The daily firehose — everything the toolchain shipped today, already filtered.
// HOW THIS ISSUE IS MADE
We read every release from the 354 tools on our watchlist at the source — GitHub and GitLab release notes, vendor release pages and changelogs, project blogs and feeds, vendor press releases, and the source code behind the tag. Bug-fix-only releases and non-product newsroom noise are dropped; what's left is summarized down to the new capability, how to try it, and any screenshots or videos the release itself published. Every entry links to the sources it was built from.
VIEW
ISSUE VIEWfull issue
Do you prefer full issue?
▾
↕
filter by tool name…
$ tct list# 41 tools matched
Same issue, same prompt, two writers:
What stands out across today's releases, grouped by what it lets you do. Every tool named links to its entry below.
govern
Answer "which tool call did that?" after an agent touches production
Arcade's Execution Tool Logs record inputs and outputs for every agent tool invocation behind RBACRBACRole-Based Access Control, a permission model where rights are assigned to roles rather than individuals, letting tools enforce least-privilege by granting users only the access their role requires. and a programmatic API, so a destructive or wrong action is reconstructable instead of inferred from application logs; Anthropic's Managed Agents add permission policies on tool calls plus a CLI for watching a running session live. Tinker adds browser-based tinker auth, API key verification and an event_details audit field, so who ran which training job is a record rather than a Slack message.
Find out which part of a voice turn is the slow one
Pipecat v1.9.0 attributes latency per segment across a pipeline turn and adds observers for service metrics, speech events and errors — the difference between knowing a response took 900ms and knowing STTSTTSpeech-to-Text: a class of APIs and models that transcribe spoken audio into written text, letting cyber tools process voice input or recorded audio without building their own acoustic models. ate 600 of it. LiveKit Agents 1.8.1 adds a DuplexModel abstraction with GPTLiveModel, plus telemetry for event-loop blocking, which is the usual hidden cause of stutter in async agent runtimes.
Serve long-context and per-tenant fine-tuned models without a second stack
ONNX Runtime 1.30.0 brings paged attention, speculative XQA and FP4FP4A 4-bit floating-point numeric format used in machine-learning inference to compress model weights, reducing memory and compute cost when running AI models at scale./FP8FP8An 8-bit floating-point numeric format used in AI accelerator hardware to store and compute model weights at reduced precision, cutting memory bandwidth and compute cost versus FP16 or FP32.QMoEQMoEA model compression technique that quantizes transformer weights to sub-1-bit precision using sparse codebooks, letting researchers shrink mixture-of-experts LLMs to a fraction of their original size without dense hardware. kernels to CUDA plus LoRA adapter loading, so the memory and throughput tricks that made vLLM viable are now reachable from an ORT deployment, and Go bindings mean a Go service can call the runtime directly. Together AI adds custom LoRA adapters on dedicated endpoints and preemptible, sub-hourly-billed GPU cluster nodes for work that can tolerate eviction; Darkbloom turns on Gemma QAT prefix caching by default.
Three unrelated things stand out today: Braintrust can mine traces and run scheduled improvement loops; Pinecone can enforce document schemas and metadata indexing at index or namespace scope; Arcade gives every agent tool call a filterable, RBACRBACRole-Based Access Control, a permission model where rights are assigned to roles rather than individuals, letting tools enforce least-privilege by granting users only the access their role requires.-controlled audit trail.
evaluate
Turn production traces into recurring model-improvement work
Patterns detects repeated behaviors on a schedule, the Debugger narrows a single failing trace, and the server-side Loop runtime can automate follow-up work. Teams get a tighter path from 'something looks off in logs' to a reproducible improvement loop without manually trawling trace tables.
Constrain retrieval data before it becomes a messy index
Document-schema indexes via SchemaBuilder and index/namespace-level metadata indexing controls move structure into the vector store boundary. That lets teams decide what fields are valid and searchable up front instead of cleaning up inconsistent metadata after retrieval quality degrades.
Arcade now records per-run Execution Tool Logs with inputs, outputs, filtering, RBACRBACRole-Based Access Control, a permission model where rights are assigned to roles rather than individuals, letting tools enforce least-privilege by granting users only the access their role requires., and an API; Tinker expanded audit logging with event_details and added training-data provenance. Together they make agent and training activity inspectable after the fact, instead of relying on screenshots, ad hoc logs, or trust in the orchestrator.
Keep agent sessions available when models or workers fail
ADK adds automatic model failover, while OpenClaw adds prepared cloud-worker sessions with a configurable reserve pool and snapshot lifecycle management. The practical gain is fewer user-visible dead ends: a model outage or cold worker no longer has to turn into a failed agent run.
ⓘSourcesRelease page →1 RELEASE · seen 2026-09-11NOTES ↗RankWhy this ranktop features scored 85/80/55, +10 product media, +6 runnable example — 89.33How ranking works →
Warp's latest updates extend team-scoping across the Oz CLI and environment pickers, add first-class Grok Build support, and improve fzf-based file and history search handoff.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Warp is an AI terminal that runs commands and helps developers build software.
Warp's latest updates extend team-scoping across the Oz CLI and environment pickers, add first-class Grok Build support, and improve fzf-based file and history search handoff.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
6 FEATURES2 OF 6 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0100
Team-scoped Oz CLI and environment selectionIMPROVED85how completely this was documenteddepth30/40specificity30/30actionability25/3085 / 100
Adds team selection to oz model list (for team-specific model catalogs), oz run list, runner listing and name-based updates, and the API key list, create, and expire commands in the Oz CLI, all of which now respect the selected team.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Oz agent catalogs and environment lists/interactive pickers also follow active team selection, with environment pickers showing both personal environments and environments for the selected window team.
fzf and Fish shell search handoffIMPROVED80how completely this was documenteddepth25/40specificity25/30actionability30/3080 / 100
Extends shell widget handoff to fzf's ctrl-t file search, alongside the existing ctrl-r history-search handoff; Fish's ctrl-r now hands off to PatrickF1/fzf.fish history search, not only junegunn fzf.
Fuzzy-search files in the current directory from the terminal without leaving Warp, using the new fzf ctrl-t handoff.
📍Press ctrl-t in a Warp terminal session to invoke fzf file search and insert the selected path into the prompt.
First-class Grok Build supportNEW55how completely this was documenteddepth25/40specificity20/30actionability10/3055 / 100
Warp adds first-class support for Grok Build (SpaceX AI Grok): it detects Grok Build sessions, adds branded chrome in the footer, and enables rich input mode.
Editable keybindings for Attach file and new windowIMPROVED55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
Attach file is now a Command Palette action with an editable, unbound-by-default keybinding, and Create New Window is now an editable keyboard shortcut that can be remapped or unbound.
ⓘSourcesRelease page →1 RELEASE · seen 2026-09-11NOTES ↗RankWhy this ranktop features scored 65/60/55, +15 breaking change, +6 runnable example — 81How ranking works →
Bolt.new launched Bolt Slides, a full slide-deck builder inside bolt.new with in-preview visual editing, rich speaker notes, a speaker view for presenting, slide reordering, and PDF/JSON export.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Bolt.new is an AI-powered web development platform that builds websites and applications from prompts.
Bolt.new launched Bolt Slides, a full slide-deck builder inside bolt.new with in-preview visual editing, rich speaker notes, a speaker view for presenting, slide reordering, and PDF/JSON export.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
7 FEATURES7 OF 7 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Select tool for in-preview visual editsNEW65how completely this was documenteddepth25/40specificity20/30actionability20/3065 / 100
Adds a Select tool in the chatbox for direct visual edits — text, colors, font size, formatting, and alignment — made right in the project preview without prompting.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Edits are batched above the chatbox and only consume tokens when the user clicks 'Save changes', making preview edits free on all plans.
Iterate on a project's look and feel without spending tokens — use the Select tool to batch visual changes and only commit when satisfied.
📍In the chatbox, activate the Select tool, click elements in the preview to edit text, colors, or font size, then click Save changes when done.
— Explains mechanism and token behaviour with exact UI stepssnapshot-20260911
0202
Speaker view for presentationsNEW60how completely this was documenteddepth25/40specificity15/30actionability20/3060 / 100
Adds speaker view to Bolt Slides presentations: your tab shows speaker notes, the upcoming slide, and a timer, while the audience tab displays a fullscreen view of the slides.
Present to an audience while keeping notes and a timer visible only to you.
📍In your Bolt Slides project, click 'Present', then choose 'Speaker view' — your tab shows notes, the upcoming slide, and a timer, while the audience tab displays fullscreen slides.
— Describes exact tab layout with a concrete Present > Speaker view pathsnapshot-20260911
thinner coverage below
0303
PDF and JSON export for Slides decksNEW55how completely this was documenteddepth20/40specificity15/30actionability20/3055 / 100
Exports Bolt Slides decks as PDF or JSON directly from the slide editor's Download menu, letting users share a file or import the JSON into another Bolt project to reuse its content.
Export a finished Slides deck as JSON to reuse its content in another Bolt project.
📍In your Bolt Slides project, open the Download menu and select 'JSON' to save the deck file locally.
Export a finished Bolt Slides deck as JSON to reuse slide content in another Bolt project.
📍In your Slides project, open the deck, then select Download › JSON to save the deck file locally.
— Names both export formats and the exact menu pathsnapshot-20260911
0404
Bolt Slides deck builderNEW40how completely this was documenteddepth15/40specificity10/30actionability15/3040 / 100
Adds Bolt Slides, enabling users to build, customize, share, and present slide decks directly within bolt.new.
When you need to generate and present a slide deck without leaving bolt.new.
📍In bolt.new, open a project and select Slides from the available use cases to build, customize, and share your presentation.
— Names the product area and entry point, no deeper mechanism givenproduct docs
0505
Rich-text speaker notes editorNEW35how completely this was documenteddepth15/40specificity15/30actionability5/3035 / 100
Adds a speaker notes editor to Bolt Slides with rich text support including bold, italic, headings, and lists.
— Lists formatting options but gives no usage stepssnapshot-20260911
0606
Slide reorder, duplicate and delete controlsNEW35how completely this was documenteddepth15/40specificity10/30actionability10/3035 / 100
Adds slide organization controls — reorder, duplicate, or delete slides — from the side panel or grid overview.
— Names the controls and UI location but no worked examplesnapshot-20260911
0707
Existing decks excluded from new Slides featuresBREAKING20how completely this was documenteddepth15/40specificity5/30actionability0/3020 / 100
Existing Bolt Slides decks retain the controls they had at creation time and do not receive the new editing and export features added in this release.
— States the impact but offers no migration or workaround stepssnapshot-20260911
└──▷ BREAKING ON UPGRADE
!Existing Bolt Slides decks retain the controls they had at creation time and do not receive the new editing and export features.
ⓘSourcesRelease page →1 RELEASE · seen 2026-09-11NOTES ↗RankWhy this ranktop features scored 70/65/65, +10 product media, +6 runnable example — 82.67How ranking works →
Cursor's biggest addition this window is Projects, a coordinator agent that delegates to thousands of parallel subagents on persistent cloud compute, keeps a shared context file, can be subscribed to Slack/schedules/PRs, and can be steered mid-run; alongside it, Origin Apps bring private integrations built on the Origin Public API.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Built to make you extraordinarily productive, agents turn ideas into code. Accelerate development by handing off tasks to Cursor.
Cursor's biggest addition this window is Projects, a coordinator agent that delegates to thousands of parallel subagents on persistent cloud compute, keeps a shared context file, can be subscribed to Slack/schedules/PRs, and can be steered mid-run; alongside it, Origin Apps bring private integrations built on the Origin Public API.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
5 FEATURES5 OF 5 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Follow-up steering mid-runNEW70how completely this was documenteddepth30/40specificity20/30actionability20/3070 / 100
Adds follow-up steering: send a message to redirect the agent mid-run without interruption.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Follow-ups queue at the next tool call rather than cutting the agent off; send immediately with 'Send now' or by pressing ⏎ twice.
— Names exact controls for immediate, actionable use.snapshot-20260911
0202
Projects: coordinator agent with cloud/local subagentsNEW65how completely this was documenteddepth30/40specificity20/30actionability15/3065 / 100
Cursor launches Projects, accessible from the left-hand nav, where a coordinator agent plans and delegates work across thousands of parallel subagents without requiring the user to stay online. Projects run on cloud agents with their own persistent compute so work continues after closing your laptop, with local agents spun up on demand for machine-specific tasks.
Kick off a long-running security remediation task — e.g. eliminating all flaky auth tests — and let the coordinator run to completion without babysitting it.
$ /goal fix all flaky tests and make CI green
— Explains mechanism and scope but no config keys or API.snapshot-20260911
0303
Subscriptions for autonomous Project triggersNEW65how completely this was documenteddepth25/40specificity20/30actionability20/3065 / 100
Adds Subscriptions within Projects, letting the coordinator watch a Slack channel, run on a schedule, or follow all PRs and act autonomously on detected signals. Configured via 'Subscriptions > Connect Slack' to point the coordinator at a channel, so it delegates a fix each time a new bug report arrives.
Route incoming bug reports from a Slack channel directly to the coordinator so it delegates fixes automatically, without waiting for a manual prompt.
📍In Projects, open Subscriptions › Connect Slack, then point it at your bug-report channel so the coordinator delegates a fix each time a new report arrives.
Origin Apps for private integrationsNEW65how completely this was documenteddepth25/40specificity20/30actionability20/3065 / 100
Cursor adds Origin Apps, private integrations built on the Origin Public API and installed from repository settings via 'Repository Settings > Origin Apps > Install'. Example integrations shown include Vercel, Depot, and Buildkite apps connected to an Origin repo.
Create a private Origin App to integrate a custom workflow with Cursor via the Origin Public API.
📍In Cursor, go to Repository Settings › Origin Apps and select 'Install' to create a new private Origin App.
— Names install path and API but omits endpoint specifics.product docs
thinner coverage below
0505
Shared context files across Projects agentsNEW45how completely this was documenteddepth25/40specificity15/30actionability5/3045 / 100
Projects maintain a shared context file set that syncs across all cloud and local agents, growing automatically as agents record codebase knowledge and working preferences.
— Explains mechanism but gives no way to inspect or configure it.snapshot-20260911
OpenAI Codex CLI shipped thread-scoped instructions, easier navigation back to the command center after canceled sessions, bundled Linux voice runtimes, a new managed-daemon shutdown flag, higher Guardian action review limits, and trigger propagation for delegated agents.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
OpenAI Codex CLI runs an agent in the terminal that reads, changes, and tests code in local repositories.
