How it reads your logs

What it reads

Claude Code

~/.claude/projects/<encoded-cwd>/<session>.jsonl, plus subagent files folded into the parent session. Unknown record types are tolerated and counted; malformed lines are skipped and counted.

Codex CLI

~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl, picked automatically when you pass --root ~/.codex.

git

Commits and line counts inside the window, and the final HEAD, from the project's repository.

What it computes

• Phases: explore, build, verify, fix, idle — from which tools ran when.
• Error arcs: consecutive failures of one command until it next succeeds; the longest is the hardest moment.
• Hotspots: files by edit count and lines changed, with errors attributed to them.
• Claims: completion statements in the final message, checked against the last verify run after the last edit — verified, contradicted or unverified.
• A number validator: anything spoken or captioned in the reel must appear in facts.json, or the render fails.

What it redacts

Before any string reaches a debrief, a storyboard or a frame:

• API keys and tokens: OpenAI/Anthropic sk-, AWS AKIA, GitHub ghp_, Slack, Stripe, bearer tokens, JWTs.
• Private-key blocks, .env-style KEY=value and key: value secrets, credentials in URLs.
• Email addresses, and /home/you or /Users/you paths, replaced with ~.
• With --redact strict, prompts are replaced with a one-line generic summary.

The demo transcripts contain fake secrets on purpose; the test suite asserts that none of them survive into any output, and the events listed on the overview page went through the same redaction.

What it does not do

• No network requests. No telemetry. No accounts.
• No model calls unless you opt in with --llm ollama, which only rephrases narration and is rejected if a number changes.