v0.6.0

v0.6.0 — 2026-07-05

Three systems around one promise: frame-accurate answers you can trust, at
a fraction of the tokens.

Self-healing answers (answer/)

• Every ask_video carries a confidence score from real retrieval
  signals (top-hit strength, margin over the runner-up, strength-gated
  evidence agreement) — calibrated against measured score distributions.
• Escalation ladder, cheapest first, stops the moment confidence clears
  the bar: dense high-res re-sampling around candidate timestamps → 2× zoom-
  crop re-OCR (recovers on-screen text the full frame mangled) → model
  verify pass, cheap tier then strong. Recovered evidence is indexed
  permanently.
• Verify pass: the model is shown the exact frames about to be cited and
  must return {supported, certainty, answer}; an eyewitness rejection
  forces the honest floor regardless of retrieval strength.
• Honest floor: below the floor the answer states plainly the video does
  not clearly show it, with the closest real moments. Citation timestamps
  can only come from indexed evidence (fabrications are stripped at
  composition, test-forced).
• Structured metadata on every answer: confidence, verified,
  escalations_used, cached, budget_stopped, evidence timestamps.

Self-improve loop (lessons/) — local, never uploaded

• report_mistake (MCP + watch-skill lessons add): a wrong answer + its
  correction becomes a classified lesson (missed-ocr / wrong-timestamp /
  hallucination / language / sampling-miss) in ~/.watch-skill/lessons.db,
  shared by every agent on the machine; where the class is mechanical the
  question is re-asked immediately to confirm the lesson works.
• Relevant lessons inject into future asks under a hard ~300-token cap.
• Every mistake becomes a test: lessons export-evals + evals run
  replay all past mistakes and report the pass-rate over time.
• Adaptive profiles: per-content-type error statistics (screencast,
  talking-head, vertical, fast-cut — auto-classified from index stats)
  become data overrides: OCR-first escalation, denser sampling, stricter
  thresholds. Inspect with profiles show, reset any time.

Token economy

• Text-first responses: timestamps in prose, zero image tokens by
  default; frames attach only on request or in the genuinely-uncertain band.
• Semantic answer cache (index migration v5): repeat and near-duplicate
  questions are free and marked cached: true; invalidated on re-watch,
  cleared with clean --cache-answers.
• Savings meter: every answer ends with ~N tokens saved vs raw-frame injection; lifetime meter via watch-skill stats / the stats MCP tool.
• Telegraphic scene descriptions (≤12 words, names/numbers kept) cut
  indexing and retrieval token weight.
• Per-question token budget the escalation ladder respects and reports.

Also

• watch-skill forget <video_id> removes one video (rows, cached answers,
  frames dir) with a structured error on unknown ids (#3).
• REST: POST /v1/answer returns the structured Answer; /v1/ask unchanged.
• No breaking changes: every v0.5 MCP tool name/signature intact; index
  upgrades v4→v5 forward-only and losslessly (migration-tested).