Preserve all original tool input fields instead of whitelisting

[wjessup]

Summary

_parse_tool_input previously hardcoded a whitelist of known fields per tool, silently dropping any unrecognized parameters. For example, Read only kept file_path — if the original input had offset, limit, or any other field, those were lost. This made the exported data less useful for training and analysis, since the tool call signatures didn't match what was actually sent.

This PR replaces the per-tool whitelist approach with a single recursive anonymizer that preserves all original fields while still applying the correct anonymization:

• Path keys (file_path, targetFile, cwd, workdir, etc.) → anonymizer.path()
• Command keys (command, cmd) → redact_text() + anonymizer.text()
• All other strings → anonymizer.text()
• Nested dicts/lists → recursively anonymized

This is a net deletion of ~75 lines and passes all 301 existing tests unchanged (the previously tested fields still appear in output — extra fields are now additionally preserved).

What changes for exported data

• Tool inputs now include all parameters the agent actually sent, not just a curated subset
• No schema change — input is still a dict, just with more keys when the original had them
• Privacy guarantees unchanged — all strings still go through anonymization/redaction

Made with Cursor