Radar brief — 2026-09-08
Silent tool denials and broken path-permission rules dominate this window, alongside a burst of GPT-5.6 / Gemini 3.8 provider wiring.
What the evidence shows
GitHub-tagged releases are the load-bearing signals. Theme density is model / CLI / agent, then MCP and security (permissions, sandbox, OAuth). Several notes are truncated (bodyOmitted); those tails are not inferred below. Trusted source mix is only 49% GitHub/changelog vs HTML — treat Kiro, Tabnine, and Factory as noisier than the versioned GitHub tags.
Versioned GitHub releases
on_failure, plus PreToolUseResult and a stable tool_call_id. Tagged breaking; 98 changes omitted./diff panel of uncommitted edits. /cost and the status line now explain likely prompt-cache misses. /reload-plugins and text /advisor work in headless/Desktop/SDK sessions. Permission/sandbox fixes: path rules with parentheses were dropped (leaving “read-only” folders writable); one bad regex deny rule aborted every edit; zsh assignments (REPORTTIME / REPORTMEMORY / DIRSTACKSIZE) no longer auto-approve hidden command substitutions. 66 changes omitted.GEMINI_API_KEY. Markdown custom agents inherit ambient skills, rules, and subagents by default. MCP OAuth fixed for authorization codes >1024 chars. Windows skill path grouping, relative subagent paths, duplicate permission grants, /boost worker config, a nil-pointer crash on closed trajectories, and hardened Remote Control reverse-tunnel routing.tui.auto_recap = false keeps manual /recap. TUI reconnects after app-server drops (drafts preserved, queued work paused). Plus/Team get an earlier half-allowance warning. Full Access skips Guardian on confirmation-only actions. 95 changes omitted./ ? search in drafts. MCP server names may include : @ / .. Per-tool output_token_limit. Credential-refresh progress (including Bedrock). Automatic approval reviews keep more transcript and survive compaction. MCP tools stay available across cache refreshes and remote plugin changes. 94 changes omitted.none; Bedrock replay/reasoning reliability improved. apply_patch no longer emits an empty move path in permission metadata. Azure CLI sign-in asks for the resource name directly.HTML changelogs (lower confidence)
tools.enableWebFetchTool is on (avoids LLM-mediated hallucination). Same page also rolls 6.5.3–6.5.1./archive (and Ctrl+X); restore from Archived. Default model is GPT-5.6 Sol when none is selected. CLI starts even if org settings are slow. Desktop v0.166.0 in the same note.One source in this window is effectively error-empty (Kiro 2.20.2). Seven open improvement proposals sit beside these releases — they are suggestions, not shipped facts.
For developers
- If you pin Claude Code path rules, upgrade to v2.1.260 and re-test folders whose paths contain
()or a character class. Confirm “read-only” roots are still read-only, and that a single malformed deny pattern no longer blocks every edit. Expect more zsh approval prompts — that is the sandbox fix, not a regression to paper over. - If you use Cline through Dify, SAP AI Core, opencode, or Codex CLI, treat cli-v3.0.61 + sdk v0.0.82 as a correctness patch, not a nice-to-have. Empty catalog capabilities were an authoritative “no tools / no images.” After upgrade, inspect a live request and confirm tool definitions and image reads are actually present.
- Goose v1.48.0: if you still name GLM-5.2, you are on a replaced Z.ai routing. Wire new declarative providers only if you intend to; turn on custom cost fields if you care about spend; add PreToolUse
on_failureand rely on stabletool_call_idin hook logs. Do not assume the omitted 98 items. - Codex rust-v0.152.0 then v0.153.0: set
tui.auto_recap = falseif recaps fight you; use package-style MCP names; set per-tooloutput_token_limit; try the plugin marketplace CLI. After a dropped app-server connection, review paused queued submissions before they resume. - Antigravity 1.1.25: Markdown custom agents now inherit ambient skills/rules/subagents — audit agent files so you do not pick up workspace skills you meant to keep off.
