Skip to content
Guidance/Daily Briefing
digesteveryonedigestevidenceradarmodelcli10 min read · Oct 22, 2026

Radar brief — 2026-10-22

Kiro 1.4.10 turns memory off by default (must-know for daily users) and CLI 2.33.0 blocks untrusted workspaces from loading hooks, agents, and MCP. Same window: Kimi Code 2.3.0 breaks bare plugin skill names; Claude Code 2.1.305–2.1.307, Copilot CLI 1.0.101–1.0.102, Zed 1.25.2,…

What the evidence shows

Versioned GitHub tags outrank HTML changelogs. Several notes omit bodies (bodyOmitted); those extra items are unknown, not implied features. Trusted source mix is 54% GitHub/changelog vs HTML; 1 source reported errors.

Kiro1.4.10
Key Signal & Impact
Changelog headline only: Memory Setting Defaults to Off, workflow plan feedback, resilient Powers. No further behavior in the public note.
Kiro2.33.0
Key Signal & Impact
Interactive V3 workspace trust before loading that folder’s hooks, agents, MCP, steering, workflows, or LSP config. Clones cannot reshape the agent until you choose Trust. V3 model fallback is back via /model.
Kimi Code@moonshot-ai/kimi-code@2.3.0
Key Signal & Impact
Model-switch confirm (new session / compact / continue / cancel; cloud-gated). Breaking: plugin skills are <pluginId>:<skillName> only; bare names no longer invoke. Remote-control stream limit 10→100 MiB. +19 changes omitted.
Claude Codev2.1.307
Key Signal & Impact
/claude-api api-version → 2026-10-21. New claude_code.safety_event OTel on permission/auto/safety/hook/sandbox stops. MCP auth, sleep-timeout, double HTTP-tool retry, autocompact, and /clear resurrection fixes. +43 omitted.
Claude Codev2.1.305
Key Signal & Impact
sandbox.allowNestedUserNamespaces (Chromium, rootless podman) — notes say this lets them get around the sandbox’s file protections. Fixes Vertex/Bedrock tool-search regression, MCP server overwrite from settings, compaction/subagent model naming. +41 omitted.
Claude Codev2.1.306
Key Signal & Impact
claude -p in cloud sessions no longer reports “Not logged in” without session env. Many-MCP-tool sessions no longer mislabel API errors as “prompt is too long” / force extra compaction.
GitHub Copilot CLIv1.0.101
Key Signal & Impact
Structured output for dynamic workflow subagents. assistedApprovalInstructions. Path approval for mkdir+cd. Headless MCP skip when tokens need sign-in. macOS sandbox localhost TCP when local access is on. Team marketplace plugins. +16 omitted.
GitHub Copilot CLIv1.0.102
Key Signal & Impact
-p/stdin honors managed Auto model. Subagents inherit custom/MCP tool approvals. Sandbox commands past 10s. NO_PROXY/ALL_PROXY. Plugin git-subdir sources.
GitHub Copilot CLIv1.0.103-0
Key Signal & Impact
Prerelease. ACP honors --sandbox on resume. Legacy --remote/--no-remote retired. MCP snapshots reused across git worktrees.
Zed AIv1.25.2
Key Signal & Impact
Smaller binaries. Inlay-hint edits via double-click / editor: accept inlay hint. ACP session notices + context compaction for external agents. Bedrock: GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5. +34 omitted.
Zed AIv1.26.1-pre
Key Signal & Impact
Prerelease. Optional Pyrefly Python LS, debugger hover trees, WASM DWARF, fuzzier agent model pickers. +26 omitted.
Qwen Codecua-driver-rs-v0.26.0
Key Signal & Impact
CUA driver binaries: macOS signed+notarized; Linux unsigned; Windows unsigned UIAccess — you must sign/trust before activation. Ships @qwen-code/cua-sdk and @qwen-code/node-repl-mcp. Optional relative coords and MCP_MODEL_PAYLOAD_FILTER. +75 omitted.
Google Antigravity1.4.4
Key Signal & Impact
Auto mode on Windows 11 24H2+ (personal Google accounts; Agent still default). MCP tools/call gets root conversation id for nested subagents. Stricter agy plugin validate. Compaction keeps tool declarations for prompt-cache reuse. Auto-mode reviews sandbox-escaping commands more tightly. +11 omitted.
Google Antigravity1.4.5
Key Signal & Impact
Thought-block UX, /diff search through folded lines, cron jitter to avoid schedule spikes, idle/stream CPU cuts. +6 omitted.
Factory Droidv0.246.0 (Desktop v0.203.0)
Key Signal & Impact
HTML changelog: image gallery, faster transcripts, usage in credits not tokens, FetchUrl can read any public URL (not only URLs you supplied).

Thin or empty tags (not in the table as product news): OpenAI Codex CLI rust-v0.164.0-alpha.11 (“Release 0.164.0-alpha.11” only), Qwen sdk-typescript-v0.1.22 (CLI bundle matrix 0.26.0…0.23.x), Augment Code v1.2.0-prerelease.202610211516 (install.sh / unsigned macOS SEA packaging).

