v0.10.0 — Feature release: opt-in secret redaction, plus recall-quality and…

Feature release: opt-in secret redaction, plus recall-quality and toolchain hardening.

Added

• Secret redaction at ingest (opt-in). longhand config --set redact.enabled=true masks secret-shaped strings — Anthropic/OpenAI/AWS/GitHub/Google/Slack/Stripe keys, SSH key headers, JWTs, database URLs with passwords, SSNs, and Luhn-validated card numbers — at parse time, before they reach SQLite, the raw JSON blobs, or the vector index. Masks keep only the first/last 4 characters and the length (sk-a…wxyz (len=50)). Default is off: Longhand stays forensic unless you opt in. (#26)
• longhand redact — retroactive scan of an existing store. Reports pattern names and counts only (never the matched values); --apply masks matches in place across events, episodes, segments, and outcomes, and re-embeds changed vector documents. (#26)

Fixed

• Benign noise no longer counts as errors. The error detector now skips known-noise lines — Next.js streaming-SSR data-dgst/<!--$!--> markers, "0 failing"/"Tests: 0 failed" summaries, empty error: fields, Error: Task not found tool churn, and missing-GNU-timeout probes — while continuing to scan, so real errors in the same output still register. This was the main inflator of the unresolved-episode rate. Run longhand analyze --all to re-extract episodes with the cleaner detector. (#25)

Security

• chromadb is now capped <1.0 for all Python versions (previously unbounded below 3.14) — a chromadb 1.x release can no longer break fresh installs. (#25)

Internal

• mypy is clean (21 → 0 errors) and the CI typecheck job is now blocking. (#25)
• ruff format adopted repo-wide and gated in CI. (#24)
• 311 tests (280 → 311). The May 2026 audit report is archived under docs/audits/. (#24–#26)