What Claude Code data does Traceplain read?

Traceplain reads newline-delimited assistant, user, and result envelopes produced by Claude Code stream JSON and compatible saved session records.

Does Traceplain upload Claude Code logs?

No. Imported content is parsed in the browser and is not sent to Traceplain or Stripe.

Does a successful tool result prove the task succeeded?

No. It supports a narrower claim about that recorded tool call. The intended outcome and the completeness of the record may still require human verification.

Claude Code JSONL Viewer — Review Tool Calls Locally

1. Capture a Claude Code stream.

For a new headless run, use Claude Code’s stream JSON output and redirect standard output to a file:

claude -p --verbose --output-format stream-json \
  "run the relevant tests and explain any failures" > claude-run.jsonl

Claude Code also stores local session transcripts under ~/.claude/projects/. Those files can contain far more context than a purpose-captured run, including prompts, commands, tool results, and file contents. Work from a copy and inspect it before sharing.

2. Import the JSONL locally.

Choose claude-run.jsonl in the free reviewer or paste its contents. Traceplain reads the file on your device and does not transmit imported activity.

WHAT TRACEPLAIN DOES WITH CLAUDE EVENTS

pairs each tool_use with its matching tool_result instead of counting two actions;
shows a recorded error separately from a successful result;
surfaces file paths supplied to file-oriented tools;
labels assistant text as reported rather than observed;
flags unrecognised record and content-block types instead of silently discarding schema drift;
keeps missing results and record completeness as unknown.

3. Review the evidence boundary.

A tool result records what the tool returned. It does not prove the wider task is correct, that every side effect is represented, or that the saved transcript is complete. A final assistant statement remains a claim unless stronger supplied evidence supports it.

OBSERVED       Write returned success for docs/example.md
OBSERVED       Bash returned “12 tests passed”
REPORTED       “The change is ready to merge”
UNKNOWN        whether the diff matches the request
NEEDS HUMAN    inspect the changed file before approval

Supported Claude Code evidence.

assistant envelopes containing tool use and assistant text;
user envelopes containing matching tool results;
result envelopes containing final run status and summary;
tool name, arguments, error state, result summary, timestamp, and recognised file path;
bounded warnings for record or content-block types that the reviewer does not interpret.
newline-delimited stream JSON and compatible saved-session records.

Traceplain is retrospective. It does not intercept a running agent, enforce permissions, or establish that omitted activity did not occur.

QUICK ANSWERS

Claude Code JSONL viewer FAQ

What command produces stream JSON?

Use claude -p --verbose --output-format stream-json and redirect stdout to a JSONL file. Add your own prompt and review the resulting file before sharing it.

Can I inspect saved Claude Code sessions?

Compatible assistant, user, tool-use, and tool-result envelopes can be reviewed, but saved transcripts may contain substantial sensitive context. Use a copy and keep it local.

Does Traceplain upload my transcript?

No. The public reviewer parses imported content in your browser. Aggregate counters contain no prompts, paths, tool arguments, or results.

Is this an Anthropic product?

No. Traceplain is an independent local-first review tool. Claude Code is an Anthropic product.