fix(coders): catch doubled-prefix hallucinations in edit-block headers#5112
Open
ekrembasari wants to merge 1 commit into
Open
fix(coders): catch doubled-prefix hallucinations in edit-block headers#5112ekrembasari wants to merge 1 commit into
ekrembasari wants to merge 1 commit into
Conversation
Small editor models occasionally emit edit-block filename headers with the chat-file's own prefix duplicated -- e.g. ".claude/.claude/foo.json" when the chat file is ".claude/foo.json". The existing prepended-bogus-dir guard in WholeFileCoder.get_edits and find_filename only covers the case where the LLM emits just the basename; multi-segment doubled prefixes fall through, abs_root_path concatenates blindly, and the file lands at a doubled path while the canonical path stays empty. Extend both code paths with progressive suffix-stripping against chat_files / valid_fnames. Strict subset of behaviour: only triggers when exact and basename matches both fail, and only resolves when a deterministic suffix is itself in the chat-files list. Tests cover: - WholeFileCoder: chat_files retains a "subdir/" prefix (two files in distinct subdirs so find_common_root resolves to the tempdir root); LLM emits doubled "subdir/subdir/sample.txt"; canonical path edited, doubled path never created. - find_filename: doubled prefix where SequenceMatcher ratio falls below the 0.8 fuzzy-match cutoff (0.778 for "sub/dir/sub/dir/foo.py" vs "sub/dir/foo.py"), so neither basename nor fuzzy match recovers it.
|
Ekrem Başarı seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account. You have signed the CLA already but the status is still pending? Let us recheck it. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Catch doubled-prefix hallucinations like
.claude/.claude/foo.json(when the chat file is.claude/foo.json) by extending the existing prepended-bogus-dir guard with progressive suffix-stripping againstchat_files/valid_fnames. Closes #5111.Why
Observed in production dispatch across small editor models (
groq/llama-3.3-70b-versatile,openrouter/openai/gpt-oss-120b,gemini-2.5-flash-lite). Existing basename-only guard inWholeFileCoderand basename+fuzzy guards infind_filenamemiss multi-segment doubled prefixes; result is a 0-byte canonical file and content at a doubled path.Changes
aider/coders/wholefile_coder.py: after basename guard, try progressively shorter suffixes againstchat_files.aider/coders/editblock_coder.py:find_filename: same pattern, between basename and fuzzy match (catches cases whereSequenceMatcherratio falls below the 0.8 cutoff).Risk
Low. Strict subset of behaviour:
chat_files/valid_fnames.Test plan
pytest tests/basic/test_wholefile.py tests/basic/test_editblock.py— 38 passed (was 36 before)