Последняя live-проверка dashboard: 2026-05-28 04:17 UTC. Это snapshot; перед новым upstream comment/PR нужно перепроверить конкретную строку через gh pr view / gh issue view. Итог refresh: active PRs остаются WATCH / monitor-only, runner backlog next_status: CANDIDATE, upstream actions 0.
| Upstream PR | Что сделано | Текущий статус | Следующее действие |
|---|---|---|---|
| OpenHands/software-agent-sdk#3394 | Уменьшен tmux viewport 1000x1000 -> 256x200, добавлен regression на bounded viewport; review feedback закрыт follow-up commit 3c8eae78. |
Open, non-draft; BLOCKED/MERGEABLE/CHANGES_REQUESTED; check rollup empty. Maintainer said looks fine and triggered eval; eval result/re-review not visible yet. |
Ждать maintainer eval / re-review / workflow approval; no repeat comment. |
| anomalyco/opencode#29565 | Structured output + thinking models: loosen tool-choice handling. | Open, non-draft; live mergeability MERGEABLE/BLOCKED; visible duplicate/standards/compliance/contributor gates green; no maintainer review yet. Adjacent #29632 may interact with structured-output retry/format semantics but is not an exact cover. |
Ждать maintainer review; watch #29632; не открывать третий opencode PR без отдельного сигнала. |
| cline/cline#11087 | OpenAI-compatible GLM: не отправлять unsupported reasoning controls. | PR-OPEN / monitor only: open, non-draft; BLOCKED/MERGEABLE; visible SDK quality, Ubuntu/Windows tests and Socket checks green after follow-up e1ecc63a1; no approval. |
Ждать maintainer review; upstream comments не нужны. |
| anomalyco/opencode#29530 | Webfetch: format=null трактуется как default, а не как schema error. |
Open, non-draft; live mergeability MERGEABLE/BLOCKED; visible duplicate/standards/compliance gates green, but not full test CI; contribution bot says guidelines satisfied; no maintainer review yet. |
Ждать maintainer review. |
| CopilotKit/CopilotKit#5035 | Release bump не должен публиковать workspace deps. | Open, non-draft; BLOCKED/MERGEABLE, REVIEW_REQUIRED; Vercel docs preview green; other Vercel contexts are team-authorization gated, not proven code failures. |
Ждать maintainer review / team preview authorization. |
| pydantic/pydantic-ai#5680 | Сохранение TextContent metadata в UI adapter round-trip. |
PR-OPEN: open, CLEAN/MERGEABLE; visible CI/harness/coverage green; no maintainer review yet. Same-file overlap risk with #5602/#5605 in AG-UI adapter. |
Ждать maintainer/bot review; recheck if #5602/#5605 merge first. |
| pydantic/pydantic-ai#5678 | Instrumentation serialization не теряет binary data. | PR-OPEN: open, CLEAN/MERGEABLE; visible CI/harness/coverage green; no maintainer review yet. |
Ждать maintainer/bot review. |
| Upstream PR | Что сделано | Текущий статус | Следующее действие |
|---|---|---|---|
| NousResearch/hermes-agent#15640 | Старый Hermes gateway fix: macOS launchd plist должен регистрировать Login Item как hermes, не python. |
PR-OPEN / WATCH-stale-clean via #67 / watch: open, current live gate MERGEABLE/CLEAN, stale since 2026-04-25; no comments/reviews/checks; exact PR #15885 closed as superseded by #15640, while #29673 is a fresh gateway launchd naming overlap risk and currently MERGEABLE/UNSTABLE, not an exact replacement. |
Не закрывать/пинговать/rebase без fresh macOS launchd validation or maintainer signal; watch #29673 for supersede/rewrite pressure. |
Последний scouting heartbeat 2026-05-28 14:13 UTC: watch note, tracker #52. Итог: 2-of-6 subagent fallback из-за agent thread limit reached, parent live gates, upstream actions 0, runner actions 0, final next_status: CANDIDATE-backup. Delta: NousResearch/hermes-agent#33906 is now triaged type/bug, comp/agent, P2; duplicate #33907 is marked duplicate, and adjacent PR #32070 covers user_id propagation but not the missing state.db row. Existing carry-forward lanes remain gated: spring-ai#6197 and zed#57155/#57957 stay CANDIDATE; terraform-mcp-server#376/#377 stays WATCH / upstream PR race; probe#568 remains COMMENT-FIRST-after-repro; claude-peers-mcp#64, Agent-S#195, vercel/ai#15652, gemini-cli#27503, spring-ai#6196, apify#88, mcphub#861, avante.nvim#3105 stay CANDIDATE/WATCH without runner/source gates.
| Upstream | Что делаем | Статус |
|---|---|---|
| NousResearch/hermes-agent#33906, #33907, #32070 | Fresh Hermes session-persistence candidate: context compression rotates session_id and writes continuation JSON, but can miss the state.db row, leaving orphan WebUI sessions instead of nested parent/child lineage. Upstream has triaged #33906 as type/bug, comp/agent, P2; #33907 is duplicate; #32070 is adjacent user_id propagation, not exact cover. Evidence: watch, tracker #52. |
CANDIDATE-backup / triaged P2; no upstream comment/PR before duplicate consolidation, source checkout, current-main repro/test card, and 3-role critique. CT216 may be used only for bounded Hermes-specific smoke/repro; it was not used in this cycle. |
| hashicorp/terraform-mcp-server#376, #377 | Fresh MCP/tool-surface regression candidate now directly covered upstream: env-var streamable HTTP startup path ignores --tools / --toolsets, falls back to toolsets=all, and can expose write/destructive Terraform MCP tools despite a restricted requested set. Evidence: watch, tracker #52. |
WATCH / upstream PR race; #377 is open, non-draft, direct Fixes #376, CLA/changelog checks green, review required. No corp upstream comment/PR unless #377 closes unmerged or maintainers request alternate regression evidence. |
| spring-projects/spring-ai#6197 | Fresh OpenAI tool-calling regression candidate: Spring AI 2.0.0-M7 throws toolCallId is required after @Tool / .defaultTools() / .tools() tool result handling with OpenAiChatModel; reported as working in 2.0.0-M6. Evidence: watch, tracker #52. |
CANDIDATE-needs-repro/process-gate; no upstream comment/PR before current-main Maven/source repro, exact duplicate/PR refresh, Spring DCO gate, and 3-role critique. |
| zed-industries/zed#57155, #57957 | Fresh AI/editor worktree lifecycle candidate: archiving an agent conversation can delete a manually-created git worktree; #57957 adds duplicate/race evidence where archiving a thread from a deleted linked worktree freezes/does nothing and bot links the shared root cause to #57155. Evidence: watch, tracker #52. |
CANDIDATE-needs-source/repro/process-gate; #57957 is WATCH / duplicate-race evidence, not a separate candidate. No upstream comment/PR before exact current-main code-path proof, tests, Zed CLA gate, and positioning against draft PRs #57441/#55703 plus prior archive/worktree PRs. |
| spring-projects/spring-ai#6196 | Fresh MCP server watch/candidate: streamable sync WebMVC exposes resource templates but reports resources.subscribe: false, making resource update notifications effectively unusable. Evidence: watch, tracker #52. |
CANDIDATE-needs-repro/process-gate; no upstream comment/PR before source checkout, exact duplicate search, current-main repro and contribution gate. |
| apify/actor-rag-web-browser#88 | Fresh browser/RAG error-signaling watch: /search returns 200 and empty content on crawl/fetch failure instead of explicit error. Evidence: watch, tracker #52. |
WATCH / CANDIDATE-backup; API semantics may be intentional/backcompat-sensitive, needs maintainer preference or repro before action. |
| openai/symphony | Новый passive repo-scope autonomous-run control-plane watch: isolated autonomous implementation runs. Evidence: watch, tracker #52. | WATCH / passive; issues are disabled, no issue lane. |
| gastownhall/beads | Новый repo-scope coding-agent memory/task watch: memory upgrade for coding agents with active Claude/Codex adjacent issue surface. Evidence: watch, tracker #52. | WATCH / high-fit memory-task layer; no issue-level candidate yet because PR queue is large and duplicate/race gate is required. |
| Hmbown/CodeWhale | Новый repo-scope terminal coding-agent watch: DeepSeek + MiMo coding agent in terminal. Evidence: watch, tracker #52. | WATCH / verify-first; high duplicate velocity and terminal/runtime issues need focused source/repro gate. |
| mksglu/context-mode | Новый repo-scope context-window optimization watch for AI coding agents. Evidence: watch, tracker #52. | WATCH / license-gated reference; license metadata Other, no issue-level candidate selected. |
| get-convex/chef | Новый low-priority repo-scope app-builder watch: AI app builder with backend. Evidence: watch, tracker #52. | WATCH / low-priority app-builder; lower actionability than Dyad/Onlook and issue surface is limited/noisy. |
| dyad-sh/dyad | Новый repo-scope app-builder watch: local open-source AI app builder / Lovable-Bolt-Replit alternative. Evidence: watch, tracker #52. | WATCH / high-fit app-builder surface; no issue-level candidate yet because license metadata is Other, PR queue is active, and fresh bugs need source/repro gates. |
| onlook-dev/onlook | Новый repo-scope visual app/design builder watch: AI-first tool to visually build/style/edit React apps. Evidence: watch, tracker #52. | WATCH / app-builder-scope; no issue-level candidate yet because open bugs are older/noisy and need source/repro/duplicate pass. |
| MCP-UI-Org/mcp-ui | Новый repo-scope protocol-UI watch: UI over MCP / generative UI protocol and SDK. Evidence: watch, tracker #52. | WATCH / protocol-ui-scope; no cold PR while bridge/rendering issues overlap active PR #201 and older transport/frame PRs. |
| zgsm-ai/costrict | Новый repo-scope coding-agent watch: strict AI coder for enterprise code review/completion/agent workflows. Evidence: watch, tracker #52. | LEAD/WATCH / verify-first; issue surface is small, #1249 is assigned, and MCP allow-list/terminal lanes need source gate. |
| modelcontextprotocol/typescript-sdk#2166, #2165 | Fresh MCP SDK watch: UriTemplate.match() multi-variable path expressions and AbortSignal cancellation error-code semantics. Evidence: watch, tracker #52. |
WATCH / assignment-first; no PR/comment because reporter asked to be assigned on both issues and #2165 needs API-design confirmation. |
| zilliztech/memsearch | Новый repo-scope memory/context watch: persistent unified memory layer for Claude Code/Codex and other agents, backed by Markdown and Milvus. Evidence: watch, tracker #52. | WATCH / high-fit agent-memory surface; no issue-level candidate yet because plugin/config/Windows lanes need source/repro/duplicate gate and PR queue is noisy. |
| gsd-build/get-shit-done | Новый passive repo-scope context-engineering watch: high-star Claude Code meta-prompting/spec-driven development system. Evidence: watch, tracker #52. | WATCH / passive ecosystem reference; no issue/PR surface currently exists. |
| Kilo-Org/kilocode | Новый repo-scope coding-agent platform watch: all-in-one agentic engineering platform for VSCode/JetBrains/CLI with Claude/Gemini/Codex-adjacent surface. Evidence: watch, tracker #52. | WATCH / high-fit coding-agent platform; no issue-level candidate yet because fresh issues are assigned/process-bound and PR queue is active. |
