Background

This replaces #889, which proposed runtime decoding of \uXXXX escape sequences in the edit/write pipeline. The Codex review on that PR (inline comment on normalize.ts:237) correctly identified that runtime decoding is unsound: it makes it impossible to write a single literal 6-char \u2192 source-code sequence to disk, because every viable JSON tool argument either decodes to the character or ends up with extra backslashes.

That breaks JS regex source (/\u2192/), Python raw strings (r"\u2192"), JSON fixtures, and any code that contains \uXXXX literally.

Approach

Document the convention in tool prompts instead of decoding at runtime. JSON natively decodes \uXXXX already, so:

• For the character → — emit "\u2192" (one backslash) in the JSON, or the literal → character. Both arrive at the tool as →.
• For the literal 6-char \u2192 (regex source, raw strings, fixtures) — emit "\\u2192" (two backslashes) so the JSON parser delivers the 6 chars verbatim.
• NEVER emit "\\u2192" (two backslashes) when you intend the character. That writes literal text, not a Unicode character.

Add a <unicode-content> section to four tool prompts spelling out both directions and the negative example:

• write.md (content)
• replace.md (old_text / new_text)
• patch.md (diff — + lines and op:create payloads)
• hashline.md (content)

Files

packages/coding-agent/CHANGELOG.md                  | 3 +++
packages/coding-agent/src/prompts/tools/hashline.md | 7 +++++++
packages/coding-agent/src/prompts/tools/patch.md    | 7 +++++++
packages/coding-agent/src/prompts/tools/replace.md  | 7 +++++++
packages/coding-agent/src/prompts/tools/write.md    | 7 +++++++
5 files changed, 31 insertions(+)

Why this is better than runtime decoding

• No silent corruption of valid source code. /\u2192/ written to a .ts file stays as 6 chars; the tool's job is to write what JSON delivered, not to second-guess it.
• No new failure mode. Runtime decoding meant LLMs that already emitted characters correctly would now have a hidden transformation between their intent and the file.
• Teaches the convention. Anthropic-style tool descriptions are read attentively by the LLM. A clear "do this, not that" instruction trains future calls; a runtime decoder hides the rule.
• Reversible. If experience shows LLMs ignore the guidance and persistently double-escape, we can revisit. A documentation change is cheap to evolve.

Verification

• bun run format-prompts clean (no formatting changes needed)
• All four prompts use the actual U+2192 character in prose where the character is intended, and the literal 6-char escape sequence in code spans where the literal text is intended (verified via xxd)