OpenAI Codex CLI shipped thread-scoped instructions, easier navigation back to the command center after canceled sessions, bundled Linux voice runtimes, a new managed-daemon shutdown flag, higher Guardian action review limits, and trigger propagation for delegated agents.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
5 FEATURES2 OF 5 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Thread-scoped instructions via ThreadInstructionsProviderNEW75how completely this was documenteddepth30/40specificity25/30actionability20/3075 / 100
Adds ThreadInstructionsProvider via StartThreadOptions to supply per-thread instructions that compose after global instructions and before repository instructions, enforced with a 10,000 estimated-token cap independent of the repository instruction budget.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Bundled Linux voice runtimes and audio buffer tuningIMPROVED55how completely this was documenteddepth25/40specificity20/30actionability10/3055 / 100
Bundles GNU voice helpers and runtimes with primary Linux musl release archives and increases ALSA capture/playback buffer sizes to support larger PipeWire graph cycles, improving audio reliability.
Turn trigger propagation to delegated agentsIMPROVED55how completely this was documenteddepth25/40specificity20/30actionability10/3055 / 100
Propagates turn_trigger through agent spawning and follow-up messages so delegated agents retain the trigger (composer input, scheduled automation) of the initiating turn.
Command center navigation after session cancellationIMPROVED50how completely this was documenteddepth25/40specificity15/30actionability10/3050 / 100
Returns the user to the command center after session deletion, a canceled unarchive/fork, or pressing Esc in a conversation open in another app, instead of exiting the TUI.
Cline desktop v0.0.26 brings live GitHub PR and CI status directly into the composer, streamlines provider selection with Cline Pass and a quick-add row, and reworks the Customize view's navigation and toggles.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Autonomous coding agent as an SDK, IDE extension, or CLI assistant.
Cline desktop v0.0.26 brings live GitHub PR and CI status directly into the composer, streamlines provider selection with Cline Pass and a quick-add row, and reworks the Customize view's navigation and toggles.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
4 FEATURES2 OF 4 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Live GitHub PR and CI status in composerNEW76how completely this was documenteddepth32/40specificity26/30actionability18/3076 / 100
The composer now shows live GitHub PR status — PR number, merge status, changed-line totals, and CI checks with log inspection — refreshing every 30 seconds, on window focus, and on demand.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Requires the gh CLI installed and signed in, with a GitHub.comorigin remote.
— Names exact requirements and refresh triggers, missing only a demo.desktop-v0.0.26
0200
Customize view redesign and controlsIMPROVED65how completely this was documenteddepth28/40specificity22/30actionability15/3065 / 100
Unifies the Customize view's Tools, Skills, and Rules tabs into a consistent list style; adds a search bar to the Tools tab that filters both sections, per-section Enable all/Disable all controls, and an in-place enable/disable toggle for Skills with Copy path / Uninstall menu options. Tabs are reordered to Tools, Plugins, Skills, Rules, MCP, Hooks, and the sidebar's 'New Task' button is renamed to 'New Session'.
— Names UI elements and controls precisely; navigation path given but no runnable command.desktop-v0.0.26
thinner coverage below
0302
Provider picker additions: Cline Pass and quick addNEW51how completely this was documenteddepth18/40specificity18/30actionability15/3051 / 100
The composer's provider picker gains a Set up another provider row that opens Settings → Models to reach the full model catalog without leaving the composer, and a new Cline Pass entry alongside Cline, configured under the same Cline account credentials.
— Names the two picker additions and where they lead, but thin on mechanism.desktop-v0.0.26
0400
Recurring schedules default to local timezoneIMPROVED35how completely this was documenteddepth15/40specificity10/30actionability10/3035 / 100
New recurring schedules now default to the local timezone instead of UTC.
— States the before/after but no further detail on configuration.desktop-v0.0.26
Letta Code is a terminal coding agent with persistent memory and identity across development sessions.
Letta Code v0.32.2 adds host-provided Slack conversation binding and an experimental unified toolset, while removing AgentFile import/export support entirely.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
3 FEATURES3 OF 3 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
AgentFile import/export removedBREAKING80how completely this was documenteddepth25/40specificity30/30actionability25/3080 / 100
The --import and --from-af CLI flags and the /export and /download slash commands are removed; AgentFile (.af) import and export are no longer supported, including imports from the agent registry.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Names exact removed flags and commands with clear migration impact.v0.32.2
thinner coverage below
0202
Slack conversation binding actionsNEW38how completely this was documenteddepth15/40specificity15/30actionability8/3038 / 100
Adds host-provided Slack conversation binding actions via the channels integration, enabling agents to bind and interact with Slack conversations.
— Names the integration but no command or config to invoke it.v0.32.2
0303
Experimental unified Letta toolsetNEW23how completely this was documenteddepth10/40specificity8/30actionability5/3023 / 100
Adds an experimental unified Letta toolset, consolidating agent tools into a single integrated set.
— Vague description with no naming of which tools are consolidated.v0.32.2
└──▷ BREAKING ON UPGRADE
!The --import and --from-af CLI flags and the /export and /download slash commands are removed; AgentFile (.af) import and export are no longer supported, including imports from the agent registry.
Graphify's core CLI now exposes graph query, path, and explain commands with per-edge provenance, adds an MCP server and assistant-skill installer for AI integration, and extends graph extraction to SQL indexes and Rust statics/consts.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Graphify converts code, docs, SQL schemas, configs, and PDFs into a queryable knowledge graph with explainable relationships.
Graphify's core CLI now exposes graph query, path, and explain commands with per-edge provenance, adds an MCP server and assistant-skill installer for AI integration, and extends graph extraction to SQL indexes and Rust statics/consts.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
9 FEATURES9 OF 9 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Core graph query, path and explain commandsNEW90how completely this was documenteddepth32/40specificity28/30actionability30/3090 / 100
graphify query <symbol> retrieves everything connected to a named symbol in the code graph, and also accepts natural-language queries with a --dfs flag to switch traversal to depth-first search.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ graphify path <symbol1> <symbol2> finds the route or shortest hop chain between two named nodes (e.g. graphify path "UserService" "DatabasePool"), returning extracted edges with file and line references. graphify explain <edge> explains why a specific edge or node exists, including its provenance label.
Trace every caller and dependency of an auth service before refactoring it, to understand blast radius.
$ graphify query AuthService
Find the call path between an authentication symbol and the database layer to audit privilege escalation routes.
$ graphify path AuthService Db
Investigate a suspicious or unexpected edge to see whether it was derived from source code or inferred by the model.
$ graphify explain <edge>
Trace how a request flows to the cache — useful for auditing lateral paths through your architecture.
$ graphify query "how does a request reach the cache?" --dfs
Find the explicit dependency path between two services to understand blast radius before a change.
$ graphify path "UserService" "DatabasePool"
Trace the dependency path between two services to understand blast radius during an incident or code review.
$ graphify path "UserService" "DatabasePool"
— Three named, flag-bearing commands with runnable examples across releases.product docs
0202
MCP server for the code graphNEW78how completely this was documenteddepth26/40specificity26/30actionability26/3078 / 100
python -m graphify.serve exposes the graph over MCP using stdio transport with 10 tools, writing no files and touching no code, so an assistant can query the graph directly (e.g. python -m graphify.serve graphify-out/graph.json).
Expose the repo graph to an AI assistant over MCP so it can answer architecture questions without touching your codebase.
— Named command, transport, tool count and a runnable example.product docs
0303
PR triage subcommandNEW65how completely this was documenteddepth18/40specificity22/30actionability25/3065 / 100
graphify prs lists pull requests from the terminal, with --triage to surface the PR queue and --conflicts to filter for pull requests that introduce graph conflicts.
Surface PRs that introduce graph conflicts before merging — useful in CI or pre-review triage.
$ graphify prs --conflicts
— Named subcommand and flags with a runnable example.product docs
0404
Provenance labels on graph edgesNEW64how completely this was documenteddepth28/40specificity24/30actionability12/3064 / 100
Every graph edge is tagged with a provenance label — parsed (from the syntax tree), inferred (from the model), or unresolved (dynamic dispatch, reflection, runtime-built imports) — so callers can tell how each connection was established, surfaced via graphify explain.
— Mechanism and label taxonomy given, but no dedicated command of its own.product docs
0505
Graph output artifacts after scanNEW63how completely this was documenteddepth20/40specificity25/30actionability18/3063 / 100
Scanning produces graphify-out/graph.html (an interactive graph viewable directly in a browser), graphify-out/GRAPH_REPORT.md (a human-readable report covering god nodes, communities, and surprises), and graphify-out/graph.json (the full machine-readable graph for downstream tooling).
— Three named output files with clear purpose, no command to generate them shown here.product docs
thinner coverage below
0606
Assistant skill installer across 17 toolsNEW57how completely this was documenteddepth15/40specificity24/30actionability18/3057 / 100
graphify install auto-detects and registers the /graphify skill, wiring Graphify into an AI assistant, across 17 supported assistants including Claude Code, Cursor, and GitHub Copilot.
— Named command and assistant count, but no example invocation shown.product docs
0707
New graph node types from SQL indexes and Rust statics/constsNEW56how completely this was documenteddepth20/40specificity24/30actionability12/3056 / 100
SQL extraction now emits index nodes for CREATE [UNIQUE] INDEX statements, linked to the table they index, and Rust module-level static and const declarations are now extracted as nodes in the knowledge graph.
— Two named extraction targets, no command or example given.v0.9.58
0808
Local multi-language parsing across 36 grammarsNEW54how completely this was documenteddepth24/40specificity20/30actionability10/3054 / 100
Source files are parsed locally across 36 language grammars, turning functions, imports, and calls into graph edges without sending code to a model backend.
— Explains mechanism and scope but no command surface given.product docs
0909
Deep scan modeNEW40how completely this was documenteddepth10/40specificity15/30actionability15/3040 / 100
A --mode deep flag on the scan command performs a multi-pass read of the repository.
— Named flag but no detail on what the extra passes change.product docs
Diagram Design is a Claude Code skill that generates editorial HTML and SVG diagrams and imports draw.io or Mermaid sources.
Adds treemap, dumbbell, slopegraph, ridgeline, bubble, beeswarm, and ten editorial diagram types, plus native Droid plugin packaging and named client profiles.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ GET THIS VERSION
$ git clone --branch commits-2026-08-12 https://github.com/cathrynlavery/diagram-design.git
# already have the repo? check out this version:$ git checkout commits-2026-08-12
└──▷ TRY IT
Catch clipped SVG content and runtime paint errors in rendered diagram examples before shipping.
$ python3 scripts/lint-render.py
Verify that all beeswarm dots sit at their declared values and that no two dots overprint, as part of a geometry gate.
$ python3 scripts/verify-beeswarm.py
Run adversarial tests against the Sankey verifier in both pass and fail polarities to confirm the checker fires on real defects.
$ python3 scripts/test-verify-sankey.py
›Adds verify-beeswarm.py to verify that beeswarm dot positions match their declared values and that no two dots overprint, enforcing nine geometric invariants including shared value scale and packing-only swarm offset.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
›Adds lint-render.py, a headless-Chromium rendering linter that catches clipped SVG content, collapsed SVGs, and page overflow by diffing screenshots with and without overflow released across staged ancestor levels.
›Adds test-verify-sankey.py adversarial test suite for verify-sankey.py, covering ribbon narrowing, volume loss, stage carry, label mismatch, and dark/light variant drift in both polarities.
›Adds treemap diagram type for part-of-whole by area.
›Adds dumbbell as a Bar chart variant.
+9 moreshow less
›Adds slopegraph as a Line chart variant for showing change between two states.
›Adds ridgeline as a Line chart variant.
›Adds bubble chart as a Scatter variant for three-value comparisons.
›Ships ten editorial diagram types in a single release.
›Adds automatic plugin updates via native marketplaces and a version gate.
›Ships Noto TC typefaces and adds a Traditional Chinese labels section.
›Requests CJK typefaces in the exported SVG @import and gates parity.
›Adds semantic patterns and accessible motion.
›Adds verify-beeswarm.py to verify that beeswarm dot positions match their declared values and that no two dots overprint, enforcing nine geometric invariants including shared value scale and packing-only swarm offset.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
›Adds lint-render.py, a headless-Chromium rendering linter that catches clipped SVG content, collapsed SVGs, and page overflow by diffing screenshots with and without overflow released across staged ancestor levels.
›Adds test-verify-sankey.py adversarial test suite for verify-sankey.py, covering ribbon narrowing, volume loss, stage carry, label mismatch, and dark/light variant drift in both polarities.
›Adds treemap diagram type for part-of-whole by area.
›Ships ten editorial diagram types in a single release.
+4 moreshow less
›Adds automatic plugin updates via native marketplaces and a version gate.
Claude Code is Anthropic's terminal coding agent that plans, edits, and tests code in local repositories.
Claude Code's v2.1.268 release expands gateway administration (pricing sync, internal-network allowlisting, access-control warnings), adds JSON output across plugin commands and a session-state cleanup flag for self-hosted runners, and restricts task-tracking tools to a specific model list.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
11 FEATURES9 OF 11 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Session-state cleanup flag for self-hosted runnersNEW85how completely this was documenteddepth30/40specificity25/30actionability30/3085 / 100
Adds --remove-session-state flag to claude self-hosted-runner (default off) to delete each session's per-session directories under <base-dir>/_sessions/ when the session ends.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Clean up session state automatically after each run in a CI self-hosted runner, keeping disk usage bounded across many short sessions.
$ claude self-hosted-runner --remove-session-state
— Exact flag, default state, and directory path givenv2.1.268
0202
Task-tracking tools restricted to specific modelsBREAKING85how completely this was documenteddepth30/40specificity30/30actionability25/3085 / 100
Changes TaskCreate/Get/Update/List and TodoWrite task-tracking tools to be offered only on Claude 3.x, Opus 4.0-4.7, Sonnet 4.0-4.6, Haiku 4.5; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to enable elsewhere.
— Exact model list and override env var givenv2.1.268
0300
JSON output across plugin commandsIMPROVED81how completely this was documenteddepth25/40specificity28/30actionability28/3081 / 100
Adds --json output flag to claude plugin install, uninstall, update, enable, and disable, and adds errorDetails/noteDetails fields to each row of claude plugin list --json.
Parse plugin operation results in automation scripts by getting structured JSON back from install and list commands.
$ claude plugin install my-plugin --json && claude plugin list --json | jq '.[].errorDetails'
— Names every command and field, with runnable examplev2.1.268
0403
gatewayInternalNetworks managed settingNEW75how completely this was documenteddepth25/40specificity25/30actionability25/3075 / 100
Adds gatewayInternalNetworks managed setting, letting administrators allow /login to a Claude apps gateway on their organization's own public IPv4 block.
Allow org employees to authenticate via /login through a self-hosted Claude apps gateway on the company's own public IP block.
yaml
gatewayInternalNetworks:
- 203.0.113.0/24
— Named setting shown with concrete config examplev2.1.268
0504
WebFetch deadline override environment variableNEW75how completely this was documenteddepth25/40specificity25/30actionability25/3075 / 100
Adds CLAUDE_CODE_WEBFETCH_DEADLINE_MS environment variable to override the WebFetch 300-second deadline (set to 0 to disable the deadline entirely).