| supermemoryai/supermemory | Новый repo-scope memory/context watch: Memory API and app for agent memory / AI memory surfaces. Evidence: watch, tracker #52. | WATCH / high-fit agent-memory surface; no issue-level candidate yet, and retrieval/plugin lanes need source/repro/duplicate gate. |
| vectorize-io/hindsight | Новый repo-scope agent-memory watch: Hindsight agent memory that learns, with Claude Code/OpenCode/OpenClaw integration signals in issue backlog. Evidence: watch, tracker #52. | WATCH / relevant memory/integration surface; no issue-level candidate yet because active issues require focused source/repro and PR queue is moving. |
| abhigyanpatwari/GitNexus | Новый repo-scope code-intelligence watch: zero-server knowledge graph / Graph RAG agent for codebases. Evidence: watch, tracker #52. | WATCH / high-materiality code-intelligence surface; no issue-level candidate yet because open bugs are heavier and PR queue is blocked/dirty/active. |
| apify/apify-mcp-server | Новый repo-scope MCP/tool-runtime watch: MCP server exposing Apify Store scrapers, crawlers and automation tools to AI agents. Evidence: watch, tracker #52. | WATCH / high-fit MCP tool surface; no issue-level candidate yet, and protocol/output-schema lanes need focused source/repro/duplicate gate. |
| TencentCloudBase/CloudBase-MCP | Новый repo-scope MCP deployment/devtool watch: CloudBase MCP connects CloudBase to AI agents and prompt-to-live-app workflows. Evidence: watch, tracker #52. | WATCH / relevant MCP-to-live-app surface; no issue-level candidate yet, and issue/PR queue is older AI-automation-heavy. |
| ogulcancelik/herdr | Новый repo-scope terminal/control-plane watch: terminal agent multiplexer for Claude Code/Codex/coding agents with tmux/TUI/workspace-manager topics. Evidence: watch, tracker #52. | WATCH / high-fit terminal multiplexer surface; no issue-level candidate yet, and many visible issues are intends-to-pr/feature lanes. |
| microsoft/apm | Новый repo-scope agent package/skills ecosystem watch: Agent Package Manager for Codex CLI, Claude Code, Copilot and context/prompt engineering packages. Evidence: watch, tracker #52. | WATCH / strategic packaging/skills surface; no issue-level candidate yet, and active PR queue/process gates require focused source/repro pass. |
| vm0-ai/vm0 | Новый repo-scope AI runtime/sandbox watch: AI teammate/runtime with Claude Code, sandbox and computer-use topics. Evidence: watch, tracker #52. | WATCH / relevant runtime/sandbox/computer-use surface; no issue-level candidate yet, and CUA lanes need desktop/runner evidence plus duplicate search. |
| MemTensor/MemOS | Новый repo-scope memory/context watch: self-evolving memory OS for LLM/AI agents with persistent memory, hybrid retrieval, MCP and skill reuse. Evidence: watch, tracker #52. | WATCH / high-fit memory/MCP/skills surface; no issue-level candidate yet, and open plugin/memory lanes need source/repro/duplicate gate because PR queue is very active. |
| tiann/hapi | Новый repo-scope mobile/remote coding-agent control watch: app for Claude Code, Codex, Gemini CLI and OpenCode remote/vibe coding. Evidence: watch, tracker #52. | WATCH / high-fit remote-control surface; no issue-level candidate yet, and active web/agent-runtime lanes are PR-covered or need focused source/repro gates. |
| agent-of-empires/agent-of-empires | Новый repo-scope TUI/Web orchestration watch: управление Claude Code, OpenCode, Codex CLI, Gemini CLI, Hermes agent и другими coding agents. Evidence: watch, tracker #52. | WATCH / high-fit Paperclip-like control-plane surface; no issue-level candidate yet, and fresh cockpit/feature/release lanes need focused source/repro/process gate before action. |
| rullerzhou-afk/clawd-on-desk, #352, #355 | Новый repo-scope desktop companion watch: Electron/desktop-pet surface watching Claude Code/Codex/Cursor/Gemini sessions. Evidence: watch, tracker #52. | WATCH / high-fit desktop/session companion surface; no issue-level candidate yet, and Ghostty same-cwd focus slice is owner-handled in #355. |
| manaflow-ai/cmux | Новый repo-scope terminal/control-plane watch: Ghostty-based macOS terminal/control surface for AI coding agents, with Claude Code/Codex/Gemini/OpenCode/tmux topics. Evidence: watch, tracker #52. | WATCH / high-fit repo-scope surface; no issue-level candidate yet. Future pass must pick one bounded terminal/session/control-plane issue before action. |
| nexu-io/open-design | Новый repo-scope design-agent watch: Apache-2.0 TypeScript project for local-first Claude Design alternative, with Claude Code/Codex/Gemini/OpenCode/Hermes topics. Evidence: watch, tracker #52. | WATCH / high-fit repo-scope surface; no issue-level candidate yet. Future pass must pick one bounded issue before action. |
| louislva/claude-peers-mcp#64 | Auto-summary uses OpenAI gpt-5.4-nano; source check found shared/summarize.ts sends max_tokens: 100, matching the reported max_completion_tokens contract risk. Evidence: watch, tracker #52. |
CANDIDATE/fallback / source-first evidence only; not runner-backed fail-before and not upstream-ready. |
| topoteretes/cognee, #2907 | Новый repo-scope memory/control-plane watch: активный Apache-2.0 Python проект для AI-agent memory/context, с graph/RAG/context-engineering topics. #2907 (RDFLibOntologyResolver hardcodes XML для file-object ontology input) является возможной future source-local lane. Evidence: watch, tracker #52. |
LEAD/WATCH / high-fit repo, но issue-level candidate ещё нет; нужны source checkout, runner/repro и повторный duplicate search. |
| rohitg00/agentmemory, #703-#708 | New repo-scope memory watch: persistent memory for AI coding agents, topics include Claude/Codex/OpenClaw/Hermes; native-surface audit cluster needs source scan. Evidence: watch, tracker #52. | LEAD/WATCH / high-fit repo, but no issue-level candidate yet; duplicate search was rate-limited and native-surface issues need separate checkout/repro. |
| langchain-ai/langchain#37736, #37737 | EnsembleRetriever.arank_fusion async normalization now has process-valid issue and author-owned PR attempt; recorded in watch. |
WATCH / author-owned process gate: PR closed automatically because author was not assigned; do not race unless maintainers close the lane without action or ask for alternate patch. |
| volcengine/OpenViking#2256 | New repo-scope memory/context watch: FindRequest has no mode field while client/plugin schemas send it; source/repro card: watch, tracker #52. |
LEAD/WATCH / high-fit AI-agent context database, but no runner/source checkout yet; #2261/#2263 are already PR-covered, and open PR queue is review-heavy. |
| langchain-ai/langchain#37735 | EnsembleRetriever.arank_fusion async normalization differs from sync rank_fusion; source-local lead recorded in watch. |
LEAD / process-blocked: issue was auto-closed as programmatic/no issue type; no upstream PR/comment unless a process-safe issue or maintainer signal appears. |
| continuedev/continue#12334 | OpenRouter autocomplete должен прокидывать requestOptions.extraBodyProperties; regression card: watch, tracker #50. |
WATCH / needs reporter path: current main already preserves reasoning on tested raw completion path; evidence comment posted. |
| modelcontextprotocol/typescript-sdk#2115 | requestId: 0 cancel notification должен abort-ить matching request; regression card: cycle 21, duplicate correction/heartbeat: cycle 22, tracker #51. |
WATCH / duplicate-race after 23:37 UTC gates: exact competing PRs #2116 and #2138 are open/mergeable/review required with green checks, changesets, and regression tests; #2138 is the freshest duplicate lane, not an upstream-selected winner; our test-plan comment is historical only; upstream action count remains 0. |
| Upstream PR | Итог | Почему | Что трекаем дальше |
|---|---|---|---|
| e2b-dev/E2B#1354 | Closed, not merged. | Наш PR был CLA-blocked и покрывал только JS SDK; maintainer/superset #1355 покрывает JS + Python и был merged 2026-05-27; issue #1349 закрыт fixed. |
SETTLED; no further action on #1354/#1349 unless regression reopens. |
| aaif-goose/goose#9447 | Closed, not merged. | Maintainer closed as superseded/fixed in #9449; наш PR не был merged. | No action; keep lesson as duplicate-race / faster upstream follow-up. |
| agno-agi/agno#8124 | Closed, not merged. | Duplicate/auto-close: issue already had referenced PRs #8065/#8084. |
Watch existing upstream PRs; не плодить duplicate. |
| langchain-ai/deepagents#3616 | Closed, not merged. | Assignment gate plus stronger technical gate: issue #3587 is waiting-on-author for raw Qwen/LiteLLM tool-call payload details. |
Ждать author/provider evidence; не PR'ить empty-id acceptance без доказательства Deep Agents contract bug. |
| NousResearch/hermes-agent#32999 | Closed obsolete. | После #32963 и #33042 upstream убрал SDK helper path; наш downstream recovery стал не нужен. | Урок перенесён в skill: regression должен проверять public runtime boundary и не глотать app-side TypeError. |
| NousResearch/hermes-agent#32884 | Closed duplicate/obsolete. | Первый Hermes/Codex null-output PR уступил более полному upstream path. | Track итог через Hermes watch note; не переоткрывать. |
Приватный контур для системной работы с open-source: какие upstream-репозитории смотрим, какие баги проверяем, какие PR ведём, где есть конкурирующие PR и какие уроки надо переносить в skills.
- Разведка и triage идут через 6 субагентов, если есть несколько repo/PR или тема быстро меняется.
- Hard gate перед upstream comment/PR:
6-subagent scouting -> parent live gates -> 3-subagent critique -> comment/PR/tracker update. - Если критика нашла оверклейм, stale status, duplicate risk или слабый regression contract, кандидат остаётся
WATCH/NO-GO, а не идёт в PR. - Repro, targeted tests и тяжёлые checkout должны уходить в dedicated runner на
corp-server:corp-opensource-runner. - Пока runner не создан, CT
216/pc-hermes-testможно использовать только для Hermes-specific bounded smoke/repro checks. Это не общий runner. - Provisioning runner-контейнера живёт в
corp-server; этот repo хранит очередь, watch notes и решения по upstream.
LEAD: только subagent-сигнал; parent live gates ещё не пройдены. Нельзя комментировать upstream или писать код.CANDIDATE: parent live gates чистые, есть regression card, но pre-fix fail ещё не доказан. Это неPR-READY.COMMENT-FIRST: баг вероятен, но repo/social gate требует сначала вопрос, assignment, maintainer direction или confirmation.PR-READY: fresh duplicate gate чистый, contribution gate разрешает PR, regression card заполнена, pre-fix fail доказан, patch surface минимален, verification command известна.PR-OPEN: PR открыт; дальше только follow-up loop.WATCH: полезно наблюдать, но сейчас нет разрешения на PR:needs-repro,needs-reporter-path,maintainer-direction-needed,assignment-first,duplicate-race,release-followup.NO-GO: не открывать PR в этом цикле:duplicate-covered, assignment-gated with closed attempts, invitation-only without invite, upstream main already fixed, no public patch surface.
