[BUG] apply_diff fails if :start_line:N is malformed with a misleading error

[rkfg]

Problem (one or two sentences)

Sometimes apply_diff starts refusing to work even though the content matches 100%. The reason is :start_line:N sometimes being malformed as :start_line=N or :start_line>N etc.

Context (who is affected and when)

Happens after a while when the model's context fills up.

Reproduction steps

Just running the usual editing loop. Here's the message:

<error_details>
No sufficiently similar match found (92% similar, needs 100%)

Debug Info:
- Similarity Score: 92%
- Required Threshold: 100%
- Search Range: start to end
- Tried both standard and aggressive line number stripping
- 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

Search Content:
:start_line=18
-------
// MatrixConfig holds Matrix connection settings.
type MatrixConfig struct {
	HomeserverURL  string   `yaml:"homeserver_url"`
	AccessToken    string   `yaml:"access_token"`
	UserID         string   `yaml:"user_id"`
	BotDisplayName string   `yaml:"bot_display_name"`
	Rooms          []string `yaml:"rooms"`
}

Best Match Found:
16 | }
17 | 
18 | // MatrixConfig holds Matrix connection settings.
19 | type MatrixConfig struct {
20 | 	HomeserverURL  string   `yaml:"homeserver_url"`
21 | 	AccessToken    string   `yaml:"access_token"`
22 | 	UserID         string   `yaml:"user_id"`
23 | 	BotDisplayName string   `yaml:"bot_display_name"`
24 | 	Rooms          []string `yaml:"rooms"`
25 | }


Original Content:
 1 | package config
 2 | 
 3 | import (
 4 | 	"fmt"
 5 | 	"os"
 6 | 
 7 | 	"gopkg.in/yaml.v3"
 8 | )
 9 | 
10 | // Config holds all configuration for the bot.
11 | type Config struct {
12 | 	Matrix      MatrixConfig      `yaml:"matrix"`
13 | 	LLM         LLMConfig         `yaml:"llm"`
14 | 	Personality PersonalityConfig `yaml:"personality"`
15 | 	Context     ContextConfig     `yaml:"context"`
16 | }
17 | 
18 | // MatrixConfig holds Matrix connection settings.
19 | type MatrixConfig struct {
20 | 	HomeserverURL  string   `yaml:"homeserver_url"`
21 | 	AccessToken    string   `yaml:"access_token"`
22 | 	UserID         string   `yaml:"user_id"`
23 | 	BotDisplayName string   `yaml:"bot_display_name"`
24 | 	Rooms          []string `yaml:"rooms"`
25 | }
26 | 
27 | // LLMConfig holds LLM API settings.
28 | type LLMConfig struct {
29 | 	APIKey      string  `yaml:"api_key"`
30 | 	BaseURL     string  `yaml:"base_url"`      // Optional custom base URL
31 | 	Model       string  `yaml:"model"`
32 | 	MaxTokens   int     `yaml:"max_tokens"`
33 | 	Temperature float64 `yaml:"temperature"`
34 | }
35 | 
36 | // PersonalityConfig holds bot personality settings.
37 | type PersonalityConfig struct {
38 | 	Name           string   `yaml:"name"`
39 | 	Description    string   `yaml:"description"`
40 | 	Traits         []string `yaml:"traits"`
41 | 	Tone           string   `yaml:"tone"`
42 | 	LanguagePref   string   `yaml:"language_preference"`
43 | }
44 | 
45 | // ContextConfig holds chat context settings.
46 | type ContextConfig struct {
47 | 	MaxHistoryMessages int `yaml:"max_history_messages"`
48 | 	MaxContextTokens   int `yaml:"max_context_tokens"`
49 | }

</error_details>

And another one:

<error_details>
No sufficiently similar match found (83% similar, needs 100%)

Debug Info:
- Similarity Score: 83%
- Required Threshold: 100%
- Search Range: start to end
- Tried both standard and aggressive line number stripping
- 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

Search Content:
:start_line=195
-------
	// Check for /reset command
	if strings.HasPrefix(body, "/reset") {
		if evt.Sender.String() == cfg.Matrix.UserID {

Best Match Found:
193 | 	}
194 | 
195 | 	// Check for /reset command
196 | 	if strings.HasPrefix(body, "/reset") {
197 | 		if evt.Sender.String() == cfg.Matrix.UserID {

Expected result

The error message explains that the mandatory :start_line:N line is missing and proposes the correct format.

