I focus on agent development, agent safety, edge inference, and backend development.
90 merged PRs across 18 upstream projects.
| Project | Merged | What the PRs cover |
|---|---|---|
| Agent Swarm (1.0k★) | 41 | Agent runtime reliability: MCP setup retries, worker readiness, credential checks, and configurable script execution |
| Microsoft Agent Framework (14.6k★) | 16 | Agent orchestration correctness: multimodal handoff preservation and explicit mode validation |
| celery (29.2k★) | 5 | Keep the first SUCCESS in the Cassandra backend; Log connection retry traceback at debug level; Fix Manager warning output during retries |
| Qwen Code (29.2k★) | 5 | Managed-agent API error handling for missing tool publications |
| MNN (16.5k★) | 5 | Edge LLM inference and export: Omni request state, attention masks, AWQ calibration, and QNN embeddings |
| NullAway (4.1k★) | 3 | Java null-safety analysis: constructor initialization checks, Maven-injected fields, and GitHub Actions security auditing |
| nanobot (50.0k★) | 2 | Agent provider compatibility and messaging: DeepSeek reasoning semantics and Discord reply support |
| Core ML Tools (5.5k★) | 2 | Model conversion correctness: scalar squeeze values and symbolic upsample output shapes |
| Reticle (1.3k★) | 2 | Browser verification and setup: expected-failure signals and automatic plain HTML integration |
| Filament (20.7k★) | 1 | Vulkan depth/stencil render-target capability detection |
| React Spectrum (16.0k★) | 1 | Calendar year-picker correctness at constrained date-range boundaries |
| CubeSandbox (13.2k★) | 1 | OpenAI Agents SDK integration documentation in English and Chinese |
| AI-Infra-Guard (7.5k★) | 1 | Vulnerability-rule consistency across case-sensitive and case-insensitive filesystems |
| PyRIT (4.8k★) | 1 | Accessibility of selected configuration paths in the AI red-teaming interface |
| fastjson2 (4.4k★) | 1 | Dynamic ClassLoader isolation when multiple fastjson2 copies coexist |
| OSV.dev (3.1k★) | 1 | Vulnerability severity-source attribution on detail pages |
| ONNX Runtime GenAI (1.2k★) | 1 | C# chat-client access to the cached generation state |
| cuOpt (1.1k★) | 1 | Quickstart documentation link maintenance |
All 41 merged Agent Swarm PRs
| PR | What changed |
|---|---|
| #2984 | fix: reject malformed worker concurrency values |
| #2953 | feat(workflows): filter webhook deliveries by header and JSON field |
| #2983 | fix(cli): reject malformed e2b stack wizard integers |
| #2985 | fix: fall back from invalid context-preamble token budgets |
| #2928 | fix: preserve event wait subscriptions when a run resumes |
| #2920 | fix(acp): normalize ACP_TARGET values |
| #2919 | fix: detect GitLab remotes by hostname |
| #2858 | fix(worker): validate provider dependencies at task start |
| #2823 | refactor(entrypoint): move provider binary checks to Bun |
| #2817 | fix: validate Slack tree stall threshold |
| #2740 | perf(workflows): scope definition cleanup to deleted runs |
| #2730 | fix(workflows): validate deployment circuit-breaker limits |
| #2695 | feat(claude): support SDK interrupt and redirect |
| #2694 | fix(scripts): add structuredClone to QuickJS |
| #2648 | fix(runner): remove obsolete YOLO mode |
| #2650 | fix: redact swarm setup links and save credentials privately |
| #2643 | docs: remove obsolete harness verification TTL setting |
| #2618 | fix: correct pi and Codex usage recording |
| #2581 | refactor(entrypoint): move API readiness polling to Bun |
| #2577 | feat(runtime-pools): add optional persistent caches |
| #2575 | fix(extensions): let hooks block and rewrite task steering |
| #2559 | fix(ui): wait for saved model configs before leaving setup |
| #2537 | fix(runner): save the default lead profile on first boot |
| #2539 | fix(ui): use a neutral greeting on Home |
| #2542 | fix(ui): prepare the first session before leaving setup |
| #2508 | fix(codex): warn when a login duplicates an existing account |