For developers

  • Kiro 1.4.10: Expect memory off unless you turn it on. Notes do not document migration; check the IDE setting before assuming prior memory still runs.
  • Kiro CLI 2.33.0: First V3 session in a folder shows Trust this workspace? Choose Trust or exit. Home directory is never remembered. To reset: delete workspaceTrusted in the per-root file under ~/.kiro/workspace-roots/. Re-enable model fallback in /model if you used it.
  • Kimi Code 2.3.0: Rewrite every plugin skill call to <pluginId>:<skillName>. Bare names will not run. If cloud config enables it, model switches in a hot session will prompt; don’t assume silent continue.
  • Claude Code: Upgrade through 2.1.307 if you hit MCP “needs authentication” with a locked macOS keychain, sleep-during-connect timeouts, double HTTP MCP calls, [1m] alias + CLAUDE_CODE_DISABLE_1M_CONTEXT, or cloud /clear coming back. Do not enable sandbox.allowNestedUserNamespaces unless you need Chromium/podman and accept weaker file isolation. Use /claude-api api-version only when you intend API 2026-10-21.
  • Copilot CLI: On 1.0.102, -p follows managed Auto; subagents should stop double-prompting tool approval. On 1.0.101, set assistedApprovalInstructions if Assisted Permissions guesses wrong. Stay off 1.0.103-0 unless you want the remote-flag rename and ACP sandbox resume fixes.
  • Zed 1.25.2: Bedrock users can pick the three newly listed models. External ACP agents get compaction/session notices. Double-click inlay hints to apply edits.
  • Qwen CUA 0.26.0: Windows: sign and trust the UIAccess worker before turning it on. Leave CUA_DRIVER_RS_COORDINATE_SPACE and MCP_MODEL_PAYLOAD_FILTER off unless you need them (both default off).
  • Antigravity 1.4.4: Windows 11 24H2+ personal accounts can try Auto via --mode auto, /config, or Shift+Tab. Run agy plugin validate — it is stricter on plugin.json, MCP, hooks.json, and sidecars. 1.4.5 is mostly UI/CPU; upgrade if the thought block or idle CPU bothers you.
  • Factory Droid: Treat FetchUrl of any public URL as a new network surface in the agent, not just a gallery/UI release.

For teams

  • Kimi skills (open proposal, team/P5): Inventory plugins and rewrite invocations to qualified pluginId:skillName before rolling 2.3.0. Same-named skills from two plugins now stay visible only under qualified names.
  • Kiro workspace trust: A cloned repo can no longer quietly install MCP servers, hooks, or agents for V3. Document the Trust screen in onboarding; untrusted = session exits. Do not expect home-directory trust to persist.
  • Claude Code sandbox & MCP: Treat sandbox.allowNestedUserNamespaces as a policy exception, not a default. Pin MCP servers: 2.1.305 fixed a same-name settings server replacing a cloud/SDK server mid-session. Subscribe to claude_code.safety_event if you already collect OTel. HTTP MCP retries now only for read-only/idempotent tools — non-idempotent tools should not be written as if they will retry.
  • Copilot CLI org controls: 1.0.101 installs Team-assigned plugins from Team marketplaces; extension capability requests use explicit approval, not the Assisted judge. Headless/SDK without OAuth skips MCP when saved tokens need sign-in. 1.0.102 makes prompt-mode Auto honor managed model selection — verify org default vs local pin.
  • Qwen Windows CUA: Do not ship the unsigned UIAccess worker to managed endpoints without your own signing/trust step. macOS is the only pre-notarized binary in this tag.
  • Antigravity: Auto mode reviews sandbox-escaping commands, pipelines, and destructive git/file ops more aggressively — expect more prompts, not fewer. MCP servers that key off conversation id should read _meta["antigravity.google/root_conversation_id"] for nested subagents.
  • Factory Droid: If FetchUrl is allowed, update DLP/egress policy: the agent can fetch any public URL, not only user-supplied ones. Usage copy is now credits, not tokens — align cost comms.
  • Zed Bedrock: Adding GPT-6 Astra / Claude Fable 5.1 / Opus 5.5 is a provider catalog change; confirm quota and data-handling before enabling org-wide.

What to ignore

  • Codex rust-v0.164.0-alpha.11: no changelog facts. Do not infer features.
  • Qwen sdk-typescript-v0.1.22: a stacked “bundles CLI version …” list, not a user-facing feature release. Use cua-driver-rs-v0.26.0 for substance.
  • Augment v1.2.0-prerelease.*: installer/SEA packaging; macOS -sea is unsigned. Low-confidence as product news.
  • Prerelease channels (Zed v1.26.1-pre, Copilot v1.0.103-0, Codex alpha, Augment prerelease) unless you already run pre.
  • Any “+N more changes” line. Those bodies were omitted; inventing them is how this radar goes wrong.
  • HTML-only blurbs (Kiro 1.4.10, Factory) vs GitHub tags: thinner, and the mix includes 1 error-reporting source. Don’t build policy from a single “Learn more ->” line.
  • Internal scores, “impact,” and marketing adjectives. Judgment here is from the release text only.

