Heads up This site is currently under heavy development.
Subscribe Get it delivered — the daily firehose, filtered to the tools you run, plus the documentation changes vendors never announce. Compare plans →

The AI Toolchain — issue 021, September 9, 2026

THE AI TOOLCHAIN NO. 021
Tail
THE DAILY RELEASE FIREHOSE
PUBLISHED SEPTEMBER 9, 2026 · EVERY WEEKDAY
EDITIONS tail grep head diff uniq

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 VIEW full issue
Do you prefer full issue? 
$ tct list   # 28 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.

Three unrelated things stand out. mlx-serve now spills prefix caches to disk, so a long system prompt survives restarts instead of being recomputed. Braintrust runs evals concurrently, turning a serial suite into something you can leave in CI. Fireworks AI adds per-user spend limits and BYO provider keys, so a runaway agent hits a ceiling instead of an invoice.

  • deploy

    Serve a long system prompt without paying for it on every request

    mlx-serve backs prefix caching with disk and adds speculative-decoding controls, so cached prefixes outlive process restarts and drafted tokens can be tuned per model rather than left at defaults. XGrammar's cached Lark compilation removes grammar-compile latency from the first constrained request, and its NPU NPU A Network Processing Unit, a dedicated chip designed to handle packet parsing, classification, and forwarding at wire speed, offloading those tasks from the CPU so cyber tools can process traffic at high throughput with low latency. backend and Windows ARM64 Windows ARM64 A build target for Windows running on ARM64 processors, such as those in Snapdragon-powered laptops; cyber tools targeting it can run natively on that hardware without emulation overhead. wheels put structured decoding on hardware that previously had none.

    mlx-serve · XGrammar

  • evaluate

    Run a full eval suite in wall-clock time short enough to gate a merge

    Braintrust executes evals concurrently with new tuning and lifecycle APIs; the practical change is that a suite that took an hour serially can finish inside a build. LangSmith lets you define an LLM-as-judge evaluator from the CLI and page through experiment comparisons via API, so scoring config lives in the repo instead of being clicked together in the UI.

    Braintrust · LangSmith

  • govern

    Cap what one user or one team can spend before finance finds out

    Fireworks AI adds per-user spending limits, bring-your-own provider keys and custom SSO/ SCIM SCIM An open standard protocol (RFC 7642–7644) for automating user and group provisioning between identity providers and applications, letting cyber tools sync accounts without custom integration code. , which means an agent loop that runs away is bounded at the platform rather than discovered on the bill. Devin Desktop's CLI takes enterprise-wide config for model access, defaults, MCP MCP Model Context Protocol, an open standard from Anthropic that lets an AI model call external tools and data sources through a uniform interface, so cyber tools can expose capabilities directly to LLM-based agents. servers and web search, so an org sets the allowed surface once instead of trusting every developer's local settings.

    Fireworks AI · Devin Desktop

Does Opus 5 read better?
DEPTH
Build
◆  AI Coding Agents

StackBlitz bolt.new

Sources Release page → snapshot-20260909 NOTES

Bolt.new is an AI-powered web development platform that builds websites and applications from prompts.

Bolt.new adds a visual editor in the preview pane — edit text, colors, and styles before committing any tokens.

└──▷ HOW TO FIND IT
Iterate on UI appearance without spending tokens — use the Select tool to tweak colors and text, review all changes in the batch, then commit once when satisfied.
📍In the chatbox, click the Select tool, then click any element in the preview to edit its text, color, font size, formatting, or alignment. Review the collected edits above the chatbox, then click 'Save changes' to apply them to your project code.
  • ›Adds a Select tool in the chatbox for making visual edits — text, colors, font size, formatting, and alignment — directly in the preview, with changes batched and applied to project code only when 'Save changes' is clicked.

1 detail in Everything

Was this useful?

Cline

Sources Release notes →Source code → 1 RELEASE · 2026-09-09 NOTES CODE Rank Why this rank top features scored 62/58/55 — 58.33 How ranking works →

Cline's desktop app adds authenticated remote Hub SDK connectivity, default web search, imported-session summarization from other coding agents, and a string of reliability and performance fixes to checkpoints, patching, and session history.

Autonomous coding agent as an SDK, IDE extension, or CLI assistant.

Cline's desktop app adds authenticated remote Hub SDK connectivity, default web search, imported-session summarization from other coding agents, and a string of reliability and performance fixes to checkpoints, patching, and session history.

└──▷ WHAT SHIPPED · 8 FEATURES 3 OF 8 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Imported session summaries from other coding agents NEW 62

Sessions imported from Claude Code, Codex, and opencode now display their source in chat and summarize the foreign history on the first resumed turn, showing a 'Summarizing the imported <tool> history...' indicator while running.

— Names sources and exact indicator text with mechanism desktop-v0.0.24
thinner coverage below
02 00
apply_patch Add File no longer overwrites silently IMPROVED 58

apply_patch 'Add File' now refuses to overwrite an existing file instead of silently replacing it.

— Names exact tool and action, clear before/after behavior desktop-v0.0.24
03 00
Checkpoint hashing performance overhaul IMPROVED 55

Checkpoints no longer re-hash every untracked file before each message; a single snapshot index is kept per session, so from the second turn cost drops to roughly git process overhead.

— Clear before/after mechanism and performance claim, no user action needed desktop-v0.0.24
04 00
Web search on by default outside YOLO mode IMPROVED 50

Web search is now enabled by default outside YOLO mode, and tool settings fail closed if they cannot be loaded.

— Explains behavior change but no config key to toggle it desktop-v0.0.24
05 02
Custom title bar on Windows NEW 50

Adds a custom title bar on Windows with caption controls that follow the compact title-bar height in narrow windows and stay above overlays.

— Describes UI behavior across window sizes, no config surface desktop-v0.0.24
06 00
@ mention no longer indexes home or filesystem root IMPROVED 43

Typing an @ mention no longer indexes the home directory or filesystem root.

— Names exact trigger and fixed scope, minimal further detail desktop-v0.0.24
07 00
Session-history token counts and costs backfilled IMPROVED 42

Token counts and costs now fill in for every visible session-history page, hydrated on demand with reads re-run when session status changes.

— Explains mechanism but no UI path or command given desktop-v0.0.24
08 03
SDK support for authenticated remote Hubs NEW 33

Adds SDK support for connecting to authenticated remote Hubs.

— Names the capability but no connection mechanism or API detail desktop-v0.0.24
Was this useful?

OpenAI Codex CLI

Sources Release notes →Source code → 2 RELEASES · 2026-09-08 → 2026-09-09 NOTES CODE Rank Why this rank top features scored 83/69/60, +6 runnable example — 76.67 How ranking works →

Codex CLI's alpha channel added live WebRTC WebRTC A browser-native, open standard for real-time peer-to-peer audio, video, and data exchange without plugins. Cyber tools use it to tunnel traffic or enable covert communication channels through firewalls via its built-in NAT traversal. voice conversations to the terminal UI with macOS and MCP MCP Model Context Protocol, an open standard from Anthropic that lets an AI model call external tools and data sources through a uniform interface, so cyber tools can expose capabilities directly to LLM-based agents. user verification, introduced an isolated memory v2 architecture, and expanded app-server daemon update controls, alongside a raised Guardian review limit and new telemetry.

OpenAI Codex CLI runs an agent in the terminal that reads, changes, and tests code in local repositories.

Codex CLI's alpha channel added live WebRTC WebRTC A browser-native, open standard for real-time peer-to-peer audio, video, and data exchange without plugins. Cyber tools use it to tunnel traffic or enable covert communication channels through firewalls via its built-in NAT traversal. voice conversations to the terminal UI with macOS and MCP MCP Model Context Protocol, an open standard from Anthropic that lets an AI model call external tools and data sources through a uniform interface, so cyber tools can expose capabilities directly to LLM-based agents. user verification, introduced an isolated memory v2 architecture, and expanded app-server daemon update controls, alongside a raised Guardian review limit and new telemetry.

└──▷ WHAT SHIPPED · 12 FEATURES 10 OF 12 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
App-server daemon update management NEW 83

Adds an app-server daemon update subcommand to update the standalone installation and restart the managed daemon. Makes app-server daemon automatic updates and the managed app-server shutdown grace period configurable, warns when the connected Codex service is older than the CLI, makes those older-version notices configurable in the TUI, and shows the configured updater settings (auto_update_enabled, update_interval_minutes) in codex doctor output.

Trigger an immediate update of the standalone Codex installation and restart the managed daemon without waiting for the automatic update cycle.
$ codex app-server daemon update
— Names a runnable subcommand and exact config keys shown in codex doctor rust-v0.154.0-alpha.7rust-v0.154.0-alpha.10.2
02 02
Live voice conversations in the TUI NEW 69

Adds live WebRTC voice conversations to the TUI, including a dedicated composer strip for voice controls, a voice mute shortcut (now configurable) with recording activity indicators, and live voice transcripts animated with split-flap tiles. Signed voice resources are now bundled in macOS releases to support this.

— Describes multiple named UI mechanisms but no command or config path rust-v0.154.0-alpha.7rust-v0.154.0-alpha.10.2
03 03
Telemetry metrics for voice sessions and TUI startup NEW 60

Records voice session lifecycle metrics under codex.voice.session.* (starts, connections, failures, endings, active duration), and tags codex.tui.start metrics with terminal_name and multiplexer attributes (tmux, zellij, or none).

— Names exact metric identifiers and attributes but no query interface rust-v0.154.0-alpha.10.2
thinner coverage below
04 04
Worktree delete action in managed worktree browser NEW 58

Adds a worktree delete action with a confirmation dialog in the managed worktree browser, including owner thread titles, relative update times, and archived/unavailable status.

— Names a UI location but no exact navigation steps rust-v0.154.0-alpha.10.2
05 05
Memory v2 with isolated storage NEW 56

Adds configurable memory versions with isolated storage (memory v2), including dual writing, dedicated consolidation and read prompts, and summary-only extraction.

— Names the mechanism but no config key or command to enable it rust-v0.154.0-alpha.7
06 06
User verification for macOS and MCP NEW 52

Adds macOS user verification with Secure Enclave signing and MCP user verification in the TUI, restricted to the host-owned apps MCP server.

— Names the mechanism and scope but no user-facing steps rust-v0.154.0-alpha.7
07 00
Guardian action review limit raised to 200,000 bytes IMPROVED 52

Raises Guardian's action review limit from its previous cap to 200,000 bytes, unlocking review of larger tool outputs.

— States the concrete new limit but not the prior value or how to invoke it rust-v0.154.0-alpha.10.2
08 07
MCP OAuth reconnect signals in TUI NEW 35

Surfaces MCP reconnect signals in the TUI when expired OAuth tokens cannot be refreshed.

— Describes the trigger but not what the reader does next rust-v0.154.0-alpha.10.2
09 08
AWS credential export for Bedrock NEW 35

Adds AWS credential export commands for Amazon Bedrock integration.

— Names the integration but not the actual command rust-v0.154.0-alpha.10.2
10 00
TUI turn and action display improvements IMPROVED 33

Shows completion timestamps after successful TUI turns, and groups adjacent computer actions in the TUI.

— Two thin display tweaks with no further detail rust-v0.154.0-alpha.7
11 09
Configurable credential providers for network proxy NEW 30

Adds configurable credential providers to the network proxy.

— Bare description with no config key or example rust-v0.154.0-alpha.10.2
12 10
Image attachments in agents overview background tasks NEW 28

Supports image attachments in agents overview background tasks.

