## (daily refresh)

- 29 new research notes, each with its sorted matrix columns, category-index entries, and Added bullets: orchestration 62 to 66 (Symphony, herdr, AgentBox, EvoAgentX), harnesses 34 to 37 (Mini-SWE-agent, Oh My Pi, Atomic Agent), automated research 12 to 17 (Karpathy Autoresearch, DeepAnalyze, Dexter, ARIS, AutoResearchClaw), evaluation and review 10 to 13 (Kiln, PR Lens, Verbatim Citation Gate), context engines 14 to 16 (LLMLingua, WrenAI), surfaces 16 to 17 (Vibeyard), session analytics 13 to 15 (Claude Code Agent Monitor, Coucou), software factory 7 to 9 (Fabro, Zeroshot), memory 22 to 23 (deja), model access 25 to 26 (LocalAI), code review 9 to 10 (git-lrc), executions 6 to 7 (Trigger.dev), task management 7 to 8 (Epiq), spec-driven development 10 to 11 (Intent-Router), trackers and leaderboards 7 to 8 (Benchmark Radar) [glm-5.3-flash]
- kimi-code: the kimi-k3 listing moved a tenth time, with a new Wafer endpoint floor ($0.40/$15.00, later $0.37/$14.89) and Relace leaving the floor at $2.00/$14.00 [glm-5.3-flash]
- Model access: Claude Max gained new included Claude Platform credits ($100/$200); SuperPlane published its first pricing (Business $199 per organization per month); Experiential's repository and org went 404 while the product and PyPI stay live; the GLM off-peak promotion ended October 7 [glm-5.3-flash]
- Status corrections: Superpowers 7.0 corrected the fixed-by-design claim; hol-guard v3.39.0's 2M-download claim was flagged against the measured 76,527/month; okto-pulse's stars halved to 52 in three days; yolobox's release count corrected from thirty to sixty-six [glm-5.3-flash]
- Rejected: the mission-control tail (RunMaestro/Maestro, tuios, qm), kestra (data-engineering-first), the absorbed assistant-runtime family (jev-chat-jarvis, AgentVerse-OS, Coworker, ironclaw, openfang), microsoft/agent-framework (single-vendor SDK), DeepAudit (a vulnerability-audit platform), OpenBitFun and webcmd on category fit; retrieval escalated CocoIndex, Crawl4AI, and DSPy to the owner as family-boundary calls [glm-5.3-flash]
- Section roots: model-selection added Claude Haiku 5.5 ($0.10/$0.50 with the 100K repricing cliff, 1M context) and recorded terra and luna's third drop off OpenAI's pricing page; the Kimi K3 floor moved an eighth time in eight days; context-management-patterns folded the 5,000-per-skill reinjection caps and the /autocompact window [glm-5.3-flash]
- Queue: eight new `? for tom:` questions appended (OpenHuman's scale call, OpenBitFun's shelf, the four dedicated-list wirings, Skill_Seekers, the retrieval family trio, and the AMO ts-file walk) [glm-5.3-flash]
- _index.md: every essay and matrix one-liner re-dated to 2026-10-10 with the twenty-nine membership changes in their counts [glm-5.3-flash]