Stacktrace · runtime detection
What Stacktrace detects and measures for Claude Code, Devin, Codex and Cursor today, and what cannot be supported for each agent and why. Cursor is read by OpenAIDR 0.8.0's Cursor reader.
† waits on a field not yet confirmed in a real session. cost-personal-anomaly runs in Fleet on uploaded usage, so it works wherever the usage upload does. Codex grades are the built code's, verified end to end over synthetic Codex homes. Cursor grades come from the stacktrace-demo policy suite, run through the real Cursor CLI and a stand-in model: 13 of 17 cases passed. Both rm -rf cases were refused by Cursor's Delete protection, which that build always turns on.
Claude Code has no detection gaps. The Claude Code plugin and the playground are Claude-only by scope, not by a missing record.
Source of truth: docs/agent-support.md in stacktrace-ai/stacktrace. The Codex evidence is Codex CLI 0.163.0 run against a mock model and its source at rust-v0.163.0, and a live run on gpt-6.1-sol against a local Fleet. The Cursor evidence is the Cursor CLI 2026.10.19-6bbf2c4 run against a mock model and a stand-in model, and its shipped bundle. The reader designs are docs/specs/codex.md and docs/specs/cursor.md in open-agent-security/openaidr.