— Bare feature name, no mechanism or path given rust-v0.154.0-alpha.10.2
Was this useful?

SST OpenCode

Sources Release notes →Source code → 1 RELEASE · 2026-09-09 NOTES CODE Rank Why this rank top features scored 60/23 — 41.5 How ranking works →

OpenCode v1.18.30 introduces a new system prompt for GPT-6 models and adds configurable reasoning effort for GitLab-hosted GPT and Claude models.

OpenCode is an open-source AI coding agent for terminal-based development tasks.

OpenCode v1.18.30 introduces a new system prompt for GPT-6 models and adds configurable reasoning effort for GitLab-hosted GPT and Claude models.

└──▷ WHAT SHIPPED · 2 FEATURES 2 OF 2 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Reasoning effort variants for GitLab-hosted models NEW 60

Adds reasoning effort variants for GitLab-hosted GPT models via reasoningEffort and for Claude models via adaptive thinking.

— Names exact config surfaces but no usage detail or defaults v1.18.30
thinner coverage below
02 02
Astra system prompt for GPT-6 models NEW 23

Adds the Astra system prompt for GPT-6 models.

— Only named in summary, no mechanism or usage given v1.18.30
Was this useful?

Anthropic Claude Code

Sources Release notes →Source code → 1 RELEASE · 2026-09-08 NOTES CODE Rank Why this rank top features scored 85/60/55, +6 runnable example — 72.67 How ranking works →

Claude Code v2.1.265 adds directory-based plugin loading with hot-reload, a 1 GB cap on saved tool results, faster worktree startup, and improved workflow and slash-command UIs, plus VS Code session archiving.

Claude Code is Anthropic's terminal coding agent that plans, edits, and tests code in local repositories.

Claude Code v2.1.265 adds directory-based plugin loading with hot-reload, a 1 GB cap on saved tool results, faster worktree startup, and improved workflow and slash-command UIs, plus VS Code session archiving.

└──▷ WHAT SHIPPED · 7 FEATURES 2 OF 7 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Plugin directory loading with hot-reload NEW 85

Pointing --plugin-dir at a folder of plugins loads each child folder containing a manifest automatically, and plugins added or removed while Claude Code is running are picked up live without a restart.

Point Claude Code at a plugins directory so all sub-plugins load at once and any you drop in or remove are picked up without restarting.
$ claude --plugin-dir ./my-plugins/
— Named flag with clear mechanism and runnable example v2.1.265
02 00
Richer /workflows agent detail view IMPROVED 60

In /workflows, tool calls are now marked running, failed, or done, the subagent's task list is shown when available, and pressing Enter unfolds listed calls to show their inputs and results.

— Names command and exact interaction (Enter key) v2.1.265
thinner coverage below
03 00
Expanded telemetry fields for gateway sessions IMPROVED 55

Claude Desktop and Cowork now send user.email and user.groups in the telemetry pushed through a Claude apps gateway, matching what terminal sessions already report.

— Names exact fields but no user action possible v2.1.265
04 00
Faster worktree startup on large repos IMPROVED 55

--worktree startup is improved on large repositories: the new worktree is now checked out in parallel, requiring git 2.32 or later.

— Names flag, mechanism and version requirement v2.1.265
05 00
Improved slash command and skill matching IMPROVED 55

Slash commands typed mid-prompt now show matches in a list, with Tab opening the list outside fullscreen instead of only showing a single suggestion, and a plugin skill is now findable by its bare name.

— Names Tab behavior but limited mechanism detail v2.1.265
06 02
Auto-archive inactive sessions in VS Code NEW 55

A new 'Archive inactive sessions' setting in the VS Code extension automatically archives sessions inactive for a set period, defaulting to 14 days.

— Names the setting and default but no path given v2.1.265
07 00
1 GB cap on saved tool results IMPROVED 45

Tool results saved to disk are now capped at 1 GB, and the in-conversation preview indicates when a saved file was truncated due to hitting this cap.

— Concrete limit named but no configurable surface v2.1.265
Was this useful?

Diagram Design

Sources Commits → changes since 2026-08-12 CODE

Diagram Design is a Claude Code skill that generates editorial HTML and SVG diagrams and imports draw.io or Mermaid sources.

Diagram Design adds ten editorial diagram types, native Droid plugin packaging, named client profiles, accessible-SVG contract, and animated semantic-pattern examples.

└──▷ 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 SVGs that will fail screen-reader accessibility — for example before merging a PR that adds a new diagram — by running the a11y lint gate locally.
$ python scripts/lint-skin.py --all --baseline
  • ›Adds role="img", a resolving aria-labelledby, and first-child <title> / <desc> slots to template.html, template-dark.html, template-full.html, and template-terminal.html as an enforced accessible-SVG contract.
  • ›Adds an a11y category to scripts/lint-skin.py that rejects a diagram SVG without a resolvable accessible name.
  • ›Adds scripts/verify-beeswarm.py to verify that every beeswarm dot sits at its declared value on the shared linear scale and that no two dots overlap.
  • ›Adds scripts/test-verify-sankey.py with adversarial fixtures covering ribbon narrowing, node volume loss, stage-flow shrinkage, label/bar mismatches, and missing shipped variants.
  • ›Ships ten editorial diagram types in one release (feat(types): ship ten editorial diagram types).
+5 more show less
  • ›Adds native Droid plugin packaging via feat(factory): add native Droid plugin packaging.
  • ›Adds example-queue-animated.html for Semantic Pattern #1 (Fan-in Queue / Bottleneck).
  • ›Adds example-paved-road-animated.html for Semantic Pattern #5 (Secure Paved Road).
  • ›Extends CI with a multi-OS testing matrix covering Linux, Windows, and macOS.
  • ›Adds a GitHub Step Summary table generated from matrix run results.
  • ›Adds role="img", a resolving aria-labelledby, and first-child <title> / <desc> slots to template.html, template-dark.html, template-full.html, and template-terminal.html as an enforced accessible-SVG contract.
  • ›Adds scripts/verify-beeswarm.py to verify that every beeswarm dot sits at its declared value on the shared linear scale and that no two dots overlap.
  • ›Adds scripts/test-verify-sankey.py with adversarial fixtures covering ribbon narrowing, node volume loss, stage-flow shrinkage, label/bar mismatches, and missing shipped variants.
  • ›Ships ten editorial diagram types in one release (feat(types): ship ten editorial diagram types).
  • ›Adds native Droid plugin packaging via feat(factory): add native Droid plugin packaging.
+3 more show less
  • ›Adds example-queue-animated.html for Semantic Pattern #1 (Fan-in Queue / Bottleneck).
  • ›Adds example-paved-road-animated.html for Semantic Pattern #5 (Secure Paved Road).
  • ›Adds a GitHub Step Summary table generated from matrix run results.

2 more in Everything

Was this useful?

Block Goose

Sources Release notes →Source code → 1 RELEASE · 2026-09-08 NOTES CODE Rank Why this rank top features scored 50/35/30 — 38.33 How ranking works →

Goose v1.50.0 expands model provider support with GPT-6 Astra and AIML API, splits out ACP npm packages, and adds Kotlin GDK GDK GNOME Development Kit, a library maintained by the GNOME Project that sits between application code and the underlying windowing system, giving GUI tools a portable drawing and event-handling layer. configuration for Databricks AI Gateway.

Goose is an open-source AI agent with a desktop app and CLI for coding and other automated tasks.

Goose v1.50.0 expands model provider support with GPT-6 Astra and AIML API, splits out ACP npm packages, and adds Kotlin GDK GDK GNOME Development Kit, a library maintained by the GNOME Project that sits between application code and the underlying windowing system, giving GUI tools a portable drawing and event-handling layer. configuration for Databricks AI Gateway.

└──▷ WHAT SHIPPED · 4 FEATURES 2 OF 4 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 00
Separate npm packages for ACP client and binary IMPROVED 50

The goose ACP client and the acp binary are now published as separate npm packages.

— Names exact packages, no install instructions v1.50.0
02 01
New provider support: GPT-6 Astra and AIML API NEW 35

Goose adds GPT-6 Astra models as a selectable provider and integrates AIML API as a new provider option.

— Names two providers but no config details v1.50.0
03 02
Kotlin GDK configuration for Databricks AI Gateway path NEW 30

Enables Kotlin callers in the GDK to configure the Databricks AI Gateway path.

— States capability but no config key or example v1.50.0
04 00
Upgraded MCP version preference IMPROVED 10

Upgrades MCP version preference to the latest available.

— Bare statement with no version number or mechanism v1.50.0
Was this useful?

Letta Code

Sources Release notes → 1 RELEASE · 2026-09-08 NOTES Rank Why this rank top features scored 30 — 30 How ranking works →

Letta Code is a terminal coding agent with persistent memory and identity across development sessions.

Letta Code shipped a single small update in this window: the ability to cancel Monitors from connected clients via the listener.

└──▷ WHAT SHIPPED · 1 FEATURE 1 OF 1 FEATURE most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Cancel Monitors from connected clients NEW 30

Connected clients can now cancel Monitors through the listener.

— No mechanism, command, or config detail given, just a bare capability. v0.31.14
Was this useful?

Alibaba Qwen Code

Sources Release notes → 3 RELEASES · 2026-09-08 NOTES Rank Why this rank top features scored 81/81/80, +15 breaking change, +6 runnable example — 101.67 How ranking works →

An open-source AI coding agent that lives in your terminal.

Qwen Code's SDK matured with multi-modal input, session resume/export, and Agent Skills promoted to stable, while the core CLI (v0.23.1) shipped a large batch of workflow, channel, and web-shell session-management features alongside two breaking removals: the read_many_files tool and the @qwen-code/webui package.

└──▷ WHAT SHIPPED · 43 FEATURES 34 OF 43 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
`extra_body` pass-through for OpenAI-compatible providers NEW 81

Adds extra_body support so OpenAI-compatible providers can receive pass-through of provider-specific request fields, e.g. passing an extended-thinking budget via {"extra_body": {"thinking": {"type": "enabled", "budget_tokens": 8000}}}.

Pass provider-specific parameters (e.g. extended thinking budget) through to an OpenAI-compatible endpoint that supports them.
json 
{
  "extra_body": {
    "thinking": {
      "type": "enabled",
      "budget_tokens": 8000
    }
  }
}
— Named config field with a runnable JSON example sdk-typescript-v0.1.10sdk-typescript-v0.1.9
02 02
`ui.showToolCallArgs` inline tool-call arguments NEW 81

Adds ui.showToolCallArgs config setting to render tool-call arguments inline in the UI, so you can audit exactly what the agent invokes during a session.

Render tool-call arguments inline in the transcript so you can audit exactly what the agent is invoking during a session.
yaml 
ui:
  showToolCallArgs: true
— Exact config key with working YAML example v0.23.1
03 03
Environment variables via `settings.env` NEW 80

Adds a settings.env field for configuring environment variables directly in the settings file, letting provider API keys or proxy variables (e.g. OPENAI_API_KEY, HTTPS_PROXY) be set for a session without exporting them to the shell.

