code_execution parser rejects generated code embedding shell/CSS/markdown in String.raw/template literals (6 parse-time failures)

[Broccolito]

Summary

The code_execution / execute_code sandbox rejects the model-generated JS/TS at parse time ("Parse error: SyntaxError…") whenever the code embeds shell, CSS, or markdown inside a String.raw template literal or an object literal. The tool call never runs — it is thrown out before execution — so the agent silently loses a whole step and has to retry. This was the single most common tool-call failure in the audited transcripts (6 occurrences, all in the "starsign master" session), and it repeatedly blocked the website-build and skill-creation work.

How it was found

Auditing three exported session transcripts (knowledge conversation, testing conversation 2, starsign master conversation) for tool-call failures. This issue collects the parse-time rejections.

Every occurrence (session: starsign master conversation)

#	Exact error	Triggering construct in the generated code
1	Parse error: SyntaxError: expected token '}', got ':' in template literal at line 11, col 16	A bash parameter-expansion "${!v:-}" written **inside a String.raw\…`template literal** (a shellfor v in …; do … ${!v:-} …heredoc). The JS parser tries to interpret${…}as a JS template substitution and chokes on the bash:`.
2	Parse error: SyntaxError: expected one of ',' or '}', got 'prefers' in object literal at line 136, col 112	CSS @media (prefers-color-scheme: dark) text sitting where the parser expected a JS object literal.
3	Parse error: SyntaxError: unterminated string literal	Multi-line Python heredoc with embedded quotes inside String.raw.
4	Parse error: SyntaxError: unterminated string literal	Same class — heredoc/quoting inside a String.raw block.
5	Parse error: SyntaxError: expected one of ';' or 'line terminator', got 'value' in lexical declaration binding list at line 399, col 87	Large String.raw payload with stray tokens confusing the const … = … binding parse.
6	Parse error: SyntaxError: expected one of ';' or 'line terminator', got 'index' in lexical declaration binding list at line 51, col 15	A String.raw work-summary payload containing an unescaped path fragment (a/Users/wanjun/Desktop/starsign master) parsed as code.

Each of these returned [tool_error kind=tool_failure retryable=false] before any code executed.

Why this is a product bug (not just model error)

String.raw\…` exists precisely so authors can embed arbitrary text (shell, CSS, regex, markdown) without escaping. The recurring failure mode is that **${…}sequences and stray tokens insideString.rawtemplate literals are still being interpreted by the JS/TS parser** instead of treated as raw text, so any shell heredoc using${VAR}/${!v:-}or any embedded CSS/markdown breaks the parse. Becauseretryable=false`, the harness gives up rather than falling back.

Suggested investigation

• Confirm how the execute_code parser handles ${…} and back-tick content inside String.raw — raw template literals should not have their ${} substitutions validated as JS.
• Consider a safer inbound-code path for large embedded blobs (e.g. accept the shell/heredoc body as a normal string argument rather than re-parsing it), or a clearer error that points at the offending embedded construct.
• Given how often the model reaches for shell({command: String.raw\python3 - <<'PY' … PY`})`, this pattern needs to parse reliably.

Environment

• BioRouter desktop, model gpt-5.5-2026-04-24 (Azure), code_execution extension enabled.
• Evidence: exported starsign master conversation session transcript.