Repository navigation
[P0] Implement bounded verify→repair self-healing loop #9
Description
Problem
The proposal's core innovation is an iterative verification loop that repairs failing tests until pass. Reality: core/generator.py:285-322 runs tests once, records tc.diagnosis, and never applies a fix or re-runs. 'Self-healing' is currently a no-op during generation.
Tasks
- Add bounded loop in generator phase 3 (e.g.
max_repair_iters, default 3). - On
TEST_WRONGdiagnosis: applydiagnosis.suggested_fix, regenerate test file, re-run. - On
CODE_BUG: flag, stop repairing (do not mutate test to force pass). - On
INTENT_WRONG: surface for intent revision (interactive) or skip (auto). - Track iteration count + final status in
VerificationResult.
Acceptance
A deliberately-wrong generated test gets diagnosed, auto-fixed, and passes on re-run within max iters; loop terminates and is unit-tested.
Activity
pro-utkarshM commentedon Oct 15, 2026
Done on
master(code2test/core/generator.py,verify_and_fix). The loop works like this:max_fix_attemptsrounds (default 2), and are re-run each time.CODE_BUG: the test is kept withpytest.mark.xfailand the diagnosis as its reason. It is never edited to force a pass.TEST_WRONGorINTENT_WRONG: the test is dropped in--automode. Interactive mode shows the diagnosis and lets the user decide.This differs slightly from the plan: repairs use the pytest output directly rather than the diagnosis's
suggested_fix, and diagnosis comes after repair. Repair rounds are recorded per file in.code2test/metrics.json, not inVerificationResult.Unit tests in
tests/test_pipeline.pycover it:test_failing_test_is_repaired_using_failure_output,test_bad_import_is_repaired,test_unfixable_test_suspected_code_bug_is_kept_as_xfail,test_unfixable_wrong_test_is_dropped. Real-model runs are inresearch/data/.Generated by Claude Code