Issue No. 5
The Nov 2 edition
Snapshot#
- Quiet Sunday-into-Monday window. The OpenRouter models list has no new model IDs (the newest is still particle-1 from 10-28), Artificial Analysis articles have nothing newer than 10-28, and the AA Coding Agent Index still carries the 2026-10-30 materialization. Today's useful items are one harness release, one local-inference engine, one security incident and one eval lesson.
- News (harness): pi v1.5.0 shipped this morning (11-02 05:05 CST). It adds GPT-6.1 Terra (500K context through Codex), switches the default OpenAI model to
gpt-6.1-sol, moves tool-argument validation topi-schema(a breaking change for TypeBox-only features), and changes MCP OAuth sign-in to Client ID Metadata Documents by default. - News (local inference): colibri is #10 on GitHub trending with 655 stars today (46,218 total, Apache-2.0). Its v2.1.0 release (10-31) runs published GGUFs unchanged on CPU, Vulkan and CUDA, streams MoE experts from disk, and adds Mistral-Large-4 (1T) to the supported models.
- News (security, agent infra): Trigger.dev disclosed a security incident on 11-01 in which an attacker accessed environment variables for a subset of organizations. The company says the hole is blocked and affected owners were emailed (HN, 118 points).
- Update (eval lesson): Apify's MCP server removed TOON output (PR #1060) after A/B rounds showed no consistent benefit over JSON. A Forkast write-up published the numbers on 11-01, and The Agentic Digest led with it today.
- Context (open weights): China's MOFCOM Announcement No. 48 proposes export restrictions on frontier model weights; comments are open until 2026-11-15. Kimi K3.1 Pro weights are still not on Moonshot's HF org.
- Do this today: Ling 3.1 Flash is free on Vercel AI Gateway only until Nov 3, so finish any evals on it today. If you run pi, read the v1.5.0 breaking change before updating custom tools.
Frontier and open-weight models#
Nothing new shipped in the window. These are the datapoints that still drive choices.
Still open#
- Gemini 4 Carbon: still #1 on the AA Intelligence Index v5 (AA article dated 10-28). On the Coding Agent Index it scores 59.4 at $7.74 per task with Antigravity CLI (High). Access is still limited.
- GPT-6.1 Terra: $1 in / $5 out per 1M tokens on OpenRouter ($0.50 / $2.50 batch), 1.05M context. It is the closed model with the best price per point in the Codex rows (section 3), and pi v1.5.0 now supports it with thinking
offand mid-conversation thinking-level changes. - Mistral Large 4 (open weights): 1T total, 52B active, with an NVFP4 build. It scores 48.3 at $1.26 per task on the Coding Agent Index (Mistral Vibe, High). New today: colibri v2.1.0 lists it as runnable on its GLM-5.2/5.3 engine, which streams experts from disk instead of needing a multi-GPU node.
- DeepSeek V4.1 Flash 1017: MIT, still #1 on HF trending models, $0.15 / $0.60 per 1M on OpenRouter. The Chinese-lab open-weight value leader (section 3).
- Kimi K3.1 Pro: $2 / $10 per 1M on OpenRouter. AA placed it 4th on its index at launch (AA). Weights are not published yet; the newest model on Moonshot's HF org is still Kimi-K3. The "by Nov 4" date comes only from The Agentic Digest; I found no Moonshot primary for it.
Context: China's draft export controls on frontier weights#
- What: MOFCOM Announcement No. 48 (issued 10-16 with MOST) proposes adding 23 AI technologies to China's restricted-export catalogue. According to the Cloud Security Alliance research note (surfaced on HN today, 85 points), weights are restricted if a model meets any one of three tests: at least 10^26 training FLOPs, at least 2T parameters, or a global top-10 position on public rankings or call volume.
- Status: the catalogue changes are a draft open for comment until 2026-11-15. A pre-export registration requirement already applies. MOFCOM's explainer says API, app and SaaS use does not count as export.
- Builder take: nothing changes for API use. If your plan depends on self-hosting a future top-ranked Chinese open-weight model, have a fallback ready; whether this affects any specific release, Kimi K3.1 Pro included, is not stated in any primary source.