GEMINI_API_KEYusers can select Gemini 3.8 Flash. Re-auth MCP if authorization codes were truncated. - OpenCode v1.18.26: Claude 5 runs that died after a prompt or tool-definition change should continue; Bedrock GPT-5.6 can use effort
none. Recheckapply_patchpermission metadata if you key off move paths. - Factory Droid v0.209.0: new sessions with no model pin now land on GPT-5.6 Sol. Pin explicitly if that is not what you want. Use
/archiveinstead of letting old threads clutter the picker. - Zed v1.19.0-pre is a pre-release. The durable agent bits are long-session memory and OpenRouter reasoning effort — do not treat editor chrome as the AI story.
For teams
- Security review this week, not “later”: Claude Code could ignore path permission rules (parentheses) and auto-approve smuggled zsh substitutions; Antigravity was accumulating duplicate permission grants and hardened reverse-tunnel routing; Codex Full Access skips Guardian for confirmation-only actions. Map those three behaviors to your approval policy before rolling the versions.
- MCP is still a production dependency, not a demo: Antigravity OAuth failed on long codes; Cline would tear down the whole session when a remote MCP server was unreachable (now a 10s budget); Codex allows richer server names and keeps tools across cache/plugin refreshes. Inventory MCP servers, timeouts, and auth helpers as part of the upgrade.
- Cost and defaults will drift under people: Goose cost tracking now follows custom provider fields; Claude
/costexplains cache misses (tool defs / system prompt / TTL idle); Codex Plus/Team warn before the five-hour window is half gone; Factory’s default is GPT-5.6 Sol and v0.208.1’s 60% GPT/Grok cut is time-boxed. Pin models in shared configs rather than inheriting picker defaults. - Cline Hub upgrades are no longer silent. Operators get a prompt with session-interrupt count, drain-then-replace, and a refuse-to-kill path for a wedged Hub. Pair that with Authenticode-signed Windows binaries and application-control error text.
- Open team proposals to decide, not to assume shipped:
- Move Cline Agent Plugins into
~/.agents/pluginsand Hub-enable them (team, P5). - Pin Cline models after cli-v3.0.61 (everyone, P4) — the GitHub notes in this brief do not document a Claude Fable default; treat the proposal as a pin-discipline item until a release note says otherwise.
- Archive Factory Droid chats and default new sessions to GPT-5.6 Sol (team, P3) — this one does match v0.209.0.
- Move Cline Agent Plugins into
What to ignore
- Omitted changelog tails (Goose +98, Claude Code +66, Codex +95/+94, Zed +50, Cline +18/+34, Kiro +8, Tabnine +7). Absence of detail is not a feature list.
- Kiro 2.20.2 — empty patch page. Kiro v2.19.2 HTML that mixes 2.19.1 and older IDE notes.
- Zed v1.19.0-pre as if it were stable.
- Factory v0.208.1 pricing as a durable platform signal; it is a limited-time discount.
- Tabnine’s bundled 6.5.3–6.5.1 items as if they shipped today in 6.5.4.
- Marketing language and internal scores. This window tracks 14 signals / 23 assistants, not a ranking.
Watch next
-
Whether Goose’s remaining v1.48.0 tail clarifies the breaking / MCP / permissions tags, or just more provider glue.
-
Whether Claude Code’s omitted tail documents more hooks / compaction / context-file behavior beyond the permission fixes already listed.
-
Whether other gateways repeat Cline’s bug: empty capability list ⇒ strip every tool.
-
Codex plugin marketplaces vs Guardian-skip on Full Access — marketplace surface area vs review policy.
-
Open proposals still live: Grok CLI prompt/
<response_guidelines>and legacy MCP/skills cleanup (developer, P4); Windsurf Agent Command Center split window (developer, P3); OpenCode PR #38252 verification videos (everyone, P2). -
Antigravity custom-agent inheritance and Gemini 3.8 Flash uptake once more than a CLI catalog line exists.
