gemini-cli - 💡(How to fix) Fix Bug: LLM prompt builders corrupt content containing $ sequences [1 pull requests]

What happened?

Several internal prompt builders interpolate user/file content into a template with String.prototype.replace('{placeholder}', value). The replacement argument honors special patterns ($$, $&, $`, $') even when the search is a plain string, so any value containing $ is silently corrupted before the prompt is sent to the model:

For example, a file containing echo $$ > pid.txt is sent to the edit-correction model as echo $ > pid.txt, and content containing $` leaks the prompt template into the <file_content> block.

Affected call sites:

• packages/core/src/utils/llm-edit-fixer.ts (edit self-correction prompt)
• packages/core/src/utils/summarizer.ts (tool-output summarizer)
• packages/core/src/services/sessionSummaryService.ts (session summary)

These routinely receive file contents, shell output, and conversation text, all of which can contain $. The same code shape additionally has a sequential-interpolation issue: a placeholder token (e.g. {current_content}) appearing inside one input value can be re-substituted by a later replacement.

What did you expect to happen?

Content should be inserted into prompts verbatim, regardless of $ sequences or placeholder-looking substrings.

Additional context

A fix is proposed in #27552 (single-pass replacement using a replacer function).