Set provider API keys or proxy variables for a session without exporting them to the shell environment.
json 
{
  "env": {
    "OPENAI_API_KEY": "sk-...",
    "HTTPS_PROXY": "http://proxy.internal:3128"
  }
}
Set provider API keys and custom base URLs directly in settings so every team member picks them up without shell exports.
json 
{
  "settings": {
    "env": {
      "OPENAI_API_KEY": "sk-...",
      "OPENAI_BASE_URL": "https://your-proxy.example.com/v1"
    }
  }
}
— Named config field with two working JSON examples sdk-typescript-v0.1.10sdk-typescript-v0.1.9
04 04
Custom output styles from output-styles directories NEW 76

Loads custom output styles from ~/.qwen/output-styles and .qwen/output-styles directories.

Apply a custom output style to all sessions by dropping a style file into the user-level output-styles directory.
$ mkdir -p ~/.qwen/output-styles
cp my-team-style.json ~/.qwen/output-styles/
— Exact directory paths with shell example v0.23.1
05 00
Managed memory gated by `memory.enableManagedAutoMemory` IMPROVED 70

Managed memory availability now respects the memory.enableManagedAutoMemory setting, so hosts that disable managed auto memory no longer see remember/dream requests admitted or managed-memory instructions injected into the system prompt.

— Named setting with precise before/after behaviour sdk-typescript-v0.1.10sdk-typescript-v0.1.9
06 05
Session ID propagation to provider requests NEW 70

Expands ${session_id} in per-provider customHeaders configuration, and sends session ID to Routify endpoints.

Forward the active session ID to a Routify-compatible provider by embedding ${session_id} in a custom request header.
yaml 
providers:
  my-provider:
    baseUrl: https://proxy.example.com
    customHeaders:
      X-Session-ID: "${session_id}"
— Named template variable with config example v0.23.1
07 06
OSC 8 link handling in VP mode NEW 70

Enables opening OSC 8 links with a plain click in VP mode and adds a right-click menu with 'Open Link', 'Copy Link Address', and 'Copy Selection'.

— Exact interactions and menu items named v0.23.1
08 07
ACP subagent tracking and visibility NEW 68

Adds parentToolCallId and subagentType fields for ACP subagent tracking, and makes subagent execution visible via parent tool_call updates in ACP.

— Names exact fields and update mechanism sdk-typescript-v0.1.10v0.23.1
09 08
Web-shell UI: navigation, context, and workflow visibility NEW 68

Adds continuous history and compact turn navigation to the web-shell transcript, a context usage tab to the right sidebar, headless global turn navigation, the ability to bridge a browser-granted local directory into a session, visualization and management of dynamic workflow runs, live status for running subagents in the transcript, customizable artifact icons, and a display of Goal token spend against its allowed window.

— Enumerates eight UI additions, no exact paths v0.23.1
10 09
Channel and worktree lifecycle management NEW 62

Adds worktree-isolated named tasks to channels, implements named-session worktree reset for channels, adds a shared multiline instructions field to channel management, adds message prefix filtering to channels, and reaps owned worktrees automatically when daemon sessions are deleted.

— Enumerates five channel/worktree additions, no config keys v0.23.1
11 10
Replaces `read_many_files` tool with `readManyFiles` utility BREAKING 60

Removes the read_many_files tool and replaces it with a readManyFiles utility, available only as a user @-command.

— Names old and new surface with usage scope sdk-typescript-v0.1.10
thinner coverage below
12 00
`cache_control` for Anthropic content generator IMPROVED 59

Adds cache_control for the system message and the last user text message in the Anthropic content generator, to improve prompt-cache reuse.

— Names exact field and scope, no config toggle shown sdk-typescript-v0.1.9
13 11
Session resume and continue in SDK NEW 57

Adds resume and continue options plus extended authType support to the SDK, and supports resuming sessions by session ID via the query interface.

— Names options and interface but no example sdk-typescript-v0.1.10
14 00
Size-triggered microcompaction with low watermark IMPROVED 56

Size-triggered microcompaction now clears old tool results toward a low watermark, preserving provider prompt-cache reuse across long tool-heavy sessions.

— Explains mechanism and benefit, no config knob given sdk-typescript-v0.1.10
15 12
Session turn navigation protocol NEW 55

Adds session turn navigation protocol support in the daemon and SDK for bounded turn indexing, snapshots, and transcript paging.

— Explains mechanism, no API names given v0.23.1
16 13
External-context memory: Mem0 recall and daemon writes NEW 52

Adds opt-in auto recall for administrator-owned Mem0 dialects in external-context, and adds daemon memory writes and daemon memory deletion to external-context.

— Names Mem0 integration, thin on mechanism v0.23.1
17 14
`contextWindowSize` configuration key NEW 50

Adds a contextWindowSize configuration key to control the model's context window size.

— Named config key but no default or example given sdk-typescript-v0.1.10sdk-typescript-v0.1.9
18 15
Session export to Markdown and HTML NEW 50

Adds an export command for session history, supporting Markdown and HTML output formats.

— Command named but no flags or example shown sdk-typescript-v0.1.10sdk-typescript-v0.1.9
19 16
Skill manager: symlink support and `zvec-grep` installer NEW 50

The skill manager gains symlink support, and a built-in zvec-grep installer is added to the skills system.

— Names installer tool, thin on mechanism sdk-typescript-v0.1.10v0.23.1
20 17
Session concurrency and control NEW 50

Allows concurrent standalone daemons with session fencing, fences Conversations writers with mandatory session leases, and adds a user-minted controller token to drive a session without per-message review.

— Describes three control mechanisms, no API surface named v0.23.1
21 00
TPM throttling retry handling IMPROVED 46

Adds TPM throttling error handling with a 1-minute retry delay.

— Gives exact retry delay but no config surface sdk-typescript-v0.1.10
22 18
Goal auto-pause after stalled turns NEW 46

Pauses a Goal automatically after three autonomous turns that make no progress.

— Exact turn count given, no override named v0.23.1
23 19
`FORK_MODE` for ProcessTransport (Electron IPC) NEW 45

Adds FORK_MODE support to ProcessTransport for Electron IPC integration.

— Names flag but minimal explanation of behaviour sdk-typescript-v0.1.10
24 20
Retirement of `@qwen-code/webui` package BREAKING 45

The @qwen-code/webui package has been retired and is no longer available.

— Names exact retired package, no migration path given v0.23.1
25 21
Multi-modal input across content generators NEW 43

Adds multi-modal input support (image, PDF, audio) across all content generators.

— Lists modalities but no usage detail sdk-typescript-v0.1.10
26 22
Coding Plan auth mode with unified AuthDialog NEW 43

Adds a Coding Plan authentication mode with a unified AuthDialog.

— Names dialog component, no setup steps sdk-typescript-v0.1.10
27 23
Serve layer: session catalog and Skills runtime NEW 41

Adds a session resource catalog to the serve layer, and a workspace-scoped Skills runtime to the serve layer.

— Names two serve-layer additions, thin on mechanism v0.23.1
28 00
OpenTUI parity closeout IMPROVED 41

Adds OpenTUI parity closeout covering dialogs, composer, and shell mode to the CLI.

— Names three covered areas, no specifics per area v0.23.1
29 00
DingTalk channel improvements IMPROVED 41

Makes background agent aggregation optional in the DingTalk channel, and shows dynamic lifecycle tags in the DingTalk channel.

— Names two channel-specific tweaks, thin mechanism v0.23.1
30 00
TUI paste handling fixes on macOS IMPROVED 38

Adds a large-paste placeholder and fixes enter-submit behaviour on macOS for improved TUI paste handling.

— Describes fix but no reproduction steps sdk-typescript-v0.1.9
31 24
Review system audits and recall floor NEW 38

Adds prose-execution and counter-frame audits to the review system, and adds a low-effort recall floor to the review system.

— Names two additions, no mechanism depth v0.23.1
32 25
Java SDK (alpha) NEW 37

Adds a Java SDK (alpha) for programmatic access to Qwen Code.

— Names new SDK language, no API details sdk-typescript-v0.1.9
33 00
`todo_write` made opt-in IMPROVED 37

Makes todo_write opt-in via core configuration.

— Names tool, no config key or default given v0.23.1
34 26
Zed extension for Qwen Code agent server NEW 35

Adds a Zed extension for the Qwen Code agent server.

— Names the editor integration only sdk-typescript-v0.1.10
35 27
MCP tool progress updates in TUI and SDK NEW 35

Adds MCP tool progress update support in TUI and SDK mode.

— Names surfaces but no mechanism detail sdk-typescript-v0.1.10
36 28
Workflow run handles and inline script persistence NEW 35

Returns run handles and persists inline scripts from workflows.

— Brief mechanism, no API name given v0.23.1
37 29
Concurrent runner for batch CLI execution NEW 34

Adds a concurrent runner for batch CLI execution.

— No flag or command name given sdk-typescript-v0.1.10
38 30
Configurable checkpoint verifier timeout NEW 32

Lets operators configure the checkpoint verifier timeout for a full claim list.

— No config key or default named v0.23.1
39 31
Removes Smart Edit tool and ClearcutLogger BREAKING 30

Removes the Smart Edit tool and ClearcutLogger.

— Names removed components, no migration guidance sdk-typescript-v0.1.10
40 32
Configurable release download base URL NEW 28

Adds support for a configurable release download base URL.

— No config key or example named v0.23.1
41 33
Model reasoning configuration NEW 28

Configures model reasoning capabilities via core configuration.

— No config key or scope named v0.23.1
42 34
Experimental LSP support NEW 25

Adds experimental LSP support for code intelligence.

— Bare mention, no scope or commands sdk-typescript-v0.1.10
43 00
Agent Skills promoted to stable IMPROVED 25

Promotes Agent Skills from experimental to stable.

— Status change only, no new capability described sdk-typescript-v0.1.10
└──▷ BREAKING ON UPGRADE
  • !The read_many_files tool is removed; functionality moves to a readManyFiles utility available only as a user @-command.
  • !The @qwen-code/webui package has been retired and is no longer available.
Was this useful?

DeepSeek Harness

Sources Commits → 1 RELEASE · 2026-09-08 CODE Rank Why this rank top features scored 70/70/65 — 68.33 How ranking works →

DeepSeek Harness v0.1.5-alpha.1 overhauls file handling with a new resource-addressing system and byte-range reads, reworks the sidebar and docking UI, and adds profile templates plus Playwright video support in its GIF skill.

DeepSeek Harness is an open-source agent harness for building coding agents with plugins.

DeepSeek Harness v0.1.5-alpha.1 overhauls file handling with a new resource-addressing system and byte-range reads, reworks the sidebar and docking UI, and adds profile templates plus Playwright video support in its GIF skill.

└──▷ WHAT SHIPPED · 7 FEATURES 4 OF 7 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Bounded byte-range file reads in fs-local and fs-e2b NEW 70

Adds bounded byte-range reads across local and E2B sandbox filesystems (fs-local and fs-e2b), enabling partial file access without loading full file content.

— Names concrete surfaces but no API syntax given dsh-v0.1.5-alpha.1
02 02
File resource addressing system NEW 70

Adds a dual-face file API with Host-resolved resources and a Client resource registry with retained subscriptions for workspace file operations, plus Session and absolute file resource addresses under workspace-path for addressing files within agent sessions.

— Names components but no usage syntax shown dsh-v0.1.5-alpha.1
03 00
Sidebar file tree navigation IMPROVED 65

Adds a Sidebar with tab navigation, lazy workspace file tree tabs, paged file tabs with retained reader state, and fullscreen shell — replacing the former Details panel for file navigation.

— Describes mechanism and UI change, no exact navigation path given dsh-v0.1.5-alpha.1
04 00
Playwright video support in GIF recording skill IMPROVED 65

Adds support for encoding Playwright video (WebM) input — in addition to ordered screenshots — in the encode_gif.py GIF recording skill.

