Import transcripts

blackbox import converts a transcript from another tool into a cassette, which you can then inspect, fork with the built-in demo agent, diff and assert like any other.

Claude Code sessions

Claude Code stores each session as a JSONL file under ~/.claude/projects/<project>/<session>.jsonl.

npx blackbox import --from claude-code --in ~/.claude/projects/my-app/1234.jsonl --out runs/session.json
npx blackbox inspect runs/session.json

• A model message that Claude Code split over several lines (text, then each tool call) is merged back into one turn.
• Parallel tool calls become one tool_calls step followed by each call's result, in call order.
• Multi-turn sessions keep every human turn and intermediate answer. The run ends with the last answer.
• Thinking blocks, token usage, model names and provider ids are dropped.
• Subagent (sidechain) messages inside a main session are skipped; the subagent's result still appears as the tool result in the main session. A subagent's own transcript (<session>/subagents/agent-*.jsonl, where every line is a sidechain) imports as a session of its own.
• A session that stopped while a tool was running is imported as incomplete.
• A session that ends with a user message nobody answered keeps that message: the cassette ends with the unanswered request (a model_input step) and is incomplete.
• Tool definitions are not in the transcript, so the cassette lists the tool names it observed.

The cassette is written only if it passes verify. A session in which a real API key appears (pasted into the chat, or printed by a command) is refused, and so is one containing the value of ANTHROPIC_API_KEY or OPENAI_API_KEY from your environment, the same check record makes. Nothing is written, and the key is not printed.

Chat JSON

--from chat-json reads a simple OpenAI-style transcript: { tools?, messages: [...] } with user, assistant (with tool_calls) and tool messages, each carrying a timestamp. It supports one tool call per assistant turn.