[BUG] Separator after :start_line: in apply_diff is not validated

[rkfg]

Problem (one or two sentences)

The LLM sometimes messes up the separator and writes the text to search without newline. Roo doesn't detect that and doesn't provide a meaningful error message so the model starts looping.

Context (who is affected and when)

Happens randomly with smaller consumer hardware friendly models such as Qwen 3.6

Reproduction steps

Tool call:

{
        "type": "tool_use",
        "id": "YwRzYCZxTOl9OZZLfMngeKXWTpbPZwVb",
        "name": "apply_diff",
        "input": {
          "path": "src/components/controls/MaskEditor.tsx",
          "diff": "<<<<<<< SEARCH\n:start_line:7\n-------import { useTranslate } from '../../i18n/I18nContext';\n\ntype MouseMode\n=======\nimport { useTranslate } from '../../i18n/I18nContext';\nimport { MaskEditorProvider, useMaskEditor } from './MaskEditorContext';\n\ntype MouseMode\n>>>>>>> REPLACE"
        }
      }

The problematic part is -------import { useTranslate } (missing \n after -------).
Roo responds with:

{
        "type": "tool_result",
        "tool_use_id": "YwRzYCZxTOl9OZZLfMngeKXWTpbPZwVb",
        "content": "<error_details>\nNo sufficiently similar match found at line: 7 (63% similar, needs 100%)\n\nDebug Info:\n- Similarity Score: 63%\n- Required Threshold: 100%\n- Search Range: starting at line 7\n- Tried both standard and aggressive line number stripping\n- Tip: Use the read_file tool to get the latest content of the file before attempting to use the apply_diff tool again, as the file content may have changed\n\nSearch Content:\n-------import { useTranslate } from '../../i18n/I18nContext';\n\ntype MouseMode\n\nBest Match Found:\n7 | import { useTranslate } from '../../i18n/I18nContext';\n
...

Expected result

A clear error about separator missing new line should be provided in the message so the model can recover

Actual result

The model thinks the tool is broken or that file has invalid indentation etc. and keeps making the same mistake

Variations tried (optional)

After explicitly telling the model about the separator issue and providing an example of correct call it was able to make further calls correctly.

App Version

Version: 3.53.0 (ad25634) (with #12200 applied because it really helps)

API Provider (optional)

OpenAI Compatible

Model Used (optional)

Qwen3.6 35B A3B APEX

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause: In src/core/diff/strategies/multi-search-replace.ts, the main parsing regex (line 292) expects the ------- separator to be on its own line followed by a newline: ((?<!\\)-------\s*\n)?. When the LLM omits the newline (e.g. -------import { useTranslate }), the separator group silently fails to match, and the text after ------- gets absorbed into the search content. This causes a confusing "No sufficiently similar match found" error instead of pointing out the malformed separator.

The validateMarkerSequencing method (lines 101-243) only validates <<<<<<< SEARCH, =======, and >>>>>>> REPLACE markers -- it does not check for ------- at all, so the malformed separator slips through without detection.

Proposed Fix:

1. Add separator validation to validateMarkerSequencing in src/core/diff/strategies/multi-search-replace.ts:

   • In the AFTER_SEARCH state, detect lines starting with ------- that have non-whitespace content after them (i.e., missing the newline separator).
   • Return a clear, actionable error message like: "ERROR: The separator line '-------' at line N must be on its own line followed by a newline. Found content 'import { useTranslate }...' on the same line. Place the separator on its own line with search content starting on the next line."

2. Add tests in src/core/diff/strategies/__tests__/multi-search-replace.spec.ts:

   • Test that -------content (no newline) produces the specific validation error
   • Test that ------- followed by a newline and then content still works correctly
   • Test with :start_line: present before the malformed separator

This is a minimal, targeted fix that catches the specific malformation early in validation and gives the LLM a clear signal to fix its formatting, preventing the looping behavior described in the issue.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[rkfg]

@roomote approved, do it

[roomote-v0]

I've seen the approval and will implement the fix now. I'll add separator validation to validateMarkerSequencing in multi-search-replace.ts to catch malformed ------- separators early with a clear error message. Will open a PR shortly.

[roomote-v0]

Opened PR #12212 with my attempt at addressing this. The fix adds early detection of malformed ------- separators (missing newline) in validateMarkerSequencing, returning a clear error message so the model can recover. Feedback welcome! #12212