— Names exact script and format supported dsh-v0.1.5-alpha.1
thinner coverage below
05 03
Dockkit docking engine NEW 55

Adds a reversible docking engine with pointer interactions (dockkit) and a responsive right column with width concessions to the layout.

— Names the engine but not how to invoke it dsh-v0.1.5-alpha.1
06 04
Profile templates via CLI NEW 45

Adds a way to create profiles from shipped templates via the CLI, letting users bootstrap agent profiles from built-in starting points.

— Describes capability without exact command syntax dsh-v0.1.5-alpha.1
07 00
Stats UI redesign IMPROVED 30

Replaces the stats strip with two icon pills and click-open stat dialogs in the client UI.

— Thin one-line UI change with no further detail dsh-v0.1.5-alpha.1
Was this useful?
Agent
◆  AI Agent Frameworks

PydanticAI

Sources Release notes → 1 RELEASE · 2026-09-09 NOTES Rank Why this rank top features scored 50/50, +15 breaking change — 65 How ranking works →

PydanticAI v2.42.0 adds support for GitHub Copilot as a model provider and tightens validation of deferred tool approvals.

PydanticAI is a framework that builds type-safe Python agents with dependency injection, model integrations, tools, and structured outputs.

PydanticAI v2.42.0 adds support for GitHub Copilot as a model provider and tightens validation of deferred tool approvals.

└──▷ WHAT SHIPPED · 2 FEATURES 2 OF 2 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
GitHub Copilot provider integration NEW 50

Adds GitHubCopilotProvider to connect agents to GitHub Copilot's OpenAI-compatible API.

— Names the class but no usage example or config detail. v2.42.0
02 02
Stricter validation of deferred tool approvals BREAKING 50

Invalid DeferredToolResults.approvals values are now rejected, whereas previously they were accepted silently.

— Explains before/after behaviour but no migration steps. v2.42.0
└──▷ BREAKING ON UPGRADE
  • !Invalid DeferredToolResults.approvals values are now rejected (previously accepted silently).
Was this useful?

Vercel AI SDK

Sources Release notes → 1 RELEASE · 2026-09-08 NOTES Rank Why this rank top features scored 55/50 — 52.5 How ranking works →

Vercel AI SDK 7.0.94 adds per-request model overrides for batch calls and refines retry behavior for empty image results.

The Vercel AI SDK provides TypeScript APIs for model generation, structured output, tool use, and streaming application interfaces.

Vercel AI SDK 7.0.94 adds per-request model overrides for batch calls and refines retry behavior for empty image results.

└──▷ WHAT SHIPPED · 2 FEATURES 1 OF 2 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 00
Smarter retry classification for empty image results IMPROVED 55

Unclassified empty image results are now retried using provider-independent retryability classification, while Google and Google Vertex prompt blocks are marked terminal (non-retryable) so they no longer trigger useless retries.

— Explains mechanism and named providers but no code/flag to act on [email protected]
02 01
Per-request model overrides in batch calls NEW 50

The batch call API now supports per-request model overrides, letting each item within a single batch target a different model instead of sharing one model for the whole batch.

— Names the surface (`batch`) but no example or config shown [email protected]
Was this useful?

AWS Strands Agents

Sources Release notes → 2 RELEASES · 2026-09-08 NOTES Rank Why this rank top features scored 78/75/73, +15 breaking change, +6 runnable example — 96.33 How ranking works →

Strands Agents is AWS's open-source SDK for production agents in Python and TypeScript, with a model-driven loop, tool integration, and deployment targets.

Strands Agents shipped parallel Python (v1.55.0) and TypeScript (v1.17.0) releases that deepen MCP 2.x protocol support, add in-process agent-to-agent communication via a new LocalAgent protocol, introduce background task execution, and extend model caching to six providers, while dropping Node.js 20 support in favor of Node.js 22+.

└──▷ WHAT SHIPPED · 15 FEATURES 13 OF 15 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
MCP server loading controls in load_servers NEW 78

load_servers MCP calls now accept prefix_with_server_name and continue_on_error options, letting callers namespace tool names per originating server and tolerate a failed server load without aborting the whole session. Shipped identically in both the Python and TypeScript SDKs.

Prevent tool-name collisions when loading multiple MCP servers by prefixing each tool name with its originating server name.
typescript 
loadServers([
  { name: 'search', url: 'http://localhost:3001' },
  { name: 'files',  url: 'http://localhost:3002' }
], { prefix_with_server_name: true, continue_on_error: true })
Load MCP servers with a name prefix and tolerant error handling so a missing server does not abort the whole session.
python 
from strands.mcp import load_servers

tools = load_servers(
    servers=['filesystem', 'network-scanner'],
    prefix_with_server_name=True,
    continue_on_error=True
)
— Named flags with runnable examples in both SDKs typescript/v1.17.0python/v1.55.0
02 00
tools_ttl field in CacheConfig IMPROVED 75

CacheConfig gained a tools_ttl field to set a time-to-live for cached tools, superseding the now-deprecated cache_tools field.

Set a TTL on cached tools so stale tool definitions are automatically invalidated without touching the broader cache strategy.
typescript 
const agent = new Agent({
  model: new BedrockModel({
    cacheConfig: {
      tools_ttl: 300
    }
  })
})
— Names both old and new fields with a runnable example typescript/v1.17.0python/v1.55.0
03 02
MCP 2.x protocol support expansion NEW 73

MCP 2.x support was expanded with call_tool, server/discover for SEP-2575 server discovery, MRTR (MCP Request/Response) extended to prompts and resources, MRTR input-required support, automatic tool refresh on list-changed notifications, and, in Python, SEP-2663 task support.

— Enumerates every named endpoint/spec but lacks examples typescript/v1.17.0python/v1.55.0
04 03
cache_config support for new model providers NEW 67

cache_config support was added to the Ollama, Writer, SageMaker, llama.cpp, Vercel, and Mistral model providers, extending provider-level tool/response caching to backends that previously had none.

Enable caching on a SageMaker-backed agent to reduce redundant inference costs on repeated tool definitions.
python 
from strands import Agent
from strands.models import SageMakerModel
from strands.models.cache import CacheConfig

model = SageMakerModel(
    endpoint_name='my-endpoint',
    cache_config=CacheConfig(tools_ttl=300)
)
agent = Agent(model=model)
agent('Summarize the threat report.')
— Names all six providers and the config key with an example typescript/v1.17.0python/v1.55.0
05 04
web_fetch vended tool NEW 60

A new web_fetch vended tool lets agents perform HTTP retrieval directly during a session.

Use the new web_fetch vended tool so an agent can retrieve live threat intel or CVE details during a session.
python 
from strands import Agent
from strands_tools import web_fetch

agent = Agent(tools=[web_fetch])
agent('Fetch https://nvd.nist.gov/feeds/json/cve/1.1/recent.json and summarise the top 5 critical CVEs.')
— Named tool with a runnable usage example python/v1.55.0
thinner coverage below
06 00
Reconnect support for bidirectional streaming adapters IMPROVED 52

The OpenAI and Gemini Google Live bidirectional streaming model adapters gained a reconnect() method to recover dropped streaming sessions, and bidirectional sessions now support proactive reconnect ahead of drops.

— Names the method and two adapters, no example typescript/v1.17.0python/v1.55.0
07 05
Background task execution in Python and TypeScript NEW 52

Python agents gained an internal in-process task engine and manager, and the TypeScript Agent class gained a backgroundTasks option backed by a new BackgroundTaskManager interface and InProcessTaskManager implementation.

— Names classes and field but no usage shown python/v1.55.0
08 06
Context manager offloading and stash support NEW 48

The context manager gained an L1 stash for durable storage of offloaded content and session support for the L1 context manager (TypeScript); Python ported the offloading strategies and stash integration; context manager types are now exported as experimental.

— Names several mechanisms but no example or config key typescript/v1.17.0python/v1.55.0
09 07
LocalAgent protocol for in-process agent-to-agent calls NEW 45

A new LocalAgent protocol and implementation support in-process, bidirectional (bidi) streaming communication between agents, added to both the Python and TypeScript SDKs.

— Names the protocol but no usage example provided typescript/v1.17.0python/v1.55.0
10 08
Classifier routing strategy in TypeScript NEW 40

A classifier routing strategy was ported from the Python SDK to TypeScript, enabling model routing decisions based on request classification.

— Names the strategy but no mechanism or example given typescript/v1.17.0
11 09
Node.js 22+ now required BREAKING 40

Node.js 20 is no longer supported; the TypeScript SDK now requires Node.js 22+.

— Clear before/after but no migration steps given typescript/v1.17.0python/v1.55.0
12 10
QmdSearchStrategy BM25 full-text search NEW 38

A new QmdSearchStrategy adds BM25 full-text search to the storage layer.

— Names the class but no mechanism detail or example python/v1.55.0
13 11
AudioCapable interface for bidi streaming NEW 30

An AudioCapable interface and typed model config were added to the TypeScript bidirectional streaming layer.

— Bare name, no explanation of behaviour typescript/v1.17.0
14 12
Bedrock API key authentication NEW 25

Amazon Bedrock model provider now supports API key authentication.

— Single-line mention with no mechanism or example python/v1.55.0
15 13
Multiple continuation inputs in Python agents NEW 23

Python agents now support multiple continuation inputs.

— Thin one-line mention with no further detail python/v1.55.0
└──▷ BREAKING ON UPGRADE
  • !Node.js 20 is no longer supported; the TypeScript SDK now requires Node.js 22+.
  • !Node.js 20 is no longer supported; Node.js 22+ is now required.
Was this useful?

Agno (formerly Phidata)

Sources Release notes → 1 RELEASE · 2026-09-08 NOTES Rank Why this rank top features scored 83/81/58 — 74 How ranking works →

Agno v3.0.8 adds bounded full-page reads to Knowledge, PostgreSQL engine factories with pool and serialization defaults, and a shared Markdown code-fence utility.

Agno is an agent framework and runtime that orchestrates and runs multi-agent systems.

Agno v3.0.8 adds bounded full-page reads to Knowledge, PostgreSQL engine factories with pool and serialization defaults, and a shared Markdown code-fence utility.

└──▷ WHAT SHIPPED · 3 FEATURES 3 OF 3 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Full-page reads for Knowledge NEW 83

Adds Knowledge.read_full_page and aread_full_page methods that return an entire published page from a single bounded SQL read, supporting revision pinning, character limits, and deadlines; oversized pages return None.

— Named methods with clear mechanism and bounds, but no usage example. v3.0.8
02 02
PostgreSQL engine factory functions NEW 81

Adds create_postgres_engine and create_async_postgres_engine factory functions that expose Agno's connection-pool and JSON serialization defaults with configurable SQLAlchemy options; plain postgres:// and postgresql:// URLs automatically select Psycopg 3.

— Named functions and driver-selection behaviour, no runnable example given. v3.0.8
thinner coverage below
03 03
Shared Markdown code-fence tracking utility NEW 58

Exposes agno.utils.markdown.advance_code_fence to share the page chunker's fence-tracking rules with application transforms that need to preserve code examples.

— Names the exact utility path but describes purpose only briefly. v3.0.8
Was this useful?

OpenClaw

Sources Release notes →Source code → 1 RELEASE · 2026-09-08 NOTES CODE Rank Why this rank top features scored 90/86/83, +15 breaking change, +6 runnable example — 107.33 How ranking works →

Personal AI assistant platform that connects messaging channels, tools, and model providers through a gateway.