Actual result

apply_diff complains about insufficient match

Variations tried (optional)

No response

App Version

Version: 3.53.0 (44f6bba)

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 identified the root cause. Here's what's happening and my proposed fix:

Root Cause

The apply_diff regex in src/core/diff/strategies/multi-search-replace.ts expects the :start_line: marker with a colon delimiter (e.g., :start_line:18), but some models (especially smaller/quantized ones like the Qwen model mentioned) output :start_line=18 with an equals sign instead.

When the regex fails to match the :start_line=18 pattern as metadata, that line (plus the ------- separator) gets captured as part of the search content itself. So the tool ends up searching for:

:start_line=18
-------
// MatrixConfig holds Matrix connection settings.
type MatrixConfig struct {
...

...instead of just the actual code. The extra lines at the top reduce the Levenshtein similarity score (92% in the first example, 83% in the second), causing the "no sufficiently similar match" error even though the code portion matches perfectly.

Proposed Fix

1. Update the main diff-parsing regex in MultiSearchReplaceDiffStrategy.applyDiff() to accept both : and = as delimiters for :start_line: and :end_line: markers. Change \:start_line:\s*(\d+) to \:start_line[:=]\s*(\d+) (and same for :end_line:).

2. Update validateMarkerSequencing() to also recognize =-delimited markers when checking for invalid markers in REPLACE sections (lines 194-199).

3. Add test cases covering the = delimiter variant to confirm the fix works.

This is a small, targeted change (just the regex patterns) that makes the parser more tolerant of model output variations without changing any behavioral semantics.

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

[rkfg]

While this is true, the core issue wouldn't be solved. According to

Roo-Code/src/core/prompts/tools/native-tools/apply_diff.ts

Line 5 in ad25634

const DIFF_PARAMETER_DESCRIPTION = `A string containing one or more search/replace blocks defining the changes. The ':start_line:' is required and indicates the starting line number of the original content. You must not add a start line for the replacement content. Each block must follow this format:

:start_line: is a required parameter, if it's not found the tool should clearly report that as an error instead of ignoring the missing parameter and using the whole block as a code fragment to search. Such an error would let the model to quickly recover instead of getting lost. Even I didn't notice the argument format error!

[rkfg]

Here's what I propose instead, @roomote tell me what you think.

Plan: Fix :start_line: Validation in ApplyDiffTool

1. Problem Description

The apply_diff tool accepts a diff parameter that uses a multi-search-replace format with <<<<<<< SEARCH, =======, and >>>>>>> REPLACE markers. The :start_line: directive (specifying the exact line number in the original file where the search block starts) is documented as required in the tool prompt (src/core/prompts/tools/native-tools/apply_diff.ts:5: "The :start_line: is required and indicates the starting line number of the original content"), but the implementation does not enforce this requirement.

The regex in src/core/diff/strategies/multi-search-replace.ts:292 makes :start_line: optional via the group ((?:\:start_line:\s*(\d+)\s*\n))? (with trailing ?). When :start_line: is absent:

• The startLine defaults to 0 (src/core/diff/strategies/multi-search-replace.ts:310)
• The search content may be empty or match incorrect lines
• Fuzzy matching occurs across the entire file without a starting position
• No clear error message informs the user that :start_line: is missing

Note: All file references in this document use backtick format (path:line) instead of clickable links, for compatibility with GitHub issue rendering.

2. Problem Location in Code

File	Lines	Issue
src/core/prompts/tools/native-tools/apply_diff.ts	5	Prompt documents :start_line: as required, but enforcement is missing in implementation
src/core/diff/strategies/multi-search-replace.ts	292	Regex pattern has ? making :start_line: group optional
src/core/diff/strategies/multi-search-replace.ts	268-269	Comment says "Optionally matches" for :start_line:
src/core/diff/strategies/multi-search-replace.ts	310	startLine defaults to 0 when :start_line: is missing
src/core/tools/ApplyDiffTool.ts	77	parseInt(params.diff.match(/:start_line:(\d+)/)?.[1] ?? "") — silently falls back to empty string
src/core/diff/strategies/multi-search-replace.ts	101-243	validateMarkerSequencing does not track or enforce :start_line: presence

3. Fix Plan

3.1. Add Explicit :start_line: Check in src/core/tools/ApplyDiffTool.ts

In src/core/tools/ApplyDiffTool.ts:71-81, before calling task.diffStrategy?.applyDiff(), add validation that :start_line: is present in the diff content:

// Before line 74, add:
if (!/:start_line:\s*\d+/.test(diffContent)) {
    task.consecutiveMistakeCount++
    task.recordToolError("apply_diff")
    pushToolResult(await task.sayAndCreateMissingStartLineError(diffContent))
    return
}

This requires creating a new helper method sayAndCreateMissingStartLineError on the Task class (see 3.4).

