Skip to content

Latest commit

 

History

History

tomlkit-619

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

Case study: python-poetry/tomlkit issue #619 (live runs, 2026-09-30)

Context: tomlkit #619 - Time/DateTime are missing the stdlib fold argument. FixFork (dev-stage) was pointed at a checkout of tomlkit (8c959b51c7da2734662af759a7c9f95e9a78060a) with the two failing tests from the issue applied (red baseline in tests/test_items.py: 2 failed, 92 passed).

Result summary: two live runs did not produce a fully green patch on their own - the raw outputs below say exactly that. The final patch in final-patch/ was completed after reviewing these runs and verified against tomlkit's full test suite (1060 passed). The runs are published because the tool's reports, raw model replies and event logs are also evidence of where the tool still fails.

Contents

  • fixfork-run-1/ - run before the edit-application fallback existed: report, raw model reply (diagnosis.json), event log, stdout, and the apply-failure analysis (1 of 6 find blocks differed only by one dropped blank line - enough to kill the branch under byte-exact matching).
  • fixfork-run-2/ - run after the locate_edit fallback was added: edits applied, but no branch went green (branch 1 red, branch 2 hit a garbled model block, branch 3 still red after 3 repair rounds).
  • final-patch/ - the completed patch for #619 (10 added lines in tomlkit/items.py plus two tests in tests/test_items.py) and the verification output: tests/test_items.py -> 94 passed; full suite -> 1060 passed (Python 3.12, unprivileged sandbox, pytest only).

Costs (router counts, all responses incl. retries)

  • run 1: ~$0.047 in model calls; run 2: ~$0.098.
  • Model: nvidia/nemotron-3-super-120b-a12b via Nebius Token Factory.

Honesty note

The final patch was not produced end-to-end by the tool. The tool reproduced the failure, identified the right area (constructors and item()/replace() round-trips), and proposed candidate edits; the final patch was then completed and verified by hand against the repository's own suite. FixFork itself gained fixes from this exercise (blank-line-tolerant edit application, added between the two runs; its test suite stood at 156 tests at that time).