OpenClaw v2026.9.3 ships public session transcript sharing, a searchable meeting library, an optional Team Reports plugin, and default bounded recursive session spawning, alongside a wide set of CLI, config, and client platform additions plus several breaking SDK and Node version changes.

└──▷ WHAT SHIPPED · 27 FEATURES 25 OF 27 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Plugin SDK breaking API changes BREAKING 90

The execution-policy SDK moves retired exec-mode and comparator helpers from infra-runtime to execPolicy on openclaw/plugin-sdk/agent-harness-runtime; callers must use resolveExecModePolicy and select only the returned fields they need. The approval SDK requires importing approval account-resolution helpers from approval-native-runtime, adapting session filtering to the full matchesApprovalRequestFilters contract, and replacing the retired generic forwarding evaluator with native channel route gates and shared predicates. SDK alias channel-inbound.buildChannelTurnMediaPayload is replaced by buildChannelInboundMediaPayload, and the named type AbortAndDrainAgentHarnessRunResult is removed — infer the result of the retained abortAndDrainAgentHarnessRun callable instead. Search result callbacks must read bounded Find/Grep text from details.content instead of the retired details.truncation.content (the outer tool-message content is unchanged), and directory result callbacks must use LsToolDetails.content and optional nextAfter in place of the retired truncation and entryLimitReached fields, passing nextAfter as after to continue listing.

— Names every retired and replacement API surface verbatim. v2026.9.3
02 02
Workshop skills migrated to per-agent collections BREAKING 86

Workshop skills ownership moves from workspace-owned to one writable collection per agent; the config key skills.workshop.allowSymlinkTargetWrites is retired. Running openclaw doctor --fix migrates proven legacy skills automatically, leaving skills with ambiguous ownership in place for manual review.

Migrate legacy workspace-owned Workshop skills to the new per-agent collection after upgrading.
$ openclaw doctor --fix
— Names exact migration command and retired config key. v2026.9.3
03 03
Node version support requirements raised BREAKING 83

Node 22, Node 25, and earlier 24.x/26.x builds are no longer supported; upgrade to Node 24.16.0 or newer on the 24.x line, or Node 26.1.0 or newer, before upgrading OpenClaw, or SQLite text will be silently truncated. Node-based CLI/Gateway installs on macOS 11–13.4 or official Linux ARMv7 no longer have a supported host.

— Gives exact version thresholds and consequence of ignoring them. v2026.9.3
04 04
`gateway.cliAgents.enabled` config key NEW 81

New config key gateway.cliAgents.enabled (default true) controls whether supported CLI agents appear in the new-session model picker; set gateway.cliAgents.enabled: false to disable CLI agents and native CLI session creation.

Disable CLI agents on a shared Gateway where only UI-created sessions should be allowed.
yaml 
gateway.cliAgents.enabled: false
— Names exact config key, default, and effect. v2026.9.3
05 05
Team Reports plugin NEW 78

Optional Team Reports plugin generates authenticated daily, weekly, and monthly GitHub reports and explicitly selected Discord sources, with stored history, optional summaries, people timelines and calendars, coverage warnings, scheduler/source health, and theme-aware report pages.

— Rich feature list but no config keys or commands given. v2026.9.3
06 06
Searchable meeting library NEW 75

Adds a searchable meeting library: browse paginated saved notes, search full transcripts, export complete Markdown or JSONL archives, and edit capture sources in Communications settings.

— Names export formats and settings location. v2026.9.3
07 07
Code Mode sandboxed encoding and TypeScript diagnostics NEW 75

Code Mode adds sandboxed TextEncoder and TextDecoder instances that survive wait/resume for local text and byte transformations, TypeScript checks against available tools, a running JavaScript environment retained across fast tool replies, and error reporting with original-source locations and bounded console output.

— Names exact APIs and behaviours added to Code Mode. v2026.9.3
08 08
`openclaw docs --limit` flag NEW 73

Adds a --limit flag to openclaw docs to cap the number of displayed terminal or JSON documentation results returned by the CLI, e.g. openclaw docs --limit 20.

Limit CLI doc output to a manageable number of results when piping plugin documentation into another tool.
$ openclaw docs --limit 20
— Names exact flag and shows runnable command. v2026.9.3
09 00
Typed MCP GetPromptResult from getPrompt IMPROVED 65

SessionMcpRuntime.getPrompt now returns the typed MCP GetPromptResult shape, preserving messages and image content, and validates external responses instead of returning arbitrary unknown values.

— Names exact API and type, no usage example given. v2026.9.3
10 09
Public session transcript sharing NEW 65

Session owners and Gateway admins can publish existing and future conversation text to a public URL accessible to anyone, and revoke access later. The read-only public view omits tools, reasoning, files, images, and executable widgets.

— Clear scope and exclusions, no exact UI path given. v2026.9.3
11 10
Runtime identity and GC duration metrics NEW 65

Adds Prometheus runtime identity metrics so operators can verify which process and build is serving a deployment, plus garbage-collection duration metrics that identify slow phases while opening agent databases.

— Names Prometheus and metric purpose, no metric names given. v2026.9.3
12 11
Repository-backed cloud sessions NEW 63

Supports creating cloud sessions from a repository URL and ref, including checkout, setup, and recoverable checkpoints owned by the cloud node.

— Explains mechanism, no command or config key named. v2026.9.3
13 12
Native dashboard reports via `show_widget` NEW 62

Adds native dashboard report display through show_widget and dashboard authoring without an iframe.

— Names the exact `show_widget` surface used. v2026.9.3
14 13
Provider account controls in Models settings NEW 60

Models settings gains controls to add and remove individual provider accounts, manage supported account priority, and clear an agent's custom order without disconnecting accounts.

— Names UI area and controls, no exact command. v2026.9.3
15 14
Live browser tab repainting for agent pages NEW 60

Adds live browser tab repainting for agent pages, with screenshot fallback when streaming is unavailable; on macOS, external links open as native WebKit Mac tabs in the Browser panel that survive chat switches.

— Explains fallback and platform behaviour, no config surface. v2026.9.3
thinner coverage below
16 15
Model fallback picker NEW 58

A model fallback picker lets users choose ordered fallback models from a searchable dropdown, remove selections, and add custom model references.

— Describes UI mechanism without exact navigation path. v2026.9.3
17 16
Bounded update repair and consent-based failure reports NEW 58

Candidate-validation failures during updates can enter a bounded repair phase using configured inference in disposable rehearsal state, with failure or rollback outcomes retained when repair cannot recover. Update failure reports are submitted only after explicit consent, using the recorded update attempt and current owner.

— Explains mechanism in prose, no exact command or flag. v2026.9.3
18 00
Telegram photo album grouping IMPROVED 55

Groups consecutive eligible Telegram photo attachments into native albums of up to ten while preserving order, captions, reply targets, and topics.

— Concrete limit of ten, but automatic behaviour with no user action. v2026.9.3
19 17
Android folding layout support NEW 53

Keeps content clear of hinges, shows navigation beside content on book folds, and splits transcript and composer across tabletop panes on Android foldables.

— Describes behaviour but no way to trigger manually. v2026.9.3
20 18
ARM64 Linux AppImage and macOS Gateway cards NEW 52

Adds ARM64 Linux AppImage packaging and live Gateway cards with direct window opening in the macOS Gateways menu.

— Names packaging format and menu location, thin on mechanism. v2026.9.3
21 19
Bounded recursive session spawning NEW 50

Recursive session spawning is enabled by default, bounded by explicit depth and concurrency limits, with existing sandbox restrictions retained.

— Describes limits generally but no concrete numbers or flags. v2026.9.3
22 20
Chrome extension setup on Mac NEW 50

Adds Chrome extension setup on Mac: request the official Chrome Store extension from local Mac device settings or CLI once native setup is ready.

— Names two entry points but no exact command. v2026.9.3
23 21
Device aliasing in Control UI NEW 45

Adds device aliasing: rename paired devices directly from the Control UI Devices page.

— Names exact UI page for the feature. v2026.9.3
24 22
iOS 'Start Live Voice' shortcut NEW 45

Adds an iOS voice shortcut: the 'Start Live Voice' App Shortcut opens the current chat and starts Talk.

— Names the exact App Shortcut. v2026.9.3
25 23
Live activity session status display NEW 44

Shows running and queued permitted sessions in Live activity on entry and reconnect, with a visible limit notice and connection status.

— Describes UI state, no exact navigation given. v2026.9.3
26 24
Position rail for long conversations NEW 42

Adds a previewable position rail for long conversations, including hover previews and direct jumps to points in the conversation.

— Thin UI description, no exact location named. v2026.9.3
27 25
Instant path filtering in folder browser NEW 30

Adds instant path filtering in the folder browser as you type.

— Bare description of a small UI convenience. v2026.9.3
└──▷ BREAKING ON UPGRADE
  • !Node 22, Node 25, and earlier 24.x/26.x builds are no longer supported; upgrade to Node 24.16.0 or newer (on the 24.x line) or Node 26.1.0 or newer before upgrading OpenClaw, or SQLite text will be silently truncated. Node-based CLI/Gateway installs on macOS 11–13.4 or official Linux ARMv7 no longer have a supported host.
  • !Execution-policy SDK: the retired exec-mode and comparator helpers move from infra-runtime to execPolicy on openclaw/plugin-sdk/agent-harness-runtime; callers must use resolveExecModePolicy and select only the returned fields they need.
  • !Approval SDK: import approval account-resolution helpers from approval-native-runtime, adapt session filtering to the full matchesApprovalRequestFilters contract, and replace the retired generic forwarding evaluator with native channel route gates and shared predicates.
  • !SDK alias channel-inbound.buildChannelTurnMediaPayload is replaced by buildChannelInboundMediaPayload; the named type AbortAndDrainAgentHarnessRunResult is removed — infer the result of the retained abortAndDrainAgentHarnessRun callable instead of importing it.
  • !Search result callbacks: read bounded Find/Grep text from details.content instead of the retired details.truncation.content; the outer tool-message content is unchanged.
  • !Directory result callbacks: use LsToolDetails.content and optional nextAfter in place of the retired truncation and entryLimitReached fields; pass nextAfter as after to continue listing.
  • !Workshop skills: workspace ownership is replaced by one writable collection per agent; the config key skills.workshop.allowSymlinkTargetWrites is retired. Run openclaw doctor --fix to migrate proven legacy skills; ambiguous ownership is left in place for manual review.
Was this useful?
Deploy
◆  AI Model & Data Infrastructure

emisar

Sources Release notes → 1 RELEASE · 2026-09-09 NOTES Rank Why this rank top features scored 70/65/60, +15 breaking change — 80 How ranking works →

Emisar v0.47.0 adds new Stripe and Braintree packs and extends the Airflow pack to cover both Airflow 2 and Airflow 3, alongside a round of access-control changes for Owners, SSO SSO Single Sign-On: an authentication scheme where one login session grants access to multiple applications, sparing tools from managing their own credential stores and letting them delegate identity to a central provider. groups, and shared reads plus runner and console reliability fixes.

Emisar lets AI agents securely run infrastructure actions through MCP, with approval controls that protect production systems.

Emisar v0.47.0 adds new Stripe and Braintree packs and extends the Airflow pack to cover both Airflow 2 and Airflow 3, alongside a round of access-control changes for Owners, SSO SSO Single Sign-On: an authentication scheme where one login session grants access to multiple applications, sparing tools from managing their own credential stores and letting them delegate identity to a central provider. groups, and shared reads plus runner and console reliability fixes.