Последний repo-map scouting: Paperclip-adjacent repo map, 2026-05-27. Anchor: paperclipai/paperclip. Core set: OpenHands, deer-flow, multica, vibe-kanban, opencode, cline, goose, browser-use, trycua/cua, MCP SDKs.
Предыдущий Paperclip bug-target block: Popular AI-native / Paperclip-like scouting, 2026-05-27, tracker #52. Итог: 3 цикла по 6 агентов, 0 upstream PR/comment. Главная immediate-дорожка из того блока была trycua/cua#1724, но текущий актуальный bounded lane после cycles 25/27 — mini-swe-agent#826 через #53, пока WATCH из-за runner/search gates. Главные maintainer-signal дорожки: OpenHands#14563, MCP python-sdk#2702, cline#10931.
Mini-swe-agent #826 follow-up 2026-05-28: 6-role refresh + parent live gates + 3-role critique подтвердили next_status: WATCH, upstream actions 0; exact PR не найден, local.py всё ещё использует subprocess.run(..., shell=True, timeout=...), но corp-opensource-runner не provisioned по #10, ssh corp-server не резолвится, CT216 Hermes-only, поэтому upstream comment/PR запрещён до runner-backed current-main repro/regression card.
Cycle 33 fresh-discovery follow-up 2026-05-28: watch note, tracker #52. Итог: 6-role scouting + parent live gates + 3-role critique, upstream actions 0, final next_status: WATCH. browser-use/browser-harness#382 выглядел как лучший press_key candidate, но live duplicate gate нашёл exact/super-strong open PR #346 и adjacent #332/#297, поэтому candidate issue не создан; re-entry только после settle этих PR и runner-backed Chrome/CDP repro.
Candidate backlog refresh 2026-05-28 02:19 UTC: watch note, tracker #52, primary tracker #80. Итог: 6-role scouting + parent live gates + 3-role critique, upstream actions 0, final next_status: CANDIDATE. Primary runner target теперь probelabs/probe#568: open bug/external, no exact PR cover, compact Rust/CLI surface; caveat — root cause не доказан, runner должен сначала подтвердить exact CLI/API stdout contamination path. Reserve candidates: serena#1519, BrowserOS#1005; emdash#1875 prior-art gated; mastra#17118 MRE-gated; ag-ui#1635 watch due #1253 overlap.
Candidate/watch backlog refresh 2026-05-28 03:16 UTC: follow-up synthesis, tracker #52, primary tracker #80, CUA tracker #72. Итог: 6 read-only roles, parent live GitHub gates, internal critique, upstream actions 0, final next_status: CANDIDATE. Primary runner target остаётся probelabs/probe#568; второй практичный runner target теперь web-infra-dev/midscene#2544 because exact fix PR не найден and timing contract is small, but it stays WATCH until current-main longPress fail-before evidence. Reserve CANDIDATE: oraios/serena#1519; BrowserOS#1005, emdash#1875, mastra#17118 stay WATCH; ag-ui#1635 stays COMMENT-FIRST only after runner evidence and overlap positioning against #1253.
Fresh browser/CUA/terminal repo discovery cycle 36 2026-05-28 03:26 UTC: watch note, tracker #52. Итог: 6 read-only roles, parent live GitHub gates, internal critique, upstream actions 0, final next_status: CANDIDATE. Selected future runner target: simular-ai/Agent-S#195 (WindowsACI.hotkey("enter") splits key names), with no exact PR found and a small string/action contract. Reserves: nanobrowser#270, magnitudedev/browser-agent#154, browserwing#18; duplicate-covered/no-go: nanobrowser#291 -> #318, Agent-S#196 -> #197, ntegrals/openbrowser issue-disabled PR backlog, browserwing#11/#12 security lanes.
Fresh AI-native issue discovery cycle 37 2026-05-28 03:40 UTC: watch note, tracker #52, selected candidate #85. Итог: 6 read-only roles, parent live GitHub gates, 3-role status critique, upstream actions 0, final next_status: CANDIDATE. Primary internal runner target: google-gemini/gemini-cli#27503 (SearchText fallback keeps --json when invoking git grep); reserve LEAD/WATCH: anomalyco/opencode#29650; duplicate-covered/no-go for implementation: mastra#17189 -> #17140, opencode#29646 -> #29645.
Runner-candidate backlog refresh 2026-05-28 03:50 UTC: follow-up synthesis, tracker #52, Gemini tracker #85. Итог: 6 read-only roles, parent live gates, 3-role critique, upstream actions 0, final next_status: CANDIDATE; no target was closed or exact-covered. Current runner order: probelabs/probe#568 (JSON stdout pollution), google-gemini/gemini-cli#27503, oraios/serena#1519, simular-ai/Agent-S#195, web-infra-dev/midscene#2544, browseros-ai/BrowserOS#1005. Runner remains unavailable via #10; local env lacks pnpm/bun/cargo/rustc.
Runner + active PR dashboard refresh 2026-05-28 04:04 UTC: follow-up synthesis, tracker #52. Итог: 6 read-only roles, parent live gates, 3-role critique, upstream actions 0. Six runner candidates remain open with no exact PR cover; active PR dashboard had no merge/close/review transition. Operational runner order stays Probe first / Gemini second to avoid churn; Gemini is recorded as the highest strategic/race-sensitive candidate, not PR-READY. Current recheck has opencode#29565/#29530 open/non-draft MERGEABLE/BLOCKED; Hermes#15640 and overlap #29673 returned UNKNOWN/UNKNOWN, so Hermes remains conservative monitor-only.
Fresh AI-native issue discovery cycle 38 2026-05-28 04:07 UTC: watch note, tracker #52, selected candidate #86, repo-scope #66. Итог: 6 read-only roles, parent live gates, 3-role critique, upstream actions 0, final next_status: CANDIDATE. Selected internal runner target: vercel/ai#15652 (@ai-sdk/anthropic jsonTool path appears to ignore disableParallelToolUse: false); this is only a package-fixture candidate, not PR-READY. Rejected/covered lanes: Qwen#4579 -> #4580, browser-use#4801 -> #4880/#4835; reserve CANDIDATE/WATCH: modelcontextprotocol/inspector#1368.
Active PR dashboard correction 2026-05-28 04:17 UTC: follow-up synthesis, tracker #52. Итог: 6 read-only roles, parent live gates, 3-role critique, upstream actions 0, final next_status: WATCH. Материальных переходов по active PRs нет: no merge/close/approval/maintainer ask/actionable failing CI. Коррекция к 04:04: Hermes#15640 current MERGEABLE/CLEAN, adjacent #29673 current MERGEABLE/UNSTABLE; это overlap/supersede pressure, не exact replacement и не основание для ping/rebase/close.
Fresh discovery cycle 39 2026-05-28 04:22 UTC: watch note, tracker #52, refresh #52. Итог: 6 read-only roles, parent live gates, 3-role critique, upstream actions 0, runner actions 0, final next_status: WATCH. Лучший свежий lane — pydantic/pydantic-ai#5696 (InstructionPart serializes with discriminator instruction, but ModelRequestPart union does not include the tag); refresh 04:55 UTC found no exact PR, but pydanty now reports Dogfooding runtime timeout and maintainer-attention/rerun need, so this is WATCH / pydanty-runtime-timeout, not CANDIDATE/PR-READY.
Fresh Aion STT discovery 2026-05-28 05:00 UTC: watch note, tracker #52. Итог: 6 read-only roles, parent live gates, upstream actions 0, runner actions 0, final next_status: CANDIDATE-needs-runner-repro. Selected lane — iOfficeAI/AionUi#3074: WebUI STT posts /api/stt multipart field audio, while public iOfficeAI/AionCore route requires file, fileName, and mimeType; exact STT/multipart PR cover not found. Adjacent AionUi#3075 stays WATCH, not bundled into the same fix before runner evidence.
Bounded Vercel/backlog refresh 2026-05-28 05:19 UTC: follow-up synthesis, Vercel tracker #86, umbrella #52. Итог: 6 read-only roles, parent live gates, 3-role critique, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Material delta: vercel/ai#15652 remains open/unassigned with no exact PR cover, and a new upstream framing comment narrows the fixture to honoring explicit disableParallelToolUse: false on the Anthropic jsonTool request while preserving the default true; this is still needs-runner-fixture, not PR-READY. Existing core candidates (gemini-cli#27503, AionUi#3074, probe#568, AionUi#3051, langgraph#7915, pydantic-ai#5696) have no exact PR cover but remain blocked by runner, ownership, or automation gates. Fresh leads playwright-mcp#1636, modelcontextprotocol/go-sdk#977, and qwen-code#3718 stay WATCH.
Fresh LangChain4j candidate refresh 2026-05-28 05:34 UTC: watch note, tracker #52. Итог: 6 read-only roles, parent live/source gates, focused 3-role critique, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Selected lane — langchain4j/langchain4j#5313: LoopPlanner starts iterationsCounter = 1 and checks iterationsCounter > maxIterations, matching the reported maxIterations + 1 loop behavior; exact open PR cover not found. It is a Java/JDK17 runner-friendly fallback behind Probe/Gemini/Vercel, not PR-READY. Rejected/covered: mem0#5220 -> #5221/#5258; langchain4j#5296 stays watch due heavier concurrency surface; agent-browser#1394 stays watch due UI/screenshot-heavy repro.
Fresh Continue link lead 2026-05-28 05:55 UTC: watch note, tracker #52. Итог: 6-role fallback scouting + parent live gates + focused 3-role critique, upstream actions 0, runner actions 0, final next_status: LEAD. Selected lead — continuedev/continue#12507: chat @ dropdown action Add more context providers still points to stale docs URL in gui/src/components/mainInput/TipTapEditor/utils/getSuggestion.ts; exact open PR cover not found. This is a valid opportunistic UI/docs-link lead, but it does not reorder the primary runner backlog and is not PR-READY.
Fresh Vercel docs lead 2026-05-28 06:10 UTC: watch note, tracker #52, Vercel tracker #86. Итог: 6-role fallback scouting + parent live/source gates, upstream actions 0, runner actions 0, final next_status: LEAD. Selected lead — vercel/ai#15663: several docs/cookbook streamText({ ... }) examples still use deprecated onFinish after #15245 introduced onEnd; exact open PR cover not found. This is docs-only and lower priority than runtime/tool-call candidates, especially vercel/ai#15652, so no upstream comment/PR.
Vercel docs follow-up 2026-05-28 06:26 UTC: watch note. Итог: live gate refresh + 6-role fallback refresh, upstream actions 0, runner actions 0, final next_status: LEAD/WATCH. Delta: vercel/ai#15663 now has 3 author comments; author offered to open PR and expanded scope to packages/mcp/README.md, examples/ai-functions/src/**, and on-finish filename examples. Exact open PR cover still not found, but author-PR race risk is now high, so do not open competing upstream PR.