— Exact env var name, default, and disable valuev2.1.268
0605
WebFetch rules no longer gate Artifact toolBREAKING70how completely this was documenteddepth25/40specificity25/30actionability20/3070 / 100
Plain WebFetch deny and ask rules no longer apply to Artifact tool reads and updates; use an Artifact rule (or WebFetch(domain:claude.ai)) to block or gate them.
— Names exact rule migration path for affected usersv2.1.268
0706
Built-in hooks-module plugins published as sourceNEW68how completely this was documenteddepth25/40specificity28/30actionability15/3068 / 100
Publishes three built-in hooks-module plugins as source under mods/: sec-default (organization default outermost plugin), diff (/diff), and telemetry ($.telemetry).
— Names all three plugins and their paths/commandsv2.1.268
0807
Gateway pricing sync in gateway.yamlNEW65how completely this was documenteddepth25/40specificity25/30actionability15/3065 / 100
Adds pricing: support in gateway.yaml so signed-in Claude Code clients receive org-configured rates through managed settings, keeping /cost and telemetry aligned with the spend meter.
— Names config key and effect, lacks example syntaxv2.1.268
thinner coverage below
0908
Startup warnings for unrestricted gateway accessNEW50how completely this was documenteddepth20/40specificity20/30actionability10/3050 / 100
Adds startup warnings for gateways when access_control.allow_cidrs is empty, and a one-time warning the first time a request arrives from a public address.
— Describes trigger conditions, no remediation step givenv2.1.268
1000
configDirectory field in auth status JSONIMPROVED40how completely this was documenteddepth10/40specificity15/30actionability15/3040 / 100
Adds configDirectory field to the output of claude auth status --json.
— Single field addition, no example providedv2.1.268
1109
Browser-tab icons for published artifactsNEW30how completely this was documenteddepth15/40specificity10/30actionability5/3030 / 100
Adds browser-tab icons for published artifacts, chosen by Claude to match each page.
— Thin description, no mechanism or setting namedv2.1.268
└──▷ BREAKING ON UPGRADE
!TaskCreate, Get, Update, List, and TodoWrite task-tracking tools are no longer offered on models outside Claude 3.x, Opus 4.0-4.7, Sonnet 4.0-4.6, and Haiku 4.5; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to restore them on other models.
!Plain WebFetch deny and ask rules no longer apply to Artifact tool reads and updates; use an Artifact rule (or WebFetch(domain:claude.ai)) to block or gate them.
Qwen Code's TypeScript SDK v0.1.12 renames all disable* settings keys to positive enable* naming (a breaking change) and adds a memory.enableManagedAutoMemory toggle for deterministic system prompts, alongside a batch of new SDK capabilities: multi-modal input, session resume, a batch runner, MCPMCPModel Context Protocol, an open standard from Anthropic that lets an AI model call external tools and data sources through a uniform interface, so builders avoid writing custom glue code for every integration. progress reporting, Coding Plan auth, and a new settings.env config field.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
An open-source AI coding agent that lives in your terminal.
Qwen Code's TypeScript SDK v0.1.12 renames all disable* settings keys to positive enable* naming (a breaking change) and adds a memory.enableManagedAutoMemory toggle for deterministic system prompts, alongside a batch of new SDK capabilities: multi-modal input, session resume, a batch runner, MCPMCPModel Context Protocol, an open standard from Anthropic that lets an AI model call external tools and data sources through a uniform interface, so builders avoid writing custom glue code for every integration. progress reporting, Coding Plan auth, and a new settings.env config field.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
9 FEATURES9 OF 9 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Managed auto-memory control via `enableManagedAutoMemory`NEW85how completely this was documenteddepth30/40specificity25/30actionability30/3085 / 100
The memory.enableManagedAutoMemory setting controls whether managed auto-memory injects remember/dream instructions into the system prompt.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Setting it to false, as in { "memory": { "enableManagedAutoMemory": false } }, gives fully deterministic system prompts on a given host.
Prevent managed auto-memory from injecting remember/dream instructions on a host where you want fully deterministic system prompts.
Renamed negative settings keys to positive namingBREAKING65how completely this was documenteddepth25/40specificity20/30actionability20/3065 / 100
Negative settings keys of the form disable* were renamed to positive enable* naming; any config or scripts using the old disable* form must be updated.
`settings.env` config fieldNEW50how completely this was documenteddepth15/40specificity20/30actionability15/3050 / 100
A new settings.env field was added to SDK configuration for setting environment variables; the release notes give no further detail on its schema or usage.
Cotool is an AI security operations platform for investigating alerts, enriching threat intelligence, and automating response workflows.
Cotool v0.68.0 adds two new enrichment integrations (Serval and Lumos), Linear API agent sessions, Okta private key JWT auth, three new LLM models, and several smaller improvements to Tines, Slack, detections, and Jira.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
9 FEATURES7 OF 9 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
New enrichment integrations: Serval and LumosNEW57how completely this was documenteddepth20/40specificity22/30actionability15/3057 / 100
Adds a Serval integration so agents can look up users, groups, access requests, tickets, and workflow runs, and a Lumos integration so agents can look up users, apps, groups, access requests, and activity logs.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Enrich an investigation by having an agent query Serval for a user's open access requests and associated workflow runs.
📍In an agent prompt, instruct: 'Use the Serval integration to look up access requests and workflow runs for user <username>.'
— Names two integrations and their specific lookup capabilities.v0.68.0
0202
Detection accuracy rate columnsNEW50how completely this was documenteddepth20/40specificity15/30actionability15/3050 / 100
Adds true positive, benign, and false positive rate columns (combined, sortable) to the detections table.
OpenClaw v2026.9.4 focuses on cloud worker infrastructure — prepared sessions with a configurable reserve pool and full snapshot lifecycle management — alongside a unified Plugins workspace, a read-only deployment mode, and Deepgram Flux voice transcription.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
OpenClaw is a personal AI assistant platform that connects messaging channels, tools, and model providers through a gateway.
OpenClaw v2026.9.4 focuses on cloud worker infrastructure — prepared sessions with a configurable reserve pool and full snapshot lifecycle management — alongside a unified Plugins workspace, a read-only deployment mode, and Deepgram Flux voice transcription.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
9 FEATURES9 OF 9 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Prepared cloud worker sessions and reserve poolNEW92how completely this was documenteddepth32/40specificity30/30actionability30/3092 / 100
Adds cloudWorkers.profiles.<id>.readyWorkers and cloudWorkers.preparedPool.maxTotal config keys to control how many prepared cloud-worker reserves are kept per project/profile and Gateway-wide (set either to zero to disable reserves).︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Prepared cloud sessions let eligible Linux sessions start from local Git projects or public GitHub repositories using these pre-built worker snapshots, avoiding repeated environment setup.
Disable all prepared cloud-worker reserves Gateway-wide to avoid incurring provider running-machine charges for warm workers.
yaml
cloudWorkers:
preparedPool:
maxTotal: 0
Cap ready workers for a specific project profile to one instead of the default, controlling warm-worker spend at a per-project level.
— Names exact config keys with runnable example values and behavior.v2026.9.4
0202
Cloud worker snapshot managementNEW75how completely this was documenteddepth28/40specificity22/30actionability25/3075 / 100
Cloud worker snapshots can be built, rebuilt, inspected, pinned, deleted, and rolled back from Settings → Connections → Cloud workers → Snapshots in the Control UI, with support for cancelling an active build.
— Exact UI path and full set of lifecycle operations given.v2026.9.4
0303
/question command for pending promptsNEW70how completely this was documenteddepth25/40specificity20/30actionability25/3070 / 100
The /question command reopens a pending terminal prompt in Gateway-connected and local TUI sessions, supporting keyboard-driven choices, free-text answers, and multi-question prompts.
— Named command with clear supported interaction modes.v2026.9.4
0404
Read-only deployment config modeNEW65how completely this was documenteddepth22/40specificity18/30actionability25/3065 / 100
New OPENCLAW_CONFIG_READONLY=1 environment variable prevents OpenClaw from rewriting deployment-managed configuration, while keeping read-only diagnostics and runtime state functional.
— Named env var is directly runnable, but scope of effect only briefly described.v2026.9.4
thinner coverage below
0505
Deepgram Flux voice-note transcriptionNEW58how completely this was documenteddepth18/40specificity25/30actionability15/3058 / 100
Voice notes can be transcribed using Deepgram Flux via the flux-general-en or flux-general-multi models; requires ffmpeg to be installed.
— Names models and dependency but no usage steps shown.v2026.9.4
0606
Unified Plugins workspace in Control UINEW50how completely this was documenteddepth20/40specificity15/30actionability15/3050 / 100
A single Plugins workspace in the Control UI lets users discover bundled and ClawHub plugins, install them, and manage setup, settings, and access all in one place.
— Describes scope well but only a UI area, no exact path.v2026.9.4
0707
Automatic rollback for failed updatesNEW50how completely this was documenteddepth28/40specificity12/30actionability10/3050 / 100
When an update fails but is compatible, OpenClaw automatically retains the previous package and restores it along with its prior configuration and service, provided schema and configuration checks confirm the rollback is safe.
— Mechanism explained but no config or trigger named.v2026.9.4
0808
Advanced cloud worker profile editing and Windows workersNEW45how completely this was documenteddepth18/40specificity15/30actionability12/3045 / 100
The Control UI adds advanced editing of cloud worker profiles and repository defaults, plus native Windows worker selection when the backend supports it.
— Names areas but no specifics on what 'advanced' editing entails.v2026.9.4
0909
GPT Image 2.5 Flare and Sunburst variantsNEW45how completely this was documenteddepth15/40specificity20/30actionability10/3045 / 100
Adds GPT Image 2.5 Flare and Sunburst variants for image generation and editing through OpenAI or fal, usable without changing the default model.
— Names the model variants and providers but no usage detail.v2026.9.4
Pipecat is a framework for realtime voice and multimodal agents, orchestrating speech-to-text, model inference, and text-to-speech as a streaming pipeline.
Pipecat v1.9.0 ships a per-segment latency breakdown for pipeline turns, three new observability observers (service metrics, speech events, errors), a Meta streaming STT integration, and a broad set of new tuning settings across roughly a dozen STT/TTS providers.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
13 FEATURES12 OF 13 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Per-segment latency breakdown for pipeline turnsNEW90how completely this was documenteddepth35/40specificity30/30actionability25/3090 / 100
Adds LatencyBreakdown.contributions to decompose user-to-bot latency into named segments (VAD silence, transcription, LLM inference, TTS, etc.), each with a stable key, owner_kind, measured_from, and total_secs.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Print with LatencyBreakdown.turn_contribution_lines() and filter short contributions with UserBotLatencyObserver(min_contribution_secs=...).
— Names fields and methods; no full runnable example given.v1.9.0
0202
Meta streaming STT service integrationNEW90how completely this was documenteddepth30/40specificity30/30actionability30/3090 / 100
Adds MetaSTTService, a streaming STT service using Meta's muse-voice-transcribe-1.0 model over the Muse Voice realtime API; install with pipecat-ai[meta]. Settings takes language, language_bias, keywords, and mode (model endpointing, client-delimited, or diarization).
Use MetaSTTService for multilingual streaming transcription with domain-specific keyword boosting, installed via the meta extra.
— Named install extra and settings, with a runnable example.v1.9.0
0303
New observers for service metrics, speech events, and errorsNEW85how completely this was documenteddepth30/40specificity30/30actionability25/3085 / 100
Adds ServiceMetricsObserver with on_service_latency (time-to-first-byte, first audio, first answer token) and on_service_usage (STT audio seconds, TTS characters, LLM token counts), one record per metric per turn. Adds SpeakingObserver reporting user_speech_started, user_speech_stopped, user_turn_started, user_turn_stopped, bot_speech_started, bot_speech_stopped, and interruption events through on_speech_event. Adds ErrorObserver reporting every pipeline error through on_error as an ErrorEvent with message, category, exception_type, processor, and processor_usable.
Wire up an ErrorObserver to catch and categorise every pipeline error in one place — useful when running unattended agents where silent failures are hard to detect.
— Full event/field names given; runnable example only for ErrorObserver.v1.9.0
0400
Expanded pipeline startup timing reportIMPROVED75how completely this was documenteddepth30/40specificity30/30actionability15/3075 / 100
Adds StartupTimingReport.warmup with warmup.duration_secs and warmup.blocking_duration_secs to account for deferred-import warming cost, plus a new BaseObserver.on_startup_warmup() event. Also expands StartupTimingReport with setup_phase_secs, start_phase_secs, and per-processor ProcessorStartupTiming.start_duration_secs to break down the full pipeline startup span.
— Detailed named fields, no example usage shown.v1.9.0
0504
Deepgram STT and TTS setting additionsNEW73how completely this was documenteddepth25/40specificity30/30actionability18/3073 / 100
Adds profanity_filter and redact to DeepgramFluxSTTService.Settings and DeepgramFluxSageMakerSTTService.Settings for profanity masking and number redaction. Adds a version field to DeepgramSTTService.Settings (and DeepgramSageMakerSTTService.Settings) to pin transcription to a specific Deepgram model version instead of latest. Adds speed to DeepgramTTSService and DeepgramHttpTTSService for a Deepgram Aura speech-rate multiplier (0.7–1.5) passed to /v1/speak.
— Multiple named settings, classes and endpoint, no example.v1.9.0
0605
Azure STT and TTS tuning settingsNEW71how completely this was documenteddepth25/40specificity28/30actionability18/3071 / 100
Adds voice_parameters to AzureTTSService and AzureHttpTTSService, exposing SSML <voice>parameters (temperature, top_p, top_k, cfg_scale, enhancePronunciation) for HD voice tuning. Adds segmentation_silence_timeout_ms (100–5000 ms) to AzureSTTService.Settings to control how much in-phrase silence Azure tolerates before emitting a final transcript.
— Named settings and value ranges, no example code.v1.9.0
0706
AssemblyAI STT multilingual, timing, and new model supportNEW68how completely this was documenteddepth25/40specificity28/30actionability15/3068 / 100
Adds language_codes (e.g. [Language.EN, Language.ES]) and timestamps settings to AssemblyAISyncSTTService for multilingual/code-switching audio and per-word timing. Adds universal-3-6-pro as a supported model for AssemblyAISTTService, with the full U3-RT-Pro feature set (built-in turn detection, prompting, continuous partials, interruption_delay, context carryover, voice focus).