└──▷ WHAT SHIPPED · 12 FEATURES 5 OF 12 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 00
Airflow 2 and 3 support in Airflow pack IMPROVED 70

The Airflow pack now supports Airflow 2 (API v1) and Airflow 3 (API v2), with jobs, assets, and backfills requiring Airflow 3.

— States exact API versions and which features require Airflow 3 v0.47.0
02 01
Runner setup key 24-hour expiration BREAKING 65

Existing unused console-generated runner setup keys are assigned a 24-hour lifetime; any old expired key must be replaced with a new setup command.

— Gives exact lifetime and required remediation action v0.47.0
03 00
SSO group role and access management on provider page IMPROVED 60

SSO groups now show role, runner, and pack access together on the provider page, with in-place role editing and member filtering.

— Names a UI location and specific controls added v0.47.0
04 00
Approval recovery error differentiation IMPROVED 60

Approval recovery distinguishes catalog failures, unavailable actions, and changed contracts, preserving entered notes, inputs, and unsaved runbook edits during runner/pack access changes.

— Details multiple failure cases and what state is preserved v0.47.0
thinner coverage below
05 02
Stripe and Braintree billing packs NEW 50

Adds new Stripe and Braintree packs covering billing investigation, refunds, disputes, and related corrections.

— Names the two packs and use cases but no mechanism v0.47.0
06 03
Owner-wide action access normalization BREAKING 50

Owners always have workspace-wide action access; existing Owner scopes are normalized with an audit record. This normalization is audited and does not restore narrower grants on rollback.

— Explains behavior and rollback caveat but no config surface v0.47.0
07 00
Windows Hermes and Goose setup paths IMPROVED 40

Windows Hermes and Goose setup uses native configuration paths, with setup guidance now including co:op.

— Names tools and a new guidance addition, no detail on mechanism v0.47.0
08 00
Runner credential rotation and dispatch journal durability IMPROVED 40

Runner credential rotation uses the authenticated connection, and the dispatch journal appends durable transitions with periodic compaction.

— Describes mechanism but no exposed interface v0.47.0
09 00
Console runbook plan review retention IMPROVED 40

Console runbook starts retain the release and plan reviewed; changed plans require another review.

— Describes behavior change with modest specificity v0.47.0
10 04
Shared reads no longer authorize actions BREAKING 35

Shared reads no longer authorize execution, cancellation, approval, or access changes.

— Clear restriction but no migration path given v0.47.0
11 05
Operational role workspace-wide activity visibility NEW 30

Operational roles can inspect workspace-wide activity while actions remain limited by current runner and pack access.

— Single sentence with no mechanism or surface named v0.47.0
12 00
Credential-aware GCP project discovery IMPROVED 25

GCP project discovery follows the current credentials.

— One-line change with no further detail v0.47.0
└──▷ BREAKING ON UPGRADE
  • !Existing unused console-generated runner setup keys are assigned a 24-hour lifetime; any old expired key must be replaced with a new setup command.
  • !Owner access normalization is audited and does not restore narrower grants on rollback.
  • !Rolling back the new group-access constraints can fail if newer independent grants still depend on them.
Was this useful?

OpenAI TypeScript SDK

Sources Release notes → 2 RELEASES · 2026-09-08 → 2026-09-09 NOTES Rank Why this rank top features scored 45/35/30 — 36.67 How ranking works →

The OpenAI TypeScript SDK provides typed JavaScript and TypeScript clients for OpenAI APIs and streaming responses.

This window's OpenAI TypeScript SDK updates added GPT Image 2.5 model support with new image options, plus API-level additions for prompt cache diagnostics, service-account key expiration, and web search call status handling.

└──▷ WHAT SHIPPED · 4 FEATURES 3 OF 4 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 00
Incomplete status for web search calls IMPROVED 45

Recognizes incomplete as a valid web search call status in streaming and response types.

— Names the exact status value and affected types. v7.11.0
02 01
GPT Image 2.5 model support NEW 35

Adds support for GPT Image 2.5 models along with new image options in the API.

— Names the model but not the new options or usage. v7.12.1
03 02
Prompt cache diagnostics in responses NEW 30

Adds prompt cache diagnostics to API responses, exposing cache performance data.

— Describes what was added but not the response fields or format. v7.11.0
04 03
Service-account API key expiration fields NEW 25

Adds service-account API key expiration fields to the SDK.

— Only mentioned in summary, no field names or mechanism given. v7.11.0
Was this useful?

XGrammar

Sources Release notes → 1 RELEASE · 2026-09-09 NOTES Rank Why this rank top features scored 81/75/33, +6 runnable example — 69 How ranking works →

XGrammar v0.2.6 adds new grammar constructs and format controls, a cached Lark compilation API, expanded structural tag support for more model families, and new backend/platform coverage including an NPU backend and Windows ARM64 wheels.

XGrammar constrains language-model output to grammars and schemas with optimized token-mask generation.

XGrammar v0.2.6 adds new grammar constructs and format controls, a cached Lark compilation API, expanded structural tag support for more model families, and new backend/platform coverage including an NPU backend and Windows ARM64 wheels.

└──▷ WHAT SHIPPED · 6 FEATURES 6 OF 6 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Token and character limits for text/tokens formats NEW 81

Adds max_tokens and max_chars support for AnyTextFormat and AnyTokensFormat, letting callers cap structured generation output length directly via these format parameters without post-processing.

Cap structured generation output by token count using the new max_tokens parameter on AnyTokensFormat to enforce a hard budget without post-processing.
python 
import xgrammar as xgr

fmt = xgr.AnyTokensFormat(max_tokens=128)
— Named parameters and formats with a runnable example v0.2.6
02 02
Cached compile_lark API for Lark grammars NEW 75

XGrammar adds a cached GrammarCompiler.compile_lark API so that compiling the same Lark grammar across requests reuses the cached result instead of recompiling, reducing latency in high-throughput serving.

Use the cached compile_lark API to avoid recompiling the same Lark grammar on every request, reducing latency in high-throughput serving.
python 
import xgrammar as xgr

compiler = xgr.GrammarCompiler(tokenizer_info)
grammar = compiler.compile_lark(lark_grammar_str)
— Named API with runnable example, but caching mechanism only briefly described v0.2.6
thinner coverage below
03 03
Lark regex flags and structured substring expressions NEW 33

XGrammar adds Lark regex flags and structured substring expressions to grammar definitions, expanding the expressiveness of Lark-based grammar syntax.

— Named additions but no mechanism or usage shown v0.2.6
04 04
Structural tag support for Cohere XML, Kimi K3, EXAONE 4.0 NEW 33

Adds structural tag support for Cohere XML, Kimi K3, and EXAONE 4.0 model formats.

— Names three supported formats but no detail on mechanism or usage v0.2.6
05 05
Windows ARM64 wheel support NEW 28

Adds Windows ARM64 wheel support for XGrammar.

— Simple platform packaging addition, thinly described v0.2.6
06 06
NPU token-bitmask backend NEW 23

Adds an NPU token-bitmask backend to XGrammar.

— Bare mention of a new backend with no further detail v0.2.6
Was this useful?

OpenAI Python SDK

Sources Release notes → 1 RELEASE · 2026-09-09 NOTES Rank Why this rank top features scored 45/30 — 37.5 How ranking works →

The OpenAI Python SDK provides typed Python clients for OpenAI model APIs, streaming, and asynchronous requests.

OpenAI Python SDK v3.10.0 adds support for GPT Image 2.5 models with new image options and introduces service-account API key expiration fields for programmatic key lifetime management.

└──▷ WHAT SHIPPED · 2 FEATURES 2 OF 2 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Service-account API key expiration fields NEW 45

Adds service-account API key expiration fields to the API, enabling programmatic inspection and management of key lifetimes.

— Explains purpose but no field names or endpoint given v3.10.0
02 02
GPT Image 2.5 model support NEW 30

The SDK adds support for new GPT Image 2.5 models, including new image options.

— Named only in summary, no options or mechanism detailed v3.10.0
Was this useful?

HeyGen HyperFrames

Sources Release notes → 1 RELEASE · 2026-09-08 NOTES Rank Why this rank top features scored 47 — 47 How ranking works →

HyperFrames v0.8.32 expands its linting and validation checks to catch more animation and rendering pitfalls before they reach production.

HyperFrames is an open-source HTML-to-video renderer that runs in AI-agent workflows.

HyperFrames v0.8.32 expands its linting and validation checks to catch more animation and rendering pitfalls before they reach production.

└──▷ WHAT SHIPPED · 1 FEATURE 1 OF 1 FEATURE most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
New lint and check rules for animation pitfalls NEW 47

Adds lint rules flagging autoAlpha on clip elements, hidden-style opacity guards, undefined GSAP color variables, and repeated fromTo state leaks, plus a new Check that detects connectors pointing at nothing and stylesheets that leak into the frame.

— Names each rule but no usage or config detail given v0.8.32
Was this useful?

Fireworks AI

Sources Release page → 1 RELEASE · 2026-09-08 NOTES Rank Why this rank top features scored 90/80/75, +15 breaking change, +6 runnable example — 102.67 How ranking works →

Fireworks AI provides hosted inference, model fine-tuning, and deployment APIs for open-weight models.

Fireworks AI expanded its deployment and access-control surface this window, adding deployment tagging (with a new firectl CLI and a breaking custom/ annotation namespace requirement), per-user spending limits, bring-your-own provider keys, custom SSO/SCIM, and a new US-only serverless endpoint for GLM 5.3 Flash.

└──▷ WHAT SHIPPED · 8 FEATURES 8 OF 8 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Deployment tags for dedicated deployments NEW 90

Adds Deployment Tags to attach customer-defined metadata to dedicated deployments, manageable via Fireworks Nexus UI (Deployments > select deployment > 'Deployment Tags'), and via firectl 1.8.3's deployment tag set, unset, and list subcommands, including atomic batch operations for managing tags.

Attach metadata to a dedicated deployment for cost attribution or environment tagging.
📍Navigate to Deployments, open a dedicated deployment, and add customer-defined key-value pairs under 'Deployment Tags'.
— Named CLI subcommands, version, and UI path make it fully actionable. Deployment tags and annotation API changes
02 02
custom/ prefix now required for REST API annotations BREAKING 80

The REST API now requires customer-managed annotation keys to be prefixed with custom/ (e.g. custom/environment); GetDeployment and ListDeployments now return only custom/* entries for regular account users, and writes using bare keys (e.g. environment) now return HTTP 403 PERMISSION_DENIED. Existing reads relying on bare keys will silently return no data for those keys.

— Full before/after and exact error code, but no migration command given. Deployment tags and annotation API changes
03 03
US-only serverless endpoint for GLM 5.3 Flash NEW 75

Adds accounts/fireworks/routers/glm-5p3-flash-us as a new US-only serverless routing endpoint for the GLM 5.3 Flash model, for data-residency-compliant inference.

Route inference requests to GLM 5.3 Flash via the US-only serverless endpoint for data-residency compliance.
$ curl -X POST https://api.fireworks.ai/inference/v1/chat/completions \
  -H 'Authorization: Bearer <API_KEY>' \
  -H 'Content-Type: application/json' \
  -d '{"model": "accounts/fireworks/routers/glm-5p3-flash-us", "messages": [{"role": "user", "content": "Hello"}]}'
— Named model route with a runnable curl example. product docs
04 04
Per-user spending limits via Fireworks Nexus NEW 70

Adds per-user spending limits on serverless inference via Fireworks Nexus, supporting account-level defaults, group limits, and per-user overrides, configured under Usage Limits.

