npm install -g @openrig/cli
rig setup --dry-run
rig start rigspec.yaml
antigravity-preview-05-2026 is now shut down. The September preview (antigravity-preview-09-2026) became the replacement when it shipped September 17, and the 18-day migration window closed today. Two completely different migration paths exist depending on how your integration works.antigravity-preview-05-2026 with local tool execution is getting "version no longer supported" errors right now; remote-output-only callers may get a graceful redirect but should still migrate to avoid silent behavior changes.antigravity-preview-09-2026. That's it — the server handles the rest, and you may already be getting a silent redirect. Still migrate explicitly to avoid relying on redirect behavior. Path 2 (local_environment or parsing function_call yourself): this is a breaking migration. The built-in tool schemas changed from snake_case to PascalCase arguments across the board, and the file-editing tool changed its fundamental behavior: write_file (full-text replacement) is now write_to_file with StartLine/EndLine/NewContent patch blocks; read_file → view_file; list_files → list_dir. Any code that deserializes tool arguments, routes on tool name, or processes file edits must be updated before it can call the September preview. An integration test against the new model ID in a staging environment is required — do not just swap the string and ship.antigravity-preview-05-2026 anywhere — in a production agent, a CI pipeline, a scheduled workflow, or a dev script; more specifically if your integration runs in local_environment or parses function_call outputs.EffortQuick (remote-output-only: one string change) / Significant (local tool execution: schema migration, rename all tool call handlers, update file-edit logic to patch format).No new model releases in the 24-hour scan window. Claude Sonnet 5.5 (Sept 28) and Claude Opus 5.5 (Sept 22) were covered in earlier digests; no new Anthropic announcements on anthropic.com/news since October 2.
gpt-6-astra-ultrafast vs gpt-6-astra are separate endpoints and the cost difference is 6×; (2) set spend limits on your API key if you have automation that may call Ultrafast at scale; (3) if you're on the $500/month Codex plan and using Ultrafast through Codex, verify that plan inclusion covers the usage you saw during the free period. Standard Astra ($10/M input, $50/M output) remains unchanged. Batch and Flex modes cut both tiers' rates by 50% if you have latency headroom.gpt-6-astra-ultrafast that will now accumulate costs.EffortQuick (just check and set spend limits if needed).arXiv.org egress-blocked for the third consecutive day (cs.AI, cs.CL, cs.LG all unreachable). HuggingFace Daily Papers also egress-blocked. KaliBench (arXiv:2610.02206) — cybersecurity tool-use evaluation for coding agents on Kali Linux — is a near-miss for the third day running; the CITA paper (arXiv:2610.02330, NeurIPS 2026 Poster) on comparative value estimation for long-horizon tool-use agents was submitted October 1, outside the 24-hour window. See Near-misses. If arXiv remains blocked through October 6, a manual Deep Research run (Mode 2) against the accumulated backlog is worth triggering.
remote_env configured, Codex's allowlist was stripping SYSTEMROOT, TEMP, and TMP — Windows runtime variables required for Node.js and many Windows-native tools to initialize. The patch preserves these three variables when explicit remote environment variables are set. An integration test verifies the variables survive while unrequested secrets do not.SYSTEMROOT, TEMP, and TMP were stripped by Unix-default env allowlisting, breaking bundled Node and any MCP server that reads temp paths.remote_env keys — only those activate the allowlist that was filtering Windows system vars. If your Windows MCP servers were initializing but occasionally failing on file writes or temp-path operations, this is the root cause. No config changes needed — just update the package; the patch adds the three vars to the preserved set automatically.@e2b/desktop-python now requires urllib3 ≥ 2.8.0, patching CVE-2026-97687 (CVSS High) and CVE-2026-97689 (CVSS High). Both CVEs are in urllib3 versions 1.26.0–2.7.x, where HTTPS proxy TLS configuration (proxy_ssl_context, proxy_assert_hostname, ssl_context, cert_reqs, verify_mode) is incorrectly applied to the target-server TLS handshake instead of the proxy connection, allowing a network-adjacent attacker to intercept data. Additional fixes in the broader release: JS SDK .dockerignore filtering aligned to BuildKit behavior (handles ** patterns and emoji); 502 POST retry logic tightened to avoid retrying resource-creating operations that may have already executed.