fix(file_write): non-greedy tag, last-block fence, schema content fallback, debug placeholder

[voidborne-d]

Closes #241.

Summary

Issue #241 reports 3 extraction bugs + 1 design gap in file_write. Reproduced and fixed all four with a 12-line, 4-file change. Net code size change: 0 (LoC down by 1 in ga.py, +1 in agent_loop.py, +1 each in the two schema files).

Changes

Bug 1 — greedy (.*) swallows multi-tag bodies (ga.py)

<file_content[^>]*>(.*)</file_content> matched to the last </file_content> in the reply, so two adjacent <file_content> blocks (or one whose body contains the literal closing string) produced corrupt content. Switched to re.findall(... .*? ...) and take the last match — non-greedy + last-wins matches the LLM's most-recent intent.

Bug 2 — first-fence-to-last-fence span (ga.py)

The fallback text.find('\``'), text.rfind('```')returned the entire span between the first and last triple backtick — when the reply contained a prose code snippet *before* the file-content fence, the prose plus its closing fence got concatenated into the file. Replaced withre.findall(r"```[^\n\`]\n([\s\S]?)```", ...)` and take the last fence body. Single-fence behavior unchanged.

Bug 3 — <file_content> silently stripped from logs (agent_loop.py)

_clean_content removed <file_content>...</file_content> entirely, so when a write turned out wrong there was no way to confirm what the model actually emitted. Now substitutes a <file_content: N chars> placeholder — visible in turn logs, zero token overhead, debugging restored.

Bug 4 — schema lacks a content parameter (tools_schema*.json)

file_write's parameter set was just path + mode; if reply-body parsing missed the content (Bugs 1/2 territory or bad LLM formatting) there was no fallback. Added an optional content string. do_file_write now does args.get(\"content\") or extract_robust_content(response.content) — body-extraction stays the canonical path, schema arg is a tail-end safety net.

Verification (no test infra in repo, so behavior verified by direct run)

✓ multi tag: returns the LAST <file_content> body
✓ article + file-content block: returns last fence body, prose ignored
✓ single fence: unchanged
✓ <file_content lang=\"py\"> attribute: matched
✓ no content: returns None
✓ <file_content> shown inside a fence + real one outside: real one wins
✓ _clean_content emits <file_content: N chars> placeholder
✓ schema includes content param (both en + cn)
✓ args.content fallback wired into ga.py

python -m py_compile ga.py agent_loop.py clean. Both schemas parse as valid JSON.

Notes

• All four fixes are surgical and isolated to the boundary of the bug. No collateral refactor.
• Schema description for the new content field calls it [Optional fallback] so the LLM keeps preferring <file_content> (preserves token-efficient streaming).
• Net delta is neutral (3 added / 4 removed in ga.py, +1 elsewhere). Aligns with the project's "ideally negative or zero" line-count guidance.