feat: extract and publish tool metadata to AMP

[thiagomoretto]

Summary

• Extracts rich metadata from tool classes during crewai tool publish
• Sends metadata to AMP backend for enhanced tool discoverability
• Displays preview of detected tools before publishing

What's Included

Tool Metadata Extraction

Automatically extracts from BaseTool classes:

• Tool name and description
• Init parameters schema (custom Pydantic fields)
• Run parameters schema (from args_schema)
• Environment variables (name, description, required, default)

Publish Preview

Shows detected tools during publish with:

• Tool name and description
• Init parameters with types and defaults
• Required environment variables

Non-blocking Errors

Metadata extraction failures show warnings but don't block publishing.

Payload Structure

{
  "tools_metadata": {
    "tools": [
      {
        "name": "ToolClassName",
        "humanized_name": "tool_name",
        "description": "Tool description",
        "run_params_schema": { },
        "init_params_schema": { },
        "env_vars": [
          {"name": "API_KEY", "description": "...", "required": true, "default": null}
        ]
      }
    ]
  }
}

---

Note

Medium Risk
Adds dynamic module importing and Pydantic schema introspection during crewai tool publish, which could surface edge-case import side effects or schema-generation failures (though publish continues without metadata). Backend payload shape also changes by adding tools_metadata, requiring server compatibility.

Overview
crewai tool publish now extracts rich metadata for exported BaseTool classes (name/description, module, JSON schemas for run/init params, and env var definitions), prints a pre-publish preview, and sends this metadata to the Enterprise/AMP API as a new tools_metadata payload.

Internally, tool discovery utilities were extended with safer temporary module loading/cleanup and schema generation that omits non-JSON-serializable defaults; publish now treats metadata extraction errors as warnings and continues. Tests were updated/added to cover the new API parameter, preview/continue behavior, and metadata extraction edge cases.

^{Written by Cursor Bugbot for commit 6f3cfb8. This will update automatically on new commits. Configure here.}

[github-actions]

This PR is stale because it has been open for 45 days with no activity.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[iris-clawd]

PR Review: feat: extract and publish tool metadata to AMP

Overall: Solid feature — clean implementation, good test coverage, and the non-blocking error handling is the right call. A few things worth addressing before merge:

---

Issues

1. Empty metadata still sends payload to API
In plus_api.py, the condition is if tools_metadata is not None, so when metadata extraction fails and tools_metadata = [], the API receives {"package": "handle", "tools": []}. Consider using if tools_metadata instead (falsy check) to avoid sending an empty wrapper object — or confirm the backend handles this gracefully.

2. _extract_field_default return type looseness
The function returns str | list[Any] | int, and humanized_name passes the result through without str() wrapping (unlike description which does str(...).strip()). If a field default is unexpectedly an int or list, humanized_name in the payload would have the wrong type. Consider normalizing:

"humanized_name": str(_extract_field_default(fields.get("name"), fallback=tool_class.__name__)),

3. _unwrap_schema has no loop guard
The while loop could theoretically spin if a malformed schema has circular type references. A simple max_depth counter (e.g., 20 iterations) would make it defensive:

for _ in range(20):
    if result.get("type") not in {...}:
        break
    ...

---

Minor / Nits

• SystemExit(1) change on L134 (was bare SystemExit) — good improvement, but it is a behavior change for anyone catching SystemExit by code. Intentional?
• Test lines are long in test_main.py — the return_value dicts in @patch decorators are 150+ chars. Consider moving them to module-level constants for readability.
• The _IGNORED_INIT_PARAMS frozenset is a nice touch for maintainability.

---

What looks good ✅

• Context manager pattern for module loading with hash-based names (no more temp_module collisions)
• Non-blocking metadata extraction — warnings only, publish continues
• Preview output before publishing
• Comprehensive test suite covering edge cases (empty project, import errors, syntax errors, multiple files, non-tool exports)
• lru_cache on _get_schema_generator

Verdict: Approve with minor fixes (items 1-3 above). None are blockers but would improve robustness.