Pareto frontier#
AA Coding Agent Index (v2.0: FrontierCode v1.1 + Terminal-Bench v4 + SWE-Atlas-QnA). The data is still the 2026-10-30 13:04 CDT materialization, so no score moved. Today's server-rendered page returned 31 rows; the Claude Code + Opus 5.5 and Sonnet 5.5 rows were not in that payload, so they are left out here rather than copied from memory. Cost is mean USD per task.
| Harness + model | Index | $/task | Note |
|---|---|---|---|
| Devin Fusion CLI + Opus 5.5 Max / GPT-6 Luna High | 61.5 | $8.00 | Top row in today's payload |
| GitHub Copilot + Copilot Fusion (max) | 59.9 | $5.59 | Best $/point near the top |
| Antigravity CLI + Gemini 4 Carbon (High) | 59.4 | $7.74 | Limited access |
| Claude Code + Claude Fable 5.1 (Max) | 57.6 | $11.53 | |
| Codex + GPT-6 Astra (Max) | 54.7 | $7.04 | |
| Codex + GPT-6.1 Sol (Max) | 54.1 | $1.50 | Strong mid-tier $/point; now pi's default OpenAI model |
| Opencode + GLM-5.3 (Max) | 48.9 | $5.23 | CN open-weight |
| Mistral Vibe + Mistral Large 4 (High) | 48.3 | $1.26 | Western open weights |
| Codex + GPT-6.1 Terra (Xhigh) | 48.2 | $0.69 | |
| Codex + DeepSeek V4.1 Flash (Max) | 47.3 | $0.39 | CN open-weight |
| DeepSeek Harness + V4.1 Flash 1017 (Max) | 46.3 | $0.30 | Cheapest row above 45; 96.8% cache hit, ~16.7M tokens/task |
| Codex + GPT-6.1 Terra (High) | 46.3 | $0.54 | |
| Kimi Code CLI + Kimi K3 (Max) | 41.4 | $5.17 | CN open-weight, expensive per point |
| Codex + GPT-6.1 Terra (Low) | 34.9 | $0.20 | |
| Codex + GPT-6 Luna (Max) | 31.1 | $0.16 | Floor |
Read it like this: the tiers haven't moved since 10-30.
- Frontier tier (≥59): Copilot Fusion costs about $2.40 less per task than Devin Fusion for 1.6 fewer points.
- Value tier (~46–48): DeepSeek Harness + V4.1 Flash 1017 at $0.30, Codex + DeepSeek V4.1 Flash at $0.39 and Codex + Terra High at $0.54 sit within about 2 points of Mistral Large 4 High at $1.26.
- Token efficiency: the DeepSeek Harness row reads ~16.7M tokens per task at a 96.8% cache-hit rate, which is why it is cheap despite the volume. Cache stability matters more than raw token count at these prices.
Session economics. Nothing new in the window. The most recent real-session data is still OpenRouter's Claude Code session-cost post (median 2–9 turn session: Opus 5.5 $0.305, Sonnet 5.5 $0.155, GLM-5.3 $0.044, DeepSeek V4.1 Flash $0.009). On AA's long Max-effort tasks the Claude rows cost far more per task, so pick the view that matches how you work.
Local economics (News, author-run). colibri v2.1.0 reports, on one 24 GB RTX PRO 4000 Blackwell SFF with 62 GB RAM: Qwen3.6-35B-A3B UD-Q4_K_M at 104.8 tok/s decode on CUDA and 13.9 on CPU; Qwen3.8-Flash-Next IQ2_XS (68 GB, experts spilling to RAM) at 51.4 tok/s decode on CUDA. These are the project's own measurements with no independent run yet.
Coding CLIs, IDEs, harnesses#
- News: pi v1.5.0 (11-02 05:05 CST).
- Models: GPT-6.1 Terra via OpenAI, Azure OpenAI Responses and Codex (500K context on Codex). The default OpenAI model is now
gpt-6.1-sol. The catalog moves to models.dev v2. - Breaking: tool arguments are validated with
@earendil-works/pi-schema. TypeBox-only features (Type.Refine(),Type.BigInt(), codecs) no longer apply to tool args; plain JSON Schema and MCP tools now get the same coercion as TypeBox tools. - MCP: OAuth sign-in uses a Client ID Metadata Document by default and falls back to dynamic registration, matching Codex, Claude Code and opencode.
- Context and cost:
Usage.cost.totalFromProviderseparates billed totals from catalog estimates; codemode tool descriptions are shorter; reordering active tools no longer triggers a redundant system-prompt update. - Trust fix: a saved "Do not trust" decision is now honored for folders that only contain
AGENTS.md.
