What problem does it solve? You have recorded digital-twin telemetry (tag history, binding maps, property sidecars) and want a plain-language analysis of a window of it, but you want that narration done by a swappable non-Anthropic model (OpenRouter, Ollama, Hermes, or even a human pasting into a chat UI) without giving that model any access to your project files. ## Core Features & Use Cases - Deterministic bundle contract: Packs a recording plus optional binding map and property sidecar into one byte-stable, size-budgeted (~100 KB) JSON bundle with sha256 provenance for every input. - Three reviewed task templates: summarize-window, narrate-anomalies, and inspection-report, each embedding mandatory honesty rules (cite only bundle stats, name bound elements, propose no actions). - Framework-written advisory reports: The worker returns only a body string; the framework wraps it in provenance frontmatter (provider, model, bundle hash, window) and writes it to reports/analysis/ and nowhere else. - Use Case: Record a day of plant tag history, run npm run analyze -- --task narrate-anomalies --recording day.ndjson --map binding_map.json, and receive a Markdown report narrating range crossings, sequence gaps, and step deltas, naming the physical element behind each tag. ## Quick Start Ask the agent to run npm run analyze with the summarize-window task on your recording file and binding map to produce an advisory analysis report under reports/analysis/.
