devbrief reads Claude Code session history from ~/.claude/projects. These JSONL transcripts may contain sensitive information including source code, file contents, API responses, and internal tool outputs from your coding sessions.

• Do not upload, share, or commit JSONL transcripts or the local SQLite DB (~/.local/share/devbrief/sessions.db).
• Treat transcript data with the same sensitivity as source code.

Commands such as devbrief list, devbrief raw, devbrief view, devbrief estimate, devbrief doctor, and devbrief capture --hook do not call Claude, the Anthropic SDK, or any external service. No data leaves your machine.

devbrief brief SESSION_ID sends a compact, truncated evidence packet to Claude, but only after:

1. You explicitly run the command.
2. devbrief displays the packet size and approximate token estimate.
3. You confirm the prompt (or pass --yes to bypass confirmation).

The evidence packet is a filtered and truncated summary of the session transcript. It is not a raw dump of all session data.

• ANTHROPIC_API_KEY is only used for AI brief generation.
• API keys are never printed in diagnostics, logs, or doctor output.
• Do not commit API keys, .env files, or ~/.config/devbrief/config.toml to version control.

The Claude Code hook integration must be capture-only:

devbrief capture --hook

Hooks that call devbrief brief, devbrief digest, claude -p, or claude --print are unsafe and will silently spend tokens on every session. devbrief doctor will detect and warn about unsafe hooks.

If you find a bug that could cause:

• Unexpected automatic LLM calls or token expenditure
• Unsafe hook installation
• Accidental exposure of API keys or session data
• Any other privacy or security issue

Please report it through the repository issue tracker, or privately if the issue is sensitive.