Issue 022 — The AI Toolchain

Letta Code

2 RELEASES · 2026-09-09 → 2026-09-10 NOTES Rank

Letta Code's latest releases add CLI model management commands, smarter headless agent routing, proactive learning from feedback, safer message queue handling on interrupt, and live credential renewal for Desktop.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

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

Letta Code's latest releases add CLI model management commands, smarter headless agent routing, proactive learning from feedback, safer message queue handling on interrupt, and live credential renewal for Desktop.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 5 FEATURES 3 OF 5 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

Model inspection and switching CLI commands NEW 67

Adds model get, model list, and model set subcommands to the CLI for inspecting and switching models without entering the interactive session.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names exact subcommands but no example of use v0.32.1

02 02

Resume queue for interrupted messages NEW 60

Adds resume_queue to allow queued user messages parked on interrupt to be resumed without dropping them.

— Names the exact mechanism but no usage example given v0.32.0

thinner coverage below

03 00

Ambient agent inference for headless routing IMPROVED 58

Infers the ambient agent automatically when using --computer for headless routing, removing the need to specify --agent explicitly in common cases.

Send a headless message through a named remote computer without specifying an agent ID, letting Letta Code infer the ambient agent automatically.

$ letta -p --computer "work-laptop" "run the nightly audit"

— Concrete flag plus a runnable example command v0.32.1

04 03

Proactive learning from human feedback NEW 31

Teaches agents proactive learning from human feedback, enabling prompt and memory updates driven by conversational corrections.

— Described only in prose, no mechanism or surface named v0.32.1

05 00

Live Desktop credential renewal IMPROVED 28

The server now accepts renewed Desktop credentials without requiring a restart.

— Behaviour stated but no config key or command shown v0.32.0

Was this useful?

Cline

1 RELEASE · 2026-09-10 NOTES CODE Rank

Cline desktop v0.0.25 focuses on account and provider ergonomics — a dedicated session-import page, clearer provider/model pickers, expanded local-CLI authentication, and a refreshed model catalog including a breaking change to ChatGPT Subscription (Codex) model limits.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

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

Cline desktop v0.0.25 focuses on account and provider ergonomics — a dedicated session-import page, clearer provider/model pickers, expanded local-CLI authentication, and a refreshed model catalog including a breaking change to ChatGPT Subscription (Codex) model limits.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 6 FEATURES 3 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

ChatGPT Subscription (Codex) model picker overhaul BREAKING 85

The Codex model picker now lists only models your plan can actually use, with gpt-5.6-terra as the new default.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Per-model context caps of 400K/272K/128K replace the previously inherited API limit of 1.05M, and the retired models gpt-5.4 and gpt-5.4-mini are removed, so accounts using either will have their default model changed.

— Names exact model IDs, caps and retirement, but no migration command desktop-v0.0.25

02 02

Local CLI provider authentication expanded NEW 70

Providers that authenticate through a local CLI — Claude Code and Codex CLI — can now start sessions without an API key. OpenCode is now also treated as a local CLI provider, showing the local CLI notice instead of a non-functional browser sign-in button, and authenticates using credentials stored by the opencode CLI.

— Explains mechanism and names all three CLI providers desktop-v0.0.25

03 03

Model catalog additions and default refresh NEW 65

Adds NaN (nan.builders) to the model catalog and refreshes default models for 36 providers, including Bedrock, Vertex, OpenRouter, Kilo, GitHub Copilot, Gemini, Cerebras, Fireworks, Requesty, and Vercel AI Gateway.

— Names new provider and lists many affected providers, no per-provider detail desktop-v0.0.25

thinner coverage below

04 00

Session import moved to dedicated Settings page IMPROVED 55

Session import from Claude Code, Codex, and opencode now has its own dedicated page in Settings instead of being a row buried in General.

— Clear before/after and named sources, but only a UI location desktop-v0.0.25

05 00

Model picker distinguishes same-named models IMPROVED 45

The model picker now distinguishes models that share a name, and Cline Pass subscription models are listed separately from the free fallback tier.

— Describes the change but not the underlying mechanism desktop-v0.0.25

06 00

Provider picker shows configured providers IMPROVED 35

The composer's provider picker now marks which providers you have already configured.

— Bare description with no mechanism or exact location desktop-v0.0.25

└──▷ BREAKING ON UPGRADE

!ChatGPT Subscription (Codex) accounts using gpt-5.4 or gpt-5.4-mini as their model will find those models removed; the default moves to gpt-5.6-terra.

Was this useful?

Diagram Design

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 gains ten new chart types, a headless-render linter, named client profiles, and native Droid plugin packaging.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ GET THIS VERSION

$ git clone --branch commits-2026-08-12 https://github.com/cathrynlavery/diagram-design.git
# already have the repo? check out this version:
$ git checkout commits-2026-08-12

└──▷ TRY IT

Catch clipped or collapsed SVG diagrams in CI before they reach reviewers — runs headless Chromium and diffs overflow-released screenshots against the authored viewport.

$ python3 scripts/lint-render.py

Run the adversarial sankey verification suite to confirm the checker fires on real defects (ribbon narrowing, node volume loss, label mismatch) and stays quiet on clean files.

$ python3 scripts/test-verify-sankey.py

›Adds scripts/lint-render.py, a headless-Chromium linter that screenshots diagrams at authored viewport size and with overflow released, diffs the two to catch clipped SVG paint, collapsed SVGs, sideways page overflow, and runtime errors that source-level linting cannot detect.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
›Adds verify-beeswarm.py with nine geometric invariants — shared value scale, no-overprint, and more — verifying that every drawn dot sits at its declared value on the linear axis and that crowding is resolved only by swarm offset, never by moving a dot along the value axis.
›Adds test-verify-sankey.py, an adversarial test suite for verify-sankey.py that covers both polarities (fires when it should, stays quiet when it should not) across ribbon narrowing, node volume loss, stage carry-loss, label/bar mismatch, and dark-variant drift.
›Adds a treemap type for part-of-whole by area (feat(types): add treemap).
›Adds a dumbbell as a Bar variant (feat(bar): add dumbbell).

+8 more show less

›Adds a slopegraph variant for change between two states (feat(line): add slopegraph).
›Adds a ridgeline as a Line variant (feat(line): add ridgeline).
›Adds a bubble variant to the scatter type for three-value comparisons (feat(scatter): add bubble).
›Adds named client profiles (Add named client profiles).
›Adds automatic plugin updates via native marketplaces and a version gate (feat(plugin): automatic updates via native marketplaces and a version gate).
›Adds animated HTML examples for Semantic Pattern #1 (Fan-in Queue / Bottleneck) as example-queue-animated.html and Semantic Pattern #5 (Secure Paved Road) as example-paved-road-animated.html.
›Adds semantic patterns and accessible motion ([verified] add semantic patterns and accessible motion).
›Deploys a live gallery to GitHub Pages, publishing skills/diagram-design/assets/ as a static site on push to main or via workflow_dispatch.

13 details in Everything

Was this useful?

DeepSeek Harness

1 RELEASE · 2026-09-10 CODE Rank

DeepSeek Harness's rc.1 release adds a new default model, DeepSeek-V41-Flash, with text and image input support, and adds a small guide UI enhancement for short entry lists.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

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

DeepSeek Harness's rc.1 release adds a new default model, DeepSeek-V41-Flash, with text and image input support, and adds a small guide UI enhancement for short entry lists.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

DeepSeek-V41-Flash as new default model BREAKING 73

Adds a deepseek-flash model entry (DeepSeek-V41-Flash) to the default catalog with inputModalities: ['text', 'image'] and systemPromptUpdate: 'in-history', and sets it as the new default for Chat Completions, replacing the previous DeepSeek-V4-Flash default.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Any integration relying on the previous default without explicitly setting a model ID will now target the new model.

— Names exact model ID, config fields and breaking default change dsh-v0.1.5-rc.1

thinner coverage below

02 02

Optional description field in sidebar guide entries NEW 57

Adds an optional description field to SidebarRightGuideEntry for a one-line description shown beneath a capsule title in the guide, displayed when the entry list is short.

— Names the exact field and its display condition dsh-v0.1.5-rc.1

└──▷ BREAKING ON UPGRADE

!The default Chat Completions model is now DeepSeek-V41-Flash (deepseek-flash); any integration that relied on the previous default (DeepSeek-V4-Flash) without explicitly setting a model ID will now target the new model.

Was this useful?

OpenAI Codex CLI

1 RELEASE · 2026-09-09 NOTES CODE Rank

Codex CLI v0.154.0 adds experimental worktree isolation for parallel sessions, inline question answering, a new GPT-6-Astra model, a /copy command, and shared background sessions on Windows, while removing the deprecated codex mcp-server entry point.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

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

Codex CLI v0.154.0 adds experimental worktree isolation for parallel sessions, inline question answering, a new GPT-6-Astra model, a /copy command, and shared background sessions on Windows, while removing the deprecated codex mcp-server entry point.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 6 FEATURES 5 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

Experimental worktree isolation for sessions NEW 85

Adds experimental worktree support via the --worktree flag or /worktree command to create isolated checkouts for new or forked sessions, with browse and resume capabilities so risky refactors or parallel experiments don't touch the main checkout.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Run a task in an isolated worktree so the main checkout is never touched — useful for risky refactors or parallel experiments.

$ codex --worktree 'Refactor the auth module to use the new token interface'

— Named flag/command with a runnable example and mechanism. rust-v0.154.0

thinner coverage below

02 02

/copy command for status and session output NEW 55

Adds the /copy command to copy status output or individual session fields, with rich-text formatting preserved when copying responses.

— Named command gives a clear starting point. rust-v0.154.0

03 00

Windows background daemon sharing IMPROVED 45

Enables Windows sessions to share a background Codex server, with daemon lifecycle commands and managed updates.

— Names the mechanism but not specific commands. rust-v0.154.0

04 03

Inline question answering during sessions NEW 40

Enables inline question answering — you can answer questions with suggested choices or custom text while Codex continues working, without losing your main draft.

— Explains behavior but no command or UI path given. rust-v0.154.0

05 04

GPT-6-Astra model added NEW 30

Adds GPT-6-Astra to the model picker and Amazon Bedrock catalogs.

— Names the model and catalogs but no usage detail. rust-v0.154.0

06 05

Removal of codex mcp-server entry point BREAKING 30

The codex mcp-server entry point is no longer available, having been previously deprecated.

— Names the removed entry point but gives no migration path. rust-v0.154.0

└──▷ BREAKING ON UPGRADE

!The codex mcp-server entry point is no longer available (previously deprecated).

