Would JSON `input_schema` for builtin tools be in scope?

[YiRaaaan]

Hi — thanks for maintaining this repo; it's been an invaluable reference.

The repo carefully tracks each builtin tool's description (the prose text Anthropic sends to the model), but the JSON input_schema that accompanies each tool definition in the same request isn't covered. Wanted to ask before opening a PR: would adding schemas be in scope?

Why it could be useful

• Schemas describe what a valid tool call looks like — required fields, enums, value ranges, regex patterns. Today anyone building a Claude Code-compatible runtime or analysis tool either re-derives them or captures them themselves.
• They change across versions like the descriptions do, so they'd slot into the existing changelog cadence.
• Structured complement to the existing prose, not a replacement.

Why it might not be

• The stated scope is "system prompts." Strictly the schemas are part of the API request body, not the prompt string.
• A new content category means more for the extraction script to track.

Proposal if yes

A small PoC PR — one tool (Bash, since it has the most coverage here) — placed at e.g. tool-schemas/bash.json with a short README explaining source and conventions. If you like the pattern, happy to fill in the rest. If not, no harm done — I'll keep them out-of-tree.

For source: I have these captured verbatim from live Anthropic API requests via a local reverse proxy, so it's exactly what Claude Code v2.1.x sends — same kind of fidelity as your JS-extraction approach.

Happy to defer to your judgment either way. Thanks!

[bl-ue]

Hi @YiRaaaan! That's an interesting idea! It would definitely be interesting to see the changes to the tool schemas over time and it would also be very useful for anyone building a compatible runtime like you said. The only concern I have is maintenance burden—but if you use a proxy technique as you suggested instead of static analysis, I think it could end up being almost zero-maintenance.

The stated scope is "system prompts." Strictly the schemas are part of the API request body, not the prompt string.

As a matter of fact, tool definition JSON schemas actually are embedded into the system prompt on the server side just before the entire context is formatted into a single flat string and fed to the model at inference time. So keeping them with the other parts of the system prompts really makes perfect sense.

Please feel free to open a PR!

[WEIFENG2333]

+1, this would be useful.

I’ve been tracking similar prompt-bearing artifacts in Phistory (https://phistory.cc/), and schemas feel like a natural companion to the prompt text.