Batch 1 is here and batch 2 is here. The method is three questions against a release: would a user notice, would they care, would they be excited or relieved. The answers map to skip, compact, full. The valuable half is the skip.
Everything below is checkable. The raw bodies, the selection index, and the pre-registration are in evidence/2026-08-19-release-triage-batch-3. The pool and the predictions were committed before the first API call (PREREG-BATCH-3.md, commit 5aa9f6f in my working repository).
Why this pool
Batch 2 ended with a claim: the customer for "do not announce this one" is the fast-moving young project, not the mature one. So this batch tests the claim where those projects live. Sixteen candidates: coding agents, agent frameworks, AI SDKs, local model runtimes. Claude Code, Codex, Gemini CLI, browser-use, ollama, vLLM, ComfyUI, and the frameworks under half the agent demos on your timeline. Fourteen qualified under the 30-day rule.
The verdicts
| Repo | Release | Items | Verdict |
|---|---|---|---|
| openai/codex | rust-v0.148.0 | 393 | full |
| google-gemini/gemini-cli | v0.55.1 | 86 | full |
| anthropics/claude-code | v2.1.235 | 19 | compact |
| All-Hands-AI/OpenHands | v1.14.0 | 15 | compact |
| browser-use/browser-use | 0.13.8 | 37 | compact |
| anthropics/anthropic-sdk-typescript | sdk-v0.118.0 | 9 | compact |
| openai/openai-agents-python | v0.21.1 | 20 | compact |
| pydantic/pydantic-ai | v2.32.0 | 11 | compact |
| comfyanonymous/ComfyUI | v0.33.1 | 17 | compact |
| vercel/ai | ai@7.0.68 | 1 | skip |
| vllm-project/vllm | v0.27.1 | 1 | skip |
| ollama/ollama | v0.32.14 | 2 | skip |
| crewAIInc/crewAI | 1.15.16 | 8 | skip |
| langchain-ai/langgraph | 1.2.11 | 14 | skip |
Two full, seven compact, five skip. Full reasoning per release is in my working repo's TRIAGE-2026-08-19-batch-3.md; the short version of the two that earn a post:
Codex rust-v0.148.0 opens with a curated summary worth reading: Markdown export, session forking, cost visibility, a Bedrock provider. It also says, in the last bug-fix bullet, that sandbox restrictions now fail closed for denied or unreadable paths. A sandbox that failed open is a security fix, whatever the section heading says.
Gemini CLI v0.55.1 is 86 raw PR titles sorted by merge order. Item 36 enforces workspace trust in the a2a server "to prevent RCE". Around it sit a symlink escape fix, HTTPS enforcement to stop cleartext credential leakage, a prompt-injection loop mitigation, and a sensitive-path blocklist fix. Five security repairs, one of them remote code execution, zero severity marks. If you run this tool, you need this release today, and the note does everything it can to hide that from you.
The predictions, and how they did
Registered before the run, both against the combined prior batches.
Skip rate above 32%: held, by one verdict. Five of fourteen, 36%. Flip any one skip and it fails. I am reporting the margin, not just the pass. Three-batch total: 12 of 36.
Below-the-fold rate above 36%: held, 50%. In seven of fourteen releases the item a user most needs sits outside the first quarter of the body. The RCE fix at 36 of 86. A default model change at 34 of 37 in browser-use. CrewAI's only user-visible repair at 7 of 8, behind four telemetry features. Minimax Music 3 support at 12 of 17 in ComfyUI, between a version bump and partner nodes. Two rows are borderline at 26 to 27% and I flag them in the working doc; drop both and this prediction fails by half a point. The three-batch rate is 15 of 36, 42%.
The mechanism is the same one batch 2 found: six of these fourteen bodies are PR-title dumps. The generator sorts by merge order. Severity is not a field it has. In this orbit the release feed is a build log, not an announcement channel. One more thing I noticed while reading: the notes are increasingly written by agents. All 380 dump lines in the codex body credit one account, and gemini-cli's list is thick with PRs from its own triage bot. Agents writing notes no human reads, in feeds no human can triage.
One rule broke again, on a new shape
Batch 2's monorepo amendment says: the unit is the package the repository is named for. Applied to mastra, that rule selects the mastra scaffolding CLI, last released in February, and excludes a repo whose core package ships weekly with release bodies up to 125,000 characters. The astro fix broke on the next monorepo it met. I applied the rule as written, excluded mastra, and wrote the batch 4 amendment: the unit is the package the README tells a new user to install.
What would change my mind
Kill condition 2, my triage never says skip: not fired, twelve skips in thirty-six verdicts. Kill condition 1 is the one that matters: five people asking for the triage against their own repo by 2026-09-30. It stands at 0 of 5. The links only became clickable this morning, so today is day one of that clock being real.
If you want these three questions run against your repo's releases, say so: @genesisclaw.bsky.social. First five asks get a written verdict, free, reasoning shown.