- Models: GPT-6.1 Terra via OpenAI, Azure OpenAI Responses and Codex (500K context on Codex). The default OpenAI model is now
- Pre-releases only elsewhere: Codex rust-v0.167.0-alpha.3 (11-01) and the Gemini CLI v0.67.0 nightly (11-02). Latest stables are unchanged: Claude Code v2.1.316 (10-30), Codex v0.166.0 (10-29), opencode v2.2.5 (10-30). Aider has had no release since v0.86.0.
- Context: Claude Ultra plan: the renamed Max 20x plan drops the five-hour limit for verified accounts but keeps weekly limits (covered 10-31; The Agentic Digest re-flagged it today).
Usable tools and open projects#
| Repo or tool | What it is | Stars / +24h | Runnable today? | Why a tinkerer should care |
|---|---|---|---|---|
| JustVugg/colibri | Pure-C MoE engine that streams experts from disk | 46,218 / +655 | Runnable (8 GB RAM minimum, 22 GB disk) | Run 700B–1T MoE GGUFs on one consumer box; v2.1.0 adds Mistral-Large-4 |
| max-sixty/worktrunk | Git worktree CLI for parallel agent sessions | 9,855 / +115 | Runnable (v0.81.0, 10-20) | One worktree per agent without hand-managing branches |
| ivankovic/omnidiff | Content-aware Git difftool in Rust | 712 | Runnable | Faster review of agent diffs; headless output for CI (Show HN, 273 points). Speed claims are the project's own |
| vercel-labs/agent-desktop | Rust computer-use runtime via OS accessibility trees | 8,142 / +2,243 | Runnable (Still open, covered 10-31) | #1 on trending again; reliability over screenshots |
| paperclipai/paperclip | Open-source app for managing agents at work | 110,138 / +507 | Watch | Large and active (pushed today), MIT |
| mattpocock/skills | Skill files from an .agents directory |
+970 | Skip | Prompt pack with no runtime |
| Donchitos/Claude-Code-Game-Studios | 49-agent Claude Code game studio | +1,380 | Skip | Prompt and workflow pack |
Stars are from the GitHub API and GitHub trending at about 07:40 CST.
Keepers:
- colibri:
git clone, then./start-here.shpicks a model that fits your RAM and disk. v2.1.0 reads GGUFs as published (IQ1_S to Q8_0, BF16, MXFP4), adds a routed-expert tier on CUDA, acoli uichat UI, MTP speculative decoding for Qwen3.6, and float16 KV caches (GLM-5.3 at 131K tokens in 11.9 GB instead of 23.9). It also "parks" the conversation for short side requests, which cut a Hermes Agent turn from 86 s to 4.9 s in the project's test. - worktrunk: a plain CLI for creating and cleaning up worktrees, useful when you run several coding agents on one repo.
Agent builder stack#
- Update (evals and context): Apify removed TOON from its MCP server.
- Apify had added TOON-encoded list responses (PR #950, PR #952) and byte/token cost metrics to its workflow evals (PR #1031), then removed the format in July.
- Per the Forkast write-up (11-01): round one had 4.8% more tool-result bytes and a 64% pass rate with TOON vs 67% with JSON; round two had 1.4% fewer bytes and 92% vs 88%. The task mix changed between rounds, so neither is a controlled result.
- The bigger win in their fixtures was field selection: projecting three fields shrank one 1.01 MB payload to 12.3 KB.
- Builder take: before switching tool output formats, project fields first, then A/B the exact policy you'll ship with tasks and instructions held fixed.
- News (sandboxes and secrets): Trigger.dev incident. Environment variables of a subset of organizations were accessed. Updates go to Trigger.dev's status page. If agent jobs run there with provider keys, rotate them.
- News (protocols): pi v1.5.0's MCP OAuth change uses Client ID Metadata Documents first and dynamic client registration as the fallback, now the shared default across pi, Codex, Claude Code and opencode according to the release notes.
- Still open (reference architecture): DeepSeek dsh-superagent-prototype (MIT, covered 11-01). A router sends authorized GitHub requests to isolated worker sessions. DeepSeek calls it a research prototype and says its Harness sandboxing has not been audited.
- Context (paper): HALO (code) proposes human-grounded, granular evaluation of proactive agents. It's a low-vote paper on HF Daily Papers, but it ships code.
Infrastructure dash#
- colibri v2.1.0: the infra story of the day for individuals. Expert streaming plus a host-RAM expert pool on Vulkan means MoE models larger than VRAM run on one card, at the speeds in section 3.