Was this useful?

Charm Crush

1 RELEASE · 2026-09-09 NOTES CODE Rank

Crush v0.93.1 lets users disable mouse support in the TUI and shows which model served each conversation turn when using Hyper.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Terminal-based AI coding assistant that integrates LLMs, LSPs, and MCP servers into your development workflow.

Crush v0.93.1 lets users disable mouse support in the TUI and shows which model served each conversation turn when using Hyper.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Mouse support toggle in TUI NEW 80

Adds option ui mouse false to crushrc and { "options": { "tui": { "mouse": false } } } to crush.json to disable mouse support in the TUI, so terminal text selection and right-click work natively.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Also toggleable via the command palette.

Disable mouse support so terminal text selection and right-click work natively during a Crush session.

$ option ui mouse false

— Exact config keys and command given v0.93.1

thinner coverage below

02 02

Routed model display for Hyper turns NEW 30

Shows the routed model that served each conversation turn when using Hyper.

— Describes behavior but no UI path or mechanism v0.93.1

Was this useful?

mex

1 RELEASE · 2026-09-09 CODE Rank

mex 0.8.1 adds a visual Context graph to the Project Hub, an Inbox for proposing Wiki knowledge changes, open-team Relays with a new draft-save command, configurable agent logging modes, and code graph ignore globs.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Mex indexes codebases through a code graph and supplies a repo-local Markdown wiki to AI coding agents.

mex 0.8.1 adds a visual Context graph to the Project Hub, an Inbox for proposing Wiki knowledge changes, open-team Relays with a new draft-save command, configurable agent logging modes, and code graph ignore globs.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Configurable agent logging modes NEW 90

Adds mex logging command with significant (default), checkpoints, and manual modes for checkout-local agent logging preferences, also configurable in Hub Settings.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Adds AgentLoggingModeSchema (significant, checkpoints, manual), AgentLoggingPolicySchema, and AgentLoggingUpdateRequestSchema to the Hub contracts API.

Switch agent logging to checkpoint mode to capture more granular activity without full verbose output.

$ mex logging checkpoints

— Command, mode names, and API schemas all named v0.8.1

02 02

Open-team Relays and draft workflow NEW 85

Adds mex relay draft save --from <draft.json> to save open-team Relay local drafts through the existing signed workflow, with Hub and CLI showing audience, current eligibility, and the sharing boundary. Relays can now be open-to-team so eligible active Members can take them, including teammates who join later, and local drafts may leave recipients undecided.

Save an open-team Relay draft from a JSON file so any eligible active member can pick it up.

$ mex relay draft save --from handoff-draft.json

— Names exact command and describes eligibility mechanism v0.8.1

03 03

Graph ignore globs for code indexing NEW 75

Adds additive Graph ignore globs under the graph.ignore key in .mex/config.json, with repository-relative validation that behaves consistently across platforms.

Exclude generated or vendored directories from the code graph to reduce noise in who-calls results.

json

{
  "graph": {
    "ignore": ["vendor/**", "dist/**", "**/__generated__/**"]
  }
}

— Exact config key and file path given with example v0.8.1

04 04

Inbox contributions to Wiki knowledge NEW 65

Adds Inbox contributions for one addition or correction to existing architecture, component, convention, decision, pattern, or guide knowledge; local drafts publish as Git-shareable Markdown proposals and explicit approval writes the existing Wiki knowledge.

— Workflow explained but no command or UI path shown v0.8.1

05 05

Context graph in Project Hub NEW 60

Adds a Context graph in the Project Hub showing existing Wiki entities, relationships, and direct code groundings, with type filters, selection details, pan/zoom, and a list alternative.

— UI feature described but no exact navigation path v0.8.1

thinner coverage below

06 06

Member reactivation and bounded Timeline notes NEW 35

Adds member reactivation with the original identity; managed agent instructions now retrieve relevant bounded Timeline notes without automatically promoting those notes to accepted project knowledge.

— Thin description, no mechanism or interface detail v0.8.1

Was this useful?

Anthropic Claude Code

1 RELEASE · 2026-09-09 NOTES CODE Rank

Claude Code v2.1.267 adds a flag for live system-prompt iteration and a cross-provider cap on model effort level.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

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

Claude Code v2.1.267 adds a flag for live system-prompt iteration and a cross-provider cap on model effort level.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Live system prompt iteration flag NEW 85

The --system-prompt-snapshot off flag makes Claude Code re-render the system prompt fresh on every request instead of replaying the recorded snapshot, enabling live iteration on prompt text without restarting the session.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Iterate on a system prompt without restarting the session — each request re-reads the prompt as written rather than replaying the snapshot recorded at session start.

$ claude --system-prompt-snapshot off -p 'Refactor the auth module'

— Names exact flag and runnable example command. v2.1.267

02 02

maxEffortLevel cap across providers NEW 65

A new maxEffortLevel setting, configurable top-level or per model under modelSettings, caps the effort level across all providers including Bedrock, Vertex, and Foundry, while still letting users select a lower level.

— Names config key and scope, but no example usage. v2.1.267

Was this useful?

All Hands AI OpenHands

1 RELEASE · 2026-09-09 NOTES Rank

OpenHands is an open-source software-development agent that writes code, runs commands, and browses the web.

OpenHands v1.17.0 tightens automation access control with split view/manage permissions, adds LLM provider connections for cloud deployments, and rounds out Agent Canvas with conversation tagging, custom manifests, and several automation-management UI refinements.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 5 FEATURES 4 OF 5 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

View/manage permission split for automations NEW 60

Automation permissions are now split into separate 'view' and 'manage' roles, with a creator escape hatch that automatically grants automation creators full manage rights over automations they created regardless of their team-wide role.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Give automation creators full control over their own automations while restricting other team members to view-only access.

📍In the automations settings, assign 'view' permission to team members and 'manage' permission to automation owners; the creator escape hatch automatically grants creators full manage rights over automations they created regardless of team-wide role.

— Mechanism and roles named, with a worked example v1.17.0

thinner coverage below

02 00

Agent Canvas automation management enhancements IMPROVED 50

Canvas gains support for generic custom automation manifests, lets users edit a custom cron expression directly from the edit-automation modal, shows manifest-declared value statements on dashboard cards and rows, makes the automation run UI task-outcome aware, and adds a unified cloud settings entrypoint for Agent Canvas.

— Five distinct UI/manifest additions named but only briefly v1.17.0

03 02

Conversation tags and filtering NEW 30

Adds conversation tags and filtering in the conversation panel.

— Names the panel but no detail on how tags work v1.17.0

04 03

LLM provider connections on cloud deployments NEW 25

Enables LLM provider connections on cloud deployments.

— Only a one-line description, no mechanism given v1.17.0

05 04

Message timestamps on hover NEW 15

Displays message timestamps on hover.

— Single-line description with no further detail v1.17.0

Was this useful?

Superset

1 RELEASE · 2026-09-09 NOTES Rank

Superset is an agentic IDE to orchestrate 100+ coding agents in parallel. Run any agent with your own subscription.

Superset's desktop v1.28.0 release adds MCP plugin integrations for agents, cloud workspace creation, iOS Live Activities showing agent status, and a round of polish across Pages, mobile terminals, and the paywall flow.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 9 FEATURES 5 OF 9 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

Host version display and remote updates NEW 50

The desktop app now shows host versions and supports updating standalone hosts directly from the app.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Describes both viewing and updating hosts desktop-v1.28.0

02 02

MCP plugin integrations skill for agents NEW 45

Adds an integrations skill that lets the agent call plugin MCP tools directly from within its workflow.

— Names MCP mechanism but no usage steps given desktop-v1.28.0

03 00

PR summary opens as workspace pane IMPROVED 45

The PR summary can now be opened as a workspace pane directly from the Changes pane.

— Clear UI navigation path but no further mechanism desktop-v1.28.0

04 03

iOS Live Activities for agent status NEW 45

Shows agent status on the iOS Lock Screen and Dynamic Island.

— Names specific iOS surfaces but no interaction detail desktop-v1.28.0

05 00

Pages product polish: onboarding, comments, theming IMPROVED 40

Adds Pages onboarding and polishes the page creation flows, splits comment mode out from the comments panel, and adds theming for published pages.

— Groups thin Pages increments but lacks mechanism detail desktop-v1.28.0

06 04

Shareable link for automations NEW 40

Users can copy a shareable link to an automation.

— Concrete action but no detail on link scope or permissions desktop-v1.28.0

07 05

Cloud workspace creation from sidebar NEW 40

A cloud workspace can be created directly from the sidebar.

— Clear entry point but no further mechanism described desktop-v1.28.0

08 00

Mobile terminal and session persistence IMPROVED 40

On mobile, switched-to terminals are kept warm with a cached buffer, and the app now remembers the last session tab per workspace.

— Groups two mobile session behaviours, both described briefly desktop-v1.28.0

09 00

Paywall flow closed via Stripe outcomes IMPROVED 30

Closes the paywall funnel through Stripe payment outcomes.

— Names Stripe but no mechanism or flow detail desktop-v1.28.0

Was this useful?

Zed

1 RELEASE · 2026-09-09 NOTES Rank

Zed is an open-source code editor that runs collaborative development workflows.

Zed v1.19.2 adds call hierarchy navigation, a tabular data preview for CSV/TSV/SSV/PSV files, and a batch of new panel and settings-UI options, while flipping project search on type to enabled by default.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 10 FEATURES 9 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

Call hierarchy navigation NEW 85

Adds call hierarchy support via call hierarchy: show incoming calls and call hierarchy: show outgoing calls commands, with default keybindings cmd-k cmd-h on macOS and ctrl-k ctrl-h on Linux and Windows.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Navigate a codebase call graph to trace who calls a function and who it calls, without leaving the editor.

📍Open the command palette and run call hierarchy: show incoming calls (macOS: cmd-k cmd-h) to see all callers of the symbol under the cursor.

— Full commands, keybindings, and a runnable usage example. v1.19.2

02 02

New panel and file-tab settings NEW 85

Adds outline_panel.multi_buffer_hide_symbols setting and outline panel: toggle symbols action to show only files and directories in multi-buffer Outline Panel views; reveal_if_open setting to reveal buffers already open in another pane instead of opening them in the active pane; folder_indicator setting for showing disclosure chevrons and folder icons together in the Project Panel, Outline Panel, and Git Panel; close_panel_on_toggle setting added to the Settings UI; and Open File on Remote and Copy Remote File URL options added to file tabs and the Project Panel.

