[LuckVd]

Summary

I fixed dynamic MCP tool-result handling so the agent no longer loses later content blocks or structured results. I preserve the legacy single-text and single-media return shapes when no extra metadata is present, and include structured_content (including explicit JSON null), is_error, top-level meta, and text-block annotations/metadata when they are present. For multi-block results, I place readable text before structured data, metadata, and opaque media, while labeling each original content index to make the display-order change explicit. This prevents earlier opaque media from hiding later evidence that fits within the existing history preview; the preview's length cap still applies to unusually long text.

Scope note

InputRequiredResult is a separate interactive SDK result, not a CallToolResult content variant. This PR does not enable allow_input_required=True; the SDK default continues to reject that response because the dynamic scanner has no elicitation-response flow.

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?

• cd mcp-scan && .venv/bin/python -m pytest -q pytests/test_mcp_sdk_v2.py — 13 passed. Covers singleton compatibility, multiple text/media/resource blocks, structured_content including explicit null, is_error, result/block metadata, empty-success behavior, explicit rejection of InputRequiredResult, and late evidence through ToolDispatcher._format_result into BaseAgent._build_history_tool_result.
• cd mcp-scan && .venv/bin/python -m pytest -q pytests/ --ignore=pytests/test_surrogate_sanitization.py — 59 passed.
• Full MCP suite — 63 passed, 2 failed in untouched pytests/test_surrogate_sanitization.py (lines 84 and 102). I reran that file on clean origin/main; the same two tests fail before scanner code runs because os.path.join() combines a string temp directory with a bytes filename (TypeError: Can't mix strings and bytes in path components).
• go build ./... — passed.
• ruff check and ruff format --check on the changed Python files — passed; git diff --check — clean.
• cd mcp-scan && PYTHONPATH=$PWD .venv/bin/python main.py --help — passed.

Related issue

Fixes #726

Checklist

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

[boy-hack]

Verified locally on top of latest main (33768c5):

• The fix correctly addresses the truncation reported in #726: content[0] slicing is replaced by full-result serialization that preserves all content blocks, structuredContent, isError, result _meta, and per-block annotations/meta.
• Good backward-compat touch: the single-block/no-metadata fast path keeps the historical plain str return, so existing consumers see no behavioral change for simple results.
• Sensible ordering (status → text → structuredContent → result meta → metadata → media) keeps readable text ahead of opaque media blobs in the agent's bounded preview.
• Tests: new coverage passes locally (multi-block, is_error, structuredContent, result meta, input-required/unsupported types). Full-suite failure set is identical to the main baseline (9 pre-existing environmental failures in test_dispatcher / test_protocol_negotiation / test_surrogate_sanitization, unrelated to this change).
• CI green after re-running the stuck fork workflow.

Thanks for the high-quality fix and the detailed issue report!