agent-scan: recover finish when LLM omits closing tags

[chaucerj]

agent-scan/agent_scan/utils/parse.py is a copy of mcp-scan's parser, but it is missing the two finish fallbacks the mcp-scan copy already has:

1. a bare <finish>...</finish> block, and
2. a truncated <function=finish> whose closing tag the LLM omitted (e.g. output cut off by max_tokens).

finish is the only way the agent loop terminates. When the parser returns None, handle_no_tool() runs instead and the entire final report produced by the scan is silently lost while the loop idles to max_iter.

This ports both fallbacks from mcp-scan (keeping agent-scan's dict | None signature and its combined <parameter= / <parameter name= regex) and hides bare <finish> blocks in clean_content(), matching mcp-scan.

Verified on: normal tool calls, closed finish, truncated finish, bare finish, tag-free and empty inputs — all behave as intended.

[boy-hack]

Thanks @chaucerj — this is a clean, well-motivated port. finish being the only loop-termination mechanism means a parser returning None here silently drops the whole final report and idles to max_iter, so handling both the bare <finish> block and the unclosed <function=finish> is exactly right.

I verified html and re are already imported in parse.py, so the html.unescape(...) calls are safe. The chained fallbacks (function tags → bare finish → unclosed finish) are ordered correctly, and hiding bare <finish> in clean_content() keeps the conversation surface consistent with mcp-scan.

Two minor, non-blocking notes:

• The second fallback uses re.DOTALL | re.IGNORECASE with a non-greedy .*? and a trailing (?:</function|$). On very large content this is O(n) but fine; if you ever see catastrophic backtracking on pathological inputs, anchoring the end helps. Not a concern for normal LLM output.
• If mcp-scan's copy evolves further, consider whether agent-scan/agent_scan/utils/parse.py and the mcp-scan parser should converge (or at least track each other), so this class of "agent-scan is a stale copy" bug doesn't recur. Out of scope here.

Otherwise correct and ready to merge. Nice catch on the silent-report-loss.

(Note: I'm leaving this as a review comment only — not merging.)

[chaucerj]

Thanks @boy-hack for the thorough verification — and for tracing the callers to confirm the silent-report-loss path.

Both notes acknowledged:

• The regex tail (?:${'<'}/function|$) stays as-is for now; if backtracking ever shows up on pathological inputs I'll anchor the end explicitly.
• Parser convergence between agent-scan and mcp-scan — agreed it's out of scope here; happy to look at it separately if there's appetite.

Anything else needed from my side? If not, would you mind adding the formal approval so this can merge?