Hide symbols in the Outline Panel when working across a multi-buffer view to reduce noise and focus on file structure.

json

{ "outline_panel": { "multi_buffer_hide_symbols": true } }

— Names every setting key and action, one has a runnable config example. v1.19.2

03 03

Tabular data preview for delimited files NEW 75

Adds cmd-shift-v and cmd-k v on macOS and ctrl-shift-v and ctrl-k v on Linux and Windows to open a tabular data preview for CSV, TSV, SSV, and PSV files.

— Exact keybindings and supported formats given, no deeper mechanism. v1.19.2

04 04

Project search on type enabled by default BREAKING 70

Project search on type is now enabled by default, automatically triggering search as you type; set { "search": { "search_on_type": false } } in ~/.config/zed/settings.json to restore the previous manual-invocation behavior.

Suppress search triggering automatically as you type in a project-wide search, restoring manual invocation.

json

{ "search": { "search_on_type": false } }

— Exact config key and file path to revert the default behavior. v1.19.2

05 05

Modal editing selection and text objects NEW 60

Adds Visual Line selection with V in terminal vi mode, and Helix mode support for closest-pair text objects with m i m and m a m.

— Exact keys named but no further mechanism or scope described. v1.19.2

thinner coverage below

06 00

Command palette recency-based ordering IMPROVED 40

Improves command palette ordering by sorting commands by recency of use, ranking previously used commands above unused ones while filtering.

— Describes the sorting mechanism but no configurable surface. v1.19.2

07 06

LSP 3.18 Markdown diagnostics rendering NEW 35

Adds support for rendering LSP 3.18 Markdown diagnostic messages from language servers.

— Names the protocol version but is an automatic, non-actionable change. v1.19.2

08 07

Renaming Terminal Threads NEW 30

Adds support for renaming Terminal Threads from the Threads Sidebar.

— Names the UI location but no further detail. v1.19.2

09 08

Effort-based reasoning for OpenRouter NEW 25

Adds support for effort-based reasoning when using OpenRouter models.

— Names the provider but not the mechanism or configuration. v1.19.2

10 09

Automatic language detection for untitled buffers NEW 20

Adds automatic language detection for untitled buffers.

— Bare description with no mechanism or scope given. v1.19.2

└──▷ BREAKING ON UPGRADE

!Project search on type is now enabled by default; set { "search": { "search_on_type": false } } to restore the previous behavior.

Was this useful?

gridctl

1 RELEASE · 2026-09-10 NOTES Rank

Gridctl is a command-line tool that manages and routes Model Context Protocol servers, skills, and tools.

gridctl reached v1.0.0-rc.1, turning into a full stack-as-code MCP gateway: gridctl apply deploys a YAML stack while new commands add client auto-linking, catalog search, token-reduction analysis, trust-on-first-use schema pinning, a SKILL.md skill library, containerized Python MCP servers from PyPI or Git, model routing, and Sigstore -signed release provenance.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 21 FEATURES 21 OF 21 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

Client auto-linking with `gridctl link` NEW 95

gridctl link, gridctl link <client>, and gridctl link --all auto-detect installed LLM clients (Claude Desktop, Claude Code, Cursor, Antigravity, Grok) and inject gateway config pointing at the gateway's MCP endpoint, exposed at http://localhost:8180/mcp as streamable HTTP with an SSE endpoint, connectable by any MCP client.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ A link: block in stack.yaml idempotently relinks the listed clients on every gridctl apply, and gridctl destroy --unlink tears down the stack while removing injected client config entries.

— Commands, config key, and endpoint all named with mechanism. product docs

02 02

Encrypted variable store and scoped delivery NEW 93

gridctl var explain <key> shows a variable's resolution, declaration, and consumers without exposing its value, and gridctl var scan --staged checks for exact secret values before commit. Variables are referenced as ${var:<KEY>} in a secrets.sets block in stack.yaml, delivered from an encrypted-at-rest store, with reserved-name enforcement for GRIDCTL_*, OP_CONNECT_TOKEN, and OP_SERVICE_ACCOUNT_TOKEN; scoped variable delivery now enables per-scope secrets and variables across stacks.

— Commands, config block, syntax, and reserved names all named. v1.0.0-rc.1

03 03

Token-reduction findings via `gridctl optimize` NEW 93

gridctl optimize, with --format json and --severity warn,critical flags, surfaces actionable token-reduction findings — unused servers/tools, schema overhead, format-conversion shortfalls — with projected weekly token impact and paste-ready YAML remediations; --format json emits a machine-readable OptimizeReport.

Audit a running gateway for wasted tokens before a long agent run — narrow to only warn and critical findings and get machine-readable output for downstream processing.

$ gridctl optimize --severity warn,critical --format json

Surface high-impact token-reduction opportunities across the running gateway and pipe the structured report into a CI artifact.

$ gridctl optimize --format json --severity warn,critical > optimize-report.json

— Exact flags, output format, and remediation mechanism all named. product docs

04 04

Trust-on-first-use schema pinning NEW 93

gridctl pins verify (exit 1 on drift), gridctl pins diff <server> (per-tool before/after plus injection scan findings), and gridctl pins approve <server> implement trust-on-first-use pinning of tool definitions, flagging drift on apply and scanning for injection signals such as hidden instructions, sensitive-file references, hidden Unicode, and cross-server tool shadowing.

Verify that no upstream MCP server has silently changed a tool schema or injected hidden instructions since the last pin — fail CI if drift is detected.

$ gridctl pins verify

Detect prompt-injection drift in a GitHub MCP server's tool definitions before an agent ever sees the changed schema.

$ gridctl pins diff github

— Commands, exit codes, and specific injection signals all named. product docs

05 05

Output format conversion for tool results NEW 90

The output_format config key, settable at the gateway: level and per-server under mcp-servers:, converts tool call results to toon, csv, json, or text before they reach the client, cutting token consumption by 25–61% for tabular and key-value data; payloads over 1 MB pass through unchanged.

Cut token consumption on a high-volume analytics server by converting its JSON responses to CSV at the gateway level, while leaving other servers on the default format.

yaml

gateway:
  output_format: json
mcp-servers:
  - name: analytics
    output_format: csv

Reduce downstream token cost for a server returning large tabular payloads by converting its output to CSV before it reaches the client.

yaml

gateway:
  output_format: toon
mcp-servers:
  - name: analytics
    output_format: csv

— Config key, all four formats, and exact reduction numbers named. product docs

06 06

SKILL.md skill library and projection NEW 88

gridctl skill list, gridctl activate <skill>, and gridctl skill project sync manage a SKILL.md registry exposed to upstream MCP clients as prompts, with projection to Claude Code, OpenCode, Copilot, and Gemini CLI dialects tracked by a shared lockfile; gridctl skill pins adds per-file digest tracking of skill documents, holding drift for human approval with injection heuristics.

— Commands, file format, target dialects, and lockfile all named. product docs

07 07

Containerized Python MCP servers from PyPI/Git NEW 88

source.type: pypi and source.type: git config keys in stack.yaml generate containerized Python MCP servers, run as non-root on Python/uv base images, built from a public PyPI package or a Git repository (with ref pinning and optional path/runtime fields) — no Dockerfile required.

Run a PyPI package as a containerised MCP server at a pinned version, without writing a Dockerfile.

yaml

mcp-servers:
  - name: fetch
    source:
      type: pypi
      package: mcp-server-fetch
      ref: 2026.8.18

Run a Python MCP server from a specific git commit and subdirectory, pinned for reproducibility.

yaml

mcp-servers:
  - name: time
    source:
      type: git
      url: https://github.com/modelcontextprotocol/servers.git
      ref: d73f99efbfd40c3aa1b61e88728b3d49fb52608f
      path: src/time
      runtime: python

— Config keys and runnable examples pin exact package/commit. v1.0.0-rc.1

08 08

Stack-as-code deployment via `gridctl apply` NEW 85

gridctl apply stack.yaml deploys a full MCP gateway from a YAML stack spec in one command, linking downstream servers (OAuth remote, containerized stdio, or local command) to upstream LLM clients.

— Command, YAML spec, and three downstream server types named. product docs

09 09

Tool surface control via groups and whitelists NEW 81

gridctl groups lists tool groups, member counts, and their /groups/{name}/mcp endpoints. Per-server and per-group tools: whitelists, exclude: lists, groups: bundles, and clients: restrictions in stack.yaml control which tools each linked client may access.

— Command, endpoint pattern, and config keys named. product docs

10 10

Import existing MCP client configs NEW 73

gridctl import scans existing client MCP configs, dedupes discovered servers, and appends selections to stack.yaml, offering any plaintext secrets it finds into the encrypted variable store.

— Command and dedupe/secrets mechanism named, single source. product docs

11 11

OAuth authentication for remote MCP servers NEW 73

The per-server auth: type: oauth config key in stack.yaml handles OAuth-protected remote MCP servers, including encrypted-on-disk token storage, refresh, and a needs auth deployment state.

— Config key and states named, no CLI command given. product docs

12 12

Stack export for safe sharing NEW 67

gridctl export re-reads the running deployment's stack file with ${var:<KEY>} variable references preserved and inline credentials blocked, for safe sharing; the web UI's Stack spec view offers the same output via an Export YAML action.

— Command and safeguard named, UI path noted but not detailed. product docs

13 13

Per-server rate limiting NEW 67

gridctl limits displays every configured rate limit and its current state; a limits.rate_limits config block per server sets calls_per_minute and burst fields.

— Command and config fields named, no runtime behavior described. product docs

14 14

Stack diff and validation commands NEW 65

gridctl plan stack.yaml diffs a stack spec against the running deployment state before applying, and gridctl validate stack.yaml lints and schema-checks a stack spec, exiting with codes 0, 1, or 2.

— Named commands and exit codes but no deeper mechanism. product docs

15 15

Catalog search and add commands NEW 63

gridctl search <term> finds MCP servers in the built-in catalog and the MCP Registry, and gridctl add <server> appends a matched entry to stack.yaml by name.

— Commands named but minimal behavioral detail. product docs

16 16

Code mode for tool-capped clients NEW 62

The gateway.code_mode: on config key reduces the tool surface presented to clients that cap the number of MCP servers or tools they can see, such as Antigravity's 100-tool limit, letting large stacks work with those clients.

— Config key and concrete client limit named, mechanism thin. product docs

17 17

Gateway API authentication NEW 60

The gateway.auth config key enables bearer token or API-key authentication on the /mcp endpoint and grouped /groups/{name}/mcp endpoints.