Fresh Codex Desktop WSL candidate 2026-05-28 06:32 UTC: watch note, tracker #52. Итог: fresh discovery + parent live/duplicate gates + 6-role fallback + focused 3-role critique, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Selected lead — openai/codex#24873: Codex Desktop WSL codex-cli 0.133.0 cannot start sandboxed tools because bwrap is not found in system/bundled lookup, while use_legacy_landlock = true hits a permission-profile panic. Exact open PR cover not found; closed #21915 is prior art, not a duplicate. Not PR-READY: needs Windows + WSL2 + Codex Desktop repro and resource lookup evidence.
Fresh Claude Code Agent Teams TUI lead 2026-05-28 06:41 UTC: watch note, tracker #52, terminal tracker #74. Итог: fresh discovery + parent live/duplicate gates + 6-role fallback + focused 3-role critique, upstream actions 0, runner actions 0, final next_status: LEAD. Selected lead — anthropics/claude-code#63019: permission prompt can corrupt terminal output while user navigates expanded Agent Teams selection UI. Exact PR cover not found; nearby #49865/#8618 are related but not exact. Not CANDIDATE: needs macOS/TUI repro or narrowed terminal-buffer evidence, and repo process looks issue-first.
Fresh Nanobot candidate pair 2026-05-28 06:53 UTC: watch note, tracker #52. Итог: fresh discovery + parent live/duplicate gates + 6-role fallback + focused 3-role critique, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Selected pair — HKUDS/nanobot#4040/#4039: /stop may fail to cancel active tasks under unified_session, and context snipping may ignore tool-schema tokens in retained-history budget. Exact PR cover not found; distinct from prior duplicate-covered #4006/#4013. #4040 is the stronger runner target; #4039 needs source confirmation.
Fresh runtime batch 2026-05-28 07:08 UTC: watch note, tracker #52. Итог: fresh discovery + parent live/duplicate gates + 6-role fallback + focused 3-role critique, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Selected candidate — NousResearch/hermes-agent#33712: hermes dashboard fails on NixOS/container with missing hermes_cli.dashboard_auth; exact cover not found and adjacent #33311 is not a cover. Watch lanes: openai/codex#24879 is covered-risk via #24877/#21928/#23767, claude-code#63025 needs Desktop/SSH evidence, and TensorRT-LLM#14676 is GPU/CI-heavy.
Fresh Candle correctness candidate 2026-05-28 07:21 UTC: watch note, tracker #52. Итог: fresh discovery + parent live/duplicate gates + 6-role fallback + focused 3-role critique, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Selected candidate — huggingface/candle#3567: Qwen2 prepare_attention_mask appears bidirectional when attn_mask is provided, causing silent wrong outputs in batched eval/training. No exact PR cover found. #3568/#3569 are valid autograd no-bwd watch lanes, but coordination-first because open #3526 already covers the same pattern for RmsNorm.
Hermes/Claude live delta 2026-05-28 07:34 UTC: runtime batch note, tracker #52. Итог: live gates + 6-role fallback refresh, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Delta: NousResearch/hermes-agent#33712 now has labels type/bug, comp/cli, area/nix, P3; still open/unassigned/no comments/no exact PR cover, so runner/package repro remains required. Claude watch: #63028 Web/plugin first-session bug has has repro, #63025 got a diagnostic SSH Remote comment with intact projects, #63029 stays weak duplicate-risk watch.
CLIProxyAPI Claude tool-name candidate 2026-05-28 07:45 UTC: watch note, tracker #52. Итог: fresh discovery + parent live gates + reused 6-agent scouting under thread-limit fallback, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Selected candidate — router-for-me/CLIProxyAPI#3592: popular Claude Code proxy issue, где Anthropic Messages response возвращает tool_use.name: "glob" для advertised Glob, ломая case-sensitive tool dispatch. Adjacent PR #3214 покрывает request-side/Bifrost remapping, а не exact response-side reverse mapping; перед upstream action нужен repeat search после reset GitHub Search API rate-limit.
CLIProxyAPI/opencode refresh 2026-05-28 07:55 UTC: CLIProxyAPI note, opencode #29694 note, tracker #52. Итог: 6-agent refresh + parent duplicate gates, upstream actions 0, runner actions 0, final next_status: WATCH. Delta: CLIProxyAPI#2746 — open conflicting but strong per-request reverse-map PR, поэтому #3592 downgraded до duplicate-covered-risk watch. Fresh anomalyco/opencode#29694 — high-value assigned watch для unwired tool-output spill cleanup; exact PR cover не найден, но assignee nexxeln blocks cold PR.
Assigned opencode/CLIProxy refresh 2026-05-28 08:05 UTC: follow-up synthesis, new watch opencode#29697, tracker #52. Итог: reused 6-agent cycle + parent live gates + critique, upstream actions 0, runner actions 0, final next_status: WATCH. opencode#29697 is a strong macOS Terminal.app IME/TUI stale-cell watch but already assigned and macOS/IME-repro heavy; opencode#29694 remains high-value assigned watch; CLIProxyAPI#3592 stays WATCH / duplicate-covered-risk because #2746 is strong adjacent cover but may not fully restore TitleCase response names.
Claude Code agent/MCP fresh watch 2026-05-28 08:12 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate search, upstream actions 0, runner actions 0, final next_status: WATCH. New product-owned watch signals: claude-code#63042 stale local git identity in Agent worktrees, #63041 cleartext MCP env output from claude mcp get, and secondary #63031 missing workflow slash-command completion in claude agents. No upstream comment/PR because Claude Code is product-owned and no external patch lane is available.
Covered PR refresh 2026-05-28 08:16 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. pydantic-ai#5663 is now covered by draft PR #5670; vercel/ai#15664 is covered by exact PR #15665; openai/codex#24475 got a fresh confirming subagent reconnect comment but remains product/app-server watch, not an OSS patch lane.
Claude/opencode post-cutoff watch 2026-05-28 08:20 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. claude-code#63043 is a high-signal Windows Grep rg.exe/conhost.exe leak watch, #63044 is a weaker stats timezone watch, opencode#29699 is an upstream-owned websocket timeout PR cover, and opencode#29692 cleared compliance but remains assigned to Hona.
Claude peers MCP candidate 2026-05-28 08:25 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live/source gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: CANDIDATE. Selected lane — louislva/claude-peers-mcp#64: auto-summary silently fails because shared/summarize.ts sends max_tokens to gpt-5.4-nano; source gate confirms the field and silent !res.ok -> null path. Needs repeat PR search and runner-backed Bun validation before upstream action.
Codex/Claude product watch 2026-05-28 08:30 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. Fresh product-watch deltas: openai/codex#24884 Windows Codex App WSL mode requires danger-full-access because sandbox blocks WSL access, and claude-code#63047 re-files Desktop/Cowork plugin PostToolUse hook skip with production evidence. claude-peers-mcp#64 remains the prior internal CANDIDATE-needs-runner-validation.
Claude/opencode follow-up watch 2026-05-28 08:35 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. Delta: opencode#29699 is now merged, closing the upstream-owned websocket timeout cover lane; claude-code#63048 is a fresh product-owned /code-review skill watch for silent fallback from empty @{upstream}...HEAD to main...HEAD and raw JSON-only output. claude-peers-mcp#64 remains the only current small source-level CANDIDATE-needs-runner-validation.
Fresh AI-native watch 2026-05-28 08:42 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. Delta: codex#24884 now has external contributor intent, so race risk is high; claude-code#63050 is a product-owned Monitor shell-snapshot/PATH bug watch; CLIProxyAPI#3594 is a WebSocket/Rivet passthrough LEAD; gemini-cli#27517 is weak node-pty resize watch; opencode#29703 is compliance/duplicate/assignee gated. claude-peers-mcp#64 remains the only current small source-level CANDIDATE-needs-runner-validation.
Fresh AI-native watch 2026-05-28 08:48 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. Delta: gemini-cli#27517 now has duplicate-bot possible duplicates; opencode#29704 is assigned and duplicate-gated against animation controls #21939; mem0#5284 is already covered by an external PR; ag-ui#1801 has author PR intent; claude-code#62959 and codex#24457 received product-watch comments. claude-peers-mcp#64 remains the best compact source-level CANDIDATE-needs-runner-validation.
Opencode/Codex watch 2026-05-28 08:55 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. Delta: opencode#29704 closed as duplicate-covered by #21939; opencode#28150 is covered by open PR #28689; opencode#16962 got fresh Warp/Mac confirmation but stays assigned/duplicate-linked watch; codex#24884 now looks product-private-surface because the WSL bootstrap path may not be in public repo. claude-peers-mcp#64 remains the best compact source-level CANDIDATE-needs-runner-validation.
Gemini/opencode covered watch 2026-05-28 09:02 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. Delta: gemini-cli#27518 is another duplicate-risk EBADF resizePty report; opencode#29706 is assigned/compliance-gated server-500 watch; mcp/python-sdk#2705 and opencode#29705 are already covered by external PRs; claude-code#63051 is product-transient API overload. claude-peers-mcp#64 remains the best compact source-level CANDIDATE-needs-runner-validation.
Claude/MCP/opencode watch 2026-05-28 09:08 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. Delta: claude-code#63051 is now duplicate/product labeled; claude-code#63052 is duplicate-labeled TUI long-line truncation watch; mcp/python-sdk#2695 has follow-up test proposal after #2700; opencode#26625 has PR-linked /timestamps ambiguity; vercel/ai#15667 is docs-only PR; JetBrains/koog#2079 is low-clarity token type/semantics lead. claude-peers-mcp#64 remains the best compact source-level CANDIDATE-needs-runner-validation.
Claude/opencode/AG-UI watch 2026-05-28 09:16 UTC: watch note, tracker #52. Итог: reused 6-agent cycle + parent live gates + duplicate/PR search, upstream actions 0, runner actions 0, final next_status: WATCH. Delta: opencode#29707 is now covered by PR #29710; ag-ui#1801 is covered by PR #1802; opencode#29709 is maintainer-owned ACP startup PR; opencode#29708 is assigned/duplicate-hinted; claude-code#63053 is product-owned VS Code markdown feature; claude-code#63052 is duplicate no-go. claude-peers-mcp#64 remains the best compact source-level CANDIDATE-needs-runner-validation.
