## (owner-prompted, GitHub-stars entrant scan added)

- Owner instruction in chat: added a rule to AGENTS.md to fetch his GitHub stars and cross-check them against category membership as part of the daily entrant scan [glm-5.3-flash]
- Baseline pass: 2,823 starred repositories, 71 already tracked or cited, 2,752 untracked, with the agentic subset enumerated for the next refresh [glm-5.3-flash]
- Surfaced candidates by category fit, among them openinterpreter, MetaGPT, AutoGPT, autogen, lobehub, litellm, autoarena, pair-review, dotskills, library-mcp, docling, open-webui, anything-llm, private-gpt, ollama, and browser-use [glm-5.3-flash]

## (owner-prompted, stars-scan candidates resolved)

- All 23 candidates resolved in one run: 11 notes and 12 logged rejections, no carryover [glm-5.3-flash]
- Added notes: litellm and ollama (model access), openinterpreter (harnesses), docling (retrieval), metagpt, autogpt, autogen, lobehub (orchestration), open-webui, anything-llm, private-gpt (assistant runtimes), each with its matrix column and index entry [glm-5.3-flash]
- Lead corrections: Ollama's free-only premise was falsified (it now sells cloud tiers, Pro $20, Max $100, Team $500); openinterpreter's 1.0 line is a Codex CLI fork [glm-5.3-flash]
- Rejected: autoarena (dormant since 2024-12), pair-review (no coverage), claude-code-viewer (single author), llm-conversations-viewer and tps-viewer (the owner's own days-old tools), chat-history (stale since February), dotskills (two days old), library-mcp (dormant five months), hubproxy (stale nine months), octopool (thin docs), metacoder (stale nine months) [glm-5.3-flash]
- browser-use rejected on category fit, not quality; a `? for tom:` question asked whether to open a browser-agents category [glm-5.3-flash]
- Category consequences: orchestration, assistant-runtimes, and model-access scopes widened; five matrices extended (model-access 16 to 18, orchestration 16 to 20, assistant-runtimes 9 to 12, retrieval 6 to 7, harnesses 28 to 29) [glm-5.3-flash]

## (owner answer, browser automation stays out)

- Owner answered the browser-use question "No browser use": no browser-agents category, browser automation stays out of scope, and the rejection is final [glm-5.3-flash]

## (defect, matrix separator rows)

- Owner reported the Assistant Runtimes matrix rendering broken; this run's column-extension script had left the separator row at its pre-extension width, so GFM dropped the table [glm-5.3-flash]
- The same defect appeared in every touched table (assistant-runtimes, harnesses, model-access, orchestration, retrieval); all five separators were regenerated to match their headers [glm-5.3-flash]

## (owner-prompted, queue questions answered)

- Owner walked the six open queue questions; all six carry resolution lines in queue.md [glm-5.3-flash]
- Q1: the untracked ADE tracker was judged finished, moved from the repo root into agents/, and published as a section-root tracker [glm-5.3-flash]
- Q2: sepia and headcount cleared the bar and joined skills (matrix six to eight columns) [glm-5.3-flash]
- Q3: experiential cleared the bar and joined model access (matrix eighteen to nineteen columns); my-free-code rejected [glm-5.3-flash]
- Q4: the b32de458 sweep and two self-resolved questions closed, and the whole-table width check was adopted after the separator defect [glm-5.3-flash]
- Post-commit fixes: experiential was mis-alphabetized in the model-access matrix and index, sepia sat one slot late in the skills matrix, and three of the moved ADE tracker's cross-links carried wrong depths [glm-5.3-flash]
- Commit 450a634f also swept in the concurrent session's MiMo Code harness note [glm-5.3-flash]