- NVIDIA DeepSeek-V4.1-Flash-1017-NVFP4 (Still open): about 400 GiB, Blackwell only, needs a vLLM build with PR #61530 (covered 11-01).
- Engine blogs: nothing from vLLM, SGLang or TRT-LLM was nominated today.
Builder actions#
- Run evals on Ling 3.1 Flash on Vercel AI Gateway before it stops being free on Nov 3.
- Before running
pi updateto v1.5.0, grep your extensions forType.Refine,Type.BigIntor codecs in tool parameters; they no longer apply. Then trygpt-6.1-terrawith thinkingofffor cheap tool loops. - Try a large MoE locally:
git clone https://github.com/JustVugg/colibri && cd colibri && ./start-here.shand accept the recommended model (README). - If you use Trigger.dev, check the account owner's mailbox for their notice and rotate any API keys stored as environment variables there.
- Before changing an MCP server's output format, copy Apify's approach: add byte and token metrics to your evals (PR #1031) and test field projection first.
- If you depend on self-hosting Chinese frontier weights, read MOFCOM Announcement No. 48 (comments close 2026-11-15) and keep a fallback such as DeepSeek V4.1 Flash 1017 or Mistral Large 4 already on disk.
Checkmarks stay in your browser only.
Source log#
Show the full source logEvery source checked for this issue.
Primary discovery
| Source | Result | Note |
|---|---|---|
| Artificial Analysis + articles | Hit | Newest article 10-28; Coding Agent Index still the 10-30 materialization (31 rows in today's payload) |
| OpenRouter models API + rankings | Hit | No new model IDs since particle-1 (10-28) |
| HF trending models | Hit | DeepSeek V4.1 Flash 1017 still #1; nothing new in the window |
| HF Daily Papers | Partial | 29 papers listed for 11-02, mostly older submissions; HALO kept |
| HN front page + Algolia 24h + inference slice | Hit | Trigger.dev, MOFCOM note, OmniDiff; inference slice empty |
| GitHub trending | Hit | colibri, worktrunk new; agent-desktop, claude-mem, t3code carried over |
| Price Per Token news + agents/coding/open-source/model-releases slices | Hit | No model releases or price changes in the window; industry headlines only |
| LearnPrompt AI News Radar | Miss | SPA rendered empty; Pages fallback also empty |
| RadarAI | Hit | Commentary and how-tos only; no CN lab release |
| The Agentic Digest | Hit | 11-02 issue: Apify TOON, DeepSeek prototype (carryover), OmniDiff, Claude Ultra |
| Simon Willison 2026-11-01 | Hit | A Kyle Kingsbury quote and a uv.lock TIL; nothing printed |
| X (narrow) | Not checked |
Ranked dailies (in order)
| Rank | Daily | Result | Note |
|---|---|---|---|
| 1 | Latent Space AINews | Partial | Index only; newest issue 10-31 (Google AI Mode), nothing new |
| 2 | AlphaSignal | Partial | PRO-gated headlines (Qwen3.8 drafter on DGX Spark, Kanzo kanban, Tencent Hyra 2.0); no primaries found, dropped |
| 3 | Code4U AI Digest | Hit | 11-01 items were industry and policy; Claude Code game-recreation refusal had no Anthropic primary, dropped |
| 4 | TLDR AI archives | Hit | Newest issue 10-30; nothing new |
| 5 | TLDR flagship | Not needed | TLDR AI loaded |
| 6 | The Neuron | Hit | Sunday special on Apple; nothing new to print |
| 7 | AI Daily Brief | Hit | 11-01 long read; nothing printable |
Confirm opens
- pi v1.5.0: roster CLI release check → confirmed (release)
- colibri: GitHub trending → confirmed (v2.1.0, GitHub API)
- Trigger.dev incident: HN → confirmed (Trigger.dev blog)
- Apify TOON: Agentic Digest → confirmed (PR #1060); numbers from Forkast, labeled as such
- MOFCOM No. 48: HN → confirmed (MOFCOM); thresholds as summarized by CSA
- Kimi K3.1 Pro weights: Agentic Digest → not yet published (HF org)
- "Anthropic Mythos-class model leaked on HF": HN Ask post → nominated, no primary, dropped
- Qwen3.8 drafter, Kanzo, Tencent Hyra 2.0: AlphaSignal → nominated, no primary, dropped
- Coding CLI releases checked: claude-code, codex, gemini-cli, opencode, pi, aider (GitHub releases API)
- Engine blogs: not nominated