Предыдущий repo-universe aggregate: AI SDK / Agents repo scope check, Docker MCP scope check, Morph SDK generic scope closure, OpenClaw/OpenClaude identity scan, DeepAgentsJS scan, HKUDS/OpenHarness scan, Infisical/agent-vault scan, BloopAI/vibe-kanban scan, mini-swe-agent #826 follow-up, cycle 29, Mastra issue-level scan, mcp-use issue-level scan, AG-UI issue-level scan, FastMCP issue-level scan, ChromeDevTools issue-level scan, cycle 28, cycle 27, cycle 25, cycle 24 и repo-card queue, tracker #52. Итог AI SDK / Agents блока: 6 scouting subagents, parent repo/issue/PR/community/license gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; vercel/ai добавлен как high-value exact provider/schema/tool-call/streaming validation watch via #66, а cloudflare/agents добавлен как issue-first runtime/MCP/streaming/compaction reference watch because upstream currently does not accept external PRs. Итог Docker MCP блока: 6 scouting subagents, parent repo/issue/PR/community/license gates, 3-subagent critique, 0 upstream PR/comment, final combined next_status: WATCH; docker/mcp-gateway оставлен только как exact-validation/config watch via #65, а docker/mcp-registry закрыт как generic scouting scope и остаётся reference / exact metadata-validation only. Итог Morph SDK generic-scope блока: 6 scouting subagents, parent repo/issue/PR/community/license gates, 3-subagent critique, 0 upstream PR/comment, final next_status: NO-GO for generic scope; morph-labs/morph-python-sdk and morph-labs/morph-typescript-sdk are closed as stale/low-signal generic targets via #64, with revisit only for Morph-specific request or fresh current-main bug card plus runner-safe secret handling. Итог OpenClaw/OpenClaude identity блока: 6 scouting subagents, parent repo/issue/PR/community/license gates, 3-subagent critique, 0 upstream PR/comment, final next_status: NO-GO; openclaw/openclaw не доказан как неканонический, но требует external canonical validation из-за аномальной активности и duplicate-race risk, а Gitlawb/openclaude заблокирован license/provenance текстом про proprietary-derived code через #63. Итог DeepAgentsJS блока: 6 scouting subagents, parent repo/issue/PR/contribution/security/test gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; langchain-ai/deepagentsjs#547 is the strongest uncovered lead via #62, but LangChain approval-first process, missing runner/current-main repro, and high duplicate pressure block upstream action. Итог HKUDS/OpenHarness блока: 6 scouting subagents, parent repo/issue/PR/contribution/test gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; repo is a strong Python agent-harness watch via #61, but current open issues are feature/discussion-heavy and PR queue is conflict-heavy, so promotion requires a reproducible current-main harness bug plus duplicate gate. Итог Infisical/agent-vault блока: 6 scouting subagents, parent repo/issue/PR/contribution/license/security gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; #194 is the best Codex/Agent Vault WebSocket proxy compatibility lead via #60, but Agent Vault-owned patch surface is unproven until secret-free v0.22.0/current-main runner repro and owner-boundary gate. Итог BloopAI/vibe-kanban блока: 6 scouting subagents, parent repo/issue/PR/contribution gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; repo remains a strong Paperclip-like control-plane reference but not a PR target while README says the project is sunsetting, asks for discussion before PR, and maintainer merge rhythm is weak. Итог mini-swe-agent #826 follow-up: повторный 23:30 UTC heartbeat после 6 scouting subagents, parent live gates и 3-subagent critique оставил final next_status: WATCH, upstream actions 0; exact competing PR не найден, но #832/#803/#807/#843 остаются adjacent/overlap context, а runner/current-main repro plus regression card всё ещё блокируют upstream comment/PR. Итог cycle 29 repo-universe блока: 6 scouting subagents, parent repo gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; repo-level promoted set is ChromeDevTools/chrome-devtools-mcp, PrefectHQ/fastmcp, openai/openai-agents-python, pydantic/pydantic-ai, SWE-agent/mini-swe-agent, BloopAI/vibe-kanban, ag-ui-protocol/ag-ui, with Infisical/agent-vault added as infra-adjacent repo-card. Next heartbeat is repo-card/scope update only; no upstream target selected. Итог Mastra блока: 6 scouting subagents, parent issue/PR/contribution gates, 3-subagent critique, 0 upstream PR/comment, final next_status: CANDIDATE; mastra-ai/mastra#17118 is internal candidate via #58, blocked on runner/local fail-before regression card and fresh duplicate gate. Итог mcp-use блока: 6 scouting subagents, parent issue/PR/contribution gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; no clean PR-ready candidate via #57 because #1581/#1576/#1570/#1568/#1569/#1551/#1555/#1553/#1438/#1415 are covered by active or merged PRs. Итог AG-UI блока: 6 scouting subagents, parent issue/PR/contribution gates, 3-subagent critique, 0 upstream PR/comment, final next_status: CANDIDATE; ag-ui-protocol/ag-ui#1635 is internal candidate via #56, blocked on runner repro, duplicate gate over #1253/#1727/#836, and assignment-before-PR. Итог FastMCP блока: 6 scouting subagents, parent issue/PR/contribution gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; PrefectHQ/fastmcp#4143 and #4124 remain COMMENT-FIRST-after-MRE/assignment lanes via #55, blocked on current-main repro, closed process-gated PR history, and FastMCP assignment-first rule. Итог ChromeDevTools блока: 6 scouting subagents, parent issue/PR/contribution gates, 3-subagent critique, 0 upstream PR/comment, final next_status: WATCH; ChromeDevTools/chrome-devtools-mcp#2138 raised only to internal CANDIDATE via #54, blocked on call-site scan, duplicate gate, and test card. Итог cycle 28 repo-universe блока: promote/create cards for ChromeDevTools/chrome-devtools-mcp and PrefectHQ/fastmcp; active watch cards for ag-ui-protocol/ag-ui and mcp-use/mcp-use; demote/passive watch for openai/evals, gptme/gptme, and anthropics/claude-code. Один subagent drifted into legacy Rails/Paperclip attachment repos; that output is recorded as scope drift and excluded from AI-native ranking.
Свежий repo-universe cycle 30: browser/terminal/sandbox agent repo update. Итог: 6 scouting subagents, parent live GitHub gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH / repo-universe update; promoted repo-card/watch targets are browseros-ai/BrowserOS (#1005 next MCP flood lane), charmbracelet/crush (#3011/#3024 provider/retry lanes), JetBrains/koog (#2020 Anthropic enum casing lane), and alibaba/OpenSandbox (#949 signed endpoint/proxy lane). ComposioHQ/agent-orchestrator and can1357/oh-my-pi stay lower-priority lead/watch because current best lanes need Windows/Bun runner gates or are already PR-covered.
BrowserOS issue-level follow-up: BrowserOS #1005 MCP list_changed flood, tracker #81. Итог: 6-role fallback scouting, parent live gates, internal 3-role critique, 0 upstream PR/comment, final next_status: CANDIDATE; exact PR duplicate не найден, вероятный path is packages/browseros-agent/apps/server/src/api/routes/mcp.ts -> services/mcp/mcp-server.ts -> register-mcp.ts, blocked on dedicated runner fail-before for repeated notifications/tools/list_changed on unchanged MCP registry.
Свежий repo-universe cycle 34: workflow / agent platform update, tracker #52. Итог: 6 read-only scouting roles, parent live GitHub/source gates, internal 3-role critique, 0 upstream PR/comment, final next_status: LEAD; selected lead is coze-dev/coze-studio#2683 Signal Arena / Agent World auth failure, but not CANDIDATE because parent code search for world.coze.site in the repo returned 0 and OSS ownership is unproven. HKUDS/nanobot#4006/#4013, lastmile-ai/mcp-agent#671 and PySpur#294 are duplicate-covered by open PRs; trigger.dev#3672/#3674/#3726 stays COMMENT-FIRST / WATCH because the repo requires vouch and exact attempts were closed unmerged.
Свежий CUA / UI automation repo watch cycle 35: midscene / Windows-MCP / UI-TARS refresh, tracker #72. Итог: 6 read-only scouting roles, parent live GitHub gates, internal 3-role critique, upstream actions 0, final next_status: WATCH. Лучший следующий runner target — web-infra-dev/midscene#2544 (longPress capped to 600ms), но не CANDIDATE без current-main browser repro, contract check and fresh duplicate gate. CursorTouch/Windows-MCP добавлен как Windows/DPI/UAC duplicate-sensitive watch; UI-TARS-desktop, openagent, himarket, open-computer-use и microsoft/agent-framework#6120/#6128 остаются watch/passive/duplicate-covered.
Свежий repo-universe cycle 31: coding-agent control-plane / MCP infra update, затем follow-up refresh. Итог исходного цикла: 6 scouting subagents, parent live GitHub/source gates, internal 3-role critique, 0 upstream PR/comment, final next_status: CANDIDATE; stablyai/orca#2930 был promoted via #82 / watch. Follow-up live gate нашёл exact open non-draft upstream PR stablyai/orca#2948, поэтому локальный статус #82 демотирован до WATCH / duplicate-covered, upstream actions 0, runner не запускался. New watch targets from this cycle: generalaction/emdash, asheshgoplani/agent-deck, agentgateway/agentgateway, googleapis/mcp-toolbox; proxysoul/soulforge remains low-priority due BUSL/no issue surface.
Свежий repo-universe fallback refresh 2026-05-28 01:45 UTC: follow-up synthesis. Попытка 6-agent scouting остановилась на 3 ролях из-за agent thread limit reached, затем parent live gates и 3-role critique выбрали upstream actions 0. Новые/обновлённые статусы: generalaction/emdash#1875 поднят до internal CANDIDATE via #83 / watch; modelcontextprotocol/inspector#1368, google-gemini/gemini-cli#27431 и vercel/ai#15652 отмечены как internal CANDIDATE lanes в scope notes; browser-use#4846, vercel/ai#15613, PrefectHQ/fastmcp#4154 и charmbracelet/crush#3024 оставлены WATCH из-за overlap/process/runner gates. Ни один lane не PR-READY; следующий практичный runner target: emdash#1875 или BrowserOS#1005, если corp-opensource-runner станет доступен.
Candidate readiness refresh 2026-05-28 01:08 UTC: follow-up synthesis. Проведены parent live gates, 6 read-only scouting roles и 3-role critique по emdash#1875, BrowserOS#1005, inspector#1368, gemini-cli#27431, vercel/ai#15652; upstream actions 0, runner недоступен (corp-server не резолвится, local Node 20/no pnpm/no bun), ни один lane не PR-READY. Все пять остаются CANDIDATE; emdash#1875 теперь имеет mandatory prior-art note по закрытым unmerged PR #1907/#1908/#1909/#1910/#1911, а лучший runner target остаётся emdash#1875 with vercel/ai#15652 as package-fixture fallback.
Candidate correction refresh 2026-05-28 01:14 UTC: follow-up synthesis. Проведены parent live gates, 6 read-only scouting roles и 3-role critique по mini-swe#826, ag-ui#1635, mastra#17118, probe#568, emdash#1875, BrowserOS#1005; upstream actions 0, runner недоступен. Главное изменение: mini-swe#826 держать как WATCH до post-#832 current-main re-test, потому что merged timeout PR #832 может частично закрывать пользовательский симптом. probe#568 остаётся compact fallback CANDIDATE; mastra#17118 теперь обязан учитывать merged partial precedent #17054; остальные lanes не PR-READY.
Свежий repo-universe cycle 32: browser/control-plane/code-context/sandbox update, selected tracker #84. Итог: 6 read-only scouting roles, parent live gates, internal 3-role critique, 0 upstream PR/comment, final next_status: CANDIDATE; selected lane is oraios/serena#1519 dashboard host-port 403. Follow-up 2026-05-28 01:46 UTC reran duplicate PR/issue search after rate-limit cleared and found no covering PR; status stays CANDIDATE, not COMMENT-FIRST/PR-READY, because runner-backed repro is still missing and security guardrail says not to wildcard/skip dashboard Host validation. colbymchenry/codegraph#507 demoted to WATCH / duplicate-covered by open PR #345; vercel-labs/agent-browser, multica-ai/multica, epiral/bb-browser, and agent-infra/sandbox remain watch/lead lanes.