— Config key and endpoints named, minimal further detail. product docs

thinner coverage below

18 18

Signed release provenance with Sigstore NEW 59

Binary releases are signed with a provenance.sigstore.json bundle, making v1.0.0-rc.1 the first provenance-covered production release, verifiable via the release verification procedure.

— Artifact name and version given, verification steps not detailed here. v1.0.0-rc.1

19 19

Token-usage observability in Metrics workspace NEW 59

A token-usage observability layer counts tokens per server, replica, client, and tool, and charts throughput and format-conversion savings in the Metrics workspace.

— Metrics dimensions named but no CLI or API access point. product docs

20 20

Model routing policy via `gridctl models` NEW 54

The gridctl models subcommand projects model routing policy from a running gateway, and the web UI adds a model routing surface for viewing and managing that policy.

— Command named but routing mechanism itself not described. v1.0.0-rc.1

21 21

Web UI with Tools, Metrics, and Library workspaces NEW 48

A web UI accessible via gridctl includes a Tools Workspace, Metrics Workspace, a Stack spec view, and a Library workspace for skill authoring.

— Workspace names given but no navigation or mechanism detail. product docs

Was this useful?

Skyvern

1 RELEASE · 2026-09-09 NOTES Rank

Skyvern automates browser-based workflows with vision models and agents, completing form-driven tasks on sites it has not seen before.

Skyvern v1.0.53 adds Parquet workflow exports and extraction-block data export, expands Copilot with web search, in-chat credential requests, and reasoning visibility on Task V3 action rows, and exposes the Skyvern 3.0 engine in the workflow editor.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Copilot credential request and management NEW 60

Lets Copilot ask for a login credential with a tool that opens an in-chat credential card, shows each saved credential's tested site and authenticator in Copilot account state, and lets Copilot use explicit access-link destinations.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Groups three related Copilot credential increments with names v1.0.53

thinner coverage below

02 00

Skyvern 3.0 engine on validation blocks and editor IMPROVED 50

Exposes the Skyvern 3.0 engine on validation blocks and the default engine list in the workflow editor.

— Names surfaces but not selection mechanism v1.0.53

03 02

Reasoning visibility on Task V3 action rows NEW 50

Surfaces the provider's reasoning summary on Task V3 action rows where available, and persists the model's turn reasoning on Task V3 action rows.

— Names surface twice but no usage path v1.0.53

04 03

Billing error surfaced on Discover 402 NEW 50

Shows a billing error and a 'Go to Billing' link when Discover hits a 402 response.

— Names status code and UI action clearly v1.0.53

05 04

Parquet and extraction-block data exports NEW 45

Adds schema-directed Parquet workflow exports, and adds data export to the extraction block.

— Names formats but no mechanism or UI path given v1.0.53

06 05

Web search helper for Copilot and code blocks NEW 45

Adds a web search helper for code blocks and the Copilot to support lead-discovery workflows without scraping Google or Bing.

— Purpose clear, no exact invocation shown v1.0.53

07 06

Getting started track for new organizations NEW 35

Shows new organizations a 'Getting started' track in the cloud sidebar.

— Clear UI location, thin on content v1.0.53

08 07

Keypress recording in Record Browser NEW 25

Records keypress gestures in the Record Browser.

— Bare description with no further detail v1.0.53

Was this useful?

CrewAI

1 RELEASE · 2026-09-09 NOTES Rank

Open-source Python framework for building multi-agent workflows using role-based AI agent Crews and event-driven Flows.

CrewAI's only shipment this window was internal telemetry to track checkpoint runtime and CLI usage.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Telemetry for checkpoint runtime and CLI usage NEW 20

CrewAI now adds telemetry to track checkpoint runtime and CLI usage.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Only a bare description; no mechanism, data collected, or opt-out named 1.15.21

Was this useful?

AWS Strands Agents

1 RELEASE · 2026-09-09 NOTES Rank

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 Python v1.55.1 adds background task scheduling to the Agent class, ports the notebook tool to Python, and simplifies bidirectional streaming model configuration for Google and OpenAI.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 3 FEATURES 2 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

Background tasks on Agent class NEW 45

Adds backgroundTasks to the Agent class, enabling agents to schedule and run background tasks.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names the API surface but no usage detail or example python/v1.55.1

02 02

Notebook vended tool in Python SDK NEW 35

Ports the notebook tool to Python as a vended tool, making it available in the Python SDK's tool library.

— States what changed but no mechanism or usage python/v1.55.1

03 00

Simplified bidi model configs for Google and OpenAI IMPROVED 35

Simplifies Google and OpenAI model configurations for bidirectional streaming (bidi) mode.

— Names the feature area but not the specific config changes python/v1.55.1

Was this useful?

CopilotKit

1 RELEASE · 2026-09-09 NOTES Rank

CopilotKit connects AI agents to application interfaces with shared state, tools, and generative UI components.

CopilotKit's v1.71.0 release adds provider-level agent configuration and unifies React Native's render-tool hooks with the shared react-core hook surface.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

React Native render-tool hooks unified with react-core IMPROVED 45

Converges React Native render-tool hooks onto react-core, unifying the hook surface across React and React Native.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names the module but no API or migration detail v1.71.0

02 01

Provider-level agent configuration NEW 20

Adds the ability to configure agents at the provider level, per the v1.71.0 release summary; no further mechanism or config keys were specified.

— Only mentioned in summary, no mechanism or config named v1.71.0

Was this useful?

Vercel AI SDK

1 RELEASE · 2026-09-09 NOTES Rank

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

Vercel AI SDK 7.0.96 adds batch cancel and list APIs for managing asynchronous AI job batches.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Batch cancel and list APIs NEW 30

Adds batch cancel and list APIs for managing batches of AI requests.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— No endpoint names, params, or usage shown. [email protected]

Was this useful?

Perplexity API

1 RELEASE · seen 2026-09-10 NOTES Rank

Perplexity API shipped a hosted remote MCP server with OAuth 2.1 and API-key authentication, a new gateway router endpoint exposing moderation and reasoning data, and added Google's Gemini 3.8 Flash model to the Agent API at promotional pricing.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Perplexity API Platform - build with the Router, Agent, Search, and Embeddings APIs. Real-time, web-wide research and Q&A capabilities for your products.

Perplexity API shipped a hosted remote MCP server with OAuth 2.1 and API-key authentication, a new gateway router endpoint exposing moderation and reasoning data, and added Google's Gemini 3.8 Flash model to the Agent API at promotional pricing.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Hosted remote MCP server with OAuth and API-key auth NEW 91

Adds a hosted remote MCP server over Streamable HTTP at https://api.perplexity.ai/mcp requiring no installation, working with any client that supports remote MCP servers (Claude Code, Cursor, VS Code, claude.ai).⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Supports OAuth 2.1 with PKCE and dynamic client registration, opening a browser sign-in on first use with billing organization selectable at sign-in, or API-key authentication via --header "Authorization: Bearer YOUR_API_KEY" for clients without OAuth support or server-side setups such as the Anthropic API MCP connector. Run /mcp inside Claude Code to connect and follow the sign-in prompt, with optional API-key override via the bearer-token header.

Connect to the Perplexity MCP server from Claude Code using an API key instead of OAuth.

$ /mcp --header "Authorization: Bearer YOUR_API_KEY"

Connect Cursor to the Perplexity MCP server via OAuth by navigating to the Tools & MCP settings.

📍1. Open Cursor Settings › Tools & MCP 2. Add the Perplexity MCP server URL 3. Click Connect next to the server entry and complete browser sign-in

Use the Perplexity remote MCP server with OAuth in an MCP client without managing API keys — just point the client at the server URL and sign in.

📍Add the remote MCP server URL to your MCP client (e.g. Claude Code, Cursor, VS Code) and sign in with your Perplexity account when prompted to authenticate via OAuth.

Connect an OAuth-capable MCP client to the Perplexity MCP Server without managing API keys — billing is tied to the org you select at sign-in.

📍Add https://api.perplexity.ai/mcp as the MCP server URL in your client (claude.ai, Claude Code, Cursor, or VS Code), then follow the OAuth prompt to sign in with your Perplexity account and select a billing organization.

— Names protocol, server URL, auth flows and a runnable connect command snapshot-20260910

02 02

Gateway router endpoint with moderation and reasoning fields NEW 90

Adds a POST /router/v1/chat/completions gateway endpoint for chat completions, whose responses include a moderation object with input and output moderation results — flagged, categories, category_scores, and category_applied_input_types — for both prompt and completion sides. Also adds a reasoning_content field on each choice's message object surfacing chain-of-thought/scratchpad reasoning, and a cache_write_tokens field under usage.prompt_tokens_details for tracking tokens written to the prompt cache.

Check whether the model flagged your prompt or its own output for policy violations using the new moderation fields in the gateway endpoint response.

$ curl --request POST --url https://api.perplexity.ai/router/v1/chat/completions --header 'Authorization: Bearer <token>' --header 'Content-Type: application/json' --data '{"model": "sonar", "messages": [{"role": "user", "content": "<prompt>"}]}' | jq '{input_flagged: .moderation.input.moderation_results[0].flagged, output_flagged: .moderation.output.moderation_results[0].flagged, categories: .moderation.output.moderation_results[0].categories}'

Extract chain-of-thought reasoning alongside the final answer to audit model decision-making in security analysis workflows.

$ curl --request POST --url https://api.perplexity.ai/router/v1/chat/completions --header 'Authorization: Bearer <token>' --header 'Content-Type: application/json' --data '{"model": "sonar-reasoning", "messages": [{"role": "user", "content": "<prompt>"}]}' | jq '{reasoning: .choices[0].message.reasoning_content, answer: .choices[0].message.content}'

— Endpoint path and every response field named with runnable curl+jq examples product docs

03 03

Gemini 3.8 Flash model in Agent API NEW 65

Adds google/gemini-3.8-flash to the Agent API, priced at $0.75/M uncached-input tokens, $0.075/M cached-input tokens, and $3.75/M output and reasoning tokens through December 31, 2026.

Call google/gemini-3.8-flash via the Perplexity API for fast, cost-efficient inference with promotional pricing through December 31, 2026.

$ curl https://api.perplexity.ai/chat/completions \
  -H 'Authorization: Bearer <your_api_key>' \
  -H 'Content-Type: application/json' \
  -d '{"model": "google/gemini-3.8-flash", "messages": [{"role": "user", "content": "Summarize the OWASP Top 10."}]}'

Use the cost-efficient Gemini 3.8 Flash model for Agent API calls during its promotional pricing window.