Watch next

  • Full Kimi 2.3.0 remainder (19 omitted) and whether bare skill names get a compatibility window.
  • Qwen CUA’s 75 omitted items — especially MCP filtering and Windows signing in real deployments.
  • Codex 0.164 when notes exist; open proposal still points at 0.163.0 /config live settings (developer/P5).
  • Copilot 1.0.103 stable after --remote flag retirement and ACP sandbox-on-resume.
  • Claude Code callers moving to API 2026-10-21 and whether safety_event becomes the audit stream teams actually query.
  • Kiro follow-up that explains memory default off beyond the 1.4.10 headline.
  • Other open proposals (not today’s evidence): xAI Grok CLI harness/TUI open-source (P2); OpenCode PR #38252 verification videos (P2).

How to use DevAgentRadar

This brief is a public, evidence-only cut of what actually shipped, ranked for people who live in these assistants. Use it as a filter, not a vendor blog:

  1. Start with must-know and breaking lines (Kiro memory/trust, Kimi skill names, Claude nested-namespace sandbox, Copilot managed Auto). Then skim the table for your tools only.
  2. Prefer GitHub/changelog versions in the table over HTML blurbs and empty alphas. If bodyOmitted is set, stop at the listed bullets.
  3. Developers apply the concrete upgrades and flag flips above on a single machine first. Teams treat MCP, sandbox, plugin, and trust changes as policy, not release notes.
  4. Ignore prereleases, unsigned binaries, and bundle-matrix SDK tags unless you have a reason to be on that channel.
  5. Recheck this radar after the next 36-hour window — omitted bodies and empty alphas often land as real notes on the following tag.

DevAgentRadar tracks 29 assistants; this brief used 40 signals. If a tool you run is missing, it did not have citable evidence in this window.


Evidence appendix

  • New in the last 36 hours: 22 release(s) from 12 tool(s)
  • Evidence window: new releases first, then top curated signals (versioned releases preferred)
  • Assistants tracked: 29
  • Signals in this brief: 40
  • Trusted source mix (GitHub/changelog vs HTML): 54%
  • Open improvement proposals: 4
  • Sources reporting errors: 1

High-signal releases used

  • Kimi Code @moonshot-ai/kimi-code@2.3.0: #4294 3f80896 — source

  • Qwen Code cua-driver-rs-v0.26.0 — cua-driver-rs v0.26.0: Qwen CUA Driver prebuilt binaries · (vendored under packages/cua-driver). · macOS: codesigned + notarized universal binary + QwenCuaDriver.app · **Linux — source

  • Zed AI v1.25.2: This week's release includes smaller Zed binaries across macOS, Windows, and Linux, the ability to apply inlay hint text edits by double-clicking, context compa — source

  • Claude Code v2.1.307: Added /claude-api api-version to move Claude API code to API version 2026-10-21 · Added the claude_code.safety_event OpenTelemetry event for each tool cal — source

  • Qwen Code sdk-typescript-v0.1.22: This SDK release bundles CLI version: 0.26.0 · Source: CLI built from source (same branch/ref as SDK) · --- · This SDK release bundles CLI version: 0.25.3 · Sou — source

  • Google Antigravity 1.4.4: Added Auto mode on Windows 11 version 24H2 and later for personal Google Accounts, selectable via --mode auto, Agent Mode in /config, Shift+Tab, or Remo — source

  • Claude Code v2.1.305: Added sandbox.allowNestedUserNamespaces so sandboxed commands on Linux can create user namespaces (Chromium, rootless podman); this lets them get around the s — source

  • Zed AI v1.26.1-pre: This week's release includes Pyrefly as an optional built-in Python language server, the ability to expand and inspect nested variable values directly in debugg — source

  • GitHub Copilot CLI v1.0.101: 2026-10-20 · Add native structured output for dynamic workflow subagents on supported models · Add assistedApprovalInstructions setting to guide Assisted Permis — source

  • GitHub Copilot CLI v1.0.102: 2026-10-21 · Prompt mode now honors managed Auto model selection for -p and piped stdin runs · Subagents honor custom and MCP tool approvals, avoiding duplicate — source

  • GitHub Copilot CLI v1.0.103-0: Added · Clicking a completed response code block copies its original code when enabled. · Improved · Reuse cached local MCP tool snapshots across regist — source

  • Google Antigravity 1.4.5: Improved how agent responses read: a blank line now separates the collapsed thought block from the response, and once the response starts, the collapsed thought — source

  • Claude Code v2.1.306: Fixed claude -p inside cloud sessions reporting "Not logged in" when started without the session's environment variables · Fixed an API error being replaced b — source

  • Augment Code v1.2.0-prerelease.202610211516 — auggie-v2 v1.2.0-prerelease.202610211516: Install the latest release (macOS/Linux): · curl -fsSL https://github.com/augmentcode/auggie/releases/latest/download/install.sh | bash · Install a specific ver — source

Open proposals

Was this guidance useful?

Anonymous · one vote per visitor · re-click to clear. Helps improve digests—not a product ranking.