fix(skill-scan): recover missing tool-call args during discovery

[li872]

Summary

Skill-scan info collection (AIG three-stage mode) often finished with an empty report after models such as qwen3-7 / glm5-2 / qwen3.7-plus emitted tool calls that the parser only partially understood. The dispatcher then called tool_func(**{}), which raised TypeError: missing 1 required positional argument: 'thought' / 'command' / 'path'. The agent treated the tools as broken, skipped the real project tree, and finished with 0 findings.

Root cause (parser + dispatcher, not the filesystem):

1. Parser (skill_scan/utils/parse.py) only extracted parameter tags inside function tags. Qwen/GLM-style JSON bodies, child tags, inline attributes, leftover free text, and tool_call wrappers produced a tool name with an empty args dict.
2. Dispatcher (skill_scan/tools/dispatcher.py) forwarded those kwargs unchanged. Missing required Python parameters became a raw TypeError. Legacy execute_shell (from mcp-scan prompts / model memory) is not a skill-scan tool, so listing attempts never recovered to dir_tree/ls.
3. Stage 1 did not inject the on-disk repo tree (stage 2 already did), so a failed discovery tool sequence had no fallback file listing.

Fix (minimal, recoverable):

• Resilient parsing of the issue-shaped payloads plus common Qwen/GLM wrappers.
• call_compat normalizes aliases (path to file_path, cmd to command, list_dir to ls), defaults dir_tree/ls path to the uploaded project folder, remaps listing execute_shell calls onto dir_tree, and otherwise returns a retryable error that names the missing argument and shows the correct function/parameter format.
• Info Collection now injects the real directory tree as a safety net.

No shell execution was added. Arbitrary execute_shell commands are rejected with a clear “use dir_tree/ls/read_file/grep” error.

Type of change

• Bug fix
• New feature
• Fingerprint rule (data/fingerprints/)
• Vulnerability rule (data/vuln/)
• MCP security plugin (data/mcp/)
• Jailbreak evaluation dataset (data/eval/)
• Documentation
• Other (please describe)

How was this tested?

Run:

cd skill-scan
python3 -m pytest pytests/test_tool_call_resilience.py -v
python3 -m pytest pytests/ -v

Result: 47 passed.

Regression tests feed the #629 shapes (missing thought, missing command, dir_tree with no path / child-tag path / Qwen tool_call JSON) and assert the scanner recovers a real SKILL.md tree or returns a retryable error — never a TypeError / empty discovery.

Tests

• skill-scan/pytests/test_tool_call_resilience.py (24 new cases)
• Full skill-scan/pytests/ suite: 47 passed

Demo

Passing regression run:

https://raw.githubusercontent.com/li872/AI-Infra-Guard/cursor/fix-skill-scan-tool-args-a83f/docs/pr-629-demo/pytest-issue-629.log

Related issue

Fixes #629

Checklist

• go build ./... succeeds locally (Go sources unchanged)
• I followed the existing file/rule format conventions for this contribution type
• I have not introduced hardcoded secrets or credentials