Transcript-analyzer
The transcript-analyzer is a focused Stage-1 sidecar generator. It takes one Claude Code session ID, fetches that session’s encrypted ciphertext from object storage, decrypts it, parses the redacted jsonl events, and writes a single structured analysis JSON file alongside the session’s metadata.
What it does
The pipeline is fixed: read the session sidecar, confirm the ciphertext was uploaded, fetch and decrypt it, parse the redacted event stream, summarize the session’s structure (tool calls, turn-shapes, durations, notable transitions), and emit one analysis.json file. It does not update downstream surfaces and holds no operational-memory write authority — the orchestrator (or an interactive caller) folds the output into adoption trackers and session logs.
When it’s commissioned
Spawned by the Night’s Watch — the nightly cron job that walks recently-uploaded sessions — or by an interactive caller asking for one session’s structured read. One transcript, one analysis JSON, one return message.
Why it exists
A session log is unstructured prose; an analysis JSON is queryable. This role is what makes the transcript record analytically tractable without exposing raw conversation content.
Links to this page
- Project historian — Reads the corpus impartially and takes a defended position; preserves refusals.
- Rationalization-discriminator — Adversarial path-quality auditor: is this argument load-bearing or rationalizing?
- Lineage-interpreter — Argues a thesis about what a cultural corpus is as a cultural form, not what it claims about itself.
- …