rule matched only the legacy `AIza…` format; Google's current AI
+  Studio keys (`AQ.` + ~40–80 chars, rolled out 2024–2025) passed through
+  unredacted — a credential-leak vector for any content sent to an LLM. The rule
+  now matches both formats (additive; same `cloud_rule_id`, legacy `AIza…`
+  redaction preserved verbatim).
+- **Removed real API keys from the redaction test fixtures.** `sanitize.test.ts`
+  had used two real keys (Google + OpenAI) as examples; replaced with synthetic
+  fixtures. The exposed keys have been rotated. (Git history still contains the
+  revoked values — rotation is the remediation; an optional history scrub is
+  tracked separately.)
+
+### Added
+- **Multi-LLM provider abstraction (Option B, single-pass).** A `complete()`-level
+  `LLMProvider` interface with OpenAI, Gemini, and Ollama wrappers alongside the
+  Anthropic provider, wired into all callers via `runtimeAwareComplete`. The
+  single-pass ingestion architecture is preserved (no agent-loop reintroduced;
+  the no-agent-SDK-import invariant test stays green).
+- **Gold-fact regression harness + cassette CI** (verify-and-harness arc). A
+  semantic fact-level precision/recall scorer, a 20-fixture / 105-fact gold
+  corpus, regression-from-baseline gating, accepted-delta normalization, and a
+  committed-cassette offline PR-gate (zero API calls in CI) plus a scheduled,
+  key-gated live-drift workflow.
+- **GC1 — token-estimation accuracy CI.** A ±15% offline PR-gate that compares
+  the `estimate_query_cost` 4-char heuristic against recorded provider-native
+  token counts (Anthropic + Gemini), gating on regression-from-baseline (compliant
+  fixtures held to ±15%, known exceedances pinned against worsening), plus a
+  scheduled live-drift tier and a committed accuracy report. Measured result:
+  30/40 within ±15% (the heuristic under-estimates dense/code content; documented).
+  OpenAI / Ollama have no native tokenizer — an explicit, documented gap.
+- **GC2 — per-provider fact-extraction quality report.** An operator-facing
+  per-provider quality summary with a recall-keyed recommend-disable signal
+  (precision is expected-low by design; the gold set is a curated subset). All
+  three cloud providers score healthy; Ollama is reported untested.
+
+### Fixed
+- **Per-provider API-key wiring in the harness recorder.** API mode read a single
+  `execution.api_key_env` for every provider, so non-Anthropic recordings were
+  handed the Anthropic key (401). The recorder now maps each provider to its own
+  env var — the reason non-Anthropic recording had never been exercised.
+- Two redaction-security defects (leaked test keys; `AQ.`-format gap) — see
+  **Security** above.
+
+---
+
 ## [0.8.4] — 2026-05-26 — Public-launch readiness (PASS-023)
 
 ### Added