$ curl https://api.perplexity.ai/chat/completions \
  -H 'Authorization: Bearer <your_api_key>' \
  -H 'Content-Type: application/json' \
  -d '{"model": "google/gemini-3.8-flash", "messages": [{"role": "user", "content": "Summarize the latest threat intel feeds."}]}'

— Model name and pricing given, but no mechanism beyond model swap snapshot-20260910

Was this useful?

Microsoft ONNX Runtime

1 RELEASE · 2026-09-10 NOTES Rank

ONNX Runtime v1.29.1 extends attention-related operators with bidirectional GroupQueryAttention, a new BNHS KV-cache layout, and CPU sliding-window attention_bias support, alongside a performance-test tooling update that removes redundant memory copies.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

ONNX Runtime runs machine-learning models across hardware and operating systems with an open-source inference engine.

ONNX Runtime v1.29.1 extends attention-related operators with bidirectional GroupQueryAttention, a new BNHS KV-cache layout, and CPU sliding-window attention_bias support, alongside a performance-test tooling update that removes redundant memory copies.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Plugin EP allocators in perf test tool IMPROVED 80

Updates onnxruntime_perf_test to use plugin Execution Provider device allocators for generated inputs, loaded test data, and pre-allocated outputs, eliminating unnecessary per-run host/device copies.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names the tool and exact behavior change, directly usable in benchmarking. v1.29.1

02 01

BNHS Value KV-cache layout support NEW 70

Adds a session option and Execution Provider metadata contract for using the BNHS Value KV-cache layout, with graph transformations preserving compatibility with the existing BNSH operator schema.

— Mechanism and compatibility scope described, no exact config key given. v1.29.1

03 02

CPU attention_bias with sliding-window cache NEW 70

Adds CPU support for attention_bias with a sliding-window KV cache, including explicit position IDs and post-eviction bias indexing.

— Names the exact operator feature and mechanism but no runnable snippet. v1.29.1

04 03

Bidirectional GroupQueryAttention support NEW 60

Adds bidirectional GroupQueryAttention support on CPU and CUDA via a backward-compatible causal attribute.

— Names the attribute and providers but no usage example. v1.29.1

Was this useful?

Hugging Face TRL

1 RELEASE · 2026-09-10 NOTES Rank

TRL v1.13.0 adds support for training on million-token contexts and vendors fused linear losses into a new trl.losses module, while removing the legacy PPO trainer stack and raising several minimum dependency versions.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

TRL trains transformer language models with reinforcement learning, covering supervised fine-tuning, reward modeling, PPO, DPO, and GRPO.

TRL v1.13.0 adds support for training on million-token contexts and vendors fused linear losses into a new trl.losses module, while removing the legacy PPO trainer stack and raising several minimum dependency versions.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 5 FEATURES 5 OF 5 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

1M-token long-context training for Qwen3-8B NEW 93

Adds a long-context training guide and runnable example demonstrating 1,048,576-token training of Qwen3-8B on a single 8×H100 node, using loss_type="chunked_nll", gradient-checkpointing offload, and YaRN RoPE levers; requires transformers >= 5.16.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Train a model on million-token sequences using chunked NLL loss and gradient-checkpointing offload to fit a single 8xH100 node.

$ trl sft \
  --model_name_or_path Qwen/Qwen3-8B \
  --dataset_name <your-long-context-dataset> \
  --loss_type chunked_nll \
  --per_device_train_batch_size 1 \
  --output_dir Qwen3-8B-1M-SFT

— Names exact loss type, config levers, dependency and includes runnable command. v1.13.0

02 02

PPO trainer stack removed BREAKING 60

PPOTrainer, PPOConfig, PreTrainedModelWrapper, AutoModelForCausalLMWithValueHead, and AutoModelForSeq2SeqLMWithValueHead are removed from TRL; from trl import PPOTrainer now fails at import time.

— Names removed classes exactly but gives no migration path. v1.13.0

thinner coverage below

03 03

Minimum dependency versions raised BREAKING 58

TRL now requires peft >= 0.13.0 (breaking installs on peft 0.12.x) and deepspeed >= 0.18.6 (breaking installs on deepspeed 0.16.4); vLLM 0.19.0 is no longer supported.

— Names exact version floors but no migration guidance. v1.13.0

04 04

Fused linear losses vendored into trl.losses NEW 35

TRL vendors fused linear losses into a new trl.losses module, with chunked cross-entropy reported to run 6× faster via bf16 tensor cores.

— Names module and speedup but no usage detail given. v1.13.0

05 05

vLLM 0.28.0 support NEW 31

Adds support for vLLM 0.28.0.

— One-line addition with no further detail. v1.13.0

└──▷ BREAKING ON UPGRADE

!PPOTrainer, PPOConfig, PreTrainedModelWrapper, AutoModelForCausalLMWithValueHead, and AutoModelForSeq2SeqLMWithValueHead are removed from TRL; from trl import PPOTrainer now fails at import time.
!peft >= 0.13.0 is now required; installations using peft 0.12.x will break.
!deepspeed >= 0.18.6 is now required; installations using deepspeed 0.16.4 will break.
!vLLM 0.19.0 is no longer supported.

Was this useful?

Eigen Labs Darkbloom

1 RELEASE · 2026-09-09 NOTES CODE Rank

Private Inference Network on Idle Macs

Darkbloom's v0.9.1 release focuses on coordinator operability, adding an isolated migration mode and more accurate request-outcome accounting, alongside a mention of per-model cache outcome reporting.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 3 FEATURES 2 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

Isolated migration mode for coordinator NEW 85

The coordinator binary gains a --migrate-only flag that runs database migrations in isolation without starting listeners, workers, admin seeding, or MDM clients.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ It requires EIGENINFERENCE_DATABASE_URL to be set, and can be used as a pre-deploy migration step in CI/CD pipelines, e.g. EIGENINFERENCE_DATABASE_URL=postgres://user:pass@host/db coordinator --migrate-only.

Run database migrations before a coordinator deployment without starting the full service — useful in a pre-deploy migration step in CI/CD pipelines.

$ EIGENINFERENCE_DATABASE_URL=postgres://user:pass@host/db coordinator --migrate-only

— Names exact flag, env var and runnable command. v0.9.1

thinner coverage below

02 00

Independent request-outcome recording IMPROVED 40

Each incoming inference request outcome is now recorded independently of provider-attempt profiles, so a request that survives multiple provider refusals before completing is counted as one outcome rather than several failures.

— Explains the behavior change but no config surface or command. v0.9.1

03 02

Per-model cache outcome reporting NEW 15

Darkbloom adds per-model cache outcome reporting, per the release summary, though no further mechanism or fields are described.

— Only named in summary; no detail on mechanism or fields. v0.9.1

Was this useful?

NVIDIA Model Optimizer

1 RELEASE · 2026-09-09 NOTES Rank

NVIDIA Model Optimizer added opt-in support for the TensorRT-RTX ABI Execution Provider for ONNX calibration on Windows arm64.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

NVIDIA Model Optimizer provides quantization, pruning, distillation, and other model-compression techniques for deployment.

NVIDIA Model Optimizer added opt-in support for the TensorRT-RTX ABI Execution Provider for ONNX calibration on Windows arm64.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

TensorRT-RTX ABI Execution Provider for ONNX calibration NEW 75

New --calibration_eps=NvTensorRtRtx and --trt_rtx_backend=abi flags enable the TensorRT-RTX ABI Execution Provider for ONNX calibration on Windows arm64; this is opt-in, with the legacy backend remaining the default.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names exact flags and platform but lacks broader behavioural detail. 0.46.1

Was this useful?

Hugging Face Transformers

1 RELEASE · 2026-09-09 NOTES Rank

Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.

Transformers v5.17.0 ships six new model architectures spanning a 780B-parameter MoE LLM, speech synthesis, multimodal encoders and multiple ASR models, alongside quantization, kernel-registration and generation-performance improvements, plus a breaking change to vision RoPE handling.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 13 FEATURES 9 OF 13 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

HYV4 mixture-of-experts LLM NEW 75

Adds HYV4 (Hy4-Preview), a 780B-parameter mixture-of-experts LLM with Multi-head Latent Attention (kv_lora_rank, kv_b_proj), DeepSeek Sparse Attention (index_topk, indexer_types), and Independent Hyper-Connections (hc_mult), with a 1M-token context window.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names architecture internals but no runnable usage snippet v5.17.0

02 02

Canary multilingual ASR model NEW 65

Adds Canary, a multilingual ASR and speech-to-text translation model using ParakeetEncoder/ParakeetEncoderConfig and CanaryProcessor for task-selecting prompt prefixes.

— Names concrete classes a reader could import directly v5.17.0

03 03

Vision RoPE migration to centralized utils BREAKING 65

Custom vision models relying on attention-layer-level or model-specific RoPE grid interleaving logic must migrate to the new centralized modeling_rope_utils.py implementation for vision (2D/3D) rotary embeddings.

— Names exact file and migration requirement for affected code v5.17.0

04 04

NeoMME multimodal encoders NEW 60

Adds NeoMME and NeoMME-Retriever: 260M/800M multimodal-native multilingual encoders processing text and raw image patches in a single bidirectional Transformer, with NeoMME-Retriever supporting MeanMaxSim late-interaction and dense cosine-similarity retrieval.

— Concrete sizes and retrieval methods, no code example v5.17.0

05 05

Fun-ASR-Nano ASR model NEW 60

Adds Fun-ASR-Nano, an 800M-parameter ASR model covering Chinese (including 7 dialects and 26 accents), English, and Japanese with hotword customization and native punctuation output.

— Rich detail on coverage but no invocation shown v5.17.0

06 06

KimiLinear hybrid attention architecture NEW 60

Adds KimiLinear, a hybrid linear attention architecture featuring Kimi Delta Attention (KDA) with per-channel forget gates, interleaved MLA full-attention layers, and DeepSeek-V3-style MoE feed-forward blocks.

— Mechanism-heavy but no usage example provided v5.17.0

thinner coverage below

07 07

Standalone function registration in KernelConfig NEW 55

Enables standalone functions (e.g., RoPE) to be registered in KernelConfig with optional non-inheritance of default mappings.

— Names config object and example function, no full snippet v5.17.0

08 00

AutoRound low-bit quantization IMPROVED 50

Supports 5-, 6-, and 7-bit quantization in AutoRound.

— Concrete bit-widths named but no config example given v5.17.0

09 08

VibeVoice speech synthesis model NEW 45

Adds VibeVoice, a next-token diffusion speech synthesis model for high-fidelity long-form multi-speaker audio generation such as podcasts and audiobooks.