— Named settings and model, but no usage example.v1.9.0
0807
DeepSeek reasoning toggleNEW65how completely this was documenteddepth20/40specificity25/30actionability20/3065 / 100
Adds thinking setting to DeepSeekLLMService.Settings to control DeepSeek's reasoning pass, e.g. thinking={"type": "disabled"} disables the V4 model's chain-of-thought.
— Named setting with example value, no full snippet.v1.9.0
0908
Additional STT/TTS provider tuning settingsNEW65how completely this was documenteddepth20/40specificity30/30actionability15/3065 / 100
Adds no_verbatim to ElevenLabsSTTService and ElevenLabsRealtimeSTTService to strip filler words, false starts, and non-speech sounds from transcripts. Adds hotwords to WhisperSTTService.Settings to bias Faster Whisper transcription toward specific words or phrases. Adds reduce_silence to SonioxTTSService to shorten inter-word pauses on supported models. Adds include_results to SpeechmaticsSTTService to request word-level results in transcript messages. Adds temperature to HumeTTSService.Settings to control Hume's synthesis sampling temperature. Adds reconnect_backoff_min_wait and reconnect_backoff_max_wait to WebsocketService to configure reconnection wait intervals (defaults: 4 s / 10 s).
— Six named settings across providers, no usage examples.v1.9.0
1009
Chirp 3 noise removal in Google STTNEW62how completely this was documenteddepth22/40specificity25/30actionability15/3062 / 100
Adds denoiser_config to GoogleSTTService, exposing Chirp 3 background-noise removal via denoise_audio and snr_threshold.
Server-side turn detection tuning for CartesiaNEW62how completely this was documenteddepth22/40specificity25/30actionability15/3062 / 100
Adds turn_start_threshold, turn_eager_end_threshold, turn_end_threshold, and turn_end_timeout_ms to CartesiaTurnsSTTSettings for server-side turn detection tuning.
— Four named fields, no defaults or example given.v1.9.0
thinner coverage below
1211
Progress and update handlers for eval frameworkNEW55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
Adds on_progress and on_update event handlers to the eval framework; EvalSession and EvalSuite are now BaseObjects with on_progress (async task) and on_update (synchronous) handlers respectively.
— Names handlers and classes but no example.v1.9.0
1312
Cancellable function calls across LLMSwitcherNEW55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
Adds cancellable_by_llm argument to LLMSwitcher.register_function(), forwarding it to every LLM the switcher fronts.
NeMo Agent Toolkit v1.9 adds new middleware for human-in-the-loop approval and tool fault tolerance, an MLflow OTLP telemetry exporter, and expanded authentication provider options, while removing the local_sandbox execution mode and changing how user identity is resolved.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
NeMo Agent Toolkit connects agent workflows to tools and model providers with profiling, evaluation, and observability.
NeMo Agent Toolkit v1.9 adds new middleware for human-in-the-loop approval and tool fault tolerance, an MLflow OTLP telemetry exporter, and expanded authentication provider options, while removing the local_sandbox execution mode and changing how user identity is resolved.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
12 FEATURES8 OF 12 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
HITLMiddleware for human-in-the-loop tool callsNEW70how completely this was documenteddepth30/40specificity20/30actionability20/3070 / 100
Adds HITLMiddleware for human-in-the-loop function interception, allowing agents to pause and await human approval before executing tool calls.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Names the middleware and its mechanism but no usage example.v1.9.0
0200
fail_closed config for output verifier toolsIMPROVED70how completely this was documenteddepth20/40specificity25/30actionability25/3070 / 100
Adds a fail_closed boolean config key to output_verifier_tools to control failure behavior when output verification cannot complete.
— Exact config key and target field, ready to set.v1.9.0
0302
CircuitBreakerMiddleware for tool fault toleranceNEW60how completely this was documenteddepth25/40specificity20/30actionability15/3060 / 100
Adds CircuitBreakerMiddleware for tool fault tolerance, automatically stopping calls to failing tools to prevent cascading failures.
— Explains mechanism and named class, lacks config detail.v1.9.0
thinner coverage below
0403
MLflow OTLP telemetry exporterNEW55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
Adds an MLflow OTLP telemetry exporter, enabling agent traces and metrics to be forwarded to an MLflow-compatible OTLP endpoint.
— Names the exporter and target but no config steps.v1.9.0
0500
Auth provider enhancements: redirects, preflight, WebSocket identityIMPROVED55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
Adds a redirect-based OAuth flow option for authentication providers, enables preflight authentication for applicable authentication providers, and adds WebSocket identity credential configuration via auth provider settings.
— Groups three thin auth additions, names `auth` settings.v1.9.0
0604
Plugin API exports for runtime context and HITL modelsNEW50how completely this was documenteddepth20/40specificity15/30actionability15/3050 / 100
Exports runtime context and interactive HITL models from the plugin API, and exports interactive prompt content models for third-party plugin authors.
— Names exported model types for plugin authors.v1.9.0
0705
Changed user identity resolution and LLM exposureBREAKING50how completely this was documenteddepth20/40specificity15/30actionability15/3050 / 100
The user_id parameter is no longer exposed to the LLM, and user identity resolution behavior has changed — existing setups relying on the previous resolution logic may require updates.
— Flags a behavior change requiring setup review, no exact migration steps.v1.9.0
0806
Opt-in provider hooks for IDs and timestampsNEW45how completely this was documenteddepth20/40specificity15/30actionability10/3045 / 100
Adds opt-in provider hooks for generated IDs and timestamps, letting integrators supply custom values for interaction prompt IDs and timestamps via providers.
— Describes purpose but no hook names or config.v1.9.0
0907
LangChain Runnable callback trackingNEW40how completely this was documenteddepth15/40specificity15/30actionability10/3040 / 100
Adds LangChain Runnable callback tracking for observability of LangChain-based workflows.
— Names integration but no mechanism or config detail.v1.9.0
1000
Configurable JobStore connection poolingIMPROVED35how completely this was documenteddepth10/40specificity15/30actionability10/3035 / 100
Makes JobStore connection pooling configurable.
— Bare statement with no parameters or defaults given.v1.9.0
1100
LangChain integrations split into optional extrasIMPROVED35how completely this was documenteddepth15/40specificity10/30actionability10/3035 / 100
Splits LangChain provider integrations into optional extras for more granular dependency installation.
— States the change but no extra names or install commands.v1.9.0
1208
Removal of local_sandbox execution modeBREAKING35how completely this was documenteddepth10/40specificity15/30actionability10/3035 / 100
The local_sandbox execution mode has been removed.
— Names removed mode but no migration guidance.v1.9.0
└──▷ BREAKING ON UPGRADE
!The local_sandbox execution mode has been removed.
!The user_id parameter is no longer exposed to the LLM.
!User identity resolution behavior has changed — existing setups relying on the previous resolution logic may require updates.
LiveKit Agents 1.8.1 introduces full-duplex speech models via a new DuplexModel abstraction and its first implementation, GPTLiveModel, alongside new STTSTTSpeech-to-Text: a class of APIs and models that transcribe spoken audio into written text, letting cyber tools process voice input or recorded audio without building their own acoustic models. and inference integrations, expanded plugin configuration, and event-loop blocking telemetry.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
LiveKit Agents builds realtime voice and video AI agents on LiveKit's WebRTC infrastructure, with turn detection, interruption handling, and telephony.
LiveKit Agents 1.8.1 introduces full-duplex speech models via a new DuplexModel abstraction and its first implementation, GPTLiveModel, alongside new STTSTTSpeech-to-Text: a class of APIs and models that transcribe spoken audio into written text, letting cyber tools process voice input or recorded audio without building their own acoustic models. and inference integrations, expanded plugin configuration, and event-loop blocking telemetry.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
6 FEATURES5 OF 6 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
DuplexModel for simultaneous speech I/ONEW88how completely this was documenteddepth30/40specificity28/30actionability30/3088 / 100
Adds a DuplexModel class enabling full-duplex speech models that can speak and listen simultaneously, with GPTLiveModel as the first implementation.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ GPTLiveModel accepts voice and responses_options (including model and instructions) so an agent can run a session using, for example, voice="marin" and responses_options={"model": "gpt-5.6-luna", "instructions": ...}.
Run a full-duplex voice session where the agent can speak and listen at the same time using OpenAI's GPT-Live model.
python
session = AgentSession(
llm=GPTLiveModel(
voice="marin",
responses_options={
"model": "gpt-5.6-luna",
"instructions": "Use tools when current information is required.",
},
),
)
ⓘSourcesRelease notes →1 RELEASE · 2026-09-10NOTES ↗RankWhy this ranktop features scored 75/70/70, +15 breaking change, +6 runnable example — 92.67How ranking works →
The Agent Development Kit is Google's code-first Python toolkit for building, evaluating, and deploying agents, with built-in tools, sessions, and deployment to Vertex AI.
ADK v2.9.0 adds automatic model failover, YAML-defined graph workflows, a LiveKitLiveKitAn open-source WebRTC server and SDK platform, maintained by LiveKit Inc., that handles real-time audio, video, and data routing so tools can embed low-latency media communication without building signalling infrastructure from scratch. voice/telephony runner, and MCP SDKMCP SDKA client library for Anthropic's Model Context Protocol, which standardises how AI models exchange context with external tools, letting cyber tools plug into LLM workflows without custom integration code. 2.x support, alongside several breaking changes to session, GCS and workflow-resume behavior.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
12 FEATURES10 OF 12 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
transfer_reason context on agent handoffsNEW75how completely this was documenteddepth25/40specificity20/30actionability30/3075 / 100
Adds a transfer_reason parameter to transfer_to_agent to carry context across agent handoffs and prevent delegation loops.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Prevent agent-transfer loops by supplying a transfer_reason so the receiving agent knows why it was handed the conversation.
python
from google.adk.tools import transfer_to_agent
await transfer_to_agent(
agent_name="billing_agent",
transfer_reason="User asked about invoice #4821, outside my scope.",
)
— Runnable example plus clear stated purpose.v2.9.0
0202
MCP SDK 2.x server supportBREAKING70how completely this was documenteddepth30/40specificity25/30actionability15/3070 / 100
Supports MCP SDK 2.x servers alongside existing 1.x servers, with 1.x remaining the default resolved version. Under MCP SDK 2.x, unknown fields a server sends on a CallToolResult or a tool declaration read via raw_mcp_tool are discarded during SDK validation before ADK sees them; only _meta (a declared field) still arrives, so servers relying on vendor extensions must move them under _meta.
— Names exact affected fields but gives no fix code example.v2.9.0
0303
Restricted local file access in GCS toolsBREAKING70how completely this was documenteddepth30/40specificity20/30actionability20/3070 / 100
GCS tools now restrict local file access to the directory named by local_file_root and refuse all local file access when local_file_root is absent.
— Names the exact config key and its before/after effect.v2.9.0
0404
YAML-defined graph workflows and resume behaviorBREAKING60how completely this was documenteddepth30/40specificity15/30actionability15/3060 / 100
Supports loading ADK 2.0 graph workflows from declarative YAML configuration. A workflow node that failed now re-executes when the workflow resumes, instead of replaying as though it had completed — node bodies with external side effects will repeat those side effects on every resume.
— Explains before/after resume behavior but no migration example.v2.9.0
0505
Strict session validation in InMemorySessionServiceBREAKING60how completely this was documenteddepth25/40specificity20/30actionability15/3060 / 100
InMemorySessionService now raises SessionNotFoundError when an event is appended to a session it does not hold, instead of silently discarding the event.
— Names the exception and trigger but gives no handling example.v2.9.0
thinner coverage below
0606
Pre-authorized artifact URLs in GcsArtifactServiceNEW50how completely this was documenteddepth15/40specificity20/30actionability15/3050 / 100
Adds get_authenticated_url and get_signed_url methods to GcsArtifactService for pre-authorized artifact access.
— Names both methods but no usage example or scope detail.v2.9.0
0707
Telemetry tracking and OTLP exportNEW50how completely this was documenteddepth20/40specificity20/30actionability10/3050 / 100
Adds skill-load tracking and script execution instrumentation to the telemetry subsystem, and exports logs over OTLP to telemetry.googleapis.com.
— Names the OTLP endpoint but no configuration steps shown.v2.9.0
0808
Automatic model failover with FallbackModelNEW45how completely this was documenteddepth20/40specificity15/30actionability10/3045 / 100
Adds a FallbackModel class for automatic failover to backup models when a primary model encounters errors.
— Names the class but no usage example or trigger conditions given.v2.9.0
0900
Evaluation set matching and reportingIMPROVED45how completely this was documenteddepth15/40specificity20/30actionability10/3045 / 100
Adds an ignore_args option to tool trajectory evaluation to relax argument matching in eval sets, and agent evaluations now print detailed results when all tests pass.
— Names the flag but reporting change is only described in prose.v2.9.0
1009
FunctionNode auto-inference in NodeToolNEW30how completely this was documenteddepth10/40specificity15/30actionability5/3030 / 100
Supports FunctionNode auto-inference in NodeTool.
— Single-line description with no usage shown.v2.9.0
1110
LiveKit voice and telephony runnerNEW25how completely this was documenteddepth10/40specificity10/30actionability5/3025 / 100
Adds a LiveKit runner integration for building voice and telephony agents.
— Bare description with no config or example.v2.9.0
1200
Hyphenated column names in BigQuery toolsIMPROVED25how completely this was documenteddepth10/40specificity10/30actionability5/3025 / 100
Supports hyphenated column names in BigQuery tools.
— Minimal description of the behavior change.v2.9.0
└──▷ BREAKING ON UPGRADE
!A workflow node that failed now re-executes when the workflow resumes, instead of replaying as though it had completed — node bodies with external side effects will repeat those side effects on every resume.
!GCS tools now restrict local file access to the directory named by local_file_root and refuse all local file access when local_file_root is absent.
!InMemorySessionService now raises SessionNotFoundError when an event is appended to a session it does not hold, instead of silently discarding the event.
!Under MCP SDK 2.x, unknown fields a server sends on a CallToolResult or a tool declaration read via raw_mcp_tool are discarded during SDK validation before ADK sees them; only _meta (a declared field) still arrives. Servers relying on vendor extensions must move them under _meta.
ⓘSourcesBlog post →1 RELEASE · 2026-09-10BLOG ↗RankWhy this ranktop features scored 80, +10 product media, +6 runnable example — 96How ranking works →
Arcade provides an actions runtime that lets AI agents securely call external services with authorization and governance controls.