How to use DevAgentRadar
DevAgentRadar is a public evidence window, not a vendor brief. Read GitHub tags first, HTML changelogs second, and open proposals as optional actions. Match assistant + version to what you actually run; if a change is not in the table, it was not in this window. Use For developers as an upgrade/test checklist and For teams as policy/cost/MCP review. Re-check after you bump versions — omitted tails and pre-releases are deliberately left blank rather than guessed.
Evidence appendix
- Evidence window: top curated signals (versioned releases preferred)
- Assistants tracked: 23
- Signals in this brief: 14
- Trusted source mix (GitHub/changelog vs HTML): 49%
- Open improvement proposals: 7
- Sources reporting errors: 1
High-signal releases used
-
Goose
v1.48.0: TrustedRouter declarative provider #11422 · OpenCode Zen gateway as a declarative provider [#10939](https://gi — source -
Claude Code
v2.1.260: Added a diff panel that opens beside the conversation in fullscreen mode and shows your uncommitted changes as Claude edits; toggle it with/diff· Added a li — source -
Zed AI
v1.19.0-pre: This week's release includes call hierarchy support, multi-select in the Git Panel, automatic language detection for untitled buffers, and project search on typ — source -
Google Antigravity
1.1.25: Added an opt-in workspace-grouped view to the/resumeconversation picker, allowing users to toggle between a flat list and conversations grouped by directory — source -
OpenAI Codex CLI
rust-v0.153.0: Vim mode now supports undo withuand redo withCtrl+R, preserving complete drafts including pasted content and attachments. ([#41941](https://github.com/op — source -
OpenAI Codex CLI
rust-v0.152.0: Vim mode supports/and?searches within drafts, highlighted matches, and repeat navigation withnandN. (#41586) · Rate-limit banners offer actions f — source -
Cline
cli-v3.0.61: Cline now handles a running Hub that is older than your CLI. · Instead of quietly talking to a hub executing stale code, you get a prompt showing how many activ — source -
Cline
sdk/sdk/v0.0.82— SDK v0.0.82: Fixed tool calling being silently disabled for gateway models whose catalog entry declares no capabilities. · Three separate producers built gateway model defin — source -
OpenCode
v1.18.26: Claude 5 sessions now tolerate stale thinking blocks instead of failing after prompt or tool changes. · Bedrock GPT-5.6 models now acceptnonereasoning effor — source -
Kiro
v2.19.2: 2.19.2 Aug 25, 2026 · 2.19.1 Aug 21, 2026 Spec Review Mouse Support and Automatic Stream Recovery This release brings mouse support to the spec review screen. · — source -
Tabnine
v6.5.4: v6.5.4 August 30, 2026 Bug Fixes Tabnine CLI: Provenance & Attribution : Fixed attribution feedback being dropped after tool inputs were edited via VS Code diff — source -
Factory Droid
v0.209.0— · Desktop v0.166.0 (September 1, 2026): Session archiving from chat, a new default model, faster startup, and steadier app views · Archive sessions from chat - Archive the current session with /ar — source -
Factory Droid
v0.208.1— · Desktop v0.165.1 (August 29, 2026): 60% off GPT and Grok models · GPT and Grok discount - GPT and Grok models now bill at 60% off for a limited time, with the discount shown in the model selec — source -
Kiro
2.20.2— Aug 31, 2026: 2.20.2 Aug 31, 2026 — source
Open proposals
- Move Agent Plugins into ~/.agents/plugins and Hub-enable them in Settings → Customize (audience: team, P5)
- Pin Cline models after the v3.0.61 switch to Claude Fable 5.1 (audience: everyone, P4)
- Replace <output_efficiency> with <response_guidelines> in Grok CLI prompts (audience: developer, P4)
- Drop legacy managed MCP configs and shadowed platform skills in Grok CLI (audience: developer, P4)
- Archive Factory Droid chats with /archive and default new sessions to GPT-5.6 Sol (audience: team, P3)
- Set devin.agentWindow.location to split Agent Command Center into its own window (audience: developer, P3)
- Improve with OpenCode pr-38252-videos: pr-38252-videos PR #38252 verification videos (audience: everyone, P2)
Anonymous · one vote per visitor · re-click to clear. Helps improve digests—not a product ranking.