Garfield corpus analysis tracks 3,654 Codex files and 1,499 transcripts
zeeg · x · 2026-07-23
A screenshot shows a systematic analysis of a local Garfield corpus: 3,654 Codex JSONL files were inventoried, 1,499 Garfield-related transcripts were rendered, and 1,259 severity-formatted findings were extracted.
Key observations from the corpus:
- Junior + Peated account for 93.7% of the data.
- Behavior/spec reviews produced 38 high/blocker findings; independent verification produced 30.
- Code comments generated the most findings (142), but none were high severity.
- Test-quality work produced 134 findings, including 14 high severity.
- Only 30 of 53 coordinators used the standard garfield: pass handoff; some were blocked and many lacked a standard outcome.
- 195 historically named subagents remain unclassified, suggesting stable reviewer names may improve longitudinal analysis.
The artifacts list also points to an analysis report, README, manifest, findings JSON, and collection script.
More from coding & agent
- Codex is being prepped as a phone-controlled voice assistant with worker agents — imjustnewatai · 2026-07-23
- Devin’s “unsolved problems” claim gets mocked as just Fable plus a few backend calls — basedjensen · 2026-07-23
- Supabase says AI coding agents have made it the backend for millions of apps — ycombinator · 2026-07-23
- Duke study says better memory, not a bigger model, lifted ARC-AGI-3 scores by 18 points — imjustnewatai · 2026-07-23
- Microsoft's mage-flow trends on Hugging Face with MCP-server and Gradio tags — microsoft · 2026-07-23
- User says they downgraded Claude and upgraded to Codex — AIandDesign · 2026-07-23