Arcade's sole release this window introduces Execution Tool Logs, a per-run audit trail for every agent tool call, complete with input/output inspection, filtering, RBACRBACRole-Based Access Control, a permission model where rights are assigned to roles rather than individuals, letting tools enforce least-privilege by granting users only the access their role requires. controls, and a programmatic API.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
1 FEATURE1 OF 1 FEATUREmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Execution Tool Logs for agent tool callsNEW80how completely this was documenteddepth35/40specificity25/30actionability20/3080 / 100
Arcade adds Execution Tool Logs, which record every tool and toolkit (including version) that ran in a project, the end user it ran for, timestamps (created, started, finished), retry attempts, and success or failure outcome.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Executions can be filtered by tool name, toolkit, outcome, time range, user, or error text, and individual runs can be drilled into for a full timeline. Show inputs and Show outputs toggles on execution records let project admins inspect the exact arguments a tool received and what it returned, including full provider error messages. An API is provided to list, count, and retrieve individual executions programmatically, with admin permission required to access input and output payloads. RBAC restricts execution detail so project members see the list, status, timing, and errors, while inputs and outputs are limited to project admins in both the UI and API; records are stored encrypted at rest and in transit with a configurable retention window on Arcade Cloud.
Investigate a failing tool call by filtering the execution log to failures for a specific tool and user, then reviewing the exact inputs and provider error without digging through agent-side logs.
📍In Arcade Cloud, go to your project › Tool Executions, set the outcome filter to 'failure', filter by tool name and affected user, open the run, then toggle 'Show inputs' and 'Show outputs' to see the arguments and full error the provider returned.
— Names UI toggles, RBAC scope, retention and API but no exact endpoint pathslaunch-20260910-70f5950f
Together AI's headline ship this window is preemptible compute for GPU Clusters — discounted, sub-hourly-billed nodes controllable via CLI, API, and Kubernetes primitives — alongside custom LoRA adapter serving on dedicated endpoints, broader GPU cluster management (CLI/REST/SkyPilot, Slurm), a new IAM model, FLUX.2/FLUX Kontext image models, and a wave of agent-framework and coding-tool integrations.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Run, train, and serve open-source AI models on Together AI.
Together AI's headline ship this window is preemptible compute for GPU Clusters — discounted, sub-hourly-billed nodes controllable via CLI, API, and Kubernetes primitives — alongside custom LoRA adapter serving on dedicated endpoints, broader GPU cluster management (CLI/REST/SkyPilot, Slurm), a new IAM model, FLUX.2/FLUX Kontext image models, and a wave of agent-framework and coding-tool integrations.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
6 FEATURES6 OF 6 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Preemptible compute for GPU ClustersNEW98how completely this was documenteddepth38/40specificity30/30actionability30/3098 / 100
Together GPU Clusters gain a second, discounted compute type (public preview) billed at a flat 50% of the on-demand rate, sub-hourly and metered every one to two minutes, with no spot-market price fluctuation.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Set a preemptible GPU target with --num-preemptible-gpus on tg beta clusters create or tg beta clusters update, alongside the required standard num_gpus count; preemptible nodes join the same Kubernetes cluster as standard nodes, with no separate cluster type, and capacity can be added at create time or later. The Together API exposes desired_preemptible_gpus and allocated_preemptible_gpus fields to report requested vs. live capacity; preemption notices arrive as Kubernetes events with reason=TogetherPreempted and via node_lifecycle_events in the API, giving workloads up to terminationGracePeriodSeconds (max 300s) to checkpoint before removal. Nodes carry the label together.ai/compute-class=preemptible so workloads can target them via nodeSelector while keeping coordinators and serving replicas on standard nodes.
Spin up a new cluster with both standard and preemptible GPUs to run batch fine-tuning jobs at half the cost.
Pin interruptible worker pods to preemptible nodes while keeping coordinators on standard nodes, and give pods the full drain window to checkpoint on SIGTERM.
GPU cluster management via CLI, REST API, SkyPilot, and SlurmNEW55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
Adds GPU cluster management for training, fine-tuning, and large-scale AI workloads, including scale/operate operations via the Together CLI, REST API, and SkyPilot, plus Slurm configuration support to customize cluster settings for specific workload requirements.
— Names control surfaces but no concrete commands or config keysproduct docs
0303
FLUX.2 and FLUX Kontext image modelsNEW50how completely this was documenteddepth15/40specificity20/30actionability15/3050 / 100
Adds FLUX.2 and FLUX Kontext image-model support with dedicated quickstart guides.
— Names the models and points to quickstart guidesproduct docs
0404
Agent framework and coding tool integrationsNEW50how completely this was documenteddepth15/40specificity25/30actionability10/3050 / 100
Adds integrations for CrewAI, LangGraph, DSPy, PydanticAI, Agno, AutoGen (AG2), and Composio for building and orchestrating AI agent workflows, plus support for powering Cline, OpenCode, Qwen Code, and OpenClaw coding/autonomous agents with Together AI models.
— Lists many named integrations but no setup stepsproduct docs
0505
LoRA adapter serving on dedicated endpointsNEW48how completely this was documenteddepth18/40specificity15/30actionability15/3048 / 100
Dedicated endpoints now support uploading and serving custom LoRA adapters sourced from Hugging Face or S3.
— Names sources but no flags, limits, or steps givensnapshot-20260911
0606
IAM model for users, credentials, and resourcesNEW25how completely this was documenteddepth10/40specificity10/30actionability5/3025 / 100
Adds an IAM model covering users, credentials, and resource organization across the platform.
— Bare description with no scope or mechanismproduct docs
ArcBox runs containers, microVM agent sandboxes, and full Linux or macOS virtual machines on a Mac from one Rust daemon and CLI.
ArcBox's fleet-agent now captures Docker and VM runner output per job and writes it to dedicated, locatable log files.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
1 FEATURE1 OF 1 FEATUREmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Per-job log files in fleet-agentNEW55how completely this was documenteddepth25/40specificity15/30actionability15/3055 / 100
The fleet agent now captures output from both Docker and VM runners and writes each job's output to its own dedicated log file, surfacing the log paths so practitioners can locate and inspect per-job logs.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Describes mechanism and behaviour but no exact path or flag givenfleet-agent-v0.1.5
Darkbloom v0.9.2 adds Nemotron Lightning model serving with native reasoning and tool support, turns on Gemma QAT prefix caching by default (invalidating existing checkpoints), and introduces adaptive Gemma MTP alongside a zero-downtime assistant activation mechanism.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Private Inference Network on Idle Macs
Darkbloom v0.9.2 adds Nemotron Lightning model serving with native reasoning and tool support, turns on Gemma QAT prefix caching by default (invalidating existing checkpoints), and introduces adaptive Gemma MTP alongside a zero-downtime assistant activation mechanism.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
6 FEATURES6 OF 6 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Gemma QAT SSD prefix caching by defaultBREAKING80how completely this was documenteddepth35/40specificity30/30actionability15/3080 / 100
Enables Gemma QAT SSD prefix caching by default for exact gemma-4-26b-qat-4bit, with authenticated complete paged checkpoints preserving tenant, model, prompt, binary, metallib, and numerical-state identity checks; other Gemma artifacts and GPT-OSS caching remain opt-in.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Note: existing 0.9.1 Gemma QAT prefix checkpoints are not reused after upgrading to v0.9.2 — a new binary starts a new checkpoint identity, so the prefix cache is cold on first run.
— Names exact model ID and includes concrete upgrade migration notev0.9.2
0202
Zero-downtime assistant activationNEW80how completely this was documenteddepth35/40specificity25/30actionability20/3080 / 100
Downloads and verifies the optional assistant while the current engine serves, advertises reloading status, drains accepted work before swapping, and returns transient 503 slot_state refusals to racing requests during the swap.
— Names concrete status field and HTTP response code for integratorsv0.9.2
0303
Nemotron Lightning model servingNEW70how completely this was documenteddepth35/40specificity25/30actionability10/3070 / 100
Adds serving for Nemotron Lightning, admitting the three explicitly qualified registry/Hugging Face IDs on existing network and standalone paths, with native paged KV, complete encrypted prefix reuse, native activation/KV precision, FP32 persistent Mamba state, and adaptive MTP depth up to seven.
— Detailed mechanism but no exact model IDs or commands givenv0.9.2
0404
Adaptive Gemma MTP decode selectionNEW60how completely this was documenteddepth30/40specificity20/30actionability10/3060 / 100
Adds adaptive Gemma MTP that automatically resolves the catalog assistant for the QAT target and selects ordinary decode or one draft token from measured committed output and elapsed time, with support for temperature/top-p/top-k/min-p target-prefix sampling.
— Mechanism described but nothing a reader can directly invokev0.9.2
thinner coverage below
0505
Native reasoning and tool support for NemotronNEW50how completely this was documenteddepth25/40specificity15/30actionability10/3050 / 100
Adds native reasoning and tool support for Nemotron, separating reasoning before tool parsing and validating required/named calls before publishing through the existing encrypted response stream.
— Explains behavior but no configuration surface namedv0.9.2
0606
Immutable assistant revision support with R2 fallbackNEW45how completely this was documenteddepth20/40specificity15/30actionability10/3045 / 100
Adds catalog-declared immutable Hugging Face assistant revision support with checksum-verified R2 fallback and jittered fetch retries for assistant downloads.
— Names R2 fallback and checksum verification but little further detailv0.9.2
└──▷ BREAKING ON UPGRADE
!Existing 0.9.1 Gemma QAT prefix checkpoints are not reused after upgrading to v0.9.2 — a new binary starts a new checkpoint identity, so the prefix cache is cold on first run.
The Google Gen AI JavaScript SDK provides TypeScript and JavaScript interfaces to Gemini APIs and live model sessions.
The Google Gen AI JavaScript SDK added support for initializing a client directly from environment configuration.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
1 FEATURE1 OF 1 FEATUREmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Environment-based client initializationNEW45how completely this was documenteddepth15/40specificity15/30actionability15/3045 / 100
Adds from_environment support for copying environment configuration into a GoogleGenAI client instance, enabling client initialization directly from ambient environment variables.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Names the feature and mechanism but no usage examplev2.22.0
The Google Gen AI Python SDK provides Python interfaces to Gemini model generation, streaming, tools, and embeddings.
Google Gen AI Python SDK v2.23.0 adds dynamic turn completion for live sessions and a client construction helper that copies environment configuration.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
2 FEATURES2 OF 2 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Dynamic turn completion in live sessionsNEW45how completely this was documenteddepth15/40specificity15/30actionability15/3045 / 100
Adds interaction_status field support in live.pyreceive() to enable dynamic turn completion in live sessions.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Names exact field and file but no usage examplev2.23.0
0202
Environment-copying client constructionNEW40how completely this was documenteddepth10/40specificity15/30actionability15/3040 / 100
Adds from_environment support for copying client environment configuration.
— Names the method but no further mechanism detailv2.23.0
HyperFrames is an open-source HTML-to-video renderer that runs in AI-agent workflows.
HyperFrames v0.8.34 expands its Registry with new carousel blocks while significantly cutting CPU overhead in the Studio editor, both at idle and during composition scrubbing.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
3 FEATURES1 OF 3 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0100
Idle CPU reduction in paused Studio editorIMPROVED55how completely this was documenteddepth30/40specificity20/30actionability5/3055 / 100
Reduces idle CPU in a paused Studio editor from ~10% of a core to ~3%, dropping scheduled frames from 301 per second to 6.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Concrete before/after numbers but no user actionv0.8.34
0200
Faster scrubbing via cached ancestor measurementIMPROVED50how completely this was documenteddepth30/40specificity15/30actionability5/3050 / 100
Cuts main-thread work for scrubbing a large composition by roughly half by measuring each ancestor once per redraw instead of once per element.
— Explains mechanism and scope but no reader actionv0.8.34
0301
Image carousel blocks in RegistryNEW35how completely this was documenteddepth15/40specificity15/30actionability5/3035 / 100
Adds 25 image carousel blocks to the Registry.
— Names a count and area but no further mechanismv0.8.34
The Anthropic TypeScript SDK provides typed JavaScript and TypeScript clients for Claude APIs, streaming, and tool use.
The Anthropic TypeScript SDK added a new user profiles beta field, a new error code for the web_fetch tool, and support for mounting public GitHub repos in Managed Agents sessions without authentication.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
3 FEATURES3 OF 3 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Public GitHub repo mounting in Managed AgentsNEW60how completely this was documenteddepth20/40specificity22/30actionability18/3060 / 100
Managed Agents sessions can now mount public GitHub repositories without providing an authorization_token.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Names the field removed and the feature area clearlysdk-v0.125.0
thinner coverage below
0202
User profiles beta field additionNEW52how completely this was documenteddepth15/40specificity22/30actionability15/3052 / 100
Adds the user-profiles-2026-09-04 beta value and an external_user_details field to the user profiles API.
— Names the beta value and field but no usage detailsdk-v0.125.0
0303
New error code for web_fetch toolNEW43how completely this was documenteddepth10/40specificity18/30actionability15/3043 / 100
Adds a content_too_large error code to the web_fetch tool.
— Bare addition, no handling guidance givensdk-v0.125.0
Anthropic Python SDK v1.5.0 adds a direct way to convert Claude responses back into request parameters, accepts tool objects natively in core methods, and extends Managed Agents and error handling.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
The Anthropic Python SDK provides typed Python clients for Claude APIs, streaming responses, and tool use.
Anthropic Python SDK v1.5.0 adds a direct way to convert Claude responses back into request parameters, accepts tool objects natively in core methods, and extends Managed Agents and error handling.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
6 FEATURES5 OF 6 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Message-to-param conversion methodsNEW75how completely this was documenteddepth25/40specificity20/30actionability30/3075 / 100
Message.to_param() and BetaMessage.to_param() convert message objects directly into API-compatible parameter dicts, letting a completed assistant message be reused as an entry in a messages list for the next turn without manually rebuilding the dict.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Turn a completed assistant message back into a request parameter for multi-turn conversations without manually rebuilding dicts.
python
import anthropic
client = anthropic.Anthropic()
# First turn
message = client.messages.create(
model="claude-opus-5",
max_tokens=1024,
messages=[{"role": "user", "content": "What is 2+2?"}],
)
# Convert response to param and continue the conversation
conversation = [
{"role": "user", "content": "What is 2+2?"},
message.to_param(),
{"role": "user", "content": "Now multiply that by 10."},
]
follow_up = client.messages.create(
model="claude-opus-5",
max_tokens=1024,
messages=conversation,
)
print(follow_up.content)
— Includes a runnable multi-turn example using the new method.v1.5.0
0200
Tool objects accepted in core API methodsIMPROVED60how completely this was documenteddepth20/40specificity25/30actionability15/3060 / 100
messages.create, parse, stream, and count_tokens now accept tool objects directly instead of requiring manual dict construction.
— Names all four affected methods but no usage example given.v1.5.0
thinner coverage below
0302
Managed Agents enhancementsNEW55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
Managed Agents gain auto mode tool permissions and support for mounting public GitHub repositories in a session without an authorization_token.