## (owner-prompted, MiMo Code harness note)

- Added agents/harnesses/mimo-code/index.md, a note on Xiaomi's MIT OpenCode fork built around checkpointed memory, a goal-verifying judge, and the Xiaomi Token Plan [deepseek-v4.1-flash]
- harness-feature-matrix: added the MiMo Code column (twenty-nine to thirty members), backfilled the missing Open Interpreter Changes bullet, and corrected the intro count from twenty-eight to thirty [deepseek-v4.1-flash]
- Wording fix before commit: replaced "honest case" and "tree-shaped" to keep the banned-terms rule [deepseek-v4.1-flash]

## (owner-prompted, queue cleanup)

- Removed every answered question and its resolution line from queue.md, leaving only the standing research-index item and the Done list [glm-5.3-flash]

## (daily refresh)

- Harnesses: mimo-code's latest-release endpoint returns v0.1.14 after v0.1.15, aider's stall re-confirmed, claude-code 148.3k stars, cline 5.46M installs, deepseek-harness still prerelease-only [glm-5.3-flash]
- cmux: re-packaged its Max tier onto a shared 64 GB/16 vCPU pool across up to 50 VMs (Pricing and Price history rewritten) [glm-5.3-flash]
- openwork: pricing churned a ninth time with a Price history row (Solo and the grandfather clause gone, Team $10/seat unlimited, Enterprise $20/user annual) [glm-5.3-flash]
- Assistant runtimes: github-agentic-workflows v0.89.21 is now the newest stable-marked release, retiring the v0.88.7 last-stable record [glm-5.3-flash]
- Code review: open-code-review 41.7k stars (about 80 percent in fourteen days); ellipsis's free-tier badge flipped back to ChatGPT-or-Claude [glm-5.3-flash]
- Hybrid execution: jevbench shipped v1.4.2.1 with Plumb-4B the new #1 at 65.84 [glm-5.3-flash]
- model-selection: added a Claude Opus 5.5 column ($4/$20, 1M flat) and flagged gpt-5.6-terra and gpt-5.6-luna as delisted; corrected GLM-5.3-FlashX's cached read to $0.075 [glm-5.3-flash]
- Rejected: drawgent (no GitHub mirror), golive-skill (no independent coverage), magpie (no HN footprint), jevmem (self-reported benchmarks), jev-code-reviewer (thin footprint), plus the weak tail [glm-5.3-flash]
- _index.md: one-liners re-dated to 2026-09-27 and the Spec Kit line moved to v1.0.12 [glm-5.3-flash]

## (owner-prompted, orchestrator directory triage)

- Triaged the awesome-multi-agent-orchestrators directory against the section's 233 notes, deduped seven already-covered entries, and the owner approved the four in-scope gaps [deepseek-v4.1-flash]
- Added 11 notes to orchestration (Orca, Sandcastle, Lanes, Crewplane, Helmor, Open Swarm, LoopTroop, GraphCode, The Perfect Orchestrator, Multica, oh-my-codex) with matrix columns and index entries, the category 20 to 31 members [deepseek-v4.1-flash]
- Added 6 notes to control planes (Microsoft Agent Governance Toolkit, Veto, SettleBridge, Okto Pulse, SIDJUA, Code Atelier Governance SDK) with matrix columns and index entries, the category 2 to 8 members [deepseek-v4.1-flash]
- Added 3 notes to session analytics (AgentTrace, ClawTrace, Agent Analytics) with matrix columns and index entries, the category 4 to 7 members; Agent Analytics sits at the category's edge, and that caveat is recorded [deepseek-v4.1-flash]
- Added Agon (automated research) with matrix column and index entry, the category 7 to 8 members [deepseek-v4.1-flash]
- Rules (owner-directed): added the directory and openorchestrators.org to the standing scan in AGENTS.md step 3 [deepseek-v4.1-flash]
- Rejected: Nomad Inno and Agentix Labs (implementation services), Crewlet, Augment Code Intent, and AgentGrid (closed products), Databricks Unity AI Gateway (commercial platform feature); the broad framework tier was deferred [deepseek-v4.1-flash]