filename and title fields in the activity feed — affecting enterprise compliance integrations that were parsing document names from audit logs.filename and title fields on file/document/artifact activities are now always empty retroactively; affects any enterprise code parsing document names from the Claude compliance audit logcurl -L https://coder.com/install.sh | sh
coder server --in-memory & # starts an in-memory dev server (no DB required for local testing)
coder login http://localhost:3000
coder create my-sandbox # provision a workspace from the default template
coder ssh my-sandbox # drop into it — your agent runs here, not on your machine
filename and title fields on file, project document, and artifact activity objects in the Compliance API activity feed are now always empty or omitted. Previously, these fields returned the actual filename or document/artifact name. The change applies retroactively — activities recorded before September 24 now also return empty fields when fetched via the API.filename or title from the Compliance Activity Feed now receives empty strings for those fields; retrieve names via the new read:compliance_user_data scope instead.activity.filename or activity.title to identify which documents or artifacts were accessed, those fields now return empty strings or are omitted entirely. The pipeline won't crash, but it will lose document-name context silently — which is the harder failure mode to catch. (2) Migration path. To recover filenames and titles, use a Compliance Access Key with the read:compliance_user_data scope. Make a separate lookup by the activity's file/document/artifact ID to retrieve the associated name. This is a more privileged scope than the read-only activity feed, so expect additional approval steps in your enterprise access control process. (3) Retroactive scope matters. If your compliance tooling has cached or stored activity feed responses and you replay or re-query them, you'll get empty name fields — even for activities from months ago. Plan accordingly before any historical audit or re-indexing pass. (4) Why this change: The release notes describe this as a privacy update — project document names and artifact titles were appearing in compliance audit logs accessible to compliance admins who don't have general data access. The new scope-based lookup gives teams explicit, auditable access to that level of detail.filename, title, or similar name fields from file, project document, or artifact activities.EffortModerate — the fix is straightforward (add a lookup call with the new scope), but getting read:compliance_user_data approved in your enterprise access control pipeline may take time; retroactive scope means any historical data re-fetch is also affected.https://api.meta.ai/v1 with an OpenAI-compatible endpoint, so existing code using the OpenAI client library can point at it with a credential swap. Muse Spark 1.3 (standard tier): $1.25/MTok input, $4.25/MTok output, $0.15/MTok cached read, 1M context. The contributor tier ($0.10/$0.20 per MTok input/output) lets Meta use your prompts and completions for training — read your data policy before using contributor in production. (2) Muse Code is out of beta. Previously invite-only, Muse Code is now available to all API users. It's positioned for long-horizon autonomous engineering tasks — agentic code generation, test generation, and codebase-scale refactoring. Anthropic Compliance API users: if you're running Claude and Muse Code side-by-side, the activity feed architecture is now relevant for both providers. (3) Muse Glimmer 30B is the open-weights option: Apache 2.0, optimized for single-GPU or Mac Mini deployment via vLLM, SGLang, llama.cpp, or ExecuTorch. At 30B parameters, it's a viable local fallback for MCP tool-routing tasks where latency matters more than peak reasoning. (4) Documentation at dev.meta.ai/docs — the API surface is similar to OpenAI's; tool use, streaming, vision, and audio are supported. Enterprise tier (compliance, SLA, private deployment) available via Meta's first-party enterprise program.product_surface values beginning with office_agents — covering Claude sessions in Excel, PowerPoint, Word, and Outlook.product_surface: "office_agents_*" filter.filename or title fields in the activity feed, you now have a stable endpoint but empty name fields. Migrate the name lookup to the read:compliance_user_data scope in the same sprint. (4) See docs at https://platform.claude.com/docs/en/manage-claude/compliance-sessions#retrieve-local-sessions for the GET /v1/compliance/sessions?product_surface=office_agents call signature.f shortcut, adds automatic background-server startup, and extends /import to remote and background-server sessions.sol and luna are now first-class model aliases in Codex. For teams running Codex on Amazon Bedrock: Sol and Luna no longer require cross-region inference configuration — they're available in standard Bedrock regions. Update your ~/.codex/config.toml model field to gpt-6-sol or gpt-6-luna to opt in. Migration prompts guide existing users with older model names to the new aliases on first run after upgrade. (2) Conversation forking (f shortcut). Press f on any open session to fork it — the fork inherits the full conversation history and any queued drafts, so you can explore a different approach without losing your current thread. This is the same mechanic as Claude Code's --fork flag but as a keyboard shortcut. Useful for A/B testing a refactor approach mid-session without committing to one direction. (3) Fullscreen transcripts are now default. Previously opt-in, the fullscreen transcript view is now the default for new sessions. Toggle with t. (4) /import in remote sessions. File-import was previously blocked in remote and background-server sessions. It's now available in both, enabling structured data injection (e.g., a JSON spec or test fixture) without a full local session context. (5) Automatic background-server startup for eligible interactive sessions means you no longer need to manually start codex server before running /import or /fork-based workflows that rely on the background server./skill command in a freshly opened workspace. (4) Window scaling fixed for small and high-DPI displays where content previously extended beyond screen bounds.1/5 * * * *).1/5 * * * *, 0/15 * * * *), these are now validated correctly. (3) Workspace folder toggle from the Conversations header streamlines multi-workspace sessions — bulk expand/collapse of all workspace folders in one click.See Repo of the Day for the full entry.
No confirmed new leaderboard movements for September 25. lmarena.ai egress-blocked; no ELO update confirmed. SWE-bench Pro September snapshot was in yesterday's digest (Claude Opus 5.5 at 89.9% on BenchLM; Muse Spark 1.1 at 61.5%/51.5% on Scale's public/commercial split).
Nothing cleared the quality bar this period. simonwillison.net egress-blocked (Sep 25 posts were a bird sighting and a Gruber quote about Meta Muse — neither developer-actionable). Nathan Lambert, Eugene Yan, and Sebastian Raschka: no Sep 25 posts confirmed. No HN threads with score >200 and new Sep 25 technical depth confirmed.
DecisionModel base class for Decisions-protocol routing models, RunContext.in_durable_context flag for detecting durable workflow execution in hooks, gemini-3.8-live and gemini-3.8-live-extended-thinking realtime support; fixes Anthropic one-hour cache write pricing bug (was being priced incorrectly — update to fix cost projections for cache-heavy workflows). [github.com/pydantic/pydantic-ai/releases/tag/v2.50.0]Anthropic confirmed both follow Opus 5.5 "within weeks" of the September 22 release. No pricing or specific dates announced. [anthropic.com/claude-opus-5-5]
Migrate to antigravity-preview-09-2026, rename snake_case built-in tool parameters to PascalCase, and update file-edit handling. See the September 18 digest Breaking Changes for the full migration guide. [ai.google.dev/gemini-api/docs/changelog]
Standard pricing: $10/$1 (cached)/$50 per MTok. Long context (>272K tokens): $20/$75 per request. Free billing period ends October 4. [platform.openai.com/docs/changelog]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.