Agent runtime / memory systems — small upstream fixes that survive review, plus local-first Agent tooling.
volcengine/OpenViking — external contributor (merged):
| PR | Focus |
|---|---|
| #4702 | pi-extension: watermark + drain budget (review follow-up) |
| #4643 | dsh-plugin: parallel profile + recall on pre-step |
| #4608 | opencode-plugin: isolate flushAll session failures |
| #4040 | rerank: send top_n on OpenAI-compatible path |
Also shipping deeper threads under review (e.g. pending-queue batch replay, path filters in staging) — design iteration with maintainers, not drive-by patches.
Occasional fixes elsewhere: Tencent/ncnn#6906.
| Repo | One-liner |
|---|---|
| traj-wash | Find failure onset → wash cascade fails → export clean Agent trajectories |
| handoff-pack | Compress OpenAI-compatible messages[] into a paste-ready model handoff pack |
| AIFlowHub | NL → executable AI workflows / agents (skill routing, validation, private deploy) |
| Site Agent | Guided navigation Agent (SSE, context engineering, allowlisted deep-links) |
Python · TypeScript · Agent / LLM apps · memory & tool loops · product+eng ownership
Prefer merged, reviewed work over open PR volume.