Последний candidate refresh: AG-UI / Mastra candidate refresh. Итог: 6 scouting subagents, parent live gates, 3-subagent critique, 0 upstream PR/comment; ag-ui-protocol/ag-ui#1635 переведён в COMMENT-FIRST after runner-backed test card because AG-UI is assignment-first and nearby Mastra adapter PRs #1253/#837/#1727/#1468 create collision risk; mastra-ai/mastra#17118 остаётся CANDIDATE: exact PR не найден, #17054 merged adjacent only, #17083 open broad/conflicting Mastracode overlap, blocked on runner-backed MRE/regression card with concrete terminal-width fixture and repeat duplicate gate.
Последний repo-scope watch: Qwen Code / Agent Framework scope watch, tracker #68. Итог: 6 scouting subagents, parent repo/issue/PR/community/license gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; QwenLM/qwen-code остаётся verify-first terminal-agent lane because #4450/#4466/#4363/#4364 are already covered by active PRs or nearby work, and microsoft/agent-framework остаётся exact Python/.NET subsystem watch with best later lanes #6120/#6006, blocked on runner repro and fresh duplicate search.
Последний verify-first repo watch: gptme / CAO / OpenACP scope, tracker #69. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; awslabs/cli-agent-orchestrator поднят как главный follow-up lane for inbox/tmux/status delivery #131/#78/#164, но blocked by #115/#251 overlap and missing runner repro; gptme/gptme остаётся passive low-noise terminal-agent watch, Open-ACP/OpenACP остаётся ACP ecosystem watch.
Последний browser-agent repo watch: browser-use / Stagehand / Skyvern scope, tracker #70. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; browserbase/stagehand лучший future browser-SDK lane, но blocked by active PR clusters #2162/#2166/#2156/#2062/#2159/#2048; browser-use/browser-use остаётся post-merge gap scan after #4880/#4881/#4882/#4927; Skyvern-AI/skyvern остаётся self-hosted/service-stack watch with AGPL caution.
Последний top-tier repo watch: pydantic-ai / OpenHands / E2B scope, tracker #71. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; pydantic/pydantic-ai остаётся top strategic watch, но текущие узкие lanes covered by #5694/#5681/#5693/#5680/#5678; e2b-dev/E2B#1352 лучший future runner-backed lead after managed-secret repro and Linear/maintainer gate; OpenHands/OpenHands#14563 уже comment-first and waits maintainer direction.
Последний coding/computer-use/browser-MCP repo watch: Cline / CUA / Playwright MCP scope, tracker #72. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; microsoft/playwright-mcp#1635 лучший immediate follow-up but already comment-first and approval-gated; trycua/cua#1739 лучший future computer-use lead after macOS Retina/AVF repro; cline/cline остаётся saturated watch while #11087 and nearby PR clusters are active.
Последний coding-assistant context / code-search repo watch: Graphify / OpenSpec / Tabby / CopilotForXcode / Probe scope, tracker #80. Probe follow-up 2026-05-27 23:07 UTC: 6 scouting subagents, parent live gates, internal 3-role critique, 0 upstream PR/comment, final next_status: CANDIDATE; probelabs/probe#568 остаётся selected lane for symbols JSON stdout pollution, blocked on runner-backed fail-before repro and repeat duplicate search before any upstream comment/PR.
Предыдущий orchestration / tool / eval framework repo watch: CrewAI / Composio / Promptfoo / AutoGen / DSPy scope, tracker #79. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; promptfoo/promptfoo лучший future eval/assertion/tooling deep-dive, но current best lane #9235 duplicate-covered by #9239/#9298, so no comment/PR until those PRs merge/close/stall and runner-backed repro confirms a remaining gap.
Предыдущий low-code agent/workflow builder repo watch: Dify / Flowise / Langflow / n8n / Activepieces scope, tracker #78. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; strong control-plane fit, but best fresh lanes are PR-covered (Flowise#6428 -> #6429, activepieces#13418 -> #13420) or blocked by Linear/team/release-branch/runner gates.
Последний RAG / memory / retrieval repo watch: LlamaIndex / Haystack / Mem0 / Qdrant / LanceDB scope, tracker #77. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; strong repo fit, but best fresh lanes are PR-covered (llama_index#21750 -> #21785, Haystack #11409 -> #11413/#11414, Mem0 #5220 -> #5258, Qdrant #9201 -> #9202, LanceDB #3447 -> #3454) or need runner-backed repro.
Последний observability/eval harness repo watch: Langfuse / Opik / DeepEval / OpenEvals / smolagents scope, tracker #76. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; langfuse/langfuse и comet-ml/opik остаются стратегическими observability/eval lanes только после exact issue/PR gate и runner-backed repro, confident-ai/deepeval/huggingface/smolagents лучше подходят для узких future repro, langchain-ai/openevals остаётся low-noise watch.
Последний MCP ecosystem repo watch: MCP registry / inspector / servers / spec scope, tracker #75. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; modelcontextprotocol/registry лучший later publisher/validator follow-up but current fresh bugs #1306/#1289 are PR-covered, modelcontextprotocol/inspector лучший tooling fit but V2 churn and nearby PRs block cold action, modelcontextprotocol/servers remains package-level reference-server watch only, modelcontextprotocol/modelcontextprotocol requires SEP/governance alignment and AI disclosure planning.
Предыдущий high-churn terminal/runner repo watch: Codex / opencode / Gemini CLI / Daytona scope, tracker #74. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; daytonaio/daytona лучший later follow-up after runner-backed gap scan, google-gemini/gemini-cli только help-wanted/approved issue path, openai/codex evidence/comment-only unless invited, anomalyco/opencode no new PR while current opencode PRs await review and same-day queue stays saturated.
Предыдущий agent/MCP framework repo watch: FastMCP / OpenAI Agents / DeepAgents scope, tracker #73. Итог: 6 scouting subagents, parent repo/issue/PR/community/test gates, internal 3-role critique, 0 upstream PR/comment, final next_status: WATCH; PrefectHQ/fastmcp#4246 лучший future MCP/server lead but blocked by maintainer PR #4248 and assignment-first gate; langchain-ai/deepagents#3639 лучший secret-free regression lead after current-main repro; openai/openai-agents-python остаётся post-merge gap watch because current lanes are PR-covered.
| Проект | Upstream | Наш контур | Статус | Следующее действие |
|---|---|---|---|---|
| NousResearch/hermes-agent | #32999, #32963, #33042 | #6, watch | #32963 merged outage fix; #32999 closed obsolete after #33042 removed the SDK stream helper path entirely | Accept closure; keep boundary-regression lesson in skill/watch |
| pydantic/pydantic-ai | #5678 | #14 | Open, mergeable, CI green; Codex review quota warning only | Watch maintainer/bot review |
| pydantic/pydantic-ai | #5680 | #16, watch | Open, mergeable, CI green; Codex review quota warning only | Watch maintainer/bot review |
| pydantic/pydantic-ai | #5688 | #47, #52, cycle 15, patch prep, popular scouting | WATCH; duplicate-covered by open #5694, local regression-first patch оставлен как evidence | Мониторить #5694; не делать racing PR/comment, если PR не закрыт без merge, maintainers не попросили alternate patch, и fix не оказался неполным |
| pydantic/pydantic-ai | #5681 | watch | Competing PR open, mergeable, CI green | Watch/review only; do not duplicate |
| trycua/cua | #1725, #1737 | #15, watch | WATCH / competing PR #1737 pending maintainer review and Windows smoke; not independently verified fixed |
Watch #1737; no competing PR/comment unless maintainers ask or Windows smoke still fails |
| cline/cline | #10737 | #17, watch | Strong bug, but duplicate-race with stale/conflicting PRs | Triage existing PRs; no fresh PR yet |
| openai/codex | #24704 | #18, watch | Strong subagent/cache bug; reporter has reference PR | Watch; no duplicate PR unless maintainers ask |
| openai/openai-agents-python | #3512, #3513, #3348, #3508, #1845, #1168 | #52, cycle 25 | WATCH / strong repo-fit, but no issue-level PR-ready target; narrow bugs have active competing PRs and realtime leads are old | Fresh current-main repro + duplicate gate before any comment/PR; no blind PR hunting |
| stablyai/orca | #2930, #2948, #2931, #2940, #2941 | #82, #52, watch, cycle 31, follow-up | WATCH / #2930 is exact-covered by open non-draft upstream PR #2948, which touches ssh-filesystem-stream-reader.ts and adds a short-final-chunk rejection regression. |
Watch #2948 merge/close/stall/gap; no runner or upstream comment/PR while the exact PR is live |
| SWE-agent/mini-swe-agent | #826, #843, #832, #821, #728 | #53, #52, cycle 27, follow-up | WATCH / 2026-05-28 01:14 correction: #826 is still open/no comments, but merged timeout PR #832 is a potential symptom cover even though current local.py still has subprocess.run(..., shell=True, timeout=...); do not claim clean uncovered bug before re-test. |
Wait for corp-opensource-runner; then run post-#832 current-main repro, capture PID/ps regression card, repeat duplicate gate and 3-subagent critique before upstream comment/PR |
| google-gemini/gemini-cli | #27503, #25378 | #85, cycle 37 | CANDIDATE / SearchText fallback appears to keep ripgrep-only --json when invoking git grep; issue is open P1/agent/bug and adjacent PR #25378 is Windows EFTYPE-only, dirty/conflicting, not an exact cover. |
Run current-main runner repro or focused arg-builder test, repeat duplicate PR search and contribution/CLA gate, then 3-role critique before any upstream comment/PR |
| oraios/serena | #1519, #1456, #1509, #1082, #1130 | #84, cycle 32, 10:26 watch | SETTLED / #1519 closed 2026-05-28 09:30 UTC; maintainer says fixed in 6dd3301e48. |
Remove from active runner backlog; no upstream comment/PR unless regression reopens |
| CopilotKit/CopilotKit | #5035 | #19, watch | Open, mergeable; docs preview green; other Vercel previews need team authorization | Watch maintainer review |
| langchain-ai/deepagents | #3616, #3587 | #20, watch | WATCH / issue waiting on author raw Qwen/LiteLLM payload; likely provider payload ambiguity, not safe Deep Agents patch yet | Wait author/provider evidence; no PR without raw payload comparison |
| OpenHands/OpenHands | #14563 | #26, #52, popular scouting | COMMENT-FIRST; confirmed host-global skills visibility gap, direction open between CLI mount and personal-skills roadmap | Wait maintainer direction |
| browser-use/browser-use | #4877, #4801, #4846 | scouting | Good repo, but top candidates already covered by open mergeable PRs | Watch/review only; no duplicate PR |