— Describes purpose but no API details given v5.17.0

10 09

Context parallelism support flag NEW 40

Adds supports_context_parallel property to PreTrainedModel for signalling context-parallelism compatibility.

— Named property but no usage detail beyond signalling v5.17.0

11 00

Paged attention no-cache error IMPROVED 30

Raises an error when a paged attention forward pass is called with no cache, surfacing silent failures instead of failing quietly.

— States behaviour change but no config or flag named v5.17.0

12 00

Hub-kernel fallback warning IMPROVED 30

Warns once when a hub-kernel function falls back to its reference PyTorch path instead of silently running slower code.

— Behaviour described but no configuration surface named v5.17.0

13 00

Reduced generation-step overhead IMPROVED 30

Reduces per-step generation overhead by avoiding unnecessary accelerator synchronization on every decode step.

— Mechanism named briefly, no measured benchmark given v5.17.0

└──▷ ALSO FROM THESE RELEASES

└──▷ BREAKING ON UPGRADE

!Custom vision models relying on attention-layer-level or model-specific RoPE grid interleaving logic must migrate to the new centralized modeling_rope_utils.py implementation for vision (2D/3D) rotary embeddings.

Was this useful?

OpenAI Python SDK

1 RELEASE · 2026-09-09 NOTES Rank

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

OpenAI Python SDK v3.11.0 adds expiration controls for service account keys.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Expiration controls for service account keys NEW 35

The API client adds expiration controls for service account keys.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names the capability but no flag, field, or mechanism given v3.11.0

Was this useful?

OpenAI TypeScript SDK

1 RELEASE · 2026-09-09 NOTES Rank

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

OpenAI TypeScript SDK v7.13.0 adds API key expiration controls to the API surface.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

API key expiration controls NEW 20

Adds API key expiration controls to the OpenAI API surface, allowing keys to be configured with expiration.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— No named flag, endpoint, or parameter given, only a brief description. v7.13.0

Was this useful?

Unsloth AI Unsloth

1 RELEASE · 2026-09-09 NOTES Rank

Unsloth v0.1.808-beta ships native Docker images for running the full Studio stack, adds a Vulkan backend for AMD iGPUs, expands Apple Silicon MLX support, and speeds up diffusion inference while shrinking the Python package.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Unsloth runs and fine-tunes language and diffusion models locally through a Python library and interface.

Unsloth v0.1.808-beta ships native Docker images for running the full Studio stack, adds a Vulkan backend for AMD iGPUs, expands Apple Silicon MLX support, and speeds up diffusion inference while shrinking the Python package.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 6 FEATURES 2 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

Docker images for Unsloth Studio NEW 80

Publishes native AMD64 and ARM64 unsloth/unsloth Docker images, enabling ARM-native deployments.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ The full Studio stack can be run on an NVIDIA GPU host with docker run -d --gpus all --ipc=host -p 8000:8000 -p 8888:8888 -e UNSLOTH_STUDIO_PASSWORD="mypassword" -e JUPYTER_PASSWORD="mypassword" -v "$PWD":/workspace/host unsloth/unsloth, exposing Studio on port 8000 and JupyterLab on port 8888.

Run the full Unsloth Studio stack on an NVIDIA GPU host using the new Docker image, exposing Studio on port 8000 and JupyterLab on port 8888.

$ docker run -d --gpus all --ipc=host \
  -p 8000:8000 -p 8888:8888 \
  -e UNSLOTH_STUDIO_PASSWORD="mypassword" -e JUPYTER_PASSWORD="mypassword" \
  -v "$PWD":/workspace/host \
  unsloth/unsloth

— Runnable command with exact ports and env vars given. v0.1.808-beta

thinner coverage below

02 00

PyTorch 2.11 default with new install extras IMPROVED 55

Updates default PyTorch to 2.11 across supported installers, with cu128, cu126, and cu130 extras (torch2110) and xformers 0.0.35.

— Names exact install extras and versions, no full command. v0.1.808-beta

03 00

Expanded MLX support on Apple Silicon IMPROVED 45

Supports DoRA fine-tuning and additional DPO loss types on Apple Silicon via MLX, and batched MLX generation now streams and samples each chat session independently.

— Names DoRA and DPO but no command or config shown. v0.1.808-beta

04 00

Faster diffusion inference IMPROVED 35

Accelerates diffusion inference 1.2x–1.7x for INT8/FP8 pathways across all models.

— Gives concrete speedup numbers but no usage detail. v0.1.808-beta

05 02

Text-only fine-tuning for multimodal models NEW 25

Enables more multimodal models to be fine-tuned using text-only datasets.

— One-line description with no named models or steps. v0.1.808-beta

06 00

Smaller Python package size IMPROVED 25

Reduces the Unsloth Python package size by more than 60%.

— States a number but no mechanism given. v0.1.808-beta

Was this useful?

Fireworks AI

2 RELEASES · 2026-09-09 NOTES Rank

Fireworks AI shipped new training tooling — a pre-run cost estimator and a coding-agent training skill covering research, configuration and debugging — alongside a workflow to evaluate fine-tuned models before deployment, per-user spending limits and SSO/SCIM authentication in Fireworks Nexus, and additions for routing third-party requests through FireRouter and tagging dedicated deployments for cost attribution.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

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

Fireworks AI shipped new training tooling — a pre-run cost estimator and a coding-agent training skill covering research, configuration and debugging — alongside a workflow to evaluate fine-tuned models before deployment, per-user spending limits and SSO/SCIM authentication in Fireworks Nexus, and additions for routing third-party requests through FireRouter and tagging dedicated deployments for cost attribution.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Training skill for coding agents NEW 65

Adds Agent Skills — a Fireworks training skill for Claude Code, Cursor, Codex, and other compatible coding agents — to install research, configure, and debug workflows.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Users describe a training goal in plain language to plan a run, estimate cost, and wait for approval before spend.

— Names agents and workflow steps but no exact command Training skill for coding agents

02 02

Per-user spending limits in Fireworks Nexus NEW 60

Adds Per-User Usage Limits to set per-user spending limits on serverless inference, with an account default spending limit, group limits, and per-user overrides configured under Fireworks Nexus > Usage Limits in the console.

Enforce a per-user spending cap so no single user can exhaust your serverless inference budget.

📍In the Fireworks console, go to Fireworks Nexus › Usage Limits, set an account default spending limit, then add per-user overrides or group limits as needed.

— Names the console path and limit types but no API product docs

03 03

FireRouter routing with your own provider key NEW 60

FireRouter can route requests to a third-party provider using your own API key instead of Fireworks credentials, configured by adding a provider API key under Fireworks Nexus > Provider Keys.

Route requests through FireRouter to a third-party provider using your own API key instead of Fireworks credentials.

📍In the Fireworks console, go to Fireworks Nexus › Provider Keys and add your provider API key so FireRouter can call that provider on your behalf.

— Clear mechanism and console path, no API detail product docs

04 04

Training cost estimator NEW 60

Adds a training cost estimator that previews the cost of a training job before it runs, using published per-token rates for Managed and Serverless deployments and allocated GPU-hour rates for Dedicated deployments.

Before submitting a fine-tuning job, check the estimated cost to avoid surprises on large datasets or GPU-intensive runs.

📍In the Fireworks AI console, navigate to Training, configure your job parameters, then open the Training cost estimator to view the projected cost before submitting.

— Explains rate mechanism per tier and gives UI path Training skill for coding agents Training cost estimator

thinner coverage below

05 05

Deployment Tags for cost attribution NEW 55

Dedicated deployments can be tagged with key-value metadata (e.g. team=security, env=prod) under Deployment Tags for cost attribution and filtering.

Tag a dedicated deployment with environment or team metadata for cost attribution and filtering.

📍In the Fireworks console, open your dedicated deployment, navigate to Deployment Tags, and attach key-value metadata (e.g. team=security, env=prod).

— Concrete tag example and UI path given product docs

06 06

Custom SSO and SCIM provisioning NEW 30

Adds Custom SSO support for Single Sign-On authentication and SCIM user and group provisioning.

— Named but no mechanism, path, or example given product docs

07 07

Evaluating Trained Models workflow NEW 30

Adds an Evaluating Trained Models workflow to evaluate a fine-tuned model before creating a production deployment.

— States the workflow exists with no further detail product docs

Was this useful?

Ollama

1 RELEASE · 2026-09-05 NOTES CODE Rank

Ollama v0.34.0 adds ChatGPT Desktop integration on macOS, OpenAI-compatible tool search and response compaction endpoints, and faster structured output on Apple Silicon via speculative decoding.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Local large language model runner that downloads, serves, and chats with models through a CLI and REST API.

Ollama v0.34.0 adds ChatGPT Desktop integration on macOS, OpenAI-compatible tool search and response compaction endpoints, and faster structured output on Apple Silicon via speculative decoding.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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 01

ChatGPT Desktop model selection integration NEW 60

Adds Ollama model selection to ChatGPT Desktop on macOS, accessible via the Ollama app, letting users run open models while keeping their existing chats, plugins, MCPs, and skills.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— UI path given but no config or API detail v0.34.0

02 02

OpenAI-compatible tool search support NEW 60

Adds OpenAI-compatible client tool search support, using the tsc_ prefix for tool search call IDs in both streaming and non-streaming responses.

— Names ID prefix but no endpoint or command shown v0.34.0

03 03

Response compaction endpoint NEW 60

Adds an OpenAI-compatible response compaction endpoint, with retry on context overflow and image preservation through compaction.

— Describes behavior but no exact route or flag named v0.34.0

04 00

Faster structured output on Apple Silicon IMPROVED 60

Improves structured output performance on Apple Silicon by enabling speculative decoding under structured output via xgrammar structural tags.

— Explains mechanism but no flag or benchmark given v0.34.0

Was this useful?

llmfit

1 RELEASE · 2026-09-10 NOTES CODE Rank

Hundreds of models & providers. One command to find what runs on your hardware.

llmfit v1.1.15 adds Ferrum as a supported benchmark provider and expands its community hardware benchmark database.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Ferrum provider support in benchmarking NEW 80

Adds --provider ferrum to llmfit bench and the FERRUM_HOST environment variable, enabling benchmarking against Ferrum endpoints via the existing OpenAI-compatible transport.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Benchmark a model running on a Ferrum endpoint to capture real tok/s and TTFT measurements on your hardware.

