I land narrow, tested fixes in upstream computer-vision code, and I keep the decision outside the weights when I can. Detection identifies. A versioned policy decides. A boolean from the thing that just ran is not evidence.
Outside contributor. I do not maintain the projects in the table below.
| Merged | Project | Change |
|---|---|---|
| 2026-09-25 | roboflow/inference#3052 | One run-scoped thread pool for executor-less workflow runs. Shipped in inference v1.7.2. |
| 2026-09-25 | roboflow/inference#3038 | Preserve image metadata on empty Template Matching results. Shipped in v1.7.2. |
| 2026-09-18 | roboflow/inference#2983 | Carry image dimensions through empty Florence-2 detections. |
| 2026-09-11 | roboflow/inference#2892 | Preserve image dimensions on empty VLM detection output. |
| 2026-08-28 | roboflow/inference#2834 | Clear every touched namespace in workflow cache blocks. |
| 2026-08-28 | roboflow/inference#2844 | Resolve inference.Model through the lazy package init (PEP 562). |
| 2026-08-03 | roboflow/roboflow-python#501, #502 | Relax the opencv-python-headless pin; return the single_upload result. |
| 2026-07-27 | tensorflow/tensorflow#122706 | XLA DenseBincount negative-input validation for runtime tensors. |
| 2026-07-09 | ultralytics/ultralytics#25020 | Exclude Sigmoid from TensorRT INT8 quantization. |
| 2026-08-26 | huggingface/peft#3603 | Docs only: troubleshooting for hybrid-architecture target_modules. |
Typo and comment fixes exist and are not listed. Closed, unmerged attempts are not listed either. An npm RFC for install-trust telemetry was closed; the scanner lives in trustcard instead of waiting on the registry.
- trustcard — signed manifests,
TOFU pinning, and a scanner for MCP servers. Published as
mcp-trustcard3.2.0 (2026-10-15). - apple-quality-recognition-engine —
edge grading prototype. A detector proposes regions;
grading_policy.yamlgrades; a human promotes labels and checkpoints. No trained apple checkpoint is in the tree. Not a commercial grader.
- Reproducer, then fix, then a scoped regression test. Not a full-suite claim.
- Judgment stays in config a person can diff.
- Verification has to produce evidence. It does not accept a boolean from the agent that ran the action.
Older public experiments — agent P2P drafts, a macOS cleaner, local agent-OS snapshots, benchmark notes — are unmaintained. If a README and the tree disagree, the tree wins.