TaskOutput tool is gone; agents and scripts that read from it must switch to the Read tool today. The same v2.1.277 release quietly adds AGENTS.md fallback support, meaning Claude Code now picks up OpenAI Codex's standard instruction file when no CLAUDE.md is present — an interoperability step that reduces friction on multi-agent repos. Second, Google Antigravity refreshed its coding-agent preview with a hard breaking change: all built-in tool parameters are now PascalCase instead of snake_case, and file edits now use line-range replacements instead of full rewrites. The old preview retires October 5 — 17 days. Third, OpenAI Codex v0.155.0 shipped experimental voice conversations with live transcripts and Touch ID verification for MCP tool-use requests on supported Macs — the first Apple-native trust mechanism for an AI agent's tool approvals.TaskOutput tool removed — if your agent code, CLI scripts, or hooks read background task output via the TaskOutput tool, migrate to the Read tool now; taskOutputMaxChars / TASK_MAX_OUTPUT_LENGTH also no longer functionantigravity-preview-09-2026 — breaking parameter rename — PascalCase built-in tool parameters replace snake_case; line-range file edits replace full rewrites; old preview shuts down October 5/plugin marketplace add mksglu/context-mode
npm install -g context-mode
ctx_stats # shows context bytes saved this session and per-tool breakdown
TaskOutput tool is fully removed in v2.1.277. Claude previously used TaskOutput to read the output of background tasks (processes launched with Task tool calls). That tool is now gone — Claude reads background task output using the Read tool instead. The taskOutputMaxChars setting in .claude/settings.json and the TASK_MAX_OUTPUT_LENGTH environment variable no longer have any effect and can be removed.TaskOutput tool removed in Claude Code v2.1.277 — agent code, scripts, and hooks that reference or depend on TaskOutput must switch to using the Read tool; taskOutputMaxChars / TASK_MAX_OUTPUT_LENGTH are now inert.TaskOutput references. If you call TaskOutput in your own code or check for it in tool-use response parsing, those calls will now return errors. Switch to using the Read tool to read the output file path returned by the Task tool. (2) Remove taskOutputMaxChars from your .claude/settings.json. The setting is silently ignored; leaving it in place won't break anything but creates confusion. Remove TASK_MAX_OUTPUT_LENGTH from any CI/environment configs for the same reason. (3) If you built custom tooling around the old TaskOutput behavior (chunked reading, output pagination), redesign around Read — which can read specific line ranges via offset/limit parameters. (4) Also in v2.1.277: AGENTS.md fallback support — Claude Code now reads AGENTS.md in projects that have no CLAUDE.md, configurable under /config → "Project instructions." This is not currently available on Bedrock, Vertex, or Foundry. (5) Gateway proxy improvements: CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 is a new env var for gateways whose only egress is a forward proxy (sends the hostname to the proxy rather than resolving locally); upstream configs now accept an optional headers: map for static headers to proxies in front of a provider.Task tool in agent workflows and read its output with TaskOutput; you set taskOutputMaxChars or TASK_MAX_OUTPUT_LENGTH in any environment or config; you run Claude Code behind a forward proxy with Claude apps gateway.EffortModerate — remove taskOutputMaxChars/TASK_MAX_OUTPUT_LENGTH (no-ops but confusing to leave); replace TaskOutput calls with Read (code change required for any agent or script that uses the background-task output pattern).antigravity-preview-09-2026 replaces antigravity-preview-05-2026 with breaking changes to the built-in tool surface. All built-in tool parameters have been renamed from snake_case to PascalCase. File-edit operations now use line-range replacements (specify a line range to replace) instead of full-file rewrites. Two new tools added: find_by_name() and grep_search(). The old preview — antigravity-preview-05-2026 — shuts down October 5, 2026. Developers have 17 days to migrate.snake_case built-in tool parameters to PascalCase, switches from full-file to line-range edit operations, and adds find_by_name() / grep_search() tools; antigravity-preview-05-2026 retires October 5.environment: "remote" and only consume output_text or model_output steps, you just need to update the agent string to antigravity-preview-09-2026 — nothing else changes for you. If you run tools locally (local_environment) or parse function_call steps from the response, your integration is breaking — the tool parameter names have changed. (2) Rename all snake_case parameters in your function_call parsing to PascalCase. Example: file_path → FilePath, start_line → StartLine — check the updated Antigravity agent docs for the full parameter name table. (3) Update any file-edit tooling. The previous behavior rewrote the full file content; the new behavior specifies a line range to replace — your parsing of these steps must handle the new format. (4) Add find_by_name() / grep_search() to any tool catalog you're manually maintaining for this agent. (5) Hard deadline: October 5, 2026. After that, calls to antigravity-preview-05-2026 will fail. Add this to your sprint backlog now.local_environment or parse function_call response steps; you maintain any wrapper or integration layer around Antigravity's built-in tools.EffortModerate — update the agent string (quick); rename snake_case → PascalCase in all function_call parsing (moderate code change); update file-edit handling from full-rewrite to line-range model (moderate code change).No new model releases within the 24-hour scan window.
