fix: FINAL() callable in REPL; parser ignores FINAL/FINAL_VAR inside code fences

[jkbrooks]

Problem

Two bugs caused the RLM to silently return wrong answers when a model chose to call FINAL() or FINAL_VAR() inside a ```repl ``` code block — a natural pattern given the system prompt examples.

Bug 1 – Parser (utils/parsing.py)

find_final_answer() ran its FINAL(...) / FINAL_VAR(...) regex over the raw assistant response, including inside fenced code blocks. So when the model wrote:

```repl
FINAL(final_answer)
```

the parser matched FINAL(final_answer) and returned the literal string "final_answer" as the completion response instead of the variable's value. No error was raised — the RLM just silently returned a wrong string.

Bug 2 – Runtime (environments/local_repl.py, environments/base_env.py)

The system prompt offers FINAL(value) as option 1 for submitting a final answer:

1. Use FINAL(your final answer here) to provide the answer directly

But FINAL was never injected into the REPL globals — only FINAL_VAR was. So when the model called FINAL(x) inside a repl block it got a NameError, and the REPL stderr was shown to the model, wasting an iteration. The only reason the run didn't always fail is that the parser still picked up FINAL(...) from the response text — leading to bug 1 above.

Fix

utils/parsing.py — strip all fenced code blocks from the response before running the FINAL/FINAL_VAR regex. This ensures only prose-level FINAL(...) signals termination.

environments/local_repl.py — add _final(value) method (mirrors _final_var for direct values), inject it as globals["FINAL"], and restore it in _restore_scaffold().

environments/base_env.py — add "FINAL" to RESERVED_TOOL_NAMES so it can't be overwritten by user code.

Tests added (tests/test_parsing.py)

Test	Covers
test_final_inside_repl_code_block_not_parsed_as_terminal	Parser ignores FINAL() in code fence
test_final_var_inside_repl_code_block_not_parsed_as_terminal	Parser ignores FINAL_VAR() in code fence
test_final_in_prose_still_works_alongside_repl_block	Prose FINAL() still terminates after code fence
test_final_callable_in_repl_environment	FINAL(x) callable in REPL sets final_answer

All 16 TestFindFinalAnswer tests pass.

How to reproduce (before fix)

from rlm import RLM
rlm = RLM(backend="openai", backend_kwargs={"model_name": "gpt-4o-mini"}, environment="local", max_depth=1)
result = rlm.completion("Compute the sum of integers from 1 to 10 and print only the number.")
print(result.response)  # prints "final_answer" instead of "55"

Made with Cursor

[Copilot]

Pull request overview

Fixes incorrect termination behavior when models emit FINAL(...) / FINAL_VAR(...) inside fenced REPL blocks, and makes FINAL(...) callable inside the local REPL to align runtime behavior with the system prompt.

Changes:

• Update find_final_answer() to ignore FINAL(...) / FINAL_VAR(...) that appear inside fenced code blocks.
• Inject FINAL into LocalREPL globals and restore it after executions; reserve FINAL as a non-overridable tool name.
• Add parsing + REPL runtime tests covering the regressions.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
tests/test_parsing.py	Adds regression tests ensuring fenced FINAL/FINAL_VAR are ignored and FINAL() is callable in LocalREPL.
rlm/utils/parsing.py	Strips fenced code blocks before scanning for terminal FINAL/FINAL_VAR markers.
rlm/environments/local_repl.py	Adds _final() helper and injects/restores FINAL in REPL globals.
rlm/environments/base_env.py	Adds FINAL to RESERVED_TOOL_NAMES to prevent override by custom tools/user code.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

text_no_code = re.sub(r"```[\s\S]*?```", "", text) strips all fenced code blocks from the entire response before searching for FINAL(...). This can silently corrupt a legitimate final answer if the payload inside FINAL(...) includes a markdown code fence (e.g., returning a code snippet), because the fenced content will be removed before extraction. Consider narrowing the stripping to only repl blocks (the ones the runtime executes), or performing code-fence removal only outside the matched FINAL(...)/FINAL_VAR(...) span so the final payload is preserved verbatim.

Suggested change

	text_no_code = re.sub(r"```[\s\S]*?```", "", text)
	text_no_code = re.sub(r"```repl[\s\S]*?```", "", text)