— Names both additions verbatim but no mechanism detail.v1.5.0
0403
New credentials error classesNEW45how completely this was documenteddepth15/40specificity20/30actionability10/3045 / 100
Adds CredentialsError and IdentityTokenFileError exception classes for more granular credentials error handling.
— Names the two new classes but no usage detail.v1.5.0
0504
User profiles beta with external_user_details fieldNEW40how completely this was documenteddepth10/40specificity20/30actionability10/3040 / 100
Adds the user-profiles-2026-09-04 beta value and an external_user_details field to user profiles.
— Names beta value and field but no usage shown.v1.5.0
0605
content_too_large error code for web_fetchNEW35how completely this was documenteddepth10/40specificity15/30actionability10/3035 / 100
Adds a content_too_large error code to the web_fetch tool.
— Bare naming of a single error code with no context.v1.5.0
The OpenAI TypeScript SDK provides typed JavaScript and TypeScript clients for OpenAI APIs and streaming responses.
OpenAI TypeScript SDK v7.14.0 adds support for the Live API.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
1 FEATURE1 OF 1 FEATUREmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Live API supportNEW25how completely this was documenteddepth10/40specificity10/30actionability5/3025 / 100
The SDK adds the Live API to the OpenAI TypeScript/JavaScript client.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Only names the feature, no mechanism or usage shownv7.14.0
The OpenAI Python SDK provides typed Python clients for OpenAI model APIs, streaming, and asynchronous requests.
OpenAI Python SDK v3.12.0 adds support for the Live API.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
1 FEATURE1 OF 1 FEATUREmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Live API supportNEW23how completely this was documenteddepth8/40specificity10/30actionability5/3023 / 100
The SDK adds support for the Live API, extending the typed Python client to this new API surface.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— No mechanism, endpoints, or usage details given beyond the name.v3.12.0
ONNX Runtime 1.30.0 adds Go bindings for the C API, a batch of CUDACUDAA parallel computing platform and programming model made by NVIDIA that lets software run code directly on NVIDIA GPUs, giving cyber tools access to massive parallelism for workloads like cracking or ML inference. operators for generative-AI serving (paged attention, DeepSeek Engram, GatedDeltaNet, speculative XQA), FP4FP4A 4-bit floating-point numeric format used in machine-learning inference to compress model weights, reducing memory and compute cost when running AI models at scale./FP8FP8An 8-bit floating-point numeric format used in AI accelerator hardware to store and compute model weights at reduced precision, cutting memory bandwidth and compute cost versus FP16 or FP32. QMoE kernel support, LoRA adapter loading, new CPU SIMD kernels, and several build-default changes affecting CUDA quantized-GEMM and CPU FP16 execution.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
ONNX Runtime runs machine-learning models across hardware and operating systems with an open-source inference engine.
ONNX Runtime 1.30.0 adds Go bindings for the C API, a batch of CUDACUDAA parallel computing platform and programming model made by NVIDIA that lets software run code directly on NVIDIA GPUs, giving cyber tools access to massive parallelism for workloads like cracking or ML inference. operators for generative-AI serving (paged attention, DeepSeek Engram, GatedDeltaNet, speculative XQA), FP4FP4A 4-bit floating-point numeric format used in machine-learning inference to compress model weights, reducing memory and compute cost when running AI models at scale./FP8FP8An 8-bit floating-point numeric format used in AI accelerator hardware to store and compute model weights at reduced precision, cutting memory bandwidth and compute cost versus FP16 or FP32. QMoE kernel support, LoRA adapter loading, new CPU SIMD kernels, and several build-default changes affecting CUDA quantized-GEMM and CPU FP16 execution.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
10 FEATURES7 OF 10 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
FP4/FP8 QMoE kernel supportNEW90how completely this was documenteddepth30/40specificity30/30actionability30/3090 / 100
Adds an opt-in FP8 DeepGEMM MoE decode path for fixed-shape QMoE workloads on Hopper GPUs, enabled via ORT_QMOE_FP4_DEEPGEMM=1 (default off, unavailable on Windows); FP4 QMoE kernels are now enabled by default in CUDA builds, with source builds able to opt out via -Donnxruntime_USE_FP4_QMOE=OFF.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Enable the FP8 DeepGEMM MoE decode path on a Hopper GPU to accelerate fixed-shape QMoE workloads without rebuilding from source.
$ ORT_QMOE_FP4_DEEPGEMM=1 python run_inference.py
— Exact env var, build flag, GPU target, and runnable commandv1.30.0
0202
CUDA fpA-intB GEMM default kernel set changeBREAKING85how completely this was documenteddepth30/40specificity30/30actionability25/3085 / 100
CUDA fpA-intB builds now default to a compact kernel set (FP16 activations, INT4/INT8 weights, scale-only quantization, block_size=32); set -Donnxruntime_USE_FPA_INTB_GEMM_FULL=ON at source build time to retain BF16, zero-point, bias, larger-block-size, and native Hopper variants.
— Exact build flag and default behavior change statedv1.30.0
0303
CUDA generative-AI attention operatorsNEW70how completely this was documenteddepth25/40specificity30/30actionability15/3070 / 100
Adds the VarlenCausalConvWithState CUDA operator for continuous batching with compact variable-length causal-convolution state updates; adds an is_causal attribute to PagedAttention plus INT4 paged KV caches with per-channel scales on CUDA; adds a compact GatedDeltaNet operator with BFloat16 support on CUDA; and extends paged XQA with speculative decoding, query-to-KV head group size 6, head size 256 with FP16 caches, and native block tables for 128-token pages.
— Names several CUDA ops and parameters, no runnable examplev1.30.0
thinner coverage below
0404
Custom kernel and memory-import APIsNEW55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
Adds KernelContext::GetPreallocatedOutput to access preallocated outputs from custom kernels, and extends memory importing with host-pointer support.
Baseten deploys and serves machine learning models as autoscaling inference endpoints, with a model API catalog and dedicated GPU deployments.
Baseten added DeepSeek V4.1 Flash to its Model APIs catalog and introduced regional deployments for pinning replicas to a specific region.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
2 FEATURES2 OF 2 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Regional deployments for replica placementNEW67how completely this was documenteddepth25/40specificity22/30actionability20/3067 / 100
Deployments can now be restricted to run every replica in a specific Baseten region, selectable via the dashboard, the Baseten CLI, or the Management API, to meet compliance or data residency requirements or reduce latency.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
ⓘSourcesRelease page →5 RELEASES · 2026-08-22 → 2026-09-04NOTES ↗RankWhy this ranktop features scored 95/90/90, +15 breaking change, +6 runnable example — 112.67How ranking works →
Tinker provides a managed API for training and sampling language models while users control their training algorithms and data.
Tinker's SDK and Cookbook releases this window added authentication and audit-log infrastructure (browser-based tinker auth login/logout, API key verification, expanded audit log with an event_details field), a weight-copying API, provenance tracking for training data, and new renderer support for GLM-5.3 and Qwen3.8, alongside a forecasting RL recipe and the removal of subprocess-isolated sampling.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
12 FEATURES11 OF 12 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
tinker auth CLI: login, API-key flag, logoutNEW95how completely this was documenteddepth35/40specificity30/30actionability30/3095 / 100
Adds tinker auth login to authenticate through the browser via WorkOS device authorization, mint an API key, and store it as the default credential; tinker auth login --api-key to store a manually pasted API key, which now verifies the key with the server before storing it and records it under its real key id with the server-held name and note; and tinker auth logout to remove the default credential, deleting a browser-minted key server-side or removing a manually entered one locally.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Authenticate to Tinker from a new workstation without managing API keys manually — the browser flow mints and stores the credential automatically.
$ tinker auth login
Authenticate in a headless CI environment by supplying a pre-generated API key directly.
$ tinker auth login --api-key
Authenticate with an API key and have it validated server-side before storage.
Weight copying via copy_weights()NEW90how completely this was documenteddepth30/40specificity30/30actionability30/3090 / 100
Adds ServiceClient.copy_weights() to copy weights into the client's own project, returning the tinker path of the copy; accepts optional ttl_seconds and weights_access_token parameters.
Copy weights from another account's source into your project with a 24-hour expiry, for use in a short-lived experiment.
— Full signature, params, and runnable example given.SDK 0.26.2
0303
Audit log expansion and field deprecationBREAKING90how completely this was documenteddepth35/40specificity30/30actionability25/3090 / 100
Expands RestClient.get_audit_log() with an event_type filter accepting 'projects', 'teams', or 'organizations' to cover administrative activity alongside checkpoints, and adds an event_details field to AuditLogEntry carrying the full record — actor, resource, and action with its own parameters — for every event type. The model_id, tinker_path, and purpose fields on AuditLogEntry are now deprecated: they remain set for checkpoint events but are unset for the new projects, teams, and organizations event types, so code reading them unconditionally will see missing values.
Audit all team-level administrative actions in your org to investigate a permission change.
python
entries = rest_client.get_audit_log(event_type='teams')
for entry in entries:
print(entry.event_details)
— Names filter values, new field, deprecated fields, and impact.SDK 0.26.2
0404
Subprocess-isolated sampling removedBREAKING75how completely this was documenteddepth30/40specificity30/30actionability15/3075 / 100
Subprocess-isolated sampling has been removed: TINKER_SUBPROCESS_SAMPLING=1 no longer has any effect, and passing subprocess_sampling= to a SamplingClient is now an error — sample() and compute_logprobs() always run in the calling process.
— Clear before/after and named surfaces, no migration example.SDK 0.27.1
0505
Datum provenance trackingNEW75how completely this was documenteddepth30/40specificity30/30actionability15/3075 / 100
Adds datum.with_provenance(model_input=..., loss_fn_inputs=...) to attach PromptProvenanceSpan / SampledProvenanceSpan runs that tile input tokens and loss rows, with partition validation at the call site.
— Named signature and mechanism, but no runnable example.SDK 0.26.0
0606
get_server_capabilities() trainable/sampleable fieldsNEW75how completely this was documenteddepth25/40specificity25/30actionability25/3075 / 100
Adds trainable and sampleable fields to SupportedModel, exposed via get_server_capabilities() to identify which models can host training runs and which have sampling capacity.
Before launching a training run, check which models support training vs. sampling to pick the right target.
python
capabilities = get_server_capabilities()
trainable_models = [m for m in capabilities if m.trainable]
sampleable_models = [m for m in capabilities if m.sampleable]
— Named fields and method with a working code example.SDK 0.26.0
0707
Qwen3.8 renderer and reasoning-turn enforcementBREAKING70how completely this was documenteddepth30/40specificity25/30actionability15/3070 / 100
Adds a qwen3_8 renderer family with tool calling, tool-declaration, and parsing support. Additionally, a renderer with reasoning disabled now refuses a turn that carries reasoning instead of silently dropping it — workflows that passed reasoning-bearing turns to a reasoning-disabled renderer will now fail explicitly.
— Names renderer and behavior change, no example.Cookbook 0.5.5
0808
get_console_url() on client and checkpoint responsesNEW65how completely this was documenteddepth25/40specificity25/30actionability15/3065 / 100
Adds get_console_url() on ServiceClient to return the current session's Console page URL, and on the responses from save_state() and save_weights_for_sampler() to return the saved checkpoint's Console page URL.
— Named methods and behavior, but no usage example.SDK 0.27.1
0909
GLM-5.3 renderer supportNEW60how completely this was documenteddepth25/40specificity25/30actionability10/3060 / 100
Adds glm5_3 renderer family with tool calling, reasoning, and <|observation|> / <|user|> turn terminator support for GLM-5.3 models, plus model metadata and hyperparameter lookup entries for GLM-5.3.
— Named renderer features but no usage example or command.Cookbook 0.5.6
thinner coverage below
1010
Forecasting RL recipe on Prophet ArenaNEW55how completely this was documenteddepth25/40specificity20/30actionability10/3055 / 100
Adds a forecasting recipe using RL on binary prediction markets from the Prophet Arena subset, with a chronological train/test split and a Brier-score reward.
— Describes mechanism but no runnable recipe reference.Cookbook 0.5.6
1100
loss_fn_config accepts string valuesIMPROVED45how completely this was documenteddepth15/40specificity20/30actionability10/3045 / 100
Allows loss_fn_config to accept string values alongside floats on forward(), forward_backward(), and their async variants.
— Names methods and change but no example or rationale.SDK 0.26.2
1211
Inkling-Small hyperparameter lookup entryNEW25how completely this was documenteddepth10/40specificity10/30actionability5/3025 / 100
Adds Inkling-Small to the hyperparameter lookup tables.
— Bare one-line addition with no further detail.Cookbook 0.5.6
└──▷ BREAKING ON UPGRADE
!Subprocess-isolated sampling has been removed: TINKER_SUBPROCESS_SAMPLING=1 no longer has any effect, and passing subprocess_sampling= to a SamplingClient is now an error — sample() and compute_logprobs() always run in the calling process.
!The model_id, tinker_path, and purpose fields on AuditLogEntry are deprecated: they remain set for checkpoint events but are unset for new projects, teams, and organizations event types — code that reads these fields unconditionally will see missing values for those event types.
!A renderer with reasoning disabled now refuses a turn that carries reasoning instead of silently dropping it — workflows that passed reasoning-bearing turns to a reasoning-disabled renderer will now fail explicitly.
OpenRouter is a routing gateway that exposes hundreds of models from many providers behind one OpenAI-compatible API with failover and unified billing.
OpenRouter's only release this window expanded team analytics: a new Analytics API alongside dashboard upgrades for spend visibility and log drill-down.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
1 FEATURE1 OF 1 FEATUREmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Team analytics dashboard and Analytics APINEW47how completely this was documenteddepth20/40specificity15/30actionability12/3047 / 100
OpenRouter adds an Analytics API so teams can query spend and usage data from the terminal, per-model spend breakdowns scoped to team-level usage in the analytics dashboard, saveable analytics charts so practitioners can preserve views they rebuild repeatedly, and click-through from any chart bar directly into the underlying request logs.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Describes four additions but no endpoint paths or exact API shape given.snapshot-20260911
whisper.cpp is a C/C++ port of OpenAI's Whisper speech recognition model, running transcription on CPU and GPU with no Python runtime.
whisper.cpp v1.9.4 adds language detection in the server API, NPUNPUA Network Processing Unit, a specialized processor designed to handle packet forwarding, classification, and manipulation at wire speed, offloading that work from the main CPU so cyber tools can inspect traffic without dropping packets. encoder offload via a new VitisAIVitisAIAMD's inference deployment framework for running neural network models on AMD/Xilinx FPGAs and AI accelerators, giving cyber tools a hardware-accelerated path for high-throughput ML inference. plugin, and Windows on ARM release binaries, alongside callback and determinism fixes.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
5 FEATURES2 OF 5 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0100
Detected language in server /detect responseIMPROVED55how completely this was documenteddepth15/40specificity20/30actionability20/3055 / 100
The server's /detect endpoint response now includes the detected language in its JSON payload.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Names exact endpoint and field change but no schema examplev1.9.4
0201
VitisAI NPU encoder offload pluginNEW55how completely this was documenteddepth15/40specificity25/30actionability15/3055 / 100
Adds a VitisAI plugin (vitisai) enabling encoder offload to AMD Ryzen AI NPUs.