Carry-forward: Claude Fable 5.1 and Mythos 5.1 (September 1), GPT-6 Astra (September 3), Gemini 3.8 Flash (September 2) — see those dates' digests for full entries.
GET /v1/compliance/apps/sessions/local) now return transcripts from Claude in Chrome sessions, not just terminal and IDE sessions. Previously, Chrome sessions were excluded from Compliance API coverage. The new product_surface value to filter for these is "claude_in_chrome". This is in beta for Claude Enterprise organizations only.claude_in_chrome coverage for local session transcripts (September 18 beta for Enterprise), enabling audit workflows to capture AI-assisted browser work alongside terminal and IDE sessions.GET /v1/compliance/apps/sessions/local filtering to include product_surface: "claude_in_chrome" if you want to capture Chrome-session transcripts alongside terminal/IDE ones. Existing requests without a product_surface filter will now include Chrome sessions in the result — check whether your pipeline handles this without data surprises. (2) Authentication: Uses the existing Compliance Access Key with the read:compliance_user_data scope — no new credentials needed. (3) Scope: This is a beta feature for Enterprise organizations; it is not available on Team, Pro, or Max plans. (4) Reference the updated documentation for the session structure of claude_in_chrome entries — the schema may differ slightly from terminal-session transcripts in how conversation context is represented.product_surface value.EffortQuick — add product_surface: "claude_in_chrome" filter if desired; verify existing pipelines handle the new session type without breakage.DynamicToolset resolution previously occurred once per agent unit (each individual LLM call); it now resolves once per durable run (the full agent execution lifecycle). MCP server sessions were also previously maintained per unit; they now persist for the full durable run — meaning an MCP session is opened once at the start of a run and reused across all units within it, rather than torn down and re-established on every LLM call. Additionally: GPT-5.6-sol, GPT-5.6-luna, and GPT-5.6-terra are now supported on Bedrock Converse; Bedrock now correctly passes xhigh effort levels when supported; tool history is preserved in MCPSamplingModel; a new TypeSafeModel is added for TypeSafe's Jev integration.DynamicToolset resolution and MCP session lifecycle from per-unit to per-durable-run, and adds GPT-5.6-sol/luna/terra on Bedrock Converse.DynamicToolset, behavior has changed: resolution runs once per run, not per LLM call. This means dynamic toolset logic that was implicitly "re-evaluated on every call" now runs once and caches the result for the run. If your toolset logic had side effects or was intentionally dynamic per-call, review the new behavior. For most users, this is a welcome optimization (fewer redundant resolution calls). (2) If you use MCP servers in pydantic-ai, your MCP sessions now persist for the full agent run rather than reconnecting on every unit. This is a significant latency improvement for MCP-heavy agents (removes repeated TCP handshakes / auth exchanges per LLM call) but means any MCP server state from earlier in the run is visible to later tools — verify this is acceptable for your use case. (3) Bedrock Converse: Add gpt-5.6-sol, gpt-5.6-luna, or gpt-5.6-terra as the model string to use these variants via the Bedrock Converse interface. xhigh effort level now correctly routes on Bedrock where supported. (4) Legacy httpx imports from pydantic_ai.mcp are removed — if you imported from that path directly (unlikely but possible), update to the new import paths.DynamicToolset with side-effecting or truly per-call-dynamic resolution logic; you use MCP servers in pydantic-ai agents (behavior change in session lifecycle); you use Bedrock Converse and want GPT-5.6 variants or correct xhigh effort routing.EffortQuick — update version; review DynamicToolset and MCP session behavior if your toolsets have per-call side effects.Nothing cleared the quality gate this period. arXiv cs.AI/cs.CL/cs.LG submissions from September 18 included agent-evaluation papers (AgentLSD 2609.19140, ScienceIDE 2609.19134, ToolLoop 2609.09072 / EMNLP 2026) but none combined: recognized-lab authorship OR associated code repo AND concrete benchmark numbers with a verifiable primary source.
AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY. Configuration is via the credential command interface — see the release notes for the config schema. (4) WSL sandbox escape fix: If you use Codex in WSL, update immediately — a process escape from the WSL sandbox was blocked. (5) Reasoning summaries in TUI: The status row now shows a summary of Codex's internal reasoning during generation — useful for tracking what the agent is "thinking" without waiting for it to write output. Note: v0.155.1 (Sep 18 hotfix) disables reasoning summaries by default for new local TUI sessions to fix rejection errors from providers that don't support them. Update to v0.155.1 or set reasoning summaries off if you see provider rejection errors.npm install -g @openai/codex@latest or equivalent; Touch ID and Bedrock credential management require additional one-time configuration per the release notes."syncClaudeAiSkills": false / "syncClaudeAiPlugins": false in settings; (3) __SYSTEM_PROMPT_DYNAMIC_BOUNDARY__ marker support in --system-prompt flag — text above the boundary is globally cached (matching SDK array caching behavior); (4) plugins from npm are now fetched with npm pack --ignore-scripts, preventing install scripts from running automatically. Prompt cache bug fixed: memory file age notes were causing cache misses.--system-prompt flag, and hardens npm plugin installs against script execution.__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__: If you use the --system-prompt flag in automated pipelines or scripts, place the stable/reusable content above this marker and your dynamic (per-run) content below it. The content above will be globally prompt-cached, reducing cost and latency on repeated invocations. This brings the --system-prompt flag's caching behavior in line with the SDK's array-based system prompt caching. (2) Skills/plugins sync: Terminal sessions signed in to claude.ai will now automatically see the same skills and plugins you've enabled on the web. If you have deliberate differences between your web and terminal setups (e.g., terminal should not run your project-specific skills), opt out in settings: "syncClaudeAiSkills": false. (3) Ctrl+Enter (send-now): In interactive sessions, use this to interrupt a running turn mid-stream and dispatch all queued messages immediately. Queued and sent messages appear in gray until the model receives them — useful when you want to redirect a long-running turn without waiting for it to complete. (4) npm plugin security: Plugins installed from npm no longer run install scripts. If you maintain a plugin that relies on postinstall scripts to set up native binaries or config, users on v2.1.275+ will need an alternative setup step — update your plugin's README.--system-prompt in automated pipelines and want prompt caching; you manage different skill/plugin setups between claude.ai web and terminal; you install Claude Code plugins from npm (if you're a plugin author); you want keyboard-driven interruption of long-running turns.EffortQuick — npm install -g @anthropic-ai/claude-code@latest; use __SYSTEM_PROMPT_DYNAMIC_BOUNDARY__ if you use --system-prompt and want to optimize caching.rg.exe, git.exe, or powershell.exe — a path-hijacking attack vector where a malicious repo could shadow legitimate system binaries. Context window compaction now uses actual token counts instead of character estimates, improving accuracy. Model catalog expanded from 5,788 to 6,079 models with 44 new default model assignments. Images attached to unsupported models now display a warning badge instead of silently disappearing. Separately, Cline Desktop v0.0.31 (September 17) adds SSH remote machine support — the app stays local while working on a remote machine — and parallel sub-agent execution for improved performance on delegated tasks.rg.exe, git.exe, or powershell.exe in the workspace. Cline would previously execute the workspace-resident copy when invoking these tools; now it blocks them and uses system paths only. If you work on untrusted repos on Windows, this was an active risk — update before cloning anything new. (2) Context compaction accuracy: The switch from character-count to actual token counts means compaction fires at the correct threshold now. If you previously calibrated workflows around when compaction kicked in, the trigger point may have shifted slightly (especially for codebases with lots of Unicode, code, or punctuation where char/token ratios vary). (3) SSH remote machines (Desktop v0.0.31): For developers running Cline Desktop on macOS/Windows while targeting a remote Linux machine (staging server, cloud dev box, WSL on another machine), SSH remote support means the coding agent now operates on the remote filesystem and shell without needing a separate local clone. Configure the SSH target in Desktop settings. (4) Parallel sub-agent execution (Desktop v0.0.31): Delegated tasks in Desktop now run sub-agents in parallel by default — useful for large refactors or test generation where multiple independent tasks can run concurrently. Monitor token usage carefully in early use./plugin marketplace add mksglu/context-mode in Claude Code (or npm install -g context-mode for other platforms). The core pattern: replace raw tool calls with the ctx_execute / ctx_batch_execute wrappers that run the same operations in sandboxed subprocesses and return only stdout. Run ctx_stats to see the savings. For session continuity across compaction, the ctx_index / ctx_search tools store and retrieve prior decisions via BM25 FTS5 — this is the piece that distinguishes context-mode from a simple output truncator.No new leaderboard entries or score changes confirmed within the 24-hour scan window.
