Guardrailed Edit Tool

Apply multi-line edits only when lint or syntax checks pass; on failure revert and show error type, proposed snippet, and original content.

A guardrailed edit tool replaces a line range in the open file in one action, grounded in stateful file viewer numbers, then runs a focused linter or syntax check before committing. Clean check → keep the edit and redisplay the viewer centered on the change. Failure → revert, and return a three-part observation: error types, how the edit would have looked, and the original content. Ablations show each part matters: missing error type causes misdiagnosis; missing proposed snippet encourages identical retries; missing original content leaves the model attending to stale transcript text.

Instruct the agent not to re-issue the identical failed edit (failure-derived prompt tips). Guardrails cut common argument hallucination and indentation errors but can force edit orderings when the viewer window cannot show a whole function — weigh that under agent-computer interface design and agent system-level defenses. On SWE-bench Lite, removing lint-before-commit costs ~3 absolute resolve points; removing the dedicated edit action entirely costs ~7.7 — see ACI component ablation. Edit recovery still degrades as consecutive failures accumulate (compound mistake amplification). Cap identical retries in tips and break edit thrash after a few lint failures so the agent re-localizes instead. A clean lint pass does not imply a minimal fix — steer toward idiomatic minimal patch preference after the edit applies. Prefer existing library helpers (to_native_string) over inventing a local bytes.decode branch that still calls the wrong converter — reproduction scripts can go green while the gold idiom remains unused.

---

Sources

• SWE-agent (Yang et al.), pp. 16–30