Note on error format: The existing error messages in src/core/diff/strategies/multi-search-replace.ts:121,145,160 use the ERROR: prefix pattern. However, other tools in the codebase (e.g., src/core/tools/ApplyDiffTool.ts:64, src/core/tools/EditFileTool.ts:242) use <error_details> XML tags. For consistency with the apply_diff tool's existing error format, the new error should use <error_details> tags instead of ERROR: prefix.

3.2. Add :start_line: Presence Tracking in validateMarkerSequencing

In src/core/diff/strategies/multi-search-replace.ts:101-243, update the validateMarkerSequencing method:

1. Add a new state tracking variable:

   let foundStartLine = false

2. In the State.AFTER_SEARCH case (around line 214-220), detect :start_line::

   case State.AFTER_SEARCH:
       if (SEARCH_PATTERN.test(marker)) return reportInvalidDiffError(SEARCH_PATTERN.source, SEP)
       if (marker.startsWith(SEARCH_PREFIX)) return reportMergeConflictError(marker, SEARCH)
       if (marker === REPLACE) return reportInvalidDiffError(REPLACE, SEP)
       if (marker.startsWith(REPLACE_PREFIX)) return reportMergeConflictError(marker, SEARCH)
       if (/:start_line:\s*\d+/.test(marker.trim())) foundStartLine = true
       if (marker === SEP) {
           if (!foundStartLine) {
               return reportInvalidDiffError(SEP, ":start_line: <line_number>")
           }
           state.current = State.AFTER_SEPARATOR
       }
       break

Note: Instead of a custom error message with ERROR: prefix, use the existing reportInvalidDiffError helper which already includes the "CORRECT FORMAT" section with :start_line: documentation (src/core/diff/strategies/multi-search-replace.ts:148-154).

3.3. Update Regex in applyDiff To Make :start_line: Mandatory

In src/core/diff/strategies/multi-search-replace.ts:292, change the regex pattern:

Before:

/(?:^|\n)(?<!\\)<<<<<<< SEARCH>?\s*\n((?:\:start_line:\s*(\d+)\s*\n))?((?:\:end_line:\s*(\d+)\s*\n))?((?<!\\)-------\s*\n)?([\s\S]*?)(?:\n)?(?:(?<=\n)(?<!\\)=======\s*\n)([\s\S]*?)(?:\n)?(?:(?<=\n)(?<!\\)>>>>>>> REPLACE)(?=\n|$)/g

After:

/(?:^|\n)(?<!\\)<<<<<<< SEARCH>?\s*\n\:start_line:\s*(\d+)\s*\n((?:\:end_line:\s*(\d+)\s*\n))?((?<!\\)-------\s*\n)?([\s\S]*?)(?:\n)?(?:(?<=\n)(?<!\\)=======\s*\n)([\s\S]*?)(?:\n)?(?:(?<=\n)(?<!\\)>>>>>>> REPLACE)(?=\n|$)/g

Key changes:

• Remove ? from ((?:\:start_line:\s*(\d+)\s*\n))? → make it \:start_line:\s*(\d+)\s*\n (mandatory)
• Update capture group indices: match[1] is now the start line number (was match[2])
• Update all subsequent capture group references (match[6] → match[5], match[7] → match[6])

Also update the comment at lines 268-269:

// 3. \:start_line:\s*(\d+)\s*\n
//    Matches the required ":start_line:" line. Group 1 is the line number.

3.4. Add sayAndCreateMissingStartLineError Helper Method

Add a new method to the Task class (src/core/task/Task.ts) following the pattern of existing src/core/task/Task.ts:1869 method (sayAndCreateMissingParamError):

async sayAndCreateMissingStartLineError(diffContent: string) {
    await this.say(
        "error",
        "Roo tried to use apply_diff without value for required parameter ':start_line:' in diff. Retrying...",
    )
    
    const formattedError = `<error_details>\n:start_line: is required in your diff content.\n\n` +
        "The ':start_line:' directive specifies the exact line number in the original file where the search block starts.\n\n" +
        "CORRECT FORMAT:\n\n" +
        "<<<<<<< SEARCH\n" +
        ":start_line:42\n" +
        "-------\n" +
        "[exact content to find including whitespace]\n" +
        "=======\n" +
        "[new content to replace with]\n" +
        ">>>>>>> REPLACE\n" +
        "\n" +
        "Your diff content:\n" +
        "```\n" +
        diffContent.slice(0, 500) + (diffContent.length > 500 ? "..." : "") + "\n" +
        "```\n" +
        "\n</error_details>"
    
    return formattedError
}

3.5. Update Error Message at Line 299

In src/core/diff/strategies/multi-search-replace.ts:299, update the error message to reflect that :start_line: is now required:

Before:

error: `Invalid diff format - missing required sections\n\nDebug Info:\n- Expected Format: <<<<<<< SEARCH\\n:start_line: start line\\n-------\\n[search content]\\n=======\\n[replace content]\\n>>>>>>> REPLACE\n- Tip: Make sure to include start_line/SEARCH/=======/REPLACE sections with correct markers on new lines`,

After:

error: `Invalid diff format - missing required sections\n\nDebug Info:\n- Expected Format: <<<<<<< SEARCH\\n:start_line: <line_number>\\n-------\\n[search content]\\n=======\\n[replace content]\\n>>>>>>> REPLACE\n- Tip: Make sure to include :start_line:, SEARCH, -------, SEARCH content, =======, REPLACE content, and REPLACE sections with correct markers on new lines`,

4. Expected Error Message

When :start_line: is missing, the user should see (using <error_details> format consistent with src/core/tools/ApplyDiffTool.ts:64 and src/core/tools/EditFileTool.ts:242):

<error_details>
:start_line: is required in your diff content.

The ':start_line:' directive specifies the exact line number in the original file where the search block starts.

CORRECT FORMAT:

<<<<<<< SEARCH
:start_line:42
-------
[exact content to find including whitespace]
=======
[new content to replace with]
>>>>>>> REPLACE

Your diff content:

[actual diff content snippet]

</error_details>

5. Test Coverage

Add tests in src/core/diff/strategies/__tests__/ for:

1. validateMarkerSequencing rejects diff without :start_line:
2. applyDiff rejects diff without :start_line: (regex returns no matches)
3. ApplyDiffTool emits proper error message when :start_line: is missing
4. Valid diff with :start_line: continues to work correctly

6. Files to Modify

File	Changes
src/core/tools/ApplyDiffTool.ts	Add :start_line: validation before applyDiff call
src/core/diff/strategies/multi-search-replace.ts	Update regex, add tracking in validateMarkerSequencing, update error messages
src/core/task/Task.ts	Add sayAndCreateMissingStartLineError helper (following src/core/task/Task.ts:1869 pattern)
src/core/diff/strategies/__tests__/multi-search-replace.spec.ts	Add tests for mandatory :start_line: validation

7. Reference: Tool Prompt

The tool prompt at src/core/prompts/tools/native-tools/apply_diff.ts already documents :start_line: as required:

const DIFF_PARAMETER_DESCRIPTION = `A string containing one or more search/replace blocks defining the changes. The ':start_line:' is required and indicates the starting line number of the original content. You must not add a start line for the replacement content. Each block must follow this format:
<<<<<<< SEARCH
:start_line:[line_number]
-------
[exact content to find]
=======
[new content to replace with]
>>>>>>> REPLACE`

[roomote-v0]

I've seen the plan and will implement it. I'll open a PR to share progress shortly.

[roomote-v0]

Opened a draft PR with my attempt at implementing the plan: #12200. Feedback welcome!

[rkfg]

@roomote built the extension with this patch and tried it in the session where I had this problem. After a couple of retries Roo was able to recover and use the correct format! However, this patch changes many tests where :start_line: was missing so a human review is needed to decide whether this parameter should be mandatory (as the prompt suggests) or not (as the tests suggest). In any case a botched parameter use should be handled and reported properly because it confuses even humans.

[roomote-v0]

I've seen the feedback from @rkfg and will look into improving the handling of :start_line: - making sure a missing or botched :start_line: is handled gracefully with a clear error message rather than a confusing similarity mismatch. I'll open a PR to share progress.

[roomote-v0]

Opened a draft PR with an alternative approach at #12201. Instead of making :start_line: mandatory (as in PR #12200), this takes a tolerant parsing approach -- the regex now accepts both :start_line: and :start_line= formats, and a fallback strips leaked directives from search content. This keeps backward compatibility and handles botched directives gracefully. Feedback welcome!

[rkfg]

I changed the title and description to better reflect the real issue.