$ llmfit bench --provider ferrum

— Names flag, env var and command; mechanism (OpenAI-compatible transport) given. v1.1.15

thinner coverage below

02 02

New community benchmark results NEW 35

Adds community benchmark results for nvidia-gb10, nvidia-geforce-rtx-5070-ti, nvidia-geforce-rtx-3080, intel-raptor-lake-p-iris-xe-graphics-integrated, and intel-coffeelake-s-gt2-uhd-graphics-630-integrated hardware profiles.

— Names hardware profiles but no mechanism or usage path. v1.1.15

Was this useful?

Comfy Org ComfyUI

1 RELEASE · 2026-09-09 NOTES Rank

ComfyUI is a node-graph interface, API, and backend for running diffusion models, chaining generation steps into reusable visual workflows.

ComfyUI v0.35.0 introduces a new Comfy Compiler for graph execution optimization, adds 3D and video editing nodes, AVIF image export, a sparse attention node, and a broad wave of new partner model integrations and LoRA support.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 15 FEATURES 10 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

New partner model integrations across partner nodes NEW 60

Adds MiniMax-H3 Max and Max Turbo model options to the H3 text-to-video and first-last-frame partner nodes; Recraft V4 Styles and V4 Styles Pro to the Create Style partner node; WAN3-Prime via the WAN partner node; Google Omni 1.1 support; SenseNova U1.5 support (CORE-411); Pixal3D multiview support (CORE-421); Meta Muse Image partner nodes; Claude Fable 5.1 to the Anthropic Claude partner node; GPT Image 2.5 Flare and Sunburst models to the OpenAI partner node; the GPT-6 Astra model and a reasoning parameter to the OpenAI partner node; and Comfy-Cloud partner nodes to run curated workflows on Comfy Cloud GPUs.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names every partner node and model added, each thinly described v0.35.0

thinner coverage below

02 02

3D file import node NEW 55

Adds the File3DToMesh node to parse GLB, GLTF, OBJ, and STL files into a MESH type for use in 3D workflows.

— Names node and supported formats, no example given v0.35.0

03 03

In-graph video trim and crop nodes NEW 55

Adds VideoTrim and VideoCrop nodes with VIDEO_EDIT widget inputs for in-graph video editing.

— Names nodes and widget type, no usage example v0.35.0

04 04

LTXV video generation nodes NEW 55

Adds LTXV generated-keyframe nodes and a Freeze Latent node, plus LTXVAddLatentGuide for pinning a pre-encoded latent as a guide frame.

— Names three nodes with brief mechanism for one v0.35.0

05 05

AMD GPU attention and FP8 support NEW 55

Enables PyTorch SDPA attention and FP8 ops for AMD gfx1170 and gfx1171 GPUs.

— Names GPU archs and ops but no config to enable v0.35.0

06 00

AVIF export in Save Image Advanced IMPROVED 50

Supports the avif format in the Save Image Advanced node.

— Names node and format, no limits or defaults given v0.35.0

07 00

Live resolution preview in ResolutionSelector IMPROVED 45

Adds a live resolution preview widget to the ResolutionSelector node.

— Names node and widget, clear UI location v0.35.0

08 06

LoRA support for MiniMax-H3 and HiDream O1 NEW 45

Supports PDD LoRA for MiniMax-H3, and supports MiniMax-H3 and HiDream O1 LoRAs trained with DiffSynth-Studio / ModelScope.

— Names models and training frameworks, no usage steps v0.35.0

09 00

Container memory limit respected IMPROVED 45

Now respects the container cgroup memory limit instead of host RAM when running in containers.

— States before/after behaviour, no config key named v0.35.0

10 07

Comfy Compiler for graph execution NEW 40

Introduces the Comfy Compiler (tracked as CORE-389) for graph execution optimization.

— Mechanism named but no detail on how it optimizes v0.35.0

11 00

ImageCompositor emits document canvas IMPROVED 40

Emits the document canvas from ImageCompositor for use in the layer editor.

— Names node and consumer, no further mechanism v0.35.0

12 08

cfgpp_ud10_ab sampler added NEW 35

Adds the cfgpp_ud10_ab sampler.

— Bare name with no mechanism explained v0.35.0

13 09

Sparse Attention node NEW 30

Adds a Sparse Attention node.

— Bare name, no explanation of behaviour v0.35.0

14 10

Color space conversion node NEW 25

Adds a color space conversion node.

— No node name or supported color spaces given v0.35.0

15 00

Save 3D Advanced output reporting IMPROVED 25

Reports the file saved by Save 3D (Advanced) as a standard 3D output item.

— Bare behavioural change, no further detail v0.35.0

Was this useful?

Pinecone

1 RELEASE · seen 2026-09-10 NOTES Rank

Pinecone launched Nexus, a curated retrieval layer with BYOC support and custom context manifests, and shipped a broad set of organization-level admin APIs covering service accounts, role bindings, invites, projects, organizations and API keys, alongside versioned 2026-07 OpenAPI specs for Inference and Admin.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Pinecone is a managed vector database that stores and queries embeddings for AI applications.

Pinecone launched Nexus, a curated retrieval layer with BYOC support and custom context manifests, and shipped a broad set of organization-level admin APIs covering service accounts, role bindings, invites, projects, organizations and API keys, alongside versioned 2026-07 OpenAPI specs for Inference and Admin.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Service account management APIs NEW 95

Adds POST /service-accounts to create a service account with optional initial role bindings (the client secret is returned only once), GET /service-accounts to list all service accounts in the caller's organization, GET /service-accounts/{id} to retrieve a service account by ID, PATCH /service-accounts/{id} to update a service account's name, DELETE /service-accounts/{id} to delete a service account and revoke all tokens it minted within seconds, and a rotate-secret endpoint to rotate a service account's OAuth client secret, revoking the previous secret and its tokens within seconds and returning the new secret only once.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Rotate a compromised service account OAuth secret immediately to revoke all outstanding tokens within seconds.

$ curl -X POST https://api.pinecone.io/service-accounts/<service-account-id>/rotate-secret \
  -H 'Authorization: Bearer <your-token>'

— All endpoints named with exact behavior and a runnable curl example snapshot-20260910

02 02

Nexus retrieval layer with BYOC and context manifests NEW 75

Pinecone launches Nexus, a curated retrieval layer with BYOC (bring your own cloud) support. Nexus adds a context manifest system letting you define custom artifact and edge types in the console, turning sources into queryable knowledge; in the console you navigate to Nexus > Context Design, select 'Design your own manifest', define custom artifact and edge types, then trigger curation to index your sources.

Design a Nexus context manifest with custom artifact and edge types to make your sources queryable in Pinecone Nexus.

📍In the Pinecone console, navigate to Nexus › Context Design, select 'Design your own manifest', define custom artifact types and edge types, then trigger curation to index your sources.

— Names mechanism and console steps, but no API endpoint given product docs

03 03

Versioned 2026-07 API specs for Inference and Admin NEW 65

Adds inference_2026-07.oas and admin_2026-07.oas OpenAPI specs, versioned 2026-07, for the Inference and Admin APIs, retrievable directly from the pinecone-api GitHub repo to generate a client or validate requests against the new contract.

Fetch the latest 2026-07 Inference API spec to generate a client or validate requests against the new versioned contract.

$ curl -O https://raw.githubusercontent.com/pinecone-io/pinecone-api/refs/heads/main/2026-07/inference_2026-07.oas.yaml

Fetch the latest 2026-07 Admin API spec to inspect new admin endpoints or integrate them into an API gateway.

$ curl -O https://raw.githubusercontent.com/pinecone-io/pinecone-api/refs/heads/main/2026-07/admin_2026-07.oas.yaml

— Exact filenames and fetch commands given, little behavioral detail product docs

thinner coverage below

04 04

Role binding and invite listing APIs NEW 50

Adds POST /role-bindings to grant a role to a principal at organization or project scope, and GET /invites to list pending and expired invites in the caller's organization.

— Endpoints named but no example or scope mechanism detail snapshot-20260910

05 05

Documentation index at /llms.txt NEW 40

Serves a complete documentation index at /llms.txt for programmatic discovery of all available pages.

— Exact path named but no further mechanism or example snapshot-20260910

06 06

Organization, project, and API key management APIs NEW 38

Adds organization management APIs to list organizations, get organization details, update an organization's name, and delete an organization; project management APIs to list projects, create a project, get project details, update a project's name, maximum Pod count, or CMEK, and delete a project; and API key management APIs to list all API keys in a project and delete an API key from a project.

— Described only as capabilities, no endpoint paths or examples given snapshot-20260910

Was this useful?

timgordontg engrim

2 RELEASES · 2026-09-10 NOTES Rank

The Universal Cross-Model Episodic Memory Standard. Local-first, project-scoped SQLite memory engine for Google Antigravity, Claude Code, Cursor, Windsurf, and Codex.

engrim shipped a new engrim merge subcommand for combining SQLite memory stores across machines and CI, alongside engrim prune for log retention and disk reclamation, tag-based filtering, and a --strict exit-code gate for agent review hooks.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 7 FEATURES 5 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

Prune command for log retention and vacuum NEW 95

New engrim prune command purges stale transcript logs and reclaims SQLite disk space via VACUUM.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ It requires --keep-days <N> or $ENGRIM_PRUNE_KEEP_DAYS to delete anything, or --vacuum alone can reclaim fragmented space without deleting logs.

Reclaim disk space on a schedule without risking log deletion — vacuum only, no retention window needed.

$ engrim prune --vacuum

Purge transcript logs older than 30 days and vacuum the SQLite store, driven by an env var so CI or cron jobs need no flag changes.

$ ENGRIM_PRUNE_KEEP_DAYS=30 engrim prune

— Flags, env var and runnable examples all given v1.3.1

02 02

Merge SQLite memory stores NEW 85

New engrim merge OTHER.db subcommand folds another SQLite store's records and logs into the current store. It uses idempotent monotonic status resolution so superseded/done records carry over and are never reactivated, plus content-hash deduplication, enabling parallel CI and multi-machine agent workflows.

— Names subcommand and merge mechanism, no usage example given v1.3.2

03 03

Strict exit-code gating for review NEW 85

Adds --strict flag to engrim review (and $ENGRIM_STRICT=1 env var) that exits with code 2 when uncommitted architectural decisions are detected, enabling stop-hook gating for agents and orchestrators.

Gate an agent stop hook so it exits with code 2 — blocking the agent from clearing context — if any uncommitted architectural decisions are detected.

$ engrim review --strict

— Exact flag, env var and runnable example provided v1.3.1

04 04

Tag filtering for recall and list NEW 65

Adds --tag / --tags flags to engrim recall and engrim list for filtering memories by tag, with full support across the CLI and the stdio MCP server tools for Claude Code and Cursor.