Carry-forward (September 17 snapshot): Claude Opus 5 leads SWE-bench Verified at 96.0% (78 models). Claude Fable 5.1 leads SWE-bench Pro at 81.2% (58 models). Claude Fable 5.1-max leads LMArena at 1,513.70 Elo. Note: Muse Spark 1.1 (Meta) reportedly tops Scale's standardized SWE-bench Pro public set at 61.5% (commercial: 51.5%), pushing GPT-5.4 (xHigh) and Claude Opus 4.6 to second — but the most recent confirmed date for that result is September 14, 4 days outside this window. Check the Sep 14 digest or confirm directly at labs.scale.com/leaderboard/swe_bench_pro_public before acting on it.
AGENTS.md as a fallback when no CLAUDE.md is present in a project. This makes AGENTS.md — originally the instruction file format popularized by OpenAI's Codex — a de facto shared standard across at least Claude Code and Codex. A developer who maintains one AGENTS.md in their repo now gets correct project-context behavior from both Codex and Claude Code without maintaining separate instruction files. OpenHands and other open-weight agents are converging on similar "project instruction file" patterns, though under varied filenames.AGENTS.md becomes the canonical cross-agent instruction file (the way .editorconfig is for editor configuration), project maintainers will start adding it to repos the way they add .gitignore — and the content of that file becomes a first-class engineering artifact that shapes how all AI agents interact with the codebase. The next question is whether the spec for what AGENTS.md should contain gets standardized, or whether each agent interprets it in its own way (divergence risk). Agents that support only their own proprietary file format — CLAUDE.md, CURSOR_RULES, .cline/rules — will face adoption pressure to converge or add aliases.Nothing cleared the quality bar this period. No Hacker News threads with score >200 and concrete technical depth on new items within the window were verified. Simon Willison (simonwillison.net), Nathan Lambert (interconnects.ai), Eugene Yan (eugeneyan.com): no new posts confirmed in the 24-hour window.
400 … Input tag 'advisor_20260301' when ANTHROPIC_BASE_URL points at a proxy or gateway (regression from v2.1.275). Update immediately if you saw this after updating yesterday. [github.com/anthropics/claude-code/releases/tag/v2.1.276]The old Antigravity preview retires in 17 days. If you use Antigravity, update the agent string to antigravity-preview-09-2026 and migrate tool parameter names from snake_case to PascalCase and file edits to the line-range model. See Breaking Changes above for the full migration guide. [ai.google.dev/gemini-api/docs/changelog]
GPT-6 Astra launched September 3, 2026 with free billing through October 4. Starting October 5: $10/$1 (cached)/$50 per MTok standard short-context. Long context (>272K tokens): $20/$75 per request. Batch/Flex: half price. Fast mode: double. Regional data residency: +10%. If you are running GPT-6 Astra in any pipeline, confirm your billing account and cost model before October 5. [platform.openai.com/docs/changelog]
DeepSeek confirmed V4.1 Pro is in development as the V4 Pro replacement. No launch date announced. V4 Pro continues at existing pricing until V4.1 Pro ships. [api-docs.deepseek.com/updates/]
Anthropic's Life Sciences Verification Program is initially open to teams and institutions; expansion to individual Claude Pro and Max subscribers is planned but no date given. [anthropic.com/news/life-sciences-verification-program]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.