[simurg79]

Problem

When apply_diff was called with multiple diff blocks and more than one block failed to apply, only the last failure was surfaced back to the model. Every earlier failure was silently overwritten. The model therefore received an error message describing a single block, with no indication that other blocks had also failed, and would typically retry blindly — often resending the blocks that had already applied successfully, or "fixing" the one block it could see while the real breakage lay elsewhere.

A closely related failure mode made this worse: a common way to break a diff block is to mistype the :start_line: marker (wrong spelling, missing colons, stray indentation). That produces a generic "search content not found" error that gives the model no signal about the actual cause, so it retries with the same malformed marker.

Defects fixed

1. Only the last failure reached the model.
All failed blocks are now reported together. Each is labeled with a --- Diff block N of M --- header so the model can map a failure back to the specific block it sent, with bounded per-block detail (see caps below).

2. Mistyped start-line markers produced an unhelpful generic error.
When a failed block's first line looks like a mistyped start-line marker, the error now includes a LIKELY CAUSE hint naming that as the probable problem, instead of only reporting that the search content was not found.

3. Partial-success messaging was ambiguous.
The partial-success path now reads Applied X of Y diff blocks and explicitly instructs the model to resend only the failed blocks. Previously the wording invited resending the entire diff, which risked re-applying already-applied blocks.

Bounded output

Reporting every failure risks flooding the context window, so the output is capped:

• a maximum number of individually reported failed blocks, with a summary line accounting for any remainder;
• a per-block cap on the amount of failure detail emitted;
• truncation of overlong content within a block's detail.

This keeps the worst case bounded regardless of how many blocks were sent or how large each one is.

Validation

• 90/90 tests pass, including two new suites:
  • src/core/tools/__tests__/applyDiffTool.failureReporting.spec.ts
  • src/core/diff/strategies/__tests__/multi-search-replace-marker-hint.spec.ts
• check-types clean across all packages.
• ESLint clean (--max-warnings=0).
• Prettier clean.

Known limitation

searchBlocks is counted with a regex over occurrences of the literal <<<<<<< SEARCH. If that literal appears inside a block's content (for example, a diff that is itself editing diff-handling code or documentation about the diff format), the count will be too high. This value is advisory only — it is used for the N of M labels and the Applied X of Y summary, and does not drive matching or application logic, so an overcount degrades only the wording of the message and cannot cause an incorrect edit.