Skip to content

context-safe-ratchet-repair.md

Latest commit

 

History

History
78 lines (58 loc) · 2.34 KB

context-safe-ratchet-repair.md

File metadata and controls

78 lines (58 loc) · 2.34 KB
Raw

Context-Safe Ratchet Repair

This case study uses the examples/context-safe-ratchet fixture to show the repair loop for legacy code under bounded diagnostics.

Goal

Show that Agent Maintainer gives an agent exact next commands and compact failure facts instead of requiring raw log dumps in chat.

Fixture

  • Repository: examples/context-safe-ratchet
  • Known failure: price_for_order accepts negative quantities
  • Verification profile: precommit
  • Expected failing check: pytest-coverage

Commands

Run from a temporary copy of the example:

git init
git add .
git commit -m "example baseline"
agent-maintainer ratchet baseline create --base-ref HEAD --force
agent-maintainer ratchet status --base-ref HEAD
agent-maintainer ratchet next --base-ref HEAD --limit 3
agent-maintainer verify --profile precommit --base-ref HEAD

Then expand only the failing check:

agent-maintainer context --log-dir .verify-logs/runs/<run-id> failures \
  --check pytest-coverage --limit 20
agent-maintainer context --log-dir .verify-logs/runs/<run-id> log \
  pytest-coverage --tail 120

Measured Result

The baseline command created:

.agent-maintainer/ratchet-baseline.json

The ratchet status reported no new, worsened, unchanged, improved, or resolved findings for this clean baseline. The verifier then failed exactly one check:

Metric Result
Profile precommit
Failed checks pytest-coverage
Test result 2 tests, 1 failure
Failing test tests.test_big::test_price_for_order_rejects_negative_quantity
Failure fact Failed: DID NOT RAISE ValueError
Coverage total 71%
Duration about 2 seconds

The verifier summary included expansion commands pointing at the run-scoped diagnostic directory instead of printing the entire pytest log.

Agent Lesson

Use the run id from the verifier summary, inspect the bounded failure facts, fix the behavior, rerun the focused failing test, then rerun the relevant verification profile.

Limitations

The example is intentionally tiny. It validates the context-safe repair loop and command shape, not an end-to-end production legacy migration.