@e2b/desktop-python in any non-isolated environment.pip install 'urllib3>=2.8.0' or update your lockfile to pull in the fixed version, then re-pin e2b-desktop-python==2.6.1. The CVE risk is most relevant if your E2B desktop agents make HTTPS requests through an explicit proxy (HTTPS_PROXY env var set, or proxy_ssl_context configured) — in that scenario, target-server certs may have been validated against proxy TLS settings rather than the target's own cert. The 502 retry fix is less urgent but prevents double-execution of resource-creating POST calls (e.g. sandbox creation) when a transient 502 occurs mid-request.@e2b/desktop-python for AI agent computer-use or desktop automation; you run desktop sandboxes through an HTTPS proxy.EffortQuick (dependency bump to urllib3 ≥ 2.8.0 + e2b-desktop-python==2.6.1).update_state path (not just the primary); update_state on an older checkpoint no longer bleeds into parallel branches; abandoned branches no longer replay into DeltaChannel forks; subgraph delta channels correctly hydrate from the caller-resolved saver; answered interrupts no longer appear in get_state results. Dependency bumps for notebook, tornado, PyJWT, and urllib3 included.update_state, branching checkpoints, or subgraph delta channels.update_state on non-latest checkpoints, or multi-agent subgraphs that communicate via DeltaChannels. The symptoms of the pre-fix bugs were subtle: incorrect channel counter drift in long-running graphs, or stale state appearing after a branch merge. These were hard to reproduce reliably, which is why they accumulated. If you have existing checkpoint stores, the fix is backward-compatible — existing saved checkpoints load correctly after updating. The urllib3 bump also picks up the same CVE-2026-97687 fix as E2B.update_state on older checkpoints, or multi-agent subgraphs with DeltaChannels; the urllib3 bump also applies if you run LangGraph behind an HTTPS proxy.EffortQuick (version bump, no API changes).npm install -g @openrig/cli (requires Node.js 22+, tmux, macOS/Linux). Define your team in a YAML rigspec.yaml specifying seats (provider, model, role, effort), pods, and edges, then rig start rigspec.yaml. Use rig send <seat> "<message>" to direct a seat; rig broadcast "<message>" to all seats; rig snapshot to checkpoint the fleet state. The TUI shows the topology graph with seat health in real time. Best first workflow: a three-seat rig with an orchestrator Claude seat, a coder Claude seat, and a reviewer Codex seat — rig spec validate before launch to catch schema errors without needing the daemon running./clear commands.No significant leaderboard movements in the scan window. Most recent LMArena changelog additions were October 1–2 (claude-sonnet-5.5-xhigh, GPT 6.1 Sol Max, Fable 5.1 Max to various arenas), outside the 24-hour window. lmarena.ai egress-blocked; standings remain as of last confirmed check: Claude Opus 5-Max leads text arena at ~1505 Elo, Claude Opus 5.5 leads SWE-bench Pro at 89.9%.
--effort / model_reasoning_effort flags, creating a resource-allocation problem across a fleet; (2) Claude Code's mod API (v2.1.289 agent.spawn) makes agent-spawning a first-class operation; and (3) real deployments expose the state-management problem — agents that share a repo but not a context produce contradictory edits. OpenRig is the first tool to address all three with a single abstraction (RigSpec topology + persistent seats + Slack/tmux communication layer).Nothing cleared the quality bar this period. arXiv, HuggingFace, and Simon Willison's blog (no new posts in window) all unavailable or empty. Community search (HN, Lobsters) returned no threads scoring ≥3 in the 24-hour window.
GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash, and Grok 4.5 removed from all Copilot experiences. 14 days remaining. Audit any Copilot API integrations that hardcode model IDs now. [github.blog/changelog/2026-09-18-upcoming-deprecation-of-selected-github-copilot-models-in-mid-october/]
No model card, pricing, or release date as of October 5. [Anthropic Opus 5.5 announcement, September 22, 2026]
Astra Ultrafast is live. Sol Ultrafast (the cheaper-but-fast tier) remains unshipped. Watch platform.openai.com/docs/changelog.
Formal marketing potentially beginning the week of November 9; institutional investor meeting at SF HQ on October 14. No developer-facing changes expected, but worth awareness for teams evaluating long-term API commitments. [Bloomberg, October 2, 2026]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.