| #2482 | fix(codex): identify OAuth pool slots in terminal auth errors |
| #2412 | feat(extensions): expose Slack messages in task creation hooks |
| #2402 | feat(extensions): emit post.task.blocked for rejected task creation |
| #2306 | perf(worker): cut boot API calls with a skills bundle |
| #2298 | fix(codex): keep session instructions out of AGENTS.md |
| #2281 | fix(github): parse boolean configuration flags consistently |
| #2257 | fix(workflows): recover waits resolved before a crash |
| #2220 | fix(fs): preserve file metadata on agent-fs uploads |
| #2166 | fix(scripts): respect API PORT in script SDK defaults |
| #2111 | feat(scripts): support zod/mini imports |
| #2090 | fix(pi): retry swarm MCP setup before dropping tools |
| #2037 | feat(scripts): make the QuickJS worker pool size configurable |
| #2004 | feat(config): expose script executor selection in dashboard |
| #1981 | fix(acp): validate target configuration before claiming tasks |
| #1960 | fix(credentials): honor pi and opencode presence-only credentials |
All 16 merged Microsoft Agent Framework PRs
| PR | What changed |
|---|---|
| #10204 | Python: close the owned stream in functional workflow agents |
| #9991 | Python: Honor injected Foundry embedding clients without endpoint settings |
| #9985 | Python: Prefer injected Oracle clients over connection settings |
| #9984 | Python: Skip Claude client settings when a client is injected |
| #9983 | Python: Prefer injected Postgres clients over connection settings |
| #9982 | Python: Prefer injected MongoDB clients over connection settings |
| #9981 | Python: Respect FOUNDRY_MODEL in FoundryEvals |
| #9993 | Python: Prefer injected Qdrant clients over connection settings |
| #9987 | Python: Prefer an injected Azure OpenAI deployment over environment model aliases |
| #9773 | .NET: Preserve GitHub Copilot reasoning token usage |
| #9616 | Python: Preserve OpenAI Responses SSE error metadata |
| #9619 | Python: Fix A2A dependency bounds test environment |
| #9543 | Python: Handle missing reasoning text in DevUI |
| #9269 | .NET: Preserve caller cancellation in memory providers |
| #8551 | Python: fix(orchestrations): preserve user multimodal content during handoff |
| #8536 | Python: Reject empty available_modes in mode helpers |
All 5 merged celery PRs
| PR | What changed |
|---|---|
| #10945 | Keep the first SUCCESS in the Cassandra backend |
| #10939 | Log connection retry traceback at debug level |
| #10948 | Fix Manager warning output during retries |
| #10938 | Fix MultiParser to honor its configured Node class |
| #10925 | test: assert exact task replacement nesting levels |
All 5 merged Qwen Code PRs
| PR | What changed |
|---|---|
| #14824 | fix(review): refuse the findings artifact by name and print the expected schema |
| #15178 | feat(agents): make default background execution configurable |
| #15311 | fix(web-shell): refresh standalone sessions after allocation |
| #14636 | test(cli): retry ENOTEMPTY in hosted harness teardowns |
| #13602 | fix(managed-agent): return 404 for unknown tool publications |
All 5 merged MNN PRs
| PR | What changed |
|---|---|
| #5049 | [CPU:Bugfix] Accumulate ARM82 average pooling in FP32 |
| #4894 | [LLM:Bugfix] Refresh PLE for successive Omni text requests |
| #4888 | [LLM:Bugfix] Route Omni attention masks by mode |
| #4806 | [LLM:Bugfix] Pass calibration input IDs to AWQ position IDs |
| #4555 | [LLM:Bugfix] Fix embedding QNN export paths |
All 3 merged NullAway PRs
| PR | What changed |
|---|---|
| #1929 | Avoid initialization-check crashes in @NullMarked constructors |
| #1848 | Treat Maven plugin parameters as externally initialized |
| #1816 | Set up zizmor-action |
All 2 merged nanobot PRs
| PR | What changed |
|---|---|
| #6132 | fix(deepseek): map minimal reasoning effort to low |
| #5800 | feat(discord): add replyToMessage parity with Telegram |
