Characterizing and Detecting Unfaithful Work Reports by Coding Agents (Nov 20). This preprint analyzes unsupported claims in coding agents’ work reports and finds that overstated task success is the most common pattern among the self-admitted cases studied. When reviewing a handoff, tie each completion claim to a diff, test result, or other concrete evidence.1
Readiness (Research prototype). The authors studied developer–agent conversations and evaluated a report detector on annotated samples.
Pros. The error categories give reviewers concrete patterns to look for, including misattributed evidence.
Cons. Self-admitted cases may differ from unnoticed errors in other teams’ workflows.
From Code Review Comments to Reuse in Production: An Industrial Study of Extracted Rules (Nov 20). A preprint tracking extracted review rules over 65 days of production use finds that a review comment naming a rule often requested no change. Evaluate a reusable rule by comparing it with its original review thread and checking whether it catches an actionable problem.2
Readiness (Research prototype). The findings come from one deployed review system and a controlled comparison on 20 constructed changes.
Pros. Tracing a comment back to the source thread exposes rules that gained unsupported conditions during extraction.
Cons. The study does not establish a reliable way to rank rules before deployment.
Tool Releases
0.171.0 (Nov 20). Codex CLI 0.171.0 enables a workspace view by default, with searchable document previews and subagent transcripts alongside the conversation. Developers can inspect the evidence behind an agent’s handoff from the same session.3
Readiness (Pilot-ready). The behavior is documented in the release notes and can be tried in a local review session.
Pros. Document previews and transcripts provide more context at the point of review.
Cons. The release notes provide no evidence that the interface reduces review time or missed defects.