| browser-use/browser-use | #4796, #4815 | #27 | Duplicate-covered; PR open/mergeable/green in tracked state | Watch/review only |
| OpenHands/software-agent-sdk | #3394, #2806 | #22, watch | PR open; old CHANGES_REQUESTED remains; maintainer said it looks fine and eval was triggered; no visible eval pass yet |
Wait maintainer eval / workflow approval / re-review |
| anomalyco/opencode | #29530, #29525 | #23, watch | PR open; duplicate/compliance/standards checks green; current gh mergeability MERGEABLE/BLOCKED, no maintainer review |
Watch maintainer review |
| cline/cline | #11087, #11086 | #24, scouting | PR open, mergeable; Greptile gap addressed; Quality Checks + Ubuntu/Windows SDK tests green | Watch maintainer review |
| cline/cline | #10139, #10384, #10140, #10141 | cycle 10 | Duplicate-race; #10384 has strongest configurable fix/tests but all competing PRs conflict | Watch maintainer consolidation; no PR |
| e2b-dev/E2B | #1352 | #25, scouting | Strategic sandbox reconnect candidate; may need live sandbox | Track; avoid PR until repro/test surface clear |
| langchain-ai/langgraph | #7688 | #48, watch, cycle 17 | WATCH / reproduced: installed langgraph-api==0.8.0 misclassifies TIME_WAIT as active port use; evidence comment posted |
Wait maintainer direction on public langgraph-api patch surface before PR |
| aaif-goose/goose | #9136 | #28 | Real bug; no duplicate PR; comment-first assignment ask posted | Wait maintainer assignment |
| pydantic/pydantic-ai | #5419, #5443 | #29 | Duplicate-covered; PR open/mergeable but coverage/check failing in tracked state | Watch existing PR, no duplicate |
| agno-agi/agno | #8062, #8124 | #30, cycle 8 | Duplicate-covered; our repro/patch worked, but #8124 auto-closed because #8065/#8084 already referenced #8062 | Watch #8065/#8084; offer integration-style MCP tool test if maintainers ask |
| langchain-ai/langgraph | #7684 | #31, cycle 8 | No-go; three assignment-gated PRs already closed | Watch only |
| trycua/cua | #1724 | #32, #52, cycle 8, popular scouting | WATCH / candidate only after fresh current-main macOS/AppKit repro; old local repro is not enough for PR-READY | Reproduce cua-driver mcp SIGABRT on current main before any PR |
| openai/codex | #24612 | #33, cycle 8 | Candidate/watch; no direct PR found; provider-history sanitizer bug | Watch/access-sensitive |
| aaif-goose/goose | #9332 | #34, watch | Comment-first; no duplicate PR; maintainer-alignment comment posted | Wait direction or run Linux repro on runner |
| aaif-goose/goose | #9446, #9447, #9449 | #49, cycle 18, cycle 19, watch | Closed, not merged; maintainer closed as superseded/fixed in #9449 |
No action; keep as duplicate-race lesson |
| BerriAI/litellm | #28962 | #35, cycle 9 | No-go / release-followup; mocked current branch maps Gemini 503 MaskedHTTPStatusError to ServiceUnavailableError |
Watch reporter/upstream; no PR unless exact leaking path is confirmed |
| microsoft/playwright-mcp | #1635, #1636 | #36, watch, cycle 9 | WATCH / #1635 strong CDP navigation lane but comment-first already done; no maintainer direction, runner repro, or full duplicate gate. #1636 is secondary long-run memory watch, not merged into #1635. |
Wait maintainer direction or approval; then repeat gates and run browser/CDP repro in runner before PR |
| ChromeDevTools/chrome-devtools-mcp | #2138, #2137, #1921, #2115, #1775 | #54, #52, watch, cycle 28 | WATCH / repo is strong AI-browser target; #2138 is internal CANDIDATE only, no upstream action. #2115 already has merged/released #2134; #1921 needs runner repro; #1775 has competing stale PR and active author. |
For #2138, scan validatePath call sites, repeat duplicate gate, fill test card; for browser lanes, run dedicated runner repro before any upstream comment/PR |
| PrefectHQ/fastmcp | #4143, #4124, #4192, #4246, #4161, #4205 | #55, #52, watch, cycle 28 | WATCH / assignment-first repo; #4143 and #4124 are comment-first-after-MRE lanes, not PR-ready. #4246/#4161/#4205 have active covering PRs; #4192 needs Windows current-main repro. |
Run secret-free MRE for #4143/#4124, then ask for assignment/approval before any PR; no upstream action in this cycle |
| browseros-ai/BrowserOS | #1005, #1029, #944, #864, #1031, #948 | #81, #52, watch, cycle 30 | CANDIDATE / #1005 is open, unlabeled and unassigned; exact PR search found no covering notifications/tools/list_changed fix. Source path points to per-request MCP server creation and repeated static tool registration, but broadcast source is not runner-proven. |
Run focused Bun/MCP repro in dedicated runner around packages/browseros-agent/apps/server; prove unchanged registry emits repeated notifications, then repeat duplicate gate before any upstream comment/PR |
| generalaction/emdash | #1875, #1907, #1908, #1909, #1910, #1911 | #83, watch, follow-up | CANDIDATE / maintainer says PR welcome; no open exact PR, but closed unmerged prior-art PRs #1907-#1911 attempted D-Bus / gnome-libsecret fixes and exposed robustness/build pitfalls; Linux Secret Service/non-GNOME fail-before not runner-proven. |
Run Linux runner repro with Secret Service and non-GNOME XDG_CURRENT_DESKTOP; design must avoid prior PR failures, then repeat duplicate gate and 3-role critique before PR. |
| BloopAI/vibe-kanban | #3329, #3247, #3408, #3420, #3394 | #59, #52, watch, cycle 29 | WATCH / strong control-plane reference but repo is sunsetting, README requires discussion before PR, and maintainer rhythm is weak. #3329 is a strong lead after repro, #3247 is lead/watch, #3349/#3343 are covered/overlapped. |
Watch maintainer/community signal; only after rhythm improves run issue-level repro/duplicate gate and start with discussion/comment, not PR |
| Infisical/agent-vault | #194, #196, #197, #198, #207 | #60, #52, watch, cycle 29 | WATCH / strong credential-boundary infra watch. #194 is best lead, but current evidence points to Codex/Agent Vault compatibility and does not prove Agent Vault-owned patch surface; latest v0.22.0 repro missing. |
Run secret-free v0.22.0/current-main repro in runner, recheck MITM/proxy PR overlap, then decide whether a narrow maintainer question is worth COMMENT-FIRST |
| HKUDS/OpenHarness | #47, #221, #18, #250, #93, #34 | #61, #52, watch, cycle 29 | WATCH / high-fit Python agent harness and Ohmo runtime, but open issues are mostly feature/design and current PR queue has high overlap/churn. No selected bug target or runner-backed repro. | Future issue-level cycle should start with #47 only if current-main source check proves wrapper incompatibility; otherwise wait for reproducible harness bug |
| langchain-ai/deepagentsjs | #547, #553, #550, #542, #554, #551, #541, #288 | #62, #52, watch, cycle 29 | WATCH / high-fit JS agent harness. #547 is strongest uncovered lead, but approval-first CONTRIBUTING, no runner repro, and active covering PRs for nearby fresh bugs block upstream action. |
Reproduce #547 on current main in runner, repeat duplicate gate, then ask a narrow maintainer question before any PR |
| ag-ui-protocol/ag-ui | #1635, #1253, #837, #1727, #1468 | #56, #52, refresh, watch, cycle 28 | COMMENT-FIRST after runner evidence / #1635 is still open with no exact PR, but AG-UI is assignment-first and near Mastra adapter PRs create collision risk. Not PR-ready and no upstream comment until runner-backed test card. |
Run secret-free @ag-ui/mastra unit repro in runner, recheck #1253/#837/#1727/#1468, then ask maintainers whether a narrow payload-less chunk fix should be separate |
| mcp-use/mcp-use | #1581, #1576, #1570, #1568, #1569, #1551, #1555, #1438, #1415 | #57, #52, watch, cycle 28 | WATCH / duplicate-heavy; no PR-ready candidate. #1551/#1555 already covered by merged PRs; fresh security/OAuth/session/router lanes have active covering PRs. |
Watch #1582/#1578/#1573/#1574/#1572/#1531/#1439/#1455/#1421; re-open only after unmerged gap, maintainer ask, or current-branch repro |
| mastra-ai/mastra | #17118, #17002, #17054, #17083, #17137, #16957 | #58, #52, refresh, watch, cycle 28 | CANDIDATE / #17118 remains open and unassigned with no exact PR; #17054 is merged adjacent ask_user wrapping only, while open #17083 is broad/conflicting Mastracode overlap, not an exact duplicate. |
Run secret-free Mastracode slash-command picker MRE/test in runner with concrete terminal-width fixture, then repeat duplicate gate before any upstream comment/PR |
| modelcontextprotocol/python-sdk | #2702 | #52, popular scouting | COMMENT-FIRST; fresh FastMCP Streamable HTTP + Starlette BaseHTTPMiddleware candidate, no direct duplicate PR found |
Wait maintainer ack/label; decide warning/docs guard vs runtime compatibility |
| cline/cline | #10931 | #52, popular scouting | COMMENT-FIRST; public PR gate clean, Linear-linked, broad terminal/pager execution-policy scope | Wait maintainer narrowing before code |
| modelcontextprotocol/typescript-sdk | #2155, #1925, #1926 | #37, cycle 9 | Duplicate-triage; server-side SSE escaping already covered by open mergeable PRs with tests | Watch maintainer review; validate only uncovered client fail-fast / JSON-response gap |
| modelcontextprotocol/typescript-sdk | #2108, #2111 | cycle 10 | Duplicate-covered; #2111 covers both mismatch directions plus match/absent initialize cases with green checks | Watch maintainer review; no PR |
| modelcontextprotocol/typescript-sdk | #2115, #2116, #2138 | #51, cycle 21, cycle 22 | WATCH / duplicate-race; 23:37 UTC heartbeat confirms exact competing PRs #2116/#2138 cover the requestId: 0 contract with tests, changesets, and green checks |
Watch maintainer choice/merge/close; no PR/comment unless both competing PRs close unmerged or maintainer asks |
| modelcontextprotocol/typescript-sdk | #2126, #2140, #2127 | cycle 13 | Duplicate-covered by #2140; #2127 closed duplicate; #2140 has regression + changeset but one unrelated-looking integration job red | Watch #2140; no PR |
| modelcontextprotocol/python-sdk | #2687, #2701 | #38, cycle 9, cycle 19 | Duplicate-covered by open mergeable #2701; covers AnyHttpUrl vs AnyUrl redirect mismatch with regression test |
Watch #2701; no duplicate PR |
| google/adk-python | #5864 | #46, cycle 14, patch prep | Local regression-first patch prepared; issue still assigned/request clarification, maintainer asked reporter to validate full test pass; no duplicate PR found |
Wait maintainer/reporter direction before PR |
| modelcontextprotocol/python-sdk | #2578, #2590, #2645, #2646 | cycle 10 | Duplicate-covered; all three linked PRs cover refresh resource omission + trailing slash normalization with green checks | Watch maintainer selection; no PR |