— Names plugin and hardware target, no setup steps givenv1.9.4
0300
Earlier encoder-begin callback during language auto-detectionIMPROVED55how completely this was documenteddepth20/40specificity20/30actionability15/3055 / 100
encoder_begin_callback is now called before language auto-detection runs, enabling callback-based control earlier in the pipeline than before.
— Names the callback and the behavioural fix preciselyv1.9.4
0402
Windows on ARM release binariesNEW35how completely this was documenteddepth10/40specificity15/30actionability10/3035 / 100
Adds Windows on ARM (WoA) support to the release CI job, producing WoA binaries.
— States what was added but no download or build detailv1.9.4
0500
Deterministic decoder re-seeding across callsIMPROVED35how completely this was documenteddepth15/40specificity10/30actionability10/3035 / 100
Decoder 0 is now re-seeded between calls for more deterministic repeated inference.
— Explains mechanism briefly, no numbers or reproduction stepsv1.9.4
The vMLX server runs compressed MLX models on Apple Silicon with disk caching, paged memory, continuous batching, and hybrid SSM scheduling.
vMLX 1.6.57 adds Spark-X2.5 model integration and improves handling of mflux image generation exports.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
2 FEATURES1 OF 2 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0100
mflux image generation export handlingIMPROVED55how completely this was documenteddepth25/40specificity20/30actionability10/3055 / 100
Improves local-folder discovery and launch validation for mflux image generation/editing exports, including quantization metadata, progress, cancellation, and output history.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Lists concrete behaviors but no exact commands or config keysv1.6.57
0201
Spark-X2.5 model integrationNEW25how completely this was documenteddepth10/40specificity10/30actionability5/3025 / 100
vMLX adds integration with the Spark-X2.5 model.
— Names the model but no mechanism or usage detailv1.6.57
ⓘSourcesRelease page →1 RELEASE · seen 2026-09-11NOTES ↗RankWhy this ranktop features scored 92/85/80, +15 breaking change, +6 runnable example — 106.67How ranking works →
Pinecone is a managed vector database that stores and queries embeddings for AI applications.
Pinecone shipped a new 2026-07 API version that restructures REST index responses and adds document-schema indexes via SchemaBuilder, introduced index/namespace-level metadata indexing controls, launched the Nexus knowledge-curation product, and rounded out its Admin API with full project, organization, API key, service account and document-listing endpoints.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
7 FEATURES7 OF 7 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Document-schema indexes via SchemaBuilderNEW92how completely this was documenteddepth32/40specificity30/30actionability30/3092 / 100
Index creation now supports a document schema built with SchemaBuilder in the Python and JavaScript SDKs and REST, using a schema.fields structure with typed fields such as dense_vector, sparse_vector, and string with full_text_search, alongside a deployment object specifying deployment_type: managed.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ A document index cannot accept metadata fields in its schema at creation; metadata indexing for it must instead be configured at the namespace level.
Create a serverless document-schema index with a dense-vector embedding field using the 2026-07 API version.
Build a document schema with a dense-vector field in Python and create a managed serverless index.
python
from pinecone import Pinecone, SchemaBuilder
pc = Pinecone(api_key='<your-api-key>')
schema = (
SchemaBuilder()
.add_dense_vector_field('embedding')
.build()
)
pc.indexes.create(
deployment={'deployment_type': 'managed'},
schema=schema,
)
— Runnable curl and Python examples with exact field and type names.product docs
0202
Project, organization, API key and service account management APIsNEW85how completely this was documenteddepth25/40specificity30/30actionability30/3085 / 100
Pinecone added full lifecycle endpoints across several admin surfaces: List projects, Create a new project, Get project details, Delete a project, and Update a project (including setting maximum Pod count and customer-managed encryption key/CMEK); List organizations, Get organization details, Delete an organization, and Update an organization; List API keys and Delete an API key; List service accounts, Create a service account, Get service account details, Delete a service account, Update a service account, and Rotate a service account's OAuth client secret; and a role-binding endpoint for granting scoped roles to service accounts.
Grant a service account a scoped role at the project level immediately after creating it, using the new role-binding endpoint.
Rotate a compromised service account secret to immediately revoke old tokens and retrieve the new secret — returned only once.
$ curl -X POST 'https://api.pinecone.io/organizations/<org-id>/service-accounts/<service-account-id>/rotate-secret' \
-H 'Api-Key: <api-key>'
— Every endpoint named, with runnable curl examples for role binding and secret rotation.snapshot-20260911
0303
API version 2026-07 restructures index REST responsesBREAKING80how completely this was documenteddepth33/40specificity27/30actionability20/3080 / 100
With X-Pinecone-Api-Version: 2026-07, the top-level deployment_type and read_capacity fields are removed from REST index responses and replaced by a top-level deployment object, and dense-vector metric now lives on the schema's dense-vector field rather than at the top level. SDKs return both old and new shapes for backward compatibility, but callers parsing the old top-level fields directly will break.
— Names exact fields and header but gives no migration path beyond SDK note.product docs
0404
Paginated document listing in namespacesNEW75how completely this was documenteddepth25/40specificity25/30actionability25/3075 / 100
A new List documents endpoint returns up to 100 documents per page within a namespace, with an optional prefix parameter to filter by ID prefix and a pagination token for iterating through results.
Paginate through all documents in a namespace whose IDs share a common prefix — useful for auditing or bulk-processing a logical subset of records.
$ curl -X GET 'https://api.pinecone.io/namespaces/<namespace>/documents?prefix=user-&pagination=<token>' \
-H 'Api-Key: <api-key>'
— Runnable curl example with named parameters and page limit.snapshot-20260911
thinner coverage below
0505
Metadata indexing controls at index and namespace levelNEW53how completely this was documenteddepth28/40specificity15/30actionability10/3053 / 100
Metadata indexing can now be limited at index creation or later using a per-field filterable setting. Namespace-level metadata indexing rules can override index-level rules, letting each namespace control which fields are indexed for filtering, while index-level rules act as the default for namespaces without their own configuration.
— Explains override behavior but gives no config example.product docs
0606
Nexus knowledge-curation productNEW45how completely this was documenteddepth20/40specificity15/30actionability10/3045 / 100
Pinecone launched Nexus, a new knowledge-curation and retrieval product with manifest-driven context design, letting practitioners define custom artifact and edge types in the console to control how sources become queryable knowledge.
— Names the product and concept but no concrete usage steps.product docs
0707
Inference and Admin API version 2026-07NEW30how completely this was documenteddepth10/40specificity15/30actionability5/3030 / 100
Pinecone published inference_2026-07.oas and admin_2026-07.oas API specifications, introducing a new 2026-07 versioned release of the Inference and Admin APIs.
— Only names the spec files, no behavior described.product docs
└──▷ BREAKING ON UPGRADE
!With X-Pinecone-Api-Version: 2026-07, REST responses remove the top-level deployment_type and read_capacity fields, replacing them with a top-level deployment object; callers parsing those top-level fields directly will break. (SDKs return both old and new shapes for backward compatibility.)
engrim v1.4.0 adds a live, WALWALWrite-Ahead Log: a durability mechanism in databases where changes are recorded to an append-only log before being applied to main storage, giving tools crash recovery and atomic transactions without full-page writes on every operation.-safe SQLite backup command, a session/CI teardown command for stale resume pointers, scoped per-project record-count introspection, and a reference multi-agent CI architecture using GitHub Agentic Workflows.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
The Universal Cross-Model Episodic Memory Standard. Local-first, project-scoped SQLite memory engine for Google Antigravity, Claude Code, Cursor, Windsurf, and Codex.
engrim v1.4.0 adds a live, WALWALWrite-Ahead Log: a durability mechanism in databases where changes are recorded to an append-only log before being applied to main storage, giving tools crash recovery and atomic transactions without full-page writes on every operation.-safe SQLite backup command, a session/CI teardown command for stale resume pointers, scoped per-project record-count introspection, and a reference multi-agent CI architecture using GitHub Agentic Workflows.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
4 FEATURES4 OF 4 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Live SQLite backup commandNEW88how completely this was documenteddepth30/40specificity28/30actionability30/3088 / 100
engrim backup COPY.db [--force] [--json] creates consistent online snapshots of active SQLite stores via Connection.backup(), guaranteeing zero WAL tearing even while agents are actively writing to the store.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Take a safe, consistent snapshot of the live memory store before a destructive operation, even while agents are mid-write.
$ engrim backup ~/.engrim/memory.backup.db --json
— Exact flags and underlying mechanism (Connection.backup) both given.v1.4.0
0202
Resume-pointer teardown commandNEW88how completely this was documenteddepth30/40specificity28/30actionability30/3088 / 100
engrim retire [-p P | --all] [--dry-run] [--json] monotonically closes active resume-pointer records when a session or CI job concludes, preventing stale [▶ RESUME HERE] pointers from accumulating across runs.
At the end of a CI job, close any stale resume-pointers for the current project without committing changes — useful for validating teardown in a dry run first.
$ engrim retire --dry-run--json
— Full flag set, exact marker string, and dry-run behavior specified.v1.4.0
0303
Scoped project record-count introspectionNEW76how completely this was documenteddepth20/40specificity26/30actionability30/3076 / 100
engrim project [-p P | --global | --all] [--json] returns single-project record counts, complementing the existing engrim projects whole-store listing.
Inspect memory record counts scoped to a specific project by tag, in machine-readable form for dashboards or health checks.
$ engrim project -p my-app --json
— Runnable command with flags but little mechanism detail.v1.4.0
0404
GitHub Agentic Workflows reference implementationNEW62how completely this was documenteddepth25/40specificity22/30actionability15/3062 / 100
A full multi-agent CI reference implementation at examples/gh-aw demonstrates engrim inside GitHub Agentic Workflows, covering ephemeral runner memory sync, safe SQLite backups, and multi-agent issue triage.
— Names the example path and scope but gives no runnable command.v1.4.0
ⓘSourcesRelease page →1 RELEASE · 2026-08-20NOTES ↗RankWhy this ranktop features scored 90/70/50, +15 breaking change, +6 runnable example — 91How ranking works →
LangSmith added monthly trace limits scoped to projects and users, improved code evaluator typing and image rendering in dataset tables, and began deprecating legacy feedback-formula and dataset-comparison endpoints in favor of newer APIs.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
LangSmith provides tracing, evaluation, and deployment tools for LLM applications.
LangSmith added monthly trace limits scoped to projects and users, improved code evaluator typing and image rendering in dataset tables, and began deprecating legacy feedback-formula and dataset-comparison endpoints in favor of newer APIs.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
5 FEATURES3 OF 5 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Batch add runs and threads to annotation queuesNEW90how completely this was documenteddepth30/40specificity30/30actionability30/3090 / 100
A single request can now add a mixed batch of runs and conversation threads to an annotation queue for human review, via POST https://<langsmith-host>/annotation-queues/<queue-id>/items with an items array of objects carrying item_type (RUN or THREAD), run_id, thread_id, and session_id.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Add a batch of runs and conversation threads to an annotation queue in one request for mixed human review.
— Runnable curl example with full endpoint and payload shape.snapshot-20260911
0202
Legacy dataset comparison SDK helpers removedBREAKING70how completely this was documenteddepth25/40specificity25/30actionability20/3070 / 100
Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs; code using the generated SDK methods for dataset comparison must be migrated to POST /v2/datasets/<id>/experiment-runs.
— Named replacement endpoint provided for migration.snapshot-20260911
thinner coverage below
0300
PEP 604 union return types in code evaluatorsIMPROVED50how completely this was documenteddepth15/40specificity20/30actionability15/3050 / 100
Code evaluators annotated with PEP 604 union return types (e.g., -> dict | None) are now accepted on upload.
— Concrete syntax example given but no CLI or endpoint.snapshot-20260911
0400
Base64 image thumbnails in dataset tablesIMPROVED40how completely this was documenteddepth25/40specificity10/30actionability5/3040 / 100
Dataset and experiment tables now detect and render base64 images as small thumbnails with a delayed larger preview instead of showing raw encoded strings.
Monthly trace limits per project and userNEW30how completely this was documenteddepth15/40specificity10/30actionability5/3030 / 100
Enforces user-defined monthly trace limits scoped to individual projects and users, rejecting new traces once the configured limit is exceeded.
— No config key or API surface named for setting the limit.snapshot-20260911
└──▷ BREAKING ON UPGRADE
!The legacy feedback formula endpoints (POST/GET /feedback/formulas and GET/PUT/DELETE /feedback/formulas/{feedback_formula_id}) are deprecated and scheduled for removal on 2026-08-20; migrate existing feedback formulas to composite evaluators before that date.
!Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs; code using the generated SDK methods for dataset comparison must be migrated to POST /v2/datasets/<id>/experiment-runs.
Langfuse provides tracing, evaluation, and monitoring for LLM applications.
Langfuse v4.34.0 introduces user-triggered session revocation and a new AI gateway Rust service with an organization control plane, alongside score filtering across the Scores and Experiments views, a redesigned session timeline, and marks legacy ingestion endpoints as deprecated.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
7 FEATURES4 OF 7 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Legacy ingestion API deprecation noticeDEPRECATED65how completely this was documenteddepth25/40specificity25/30actionability15/3065 / 100
Adds a _deprecation field attached to successful legacy ingestion API write responses, and documents that legacy endpoints delay data by approximately 10 minutes.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
— Names the exact field and a concrete delay figure.v4.34.0
thinner coverage below
0200
OTEL trace IDs on tRPC errorsIMPROVED50how completely this was documenteddepth20/40specificity20/30actionability10/3050 / 100
Attaches OTEL trace IDs to tRPC error responses to aid distributed tracing of server errors.
— Names the mechanism but not how to consume the trace ID.v4.34.0
0302
Score filtering in Scores and Experiments viewsNEW45how completely this was documenteddepth15/40specificity15/30actionability15/3045 / 100
Adds score filters and a search bar to the Scores view, and adds score filter display in the Experiments search bar.
— Names both surfaces but no filter syntax or fields.v4.34.0
0403
User-triggered session revocationNEW35how completely this was documenteddepth15/40specificity10/30actionability10/3035 / 100
Adds user-triggered session revocation so users can invalidate their own active sessions.
— States the capability but no mechanism or UI path.v4.34.0
0504
AI gateway Rust service with control planeNEW35how completely this was documenteddepth15/40specificity15/30actionability5/3035 / 100
Bootstraps a new AI gateway Rust service with an organization control plane.