Enforce a spending cap for a specific end-user on serverless inference to prevent runaway costs in a multi-tenant application.
📍Navigate to Fireworks Nexus › Usage Limits, set an account-level default, then override for an individual user or group under 'Per-User Usage Limits'.
— UI navigation path given but no API or config key. product docs
05 05
Bring Your Own provider keys for FireRouter NEW 65

Adds Provider Keys (Bring Your Own) support so FireRouter can call external providers using customer-supplied API keys, registered via Fireworks Nexus > Provider Keys.

Route inference requests through your own API keys for external providers via FireRouter without exposing them to end users.
📍Navigate to Fireworks Nexus › Provider Keys, select 'Bring Your Own', and enter your provider API key to register it for FireRouter use.
— UI steps given, no endpoint or key format named. product docs
thinner coverage below
06 06
Custom SSO and SCIM provisioning NEW 40

Adds Custom SSO authentication and SCIM user and group provisioning for Fireworks AI accounts.

— Named capability only; no setup steps or config given. product docs
07 07
Training cost estimator for Managed Training NEW 35

Adds a training cost estimator for Managed Training, with cost comparisons between Fireworks Serverless, Dedicated, and Tinker.

— Names comparison targets but no interface or usage detail. product docs
08 08
RL rollout cost comparison tool NEW 35

Adds an RL rollout cost comparison tool for estimating multi-turn agentic RL inference costs between Fireworks and Tinker.

— Brief description with no steps or interface named. product docs
└──▷ BREAKING ON UPGRADE
  • !REST API writes using bare annotation keys (e.g. environment) now return HTTP 403 (PERMISSION_DENIED); keys must be migrated to the custom/ prefix (e.g. custom/environment).
  • !GetDeployment and ListDeployments now omit any annotation keys outside the custom/* namespace for regular account users — existing reads relying on bare keys will silently return no data for those keys.
Was this useful?
◆  Local LLM Runtimes

MLX Serve

Sources Release notes → 1 RELEASE · 2026-09-09 NOTES Rank Why this rank top features scored 83/80/80, +10 product media, +6 runnable example — 97 How ranking works →

The mlx-serve runtime provides Apple Silicon inference with OpenAI- and Anthropic-compatible APIs, chat, agent mode, and tool calling.

mlx-serve v26.9.2 adds per-model configuration, multi-provider model registration, disk-backed prefix caching, and new speculative-decoding controls, alongside faster constrained JSON decoding and a chat UI width toggle.

└──▷ WHAT SHIPPED · 8 FEATURES 7 OF 8 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Multi-provider model registration NEW 83

Adds ~/.mlx-serve/providers.json for configuring OpenAI-compatible external providers (cloud APIs, remote Macs, local runtimes), whose models appear in the picker as <model>@<name>; the same registration is available via Settings > Providers UI to register any OpenAI-compatible chat server with its key.

— Names config file, naming convention, and UI location v26.9.2
02 02
Disk-backed prefix cache spill NEW 80

Adds --prefix-cache-disk flag to spill inactive conversation prefix caches to SSD, keeping the active model and conversation in memory while restoring others in seconds.

Run a long-context Flash Next server where inactive conversations spill to disk and speculation is cut off past 32k tokens to avoid diminishing returns.
$ mlx-serve launch flash-next --prefix-cache-disk --max-mtp-ctx 32768
— Names flag and mechanism, no size limits given v26.9.2
03 03
Per-model settings for context, cache, speculation NEW 80

Adds ~/.mlx-serve/model-settings.json for per-model configuration of context size, KV cache precision, and speculative-decoding defaults, applied at load or hot-applied to a running model; the same settings are editable via right-click on a model in My Models > Model Settings.

— Names config file and exact UI path for all fields v26.9.2
04 04
Default reply token budget flag NEW 70

Adds --max-tokens N serve-mode flag to set a default reply token budget for clients that omit max_tokens.

— Exact runnable flag but no default value stated v26.9.2
05 05
Speculative decoding context cutoff NEW 65

Adds --max-mtp-ctx <n> flag to set a hard context-length cutoff beyond which speculative decoding is disabled.

Run a long-context Flash Next server where inactive conversations spill to disk and speculation is cut off past 32k tokens to avoid diminishing returns.
$ mlx-serve launch flash-next --prefix-cache-disk --max-mtp-ctx 32768
— Exact runnable flag but no default or threshold guidance v26.9.2
thinner coverage below
06 00
Full-speed constrained JSON decoding on Flash Next IMPROVED 55

Delivers constrained (JSON schema) decoding at full model speed on Flash Next, replacing a previous ~1 token/second cap.

— Gives before/after numbers but no invocation steps v26.9.2
07 06
Adaptive speculative decoding on Flash Next NEW 50

Adds adaptive speculative decoding that automatically disables and re-enables speculation on Flash Next based on per-step cost/benefit measurement.

— Explains mechanism but nothing for reader to invoke v26.9.2
08 07
Chat UI width control NEW 50

Adds chat UI width control via Narrow/Medium/Wide settings, accessible from Settings or F1–F3 keyboard shortcuts.

— Named shortcuts, but a minor cosmetic addition v26.9.2
└──▷ ALSO FROM THESE RELEASES
26 9 1-vs-26 9 2imageimage
Was this useful?

vMLX

Sources Release notes → 1 RELEASE · 2026-09-08 NOTES Rank Why this rank top features scored 65/60/50 — 58.33 How ranking works →

vMLX 1.6.56 adds MiniCPM5 native tool/reasoning support, per-request image/video controls with diagnostics, atomic safetensors alignment repair during model preflight, and a redesigned Console Amber UI with new navigation and settings grouping.

The vMLX server runs compressed MLX models on Apple Silicon with disk caching, paged memory, continuous batching, and hybrid SSM scheduling.

vMLX 1.6.56 adds MiniCPM5 native tool/reasoning support, per-request image/video controls with diagnostics, atomic safetensors alignment repair during model preflight, and a redesigned Console Amber UI with new navigation and settings grouping.

└──▷ WHAT SHIPPED · 7 FEATURES 3 OF 7 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Atomic safetensors alignment repair in preflight NEW 65

Adds atomic on-disk safetensors alignment repair during local model preflight, including nested draft/media shards, with validation and unsafe-path protections.

— Clear mechanism and scope of the repair process v1.6.56
02 00
Console Amber UI redesign and settings grouping IMPROVED 60

Adds new Chat & Images, Servers & API, and Models navigation sections in the Console Amber UI, with an updated app icon and General Preferences; server settings now group concurrent processing, SSD prefix caching, power, generation, MCP, and speculative decoding controls together.

— Names exact new nav sections and settings groups v1.6.56
thinner coverage below
03 02
MiniCPM5 native tool and reasoning support NEW 50

Adds MiniCPM5 native tool/reasoning integration with guarded affine projection handling.

— Names model and mechanism but no usage path v1.6.56
04 03
Per-request image/video controls and diagnostics NEW 50

Adds per-request image/video controls and typed strict-mode errors across API dialects, plus effective-setting diagnostics for image/video API requests.

— Describes new API behaviour but no exact field names v1.6.56
05 00
Hybrid cache corrections and SSD cleanup IMPROVED 40

Adds hybrid partial-prefix and media-position cache corrections, idle SSD cleanup, and capacity-only notices.

— Names cache mechanisms but no config surface v1.6.56
06 00
Tool/JSON parsing and Ollama streaming fixes IMPROVED 35

Tool/JSON parsing now preserves string data and terminal punctuation; Ollama generation streams retain final usage without repeating content.

— Describes fix behaviour but no reproduction steps v1.6.56
07 00
JANG model feed filtering IMPROVED 25

JANG model feeds now filter out non-JANG repositories.

— Single-line change with no further detail v1.6.56
Was this useful?
Data
◆  Vector DBs & RAG

AWS Context Ontology Accelerator

Sources Commits → 1 RELEASE · 2026-09-08 CODE Rank Why this rank top features scored 85 — 85 How ranking works →

Context Ontology Accelerator v0.3.0 adds a Databricks connector for Athena Federation Athena Federation An AWS open-source framework that lets you query data in external sources—S3, databases, APIs—using standard SQL through Amazon Athena, without moving the data first. , letting COA query Databricks Unity Catalog Unity Catalog An open-source data governance layer maintained by Databricks that provides unified access control, auditing, and lineage tracking across tables, files, and ML models stored in a lakehouse. via federated SQL through a new CDK CDK A framework, maintained by AWS, for defining cloud infrastructure as code and synthesizing it into CloudFormation templates; tools use it to provision and manage AWS resources programmatically. construct.

Context Ontology Accelerator provides an open-source semantic layer that helps AI agents make consistent, explainable decisions using formal ontologies and knowledge graphs.

Context Ontology Accelerator v0.3.0 adds a Databricks connector for Athena Federation Athena Federation An AWS open-source framework that lets you query data in external sources—S3, databases, APIs—using standard SQL through Amazon Athena, without moving the data first. , letting COA query Databricks Unity Catalog Unity Catalog An open-source data governance layer maintained by Databricks that provides unified access control, auditing, and lineage tracking across tables, files, and ML models stored in a lakehouse. via federated SQL through a new CDK CDK A framework, maintained by AWS, for defining cloud infrastructure as code and synthesizing it into CloudFormation templates; tools use it to provision and manage AWS resources programmatically. construct.

└──▷ WHAT SHIPPED · 1 FEATURE 1 OF 1 FEATURE most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Databricks connector for Athena Federation NEW 85

A new AthenaFederationConnector CDK construct, shipped in the coa-connector-cdk package, lets you build and deploy COA connectors as Lambda-backed stacks using helper functions functionNamePrefix, optionalEnv, optionalIntEnv, queryRoleArns, and requiredEnv. It includes an InformationSchemaReader that reads Databricks Unity Catalog's information_schema to surface column comments and key constraints that standard JDBC metadata handlers omit, enabling federated SQL queries against Unity Catalog through Athena Federation.

— Names construct, package, and five helper functions but no usage example. v0.3.0
Was this useful?
Evaluate
◆  AI Observability & Evals

LangChain LangSmith

Sources Release page → 1 RELEASE · 2026-08-20 NOTES Rank Why this rank top features scored 95/90/90, +15 breaking change, +6 runnable example — 112.67 How ranking works →

LangSmith provides tracing, evaluation, and deployment tools for LLM applications.

LangSmith's release centers on evaluator tooling and annotation workflows, adding CLI LLM-as-judge evaluator creation, a new annotation-queue items endpoint supporting mixed RUN/THREAD batches, and a paginated experiment-comparison API that retires legacy dataset-comparison helpers, alongside bulk dataset-split management in the experiment UI and new permission controls.

└──▷ WHAT SHIPPED · 10 FEATURES 7 OF 10 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Annotation queue items endpoint for runs and threads NEW 95

Adds POST /annotation-queues/<id>/items endpoint to add RUN items to a single annotation queue, with the server resolving runs via ClickHouse or SmithDB and returning a standards-shaped items envelope; the endpoint also supports item_type THREAD (with thread_id + session_id) so conversation threads can be queued, with mixed RUN and THREAD batches allowed in one request. Requests are capped at a maximum of 200 items per request, returning a clear validation error when the limit is exceeded.