| BerriAI/litellm | #28971, #28970 | #39, cycle 10 | Comment-first / missing-regression-offered; #28970 is adjacent /apply_guardrail, not exact /v1/responses spend-log path |
Wait maintainer signal before PR |
| e2b-dev/E2B | #1349, #1354, #1355 | #39, cycle 10, cycle 18, follow-up | SETTLED / #1355 merged 2026-05-27, covers JS + Python SDK multi-source Dockerfile parsing, and #1349 is closed fixed |
No further action on #1354/#1349 unless issue reopens or latest SDK regresses |
| openai/codex | #24725 | #41, cycle 11 | Regression-first candidate, but upstream external PRs are invitation-only | Watch/comment-first; no PR without maintainer signal |
| anomalyco/opencode | #15226, #29565 | #44, watch, cycle 12 | PR open; duplicate/standards/compliance/contributor checks green; current gh mergeability MERGEABLE/BLOCKED, no maintainer review |
Watch maintainer review; no extra opencode PR until review or distinct gap |
| langchain-ai/deepagents | #3568 | scouting | Small prompt/schema bug, but reporter has local fix and assignment gate risk | Watch/comment only if maintainers ask |
| continuedev/continue | #12334 | #50, watch | WATCH / needs reporter path; current main at cb273098 preserves reasoning on tested raw completion path; upstream evidence comment posted |
Wait reporter version/path or exact outbound body before PR |
- #1 Track Hermes Codex null-output upstream fix
- #2 Launch open-source bug contribution program
- #3 Scout Hermes follow-up bugs after Codex null-output merge
- #4 Scout opencode for reproducible coding-agent bugs
- #5 Scout MarkItDown for fixture-friendly conversion bugs
- #6 Hermes Codex recovery may mask unrelated TypeError after #32963
- #7 MarkItDown IpynbConverter crashes on non-ASCII PDF
- #8 opencode Task subagents omit tools for Bedrock/LiteLLM
- #9 Firecrawl cancel crawl floods webhook failures
- #10 Provision corp-opensource runner container on corp-server
- #11 Weekly AI-native framework top scan
- #12 Design Hermes/Codex worker pool for parallel scouting
- #13 Scout Paperclip-like harness expansion opportunities
- #14 pydantic-ai serialize_any loses binary data
- #15 trycua Windows element-index clicks miss click.png marker
- #16 pydantic-ai TextContent metadata lost in UI adapter round-trip
- #17 cline MCP task_progress leaks into MCP args
- #18 openai/codex forked subagents lose prompt-cache lineage
- #19 CopilotKit workspace protocol deps leak into published package
- #20 deepagents subagent task fails without tool_call_id on OpenAI-compatible Qwen
- #21 OpenHands conversation metadata race hides started resolver conversations
- #22 OpenHands SDK tmux 1000x1000 viewport burns CPU
- #23 opencode webfetch nullable format mismatch
- #24 cline CLI sends unsupported reasoning field to OpenAI-compatible GLM endpoints
- #25 E2B resumed sandbox stdout stream wedges after large response
- #26 OpenHands global skills не видны в openhands serve
- #27 browser-use boolean index action schema
- #51 MCP TypeScript SDK requestId 0 cancellation candidate
- #52 Popular AI-native / Paperclip-like repo scouting 2026-05-27
- #68 Qwen Code / Agent Framework repo-scope watch
- #69 gptme / CAO / OpenACP verify-first watch
- #70 Browser-agent repo-scope watch
- #71 Pydantic AI / OpenHands / E2B top-tier repo watch
- #72 Cline / CUA / Playwright MCP repo-scope watch
- #73 FastMCP / OpenAI Agents / DeepAgents repo-scope watch
- #74 High-churn terminal runner repo watch
- #75 MCP ecosystem repo-scope watch
- #76 Observability / eval harness repo-scope watch
- #77 RAG / memory retrieval infra repo-scope watch
- #78 Low-code agent workflow builder repo-scope watch
- #79 Orchestration / tool / eval framework repo-scope watch
- #80 Coding assistant context and code-search repo-scope watch
- #81 BrowserOS #1005 MCP list_changed flood candidate
- #82 Orca #2930 short stream chunks zero-fill candidate
- #83 Emdash #1875 Linux safeStorage non-GNOME candidate
- #66 AI SDK / Agents repo scope watch
- #65 Docker MCP exact-validation scope watch
- #67 Hermes #15640 stale cleanup candidate
- #53 SWE-agent mini #826 timeout leaves child processes running
- #28 goose repeated tool calls bypass RepetitionInspector
- #29 pydantic-ai multi-MCP discovery AG-UI stalls
- #30 agno MCP tools crash when stream=True returns async iterator
- #31 LangGraph PostgresStore numeric filters assignment-gated
- #32 trycua cua-driver mcp SIGABRT before handshake
- #33 Codex web_search_call history crashes custom provider after model switch
- #34 goose PR_SET_PDEATHSIG kills stdio MCP subprocesses
- #35 LiteLLM Gemini AI Studio 5xx error mapping
- #36 Playwright MCP browser_navigate_back CDP timeout
- #37 MCP TypeScript SDK U+2028/U+2029 SSE timeout duplicate triage
- #38 MCP Python SDK AnyUrl/AnyHttpUrl OAuth redirect mismatch
- #39 Cycle 10 scouting: LiteLLM guardrail duplicate gate and E2B next repro
- #40 Cycle 11 scouting: regression-first AI-native bug targets
- #41 Codex plugin-namespaced skills exceed MAX_NAME_LEN
- #42 Cycle 12 scouting: unblock next AI-native contribution target
- #44 opencode structured output conflicts with thinking tool_choice
- #45 Cycle 13 scouting: no safe new PR after six-agent triage
- #46 google-adk vertexai optional dependency import guard
- #47 pydantic-ai MCPToolset http_client follow_redirects guard
- #48 LangGraph dev TIME_WAIT port false-positive
- Репозиторий ведём на русском: README, playbooks, watch notes, issue bodies.
- Для bugfix PR сначала оформляем regression card в issue/watch note:
контракт,test file,command,pre-fix fail,post-fix pass. Если pre-fix failure нельзя дешево доказать, candidate остаётсяWATCH / needs-repro, а не ready PR. - Разведку по новым target repo делаем через субагентов:
- repo-fit: активность, contribution rules, test surface, maintainer response;
- bug-signal: свежие баги, дубликаты, реакции, боль пользователей;
- repro-path: можно ли доказать баг локально/fixture/CLI без внешних секретов;
- patchability: где минимальный patch surface и есть targeted tests;
- duplicate-race: linked PRs, issue timeline, competing PRs, maintainer comments;
- PR-readiness: contribution guide, PR template, issue policy, test commands, draft/ready expectation.
- После синтеза и parent live gates делаем 3-subagent critique:
- fact-check: не оверклеймим freshness, maintainer signal, duplicate status и результат субагентов;
- process-gate: проверяем порядок
6-subagent scouting -> parent live gates -> 3-subagent critique -> comment/PR/tracker; - actionability: проверяем, что следующий шаг один из
CANDIDATE,COMMENT-FIRST,PR-READY,WATCH,NO-GO, и у него есть конкретный gate/unblock event.
- Не начинаем PR без воспроизведения или понятного failing/regression test.
- Для bugfix PR сначала доказываем, что regression test реально ловит старый баг: временно возвращаем старое поведение или запускаем тест на pre-fix state и записываем точный failing assertion/error.
- Для schema/runtime mismatch regression пишем через raw wire/tool boundary: передаём тот же payload, который видит runtime/модель (
null, missing default, etc.), а не только typed helper. - Для downstream recovery вокруг SDK/dependency parser failures тест обязан доказать две стороны границы: SDK failure восстанавливается через публичный runtime path, а app-side callback/handler
TypeErrorс таким же текстом не проглатывается. Helper predicate test здесь недостаточен. - Для publish/release багов проверяем не только helper, но и внешний артефакт/контракт: registry metadata, package manifest, install/runtime symptom или closest publishable fixture.
- Для optional dependency import bugs regression должен проверять публичную границу импорта и понятный missing-extra/install-extra error. Helper-only тест недостаточен, если пользовательский симптом — raw
ModuleNotFoundErrorпри импорте. - Для adapter bugs вокруг user-supplied client/callback/factory regression должен идти через публичный constructor/config API и runtime callback shape, которую реально вызывает зависимость. Helper-only тест недостаточен.
- Если есть competing PR, ведём duplicate triage публично: что сравнили, что портировали, чего не хватает.
- Duplicate scan повторяем прямо перед кодом. Если candidate уже закрыт open/mergeable PR с тестами, переключаемся в watch/review/validation.
- Реакции и лайки считаем impact-сигналом, но не approval и не CI.
- Стратегия open-source contribution
- Repo scope: AI-native frameworks and harnesses
- Разведка багов в репозиториях
- Гонка duplicate PR
- Follow-up по старым upstream PR/issues
- Upstream follow-up 2026-05-27
- Runner-контейнер на corp-server
- AI-native frameworks scouting cycle 2
- AI-native frameworks scouting cycle 7
- AI-native frameworks scouting cycle 8
- AI-native frameworks scouting cycle 9
- AI-native frameworks scouting cycle 10
- AI-native frameworks scouting cycle 11
- AI-native frameworks scouting cycle 12
- AI-native frameworks scouting cycle 12 PR-readiness
- AI-native frameworks scouting cycle 13
- AI-native frameworks scouting cycle 14
- AI-native frameworks scouting cycle 15
- AI-native frameworks scouting cycle 16 duplicate gate
- AI-native frameworks scouting cycle 17
- AI-native frameworks scouting cycle 18
- AI-native frameworks scouting cycle 19
- AI-native frameworks scouting cycle 22
- Paperclip-adjacent repo map 2026-05-27
- Popular AI-native / Paperclip-like scouting 2026-05-27
- Popular AI-native repo universe cycle 23
- Popular AI-native repo universe cycle 24
- Popular AI-native openai-agents cycle 25
- Popular AI-native SWE-agent mini cycle 27
- SWE-agent mini #826 follow-up
- Popular AI-native repo universe cycle 29
- Popular AI-native repo universe cycle 30
- Popular AI-native repo universe cycle 31
- Upstream follow-up 2026-05-28
- BloopAI/vibe-kanban scouting
- Repo cards: AI-native universe 2026-05-27
- BrowserOS #1005 MCP list_changed flood
- Orca #2930 short stream zero-fill
- Playwright MCP #1635 CDP back-navigation timeout
- google/adk-python #5864 Vertex AI optional dependency import guard
- pydantic-ai #5688 MCPToolset http_client follow_redirects patch prep
- LangGraph #7688 TIME_WAIT port false-positive
- Goose #9332 PDEATHSIG MCP subprocess lifecycle
- Goose #9446 response.completed missing output
- Continue #12334 OpenRouter autocomplete extraBodyProperties
- Retrospective: scouting cycles 18-20
repo-scouting— поиск и оценка репозиториев.candidate-bug— потенциальный баг для PR.needs-subagents— нужна параллельная разведка.needs-repro— нет подтверждённого воспроизведения.duplicate-triage— есть конкурирующие PR.release-followup— проверить, дошёл ли fix до релиза.