All 2 merged Core ML Tools PRs
| PR | What changed |
|---|---|
| #2862 | Fix squeeze value inference for scalar outputs |
| #2840 | Fix symbolic nearest1d upsample output rank |
All 2 merged Reticle PRs
| PR | What changed |
|---|---|
| #1488 | fix(engine): recognize refusal signals on expected failures |
| #1466 | fix(init): wire plain HTML projects with package.json |
All 1 merged Filament PR
| PR | What changed |
|---|---|
| #10424 | vulkan: fix depth/stencil format support query |
All 1 merged React Spectrum PR
| PR | What changed |
|---|---|
| #10536 | fix: include final constrained year in CalendarYearPicker |
All 1 merged CubeSandbox PR
| PR | What changed |
|---|---|
| #1386 | docs(integrations): add OpenAI Agents SDK guide |
All 1 merged AI-Infra-Guard PR
| PR | What changed |
|---|---|
| #569 | fix(data): prevent recurring case-colliding vulnerability rules |
All 1 merged PyRIT PR
| PR | What changed |
|---|---|
| #2643 | FIX: Improve selected configuration path contrast |
All 1 merged fastjson2 PR
| PR | What changed |
|---|---|
| #7873 | fix: use own ClassLoader as parent when the context ClassLoader loads a foreign fastjson2 |
All 1 merged OSV.dev PR
| PR | What changed |
|---|---|
| #5933 | fix(website): display severity source |
All 1 merged ONNX Runtime GenAI PR
| PR | What changed |
|---|---|
| #2475 | [C#] Expose cached Generator through IChatClient.GetService |
All 1 merged cuOpt PR
| PR | What changed |
|---|---|
| #1925 | docs: update NVIDIA Brev documentation links |
Open PRs (26)
| Project | PR | What I'm working on |
|---|---|---|
| hermes-agent | #152514 | fix(vault): pass selected page origin to code prompts |
| MNN | #5120 | [Vulkan:Bugfix] Record Attention commands after descriptor updates |
| browser | #4548 | input: skip inert subtrees during mouse hit testing |
| Agent Swarm | #2819 | feat(workflows): add opt-in terminal run retention |
| Microsoft Agent Framework | #10114 | Python: Defer experimental AgentFactory creation in WorkflowFactory |
| hermes-agent | #149018 | fix(desktop): route chat pins through their owning connection |
| browser | #4444 | parser: preserve XML namespace declarations and CDATA sections |
| OpenHands | #19295 | feat(onboarding): add an operator opt-in to skip first-run setup |
| celery | #10986 | Avoid misleading root diagnostics when group entries are missing |
| celery | #10976 | feat(cli): add --color to force colored output |
| celery | #10971 | fix(canvas): preserve arguments when cloning chains |
| lexical | #9414 | [lexical-markdown] Chore: Cover backslash emphasis imports and formatted text round trips |
| Qwen Code | #14985 | fix(daemon): separate ACP startup budgets and bound restore waits |
| Qwen Code | #14982 | fix(core): make the always-on context warning accurate for CJK and actionable |
| Qwen Code | #14844 | fix(cli): prefer the user's global npm on Windows |
| nanobot | #6194 | fix(api): release idle session locks |
| Qwen Code | #14026 | refactor(cli): share compression token count formatting |
| Qwen Code | #13954 | fix(ci): enforce Flyway migration ordering |
| swarms | #2522 | [FIX][MajorityVoting][Respect verbose for panels] |
| workers-sdk | #16090 | [deploy-helpers] Add result-only output for Preview deployments |
| volclava | #119 | fix: preserve askedHosts for submit request cleanup |
| OSV.dev | #6102 | fix(website): collapse long affected range event lists |
| nanobot | #5826 | perf(session): accelerate canonical history search with FTS5 |
| React Spectrum | #10604 | fix: support custom focus targets in DateRangePicker |
| volclava | #90 | fix: select RPC and NIS build dependencies by platform |
| lexical | #9106 | [lexical][lexical-clipboard][lexical-code-core][lexical-code-prism][lexical-code-shiki][lexical-eslint-plugin][lexical-extension][lexical-list][lexical-playground][lexical-react][lexical-table][lexical-y] Feature: flag accidental nested editor updates |