Add a batch of conversation threads to an annotation queue for human review, mixing RUN and THREAD items in a single request.
$ curl -X POST 'https://api.smith.langchain.com/annotation-queues/<queue_id>/items' \
  -H 'Content-Type: application/json' \
  -d '{"items": [{"item_type": "THREAD", "thread_id": "<thread_id>", "session_id": "<session_id>"}, {"item_type": "RUN", "run_id": "<run_id>"}]}'
— Endpoint, resolution mechanism, and item limit all named snapshot-20260909
02 02
CLI evaluator creation and in-place updates NEW 90

Adds langsmith evaluator create-llm CLI subcommand to define structured LLM-as-judge evaluator rules from a prompt, schema, and model config file, targeting a project or dataset. The evaluator upload --replace flag on the LangSmith CLI now updates existing code evaluator rules in place, avoiding a delete-before-create window if the replacement upload fails.

— Names exact CLI subcommand and flag with their behavior snapshot-20260909
03 03
Thread evaluator testing before save NEW 90

POST /runs/rules/validate now accepts test_thread_id and session_id to test a thread evaluator against a real conversation before saving. Multi-turn evaluators also gain a Test action in the UI to run the evaluator against a sample thread before committing the rule.

Test a multi-turn thread evaluator against a real conversation before committing the rule, to confirm it grades correctly on live data.
$ curl -X POST 'https://api.smith.langchain.com/runs/rules/validate' \
  -H 'Content-Type: application/json' \
  -d '{"test_thread_id": "<thread_id>", "session_id": "<session_id>", ...}'
— Endpoint parameters and matching UI action both named snapshot-20260909
04 04
Legacy feedback formula endpoints scheduled for removal DEPRECATED 80

The legacy feedback formula endpoints POST/GET /feedback/formulas and GET/PUT/DELETE /feedback/formulas/{feedback_formula_id} are scheduled for removal on 2026-08-20; existing feedback formulas must be migrated to composite evaluators before that date.

— Exact endpoints and removal date given snapshot-20260909
05 05
Paginated experiment comparison API replaces legacy helpers BREAKING 75

POST /v2/datasets/<id>/experiment-runs is now the supported public API for paginated experiment comparison; legacy dataset comparison helper endpoints are removed from the public OpenAPI spec and generated SDKs, requiring callers to migrate to the new endpoint.

— New and removed endpoints named, no migration steps given snapshot-20260909
thinner coverage below
06 06
Download datasets permission NEW 55

Introduces a separate download datasets permission, enforced in the application and APIs; the download button is disabled in the UI for users who lack it.

— Names permission but not full enforcement scope snapshot-20260909
07 00
Overflow dropdown for config chips in Experiments table IMPROVED 55

Model, prompt, and tool chips in the Experiments table config cells now feature a clickable +N overflow badge dropdown exposing filter, group-by, open-in-playground, and details actions.

— UI action named, no deeper mechanism described snapshot-20260909
08 00
Image thumbnails and text truncation in tables IMPROVED 45

Dataset and experiment tables now detect and render base64 images as small thumbnails with a delayed larger preview, and truncate long input and reference-output text, avoiding oversized hidden DOM content.

— Describes behavior, no config or limits named snapshot-20260909
09 07
User-defined monthly trace limits NEW 35

LangSmith now enforces user-defined monthly trace limits scoped to individual projects and users.

— No config surface or default numbers specified snapshot-20260909
10 00
Longer execution time for code evaluators IMPROVED 30

Code evaluators now receive more execution time per batch, reducing timeouts for evaluators that import heavy libraries such as scikit-learn.

— No concrete time limit or numbers given snapshot-20260909
└──▷ BREAKING ON UPGRADE
  • !The legacy feedback formula endpoints (POST/GET /feedback/formulas, GET/PUT/DELETE /feedback/formulas/{feedback_formula_id}) are scheduled for removal on 2026-08-20; migrate existing feedback formulas to composite evaluators before that date.
  • !Legacy dataset comparison helper endpoints are removed from the public OpenAPI spec and generated SDKs; use POST /v2/datasets/<id>/experiment-runs for paginated experiment comparison instead.
Was this useful?

Weights & Biases

Sources Release notes → 1 RELEASE · 2026-09-09 NOTES Rank Why this rank top features scored 85/80/75, +15 breaking change, +6 runnable example — 101 How ranking works →

Weights & Biases tracks experiments, sweeps, artifacts, and model versions for training runs, and hosts the registry teams promote models through.

Weights & Biases v0.30.0 adds a non-interactive login check, a new ARIA automation action, and several LEET TUI TUI A text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. usability improvements, alongside breaking removals of legacy API classes and CLI sync flags.

└──▷ WHAT SHIPPED · 7 FEATURES 7 OF 7 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Non-interactive login credential check NEW 85

Adds wandb.login(prompt=False) to check whether the user is logged in using configured credentials, returning False without prompting for an API key or switching to offline mode — useful for guarding scripts in CI or automated pipelines.

Guard a script block that needs W&B credentials without blocking on an interactive prompt — useful in CI or automated pipelines.
python 
import wandb

if not wandb.login(prompt=False):
    raise SystemExit("W&B credentials not configured. Set WANDB_API_KEY or run `wandb login`.")

with wandb.init(project="my-project") as run:
    run.log({"metric": 0.99})
— Runnable example shows exact usage and behavior v0.30.0
02 02
Console log filtering in LEET TUI NEW 80

Adds console log filtering to the W&B LEET TUI: with the logs pane focused, press / and type a regex pattern, or press Tab then type a glob, to show only matching lines; new matches are followed live and ctrl+/ clears the filter.

— Exact key bindings and behavior given, no code example v0.30.0
03 03
wandb sync CLI flag changes BREAKING 75

The wandb sync --sync-all flag has been removed; run wandb sync with no arguments instead. The wandb sync --job_type flag (underscore) has been removed; use --job-type (hyphen) instead.

— Exact flag names and migration path given v0.30.0
04 04
Custom x-axis charts in LEET TUI NEW 70

Adds support for charts with custom x-axes set via run.define_metric(...) in the W&B LEET TUI, including glob definitions such as run.define_metric('train/*', step_metric='train/step'); applies to runs opened from local .wandb files.

— Names the API call but no full walkthrough v0.30.0
05 05
Removal of legacy Api classes BREAKING 70

wandb.InternalApi, wandb.PublicApi, wandb.apis.PublicApi, wandb.apis.internal, and wandb.sdk.internal.internal_api have been removed; replace all usages with wandb.Api().

— Names all removed classes and the replacement v0.30.0
06 06
Persisted filters in LEET TUI NEW 65

Persists metrics, system metrics, and runs filters per wandb directory in .wandb-leet.json, restoring them automatically the next time the directory is opened in either the workspace or single-run view.

— Names storage file but no interaction steps v0.30.0
thinner coverage below
07 07
SendPromptToAria automation action NEW 45

Adds SendPromptToAria to the automations API, enabling sending a prompt to ARIA as an automation action.

— Named but no usage example or config shown v0.30.0
└──▷ BREAKING ON UPGRADE
  • !wandb.InternalApi, wandb.PublicApi, wandb.apis.PublicApi, wandb.apis.internal, and wandb.sdk.internal.internal_api have been removed; replace all usages with wandb.Api().
  • !The wandb sync --sync-all flag has been removed; run wandb sync with no arguments instead.
  • !The wandb sync --job_type flag (underscore) has been removed; use --job-type (hyphen) instead.
Was this useful?

Arize Phoenix

Sources Release notes → 3 RELEASES · 2026-09-08 NOTES Rank Why this rank top features scored 91/55 — 73 How ranking works →

Arize Phoenix added error/latency filtering to its trace-retrieval APIs and expanded its integration and model-provider surface with AG2, MiniMax, ATIF/Harbor tracing, and Meta's Muse Spark models.

Arize Phoenix is an open-source platform that monitors LLM application traces and evaluates their outputs.

Arize Phoenix added error/latency filtering to its trace-retrieval APIs and expanded its integration and model-provider surface with AG2, MiniMax, ATIF/Harbor tracing, and Meta's Muse Spark models.

└──▷ WHAT SHIPPED · 2 FEATURES 2 OF 2 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Error and latency filters on trace retrieval NEW 91

Adds error and latency filter parameters to get_traces in the Python and TypeScript clients for targeted trace retrieval. This was extended with error, minLatencyMs, and maxLatencyMs filter parameters on getTraces in the JS client, matching the query parameters on GET /v1/projects/{id}/traces (requires Phoenix server >= 20.8.0).

— Names exact params, method and path, and version requirement @arizeai/[email protected]arize-phoenix-client-v3.5.0
thinner coverage below
02 02
New integrations and model providers NEW 55

Phoenix added several new integration and provider surfaces: AG2 integration for tracing and observability within the platform, MiniMax as a supported LLM provider, ATIF tracing support added to the Harbor plugin, and META as a built-in OpenAI-compatible model provider enabling Meta Muse Spark models to be used directly in Phoenix without manual provider configuration.

— Names each provider/integration but gives no setup steps @arizeai/[email protected]arize-phoenix-v20.9.0arize-phoenix-client-v3.5.0
Was this useful?

PromptLayer

Sources Release page → 1 RELEASE · 2026-09-08 NOTES Rank Why this rank top features scored 75/23, +6 runnable example — 55 How ranking works →

PromptLayer added support for Google Gemini 3.8 Flash across three providers with configurable reasoning effort, and fixed video playback in the Playground's fullscreen view.

PromptLayer is a platform for managing, evaluating, and observing prompts and LLM applications.

PromptLayer added support for Google Gemini 3.8 Flash across three providers with configurable reasoning effort, and fixed video playback in the Playground's fullscreen view.

└──▷ WHAT SHIPPED · 2 FEATURES 1 OF 2 FEATURES most completely described first
what'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.

  • depth0–40 what it does and how it works or what changed
  • specificity0–30 names real surfaces — APIs, flags, formats, limits, numbers
  • actionability0–30 enough 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.

01 01
Gemini 3.8 Flash support with reasoning effort levels NEW 75

PromptLayer added support for Google Gemini 3.8 Flash across Google AI Studio, Vertex AI, and OpenRouter, with configurable reasoning effort levels. In the Playground, users can select 'Gemini 3.8 Flash' under any of the three providers and set the reasoning effort level (e.g. 'high') in the model configuration panel before running a prompt.

Select Gemini 3.8 Flash with high reasoning effort in the Playground to tackle complex agent workflows requiring long-context reasoning.
📍In the Playground, open the model selector and choose 'Gemini 3.8 Flash' under the Google AI Studio, Vertex AI, or OpenRouter provider. Set the reasoning effort level to 'high' in the model configuration panel before running your prompt.
— Names providers and exact selection steps, lacks effort level enumeration September 08, 2026
thinner coverage below
02 00
Video playback in Playground fullscreen view IMPROVED 23

Enhances the Playground fullscreen view to properly display video content and playback controls.

— Bare description of a fix, no mechanism or navigation given September 08, 2026
Was this useful?
// END OF TRANSMISSION

The AI Toolchain · curated daily · September 9, 2026
Releases & features from the open-source and commercial AI tools that practitioners actually run.

▷ Subscribe
my-toolchain — 0 tools
paste an install list to detect your tools

A brew list, a Brewfile, requirements.txt, a Dockerfile — or just the product names, free-form. Nothing leaves your browser.

    browse all tools →

    Send us feedback