— Names the service but not how to use it yet.v4.34.0
0600
Redesigned session timeline UIIMPROVED30how completely this was documenteddepth10/40specificity10/30actionability10/3030 / 100
Redesigns the session timeline UI in the Sessions view.
— Names the view but no detail on what changed visually.v4.34.0
0700
Unified Users filter UIIMPROVED30how completely this was documenteddepth10/40specificity10/30actionability10/3030 / 100
Unifies the Users filter experience to match the Traces filter UI.
— Names both areas but no specifics on filter fields.v4.34.0
Arize Phoenix is an open-source platform that monitors LLM application traces and evaluates their outputs.
Arize Phoenix added a new completeness evaluator to its evals library and let prompt versions carry arbitrary metadata via the JS client.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
2 FEATURES1 OF 2 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0100
Metadata parameter on promptVersion()IMPROVED55how completely this was documenteddepth15/40specificity20/30actionability20/3055 / 100
Phoenix client's promptVersion() now accepts a metadata parameter so prompt versions can carry arbitrary JSON metadata.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
ⓘSourcesRelease page →1 RELEASE · 2026-09-01NOTES ↗RankWhy this ranktop features scored 93/83/81, +15 breaking change, +6 runnable example — 106.67How ranking works →
Braintrust provides evaluation, tracing, and improvement workflows for AI applications.
Braintrust shipped a wave of AI-driven improvement tools — Patterns for scheduled pattern detection, a single-trace Debugger, and a server-side Loop runtime with scheduled automations — alongside a columnar metric store and async log search for faster high-volume querying, blind human reviews, dashboard sections, and broad SDK additions spanning Python, TypeScript, Java, and Go.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
22 FEATURES18 OF 22 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Java SDK eval concurrency and lifecycle APIsNEW93how completely this was documenteddepth35/40specificity30/30actionability28/3093 / 100
Adds Eval.start() to start an eval in the background, returning an EvalResult as soon as the experiment exists, with lifecycle APIs getStatus(), isDone(), awaitCompletion(), getStartedAt(), getDuration(), getCasesExecuted(), and getAbortCause().︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ Concurrency is controlled via .maxConcurrency(int) on Eval.Builder or Devserver.Builder (default: 3 concurrent cases), or a custom Executor via .executor(Executor); global defaults for concurrency and OTel batch tuning can be set via the environment variables BRAINTRUST_DEFAULT_MAX_CONCURRENCY, BRAINTRUST_OTEL_MAX_QUEUE_SIZE, BRAINTRUST_OTEL_MAX_EXPORT_BATCH_SIZE, and BRAINTRUST_OTEL_EXPORT_INTERVAL_MILLIS (Java SDK v0.3.23).
Cap Java eval concurrency and tune OTel export batching via environment variables for high-throughput eval pipelines.
— Every method, flag, and env var named with defaultssnapshot-20260911
0202
Custom iframe panels in trace spansNEW83how completely this was documenteddepth30/40specificity28/30actionability25/3083 / 100
Adds a Span iframe project setting to render custom iframe URLs as fields inside trace spans, using mustache {{parameters}} to reference span data paths, e.g. https://example.com/{{input.id}}?q={{metadata.foo.bar}}. A Post span data toggle POSTs the full span payload to the iframe on load, for cases where data is too large for URL parameters or the iframe needs to write data back to the span.
Embed a custom review tool inside every trace span by linking span fields into the iframe URL — no separate lookup needed.
📍In the project settings, go to Span iframe › Create span iframe. Set URL to your tool's endpoint using mustache parameters, e.g. https://example.com/{{input.id}}?q={{metadata.foo.bar}}, give it a unique Field name, then save.
SpanFilters for trace-level scorersNEW81how completely this was documenteddepth25/40specificity28/30actionability28/3081 / 100
Adds has_error, metadata, and duration fields to SpanFilters for trace.get_spans() in the Python SDK (v0.39.0), letting trace-level scorers filter spans by error state, metadata values, and execution time bounds.
Filter spans in a trace-level scorer to only those that errored and took longer than 2 seconds, narrowing scorer focus to slow failures.
Filter spans in a trace-level scorer to only those that errored and ran longer than 5 seconds, using the new has_error and duration SpanFilters fields.
Agno eval auto-instrumentationNEW81how completely this was documenteddepth25/40specificity28/30actionability28/3081 / 100
Adds Agno eval instrumentation to the Python SDK (v0.36.0) — AccuracyEval, AgentAsJudgeEval, ReliabilityEval, PerformanceEval, and eval suites are traced automatically via auto_instrument(); eval suite runs automatically open a Braintrust experiment.
Instrument Agno evals for automatic tracing and experiment creation in Braintrust, using the new auto_instrument() integration.
ElevenLabs tracing integrationNEW75how completely this was documenteddepth25/40specificity25/30actionability25/3075 / 100
Adds wrapElevenLabs instrumentation in the TypeScript SDK (v3.32.0) to trace ElevenLabs SDK calls — text-to-speech and speech-to-text, including streaming and timestamped audio — for debugging audio generation, evaluating models, and monitoring production usage.
Instrument all ElevenLabs TTS/STT calls in a TypeScript agent for automatic tracing, including streaming and timestamped audio.
typescript
import { wrapElevenLabs } from 'braintrust';
import ElevenLabs from 'elevenlabs';
const client = wrapElevenLabs(new ElevenLabs({ apiKey: process.env.ELEVENLABS_API_KEY }));
Loop automations with scheduled runsNEW68how completely this was documenteddepth28/40specificity22/30actionability18/3068 / 100
Adds Loop automations: scheduled Loop runs with their own instruction, model, and write permissions that produce a read-only thread you can Continue, with results deliverable to a Slack channel or webhook.
— Names delivery targets and thread action; no config keys givensnapshot-20260911
0700
Loop managed runtime upgradeIMPROVED67how completely this was documenteddepth30/40specificity25/30actionability12/3067 / 100
Upgrades Loop to run in a Braintrust-managed, server-side runtime rather than the browser, persisting threads and enabling work across logs, experiments, and datasets, with the ability to create and edit prompts, scorers, datasets, facets, custom views, dashboards, and automations; runs on built-in GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna models.
— Names models and scope of work; no direct setup stepsnapshot-20260911
0807
OpenAI Batch API instrumentationNEW66how completely this was documenteddepth20/40specificity28/30actionability18/3066 / 100
Adds OpenAI Batch API instrumentation in the TypeScript SDK (v3.30.0) via openaiFilesCreateTraced, openaiBatchesRetrieveTraced, and completeOpenAIBatchTrace for tracing asynchronous OpenAI Batch requests.
Harbor plugin verifier output uploadNEW66how completely this was documenteddepth20/40specificity28/30actionability18/3066 / 100
Adds standard verifier output upload to the Harbor plugin in the Python SDK (v0.36.0): test-stdout.txt, test-stderr.txt, and ctrf.json, with attachments and redact_patterns configuration.
Columnar metric storage for high-volume logsNEW65how completely this was documenteddepth25/40specificity20/30actionability20/3065 / 100
Adds a Metric columnstore toggle in project settings that writes trace-level metrics in a columnar format during compaction, speeding up loading, sorting, and querying traces on the Logs page for high-volume projects; enable it alongside log search optimization.
Speed up Logs page queries on a high-volume project by enabling columnar storage for trace-level metrics.
📍In the project settings, go to Metric columnstore and enable the toggle to write trace-level metrics in a columnar format.
— Named toggle and page; missing a quantified speedupsnapshot-20260911
1110
Blind human reviewsNEW65how completely this was documenteddepth25/40specificity20/30actionability20/3065 / 100
Adds Blind human reviews, which hide peer scores, comments, and aggregates until a reviewer submits their own; enabling the 'Blind human reviews' project setting extends that hiding for the full review session.
Go SDK v0.13.0 with prompt supportNEW60how completely this was documenteddepth25/40specificity20/30actionability15/3060 / 100
Ships Braintrust Go SDK v0.13.0 with a new Go API reference, adding prompt support: load a Braintrust prompt, render its variables locally, call any LLM client with the result, drive eval parameters, and link rendered calls back to their prompt in Braintrust traces.
— Names version and prompt workflow; no code samplesnapshot-20260911
1312
glm-5.3-flash built-in modelNEW60how completely this was documenteddepth20/40specificity22/30actionability18/3060 / 100
Adds glm-5.3-flash as a built-in Braintrust-served model requiring no AI provider setup, selectable under the Braintrust provider in playgrounds, prompts, and scorers, or requestable via glm-5.3-flash through the Braintrust Gateway.
— Names model and access surfaces; no request examplesnapshot-20260911
1413
Patterns: scheduled AI pattern detectionNEW60how completely this was documenteddepth30/40specificity18/30actionability12/3060 / 100
Adds Patterns: Braintrust runs Loop against a project on a schedule to surface recurring problems and trends not covered by any scorer, saving each finding with supporting traces and a suggested fix. Findings can create a scorer or classifier, continue investigation in Loop, or be copied as a prompt.
— Explains mechanism and outputs; no config surface namedsnapshot-20260911
thinner coverage below
1500
Asynchronous execution engine for log searchIMPROVED55how completely this was documenteddepth25/40specificity20/30actionability10/3055 / 100
Adds an async execution engine to Brainstore log search that streams results as segments finish scanning, speeding up full-text, phrase, and multi-clause trace searches, including negated ANY_SPAN() queries — no configuration required.
— Explains streaming mechanism; nothing for reader to configuresnapshot-20260911
1614
Debugger for single-trace failure analysisNEW55how completely this was documenteddepth28/40specificity15/30actionability12/3055 / 100
Adds Debugger, which reports likely failure modes for a single trace, citing spans, tool calls, and model outputs; its 'Analyze trace' action groups a run into Work sections explaining what the agent did.
— Describes behaviour and UI action; no config detailsnapshot-20260911
1700
Optional score name in scorer functionsIMPROVED55how completely this was documenteddepth18/40specificity22/30actionability15/3055 / 100
Allows scorer functions in the Python SDK (v0.38.0) and TypeScript SDK (v3.31.0) to return {"score": 0.5} without a name field, using the function's name as the score key.
Dashboard sectionsNEW52how completely this was documenteddepth22/40specificity18/30actionability12/3052 / 100
Adds Dashboard sections: group charts into named, collapsible, reorderable, and duplicable sections with per-browser collapse-state memory, available on Pro and Enterprise plans.
wrapGoogleGenerativeAI legacy SDK instrumentationNEW50how completely this was documenteddepth15/40specificity20/30actionability15/3050 / 100
Adds wrapGoogleGenerativeAI instrumentation in the TypeScript SDK (v3.32.0) for the legacy @google/generative-ai SDK (v0.24.x).
— Names function and SDK version; no usage example givensnapshot-20260911
2000
Span metadata enrichment for Vercel AI SDK and Pydantic AIIMPROVED50how completely this was documenteddepth15/40specificity25/30actionability10/3050 / 100
Adds toolCallId to Vercel AI SDK v6 tool call span metadata in the TypeScript SDK (v3.31.0), and adds invocation parameters (temperature, max_tokens, and similar model settings) to Pydantic AI LLM span metadata in the Python SDK (v0.37.0).
Java SDK requires thread-safe Task and ClassifierBREAKING45how completely this was documenteddepth15/40specificity15/30actionability15/3045 / 100
Breaking change: Task and Classifier implementations in the Java SDK must now be thread-safe, a consequence of the new concurrent eval execution model.
— States the exact requirement; no migration guidance givenproduct docs
2218
input_audio support in prompt templatesNEW40how completely this was documenteddepth12/40specificity18/30actionability10/3040 / 100
Adds input_audio content type support in Mustache and Nunjucks prompt template rendering in the TypeScript SDK (v3.32.0).
Mistral shipped OCR 4.1 to general availability and added block-level granularity to OCR confidence scoring.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Learn how to deploy and use Mistral AI's Large Language Models with our comprehensive documentation, guides, and tutorials.
Mistral shipped OCR 4.1 to general availability and added block-level granularity to OCR confidence scoring.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
2 FEATURES2 OF 2 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Block-level granularity for OCR confidence scoresNEW78how completely this was documenteddepth25/40specificity28/30actionability25/3078 / 100
The OCR API's confidence_scores_granularity parameter now accepts "block" in addition to "page" and "word".︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ "page" returns page-level scores only, "block" returns page-level and block-level scores, and "word" returns page-level and word-level scores.
— Names exact parameter and all three accepted values.snapshot-20260911
thinner coverage below
0202
OCR 4.1 general availabilityNEW50how completely this was documenteddepth15/40specificity20/30actionability15/3050 / 100
mistral-ocr-4-1 is now Generally Available, and both mistral-ocr-latest and mistral-ocr-4 now point to it.
— Names model ids and aliasing but no capability detail.snapshot-20260911
Anthropic extended Claude Managed Agents with new permission-policy machinery for tool calls and a CLI command for live human oversight of running agent sessions.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
Anthropic provides Claude AI models and APIs for building applications that generate, analyze, and automate work.
Anthropic extended Claude Managed Agents with new permission-policy machinery for tool calls and a CLI command for live human oversight of running agent sessions.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄
└──▷ WHAT SHIPPED ·
2 FEATURES2 OF 2 FEATURESmost completely described firstwhat's the number?
Each feature carries 0–100 for how completely the vendor documented
it — not how big or important the work is. A major capability
described in eight words scores low, and that is the finding.
depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point
Code and config rank first by construction: a documented endpoint,
flag or config key scores at the top of specificity and actionability,
so it sorts above a feature described only in prose. Under
60 the notes go thin — everything below the “thinner
coverage” line is thinner documentation, not smaller work.
Hover any meter for that feature's three sub-scores.
0101
Live terminal connect command for agent sessionsNEW80how completely this was documenteddepth25/40specificity25/30actionability30/3080 / 100
The ant CLI gains ant beta:sessions connect, attaching your terminal to a Claude Managed Agents session so you can follow it live, send messages, and allow or deny pending tool calls.︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄ A --web flag serves the Claude Console session viewer locally and opens the session in a browser instead of the terminal.
— Exact CLI command and flag given, directly runnablesnapshot-20260911
0202
Permission policy events and auto mode for Managed AgentsNEW73how completely this was documenteddepth30/40specificity28/30actionability15/3073 / 100
Claude Managed Agents gains agent.tool_use and agent.mcp_tool_use permission policy events, letting the server evaluate each agent or MCP tool call and run it, deny it, or pause for human approval. A new auto permission policy automatically runs, denies, or pauses calls for approval, and the events gain evaluation and evaluated_permission fields reporting how each call was evaluated under auto. Managed Agents permission policy support is extended to Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5.
— Names events, fields, and models but no direct config examplesnapshot-20260911
THE AI TOOLCHAIN
