Checkable Coding Agent Controls

Repository agents need a thinner interface between exploration and editing. SWE-Edit gives a practical design: a Viewer returns only task-relevant code blocks for a file and query, while an Editor applies a patch from a natural-language edit instruction. The main agent keeps the bug reasoning and fix plan, with less file-search residue in its context and fewer brittle find-and-replace failures.

The reported gains are modest on issue resolution and more concrete on cost and edit reliability. On SWE-bench Verified, SWE-Edit raises resolved rate from 69.9% to 72.0%, cuts inference cost by 17.9%, and raises edit success from 93.4% to 96.9%. The Viewer returns 39.7% of requested file content on average, which is a direct way to test whether the interface is reducing code surface without hiding the needed lines.

A team running coding agents on an internal monorepo can test this without changing the model: add a file-view endpoint that takes a path and query, add a patch endpoint that accepts an edit instruction, and compare agent runs on recent bugs. Track resolved issues, failed patches, non-cached input tokens, and reviewer complaints about unrelated edits.

Coding agents can pass lint and typecheck while missing team decisions stored outside the repo. The practical build is a PR gate that turns product and engineering decisions into explicit checks: required audit logging, approved UI components, feature flags, ORM choices, auth wrappers, and deprecated patterns. The agent can receive retrieved decisions during spec drafting and mid-build review, and the final diff can be scored against the same decision list.

Context-Augmented Code Generation reports a large gain on a small clean-room benchmark: Claude Code with codebase access alone scores 46% weighted decision compliance, while Claude Code plus Brief scores 95%. The setup also changes the workflow by adding specs, acceptance criteria, and mid-build guidance, so the safer takeaway is operational: teams should measure whether their agent PRs follow stored decisions, not only whether the code compiles.

A cheap first version can cover one product area and ten recurring decisions. Store each decision with an owner, examples, and a regex or review checklist. Run the coding agent on old tickets, then compare generated diffs against the checklist before and after retrieval.

SAST adoption is blocked by alert queues that developers stop trusting. QASecClaw points to a bounded use of a coding model: keep Semgrep as the high-recall scanner, send each finding with CWE type, file location, and source context to an LLM reviewer, and suppress only findings judged false positive. If the model times out or returns malformed JSON, keep the original Semgrep alert.

On OWASP Benchmark v1.2, QASecClaw reduces false positives from 560 to 64 and raises F1 from 78.39% to 90.93%, with recall down by 3.1%. That tradeoff is practical for security teams that already review Semgrep output, because the model is judging a known candidate alert instead of searching the full codebase.

The first deployment should run in shadow mode on existing CI alerts. Compare model-suppressed findings with human triage for a few CWE categories, keep all suppressed alerts visible in an audit view, and require structured reasons tied to nearby sanitization, validation, encoding, or parameterized APIs.