— Flags and surfaces named, no runnable example v1.3.1

thinner coverage below

05 00

Negation-aware decision capture IMPROVED 35

Capture checks now include explicit negation handling so reversed decisions (e.g. 'decided not to use X') no longer falsely match existing positive decisions.

— Describes behavior change but no command or flag v1.3.2

06 00

Fixed review status on fresh environments IMPROVED 30

In a fresh, unlogged environment, review status is now reported as unknown rather than a false-positive 'safe to clear' when no session logs exist.

— Brief bug-fix description, no surface named v1.3.2

07 05

Version flag for CLI NEW 15

Adds -v / --version flag to the engrim CLI parser.

— Bare flag name with no further detail v1.3.1

Was this useful?

PromptLayer

1 RELEASE · 2026-09-10 CHANGELOG Rank

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

PromptLayer added multimodal input support for Gemini 3.8 Flash workflows and a fullscreen view for video playback in conversations.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Multimodal inputs for Gemini 3.8 Flash NEW 40

PromptLayer now supports multimodal inputs — text, images, video, audio, and PDF files — for Gemini 3.8 Flash model workflows.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names model and input types but no config or API detail changelog-20260910-eacb230b

02 00

Fullscreen video playback view IMPROVED 25

Adds a fullscreen view for video content playback in the conversation display.

— Brief UI change with no further mechanism described changelog-20260910-eacb230b

Was this useful?

LangChain LangSmith

1 RELEASE · 2026-08-20 NOTES Rank

LangSmith's biggest change this window is deprecating its legacy feedback-formula and dataset-comparison APIs in favor of a new paginated experiment-runs endpoint and composite evaluators, alongside new tooling for testing multi-turn evaluators and managing dataset splits in bulk.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

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

LangSmith's biggest change this window is deprecating its legacy feedback-formula and dataset-comparison APIs in favor of a new paginated experiment-runs endpoint and composite evaluators, alongside new tooling for testing multi-turn evaluators and managing dataset splits in bulk.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 5 FEATURES 3 OF 5 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

Test multi-turn evaluators against real conversations NEW 85

The POST /runs/rules/validate endpoint now accepts test_thread_id and session_id parameters, letting you validate a thread evaluator against a real conversation before saving the rule.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ A matching Test action was added to multi-turn evaluators in the evaluator config UI, running the evaluator against a sample thread before it's saved.

Validate a multi-turn thread evaluator against a real conversation before saving the rule, to confirm it scores as expected.

$ curl -X POST 'https://api.smith.langchain.com/runs/rules/validate' \
  -H 'Content-Type: application/json' \
  -H 'x-api-key: <api_key>' \
  -d '{"rule": {"evaluator_type": "thread"}, "test_thread_id": "<thread_id>", "session_id": "<session_id>"}'

— Names endpoint, params, and includes a runnable curl example snapshot-20260910

02 02

Experiment comparison API replaced with paginated endpoint BREAKING 70

POST /v2/datasets/<id>/experiment-runs is now the supported public API for paginated experiment comparison. Legacy dataset comparison helpers have been removed from the public OpenAPI spec and generated SDKs, though the underlying HTTP routes continue to work only for LangSmith UI clients.

— Names replacement endpoint but no usage example snapshot-20260910

thinner coverage below

03 03

Mixed thread and run items in annotation queues NEW 50

The annotation queue items API supports adding a batch of conversation threads and runs to a queue in a single request, mixing item_type: THREAD (with thread_id/session_id) and item_type: RUN (with run_id) entries via POST /annotation-queues/<queue_id>/items.

Add a batch of conversation threads to an annotation queue for human review, mixing thread and run items in a single request.

$ curl -X POST 'https://api.smith.langchain.com/annotation-queues/<queue_id>/items' \
  -H 'Content-Type: application/json' \
  -H 'x-api-key: <api_key>' \
  -d '{"items": [{"item_type": "THREAD", "thread_id": "<thread_id>", "session_id": "<session_id>"}, {"item_type": "RUN", "run_id": "<run_id>"}]}'

— Only known via example, not described in release notes snapshot-20260910

04 00

Dataset split chips in Examples table IMPROVED 38

Each example's dataset splits now render as interactive chips in the dataset Examples table, with a clickable +N overflow menu when an example belongs to more splits than fit the column.

— UI-only description, no navigation path or config given snapshot-20260910

05 00

Deferred rendering for large experiment tables IMPROVED 30

Experiment tables now defer full payload rendering and output diff preparation until those views are requested, improving responsiveness for runs with large agent trajectories.

— States the change but no metrics or trigger detail snapshot-20260910

└──▷ BREAKING ON UPGRADE

!The legacy feedback formula endpoints (POST/GET /feedback/formulas and GET/PUT/DELETE /feedback/formulas/{feedback_formula_id}) are deprecated and scheduled for removal on 2026-08-20; migrate existing feedback formulas to composite evaluators.
!Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs (existing HTTP routes continue to work only for LangSmith UI clients).

Was this useful?

Arize Phoenix

1 RELEASE · 2026-09-09 NOTES Rank

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

Phoenix Evals adds a new evaluator for checking whether conversational requests were fully completed.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Completeness evaluator for conversations NEW 53

Adds createCompletenessEvaluator to judge whether every active user request in a conversation was actually completed.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names exact function but no usage or mechanism detail. @arizeai/[email protected]

Was this useful?

Langfuse

1 RELEASE · 2026-09-09 NOTES Rank

Langfuse provides tracing, evaluation, and monitoring for LLM applications.

Langfuse's v4.33.0 release adds a time-scoped observation lookup to the public API, new filter operators for nullable fields, and a small playground UI improvement.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 3 FEATURES 2 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

Optional startTime parameter on observation lookup NEW 80

The public API endpoint GET /observations/{id} now accepts an optional startTime query parameter, enabling time-scoped observation lookups so you can fetch a specific observation without scanning the full history when you know approximately when it occurred.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Fetch a specific observation scoped to a start time to avoid scanning the full history when you know approximately when the observation occurred.

$ curl -X GET 'https://cloud.langfuse.com/api/public/observations/{id}?startTime=2025-09-01T00:00:00Z' \
  -H 'Authorization: Basic <base64-encoded-credentials>'

— Named endpoint, param, and runnable curl example. v4.33.0

thinner coverage below

02 02

has: and -has: filter operators for nullable fields NEW 40

The filter UI now surfaces has: and -has: filter operators while typing a nullable field name, allowing filtering on presence or absence of a value.

— Named operators but no exact UI path or example. v4.33.0

03 00

Source prompt name in playground IMPROVED 20

The playground window content now displays the source prompt name.

— Very thin description, no mechanism or location detail. v4.33.0

Was this useful?

Braintrust

1 RELEASE · 2026-09-01 NOTES Rank

Braintrust provides evaluation, tracing, and improvement workflows for AI applications.

Braintrust's biggest window changes centered on Loop, which now runs in a Braintrust-managed runtime with scheduled automations, alongside new AI-driven trace analysis (Patterns, Debugger) and Java SDK concurrency controls that require Task/Scorer/Classifier implementations to become thread-safe.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Smaller additions covered new built-in models, dashboard sections, blind human reviews, and SDK-level metadata across Python, Java, and TypeScript.

└──▷ WHAT SHIPPED · 12 FEATURES 9 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

Java SDK concurrent eval execution and tuning BREAKING 95

The Java SDK now runs eval cases concurrently, so Task, Scorer, and Classifier implementations must be thread-safe.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ Concurrency and OTel batching are controlled via the BRAINTRUST_DEFAULT_MAX_CONCURRENCY, BRAINTRUST_OTEL_MAX_QUEUE_SIZE, BRAINTRUST_OTEL_MAX_EXPORT_BATCH_SIZE, and BRAINTRUST_OTEL_EXPORT_INTERVAL_MILLIS environment variables, or programmatically via .maxConcurrency(int) on Eval.Builder and Devserver.Builder (default: 3 at a time) and .executor(Executor).

Tune Java eval concurrency and OTel export behavior without touching code, useful in high-throughput CI environments.

$ export BRAINTRUST_DEFAULT_MAX_CONCURRENCY=10
export BRAINTRUST_OTEL_MAX_QUEUE_SIZE=4096
export BRAINTRUST_OTEL_MAX_EXPORT_BATCH_SIZE=512
export BRAINTRUST_OTEL_EXPORT_INTERVAL_MILLIS=2000

Cap concurrent eval case execution and tune OTel export batching without touching code, using environment variables.

$ BRAINTRUST_DEFAULT_MAX_CONCURRENCY=10 BRAINTRUST_OTEL_MAX_QUEUE_SIZE=512 BRAINTRUST_OTEL_MAX_EXPORT_BATCH_SIZE=64 BRAINTRUST_OTEL_EXPORT_INTERVAL_MILLIS=2000 your-eval-runner

— Runnable env vars and builder methods with an explicit default and migration requirement snapshot-20260910

02 02

GLM-5.3 Flash built-in model NEW 65

Adds GLM-5.3 Flash as a built-in multimodal reasoning model available under the Braintrust provider in playgrounds, prompts, and scorers, or via the Braintrust Gateway using the identifier glm-5.3-flash.

— Exact gateway identifier makes this directly usable snapshot-20260910

03 03

Eval.start() async eval execution in Java SDK NEW 60

Adds Eval.start() to the Java SDK (v0.3.23) to start an eval in the background and return an EvalResult as soon as the experiment exists, while Eval.run() still awaits full completion.

— Names the exact method and version but no runnable example snapshot-20260910

thinner coverage below

04 04

Debugger and Analyze trace view NEW 55

Adds the Debugger to the trace viewer, which reports likely failure modes for a single trace citing the spans, tool calls, and model outputs behind each one, alongside an 'Analyze trace' view that groups a run into labeled Work sections.

— Names the UI view and what it groups, giving a starting point snapshot-20260910

05 05

Blind human reviews project setting NEW 55

Adds the 'Blind human reviews' project setting, which hides peer scores, comments, and aggregates until a reviewer submits their own scores; reviewers with the Update permission are always exempt.

— Names the exact setting and permission exemption clearly snapshot-20260910

06 06

org_name parameter for name-lookup tools NEW 55

Adds org_name parameter to name-lookup tools to disambiguate patterns, prompts, and scorers across organizations.

— Names the parameter but not the tool or endpoint it modifies product docs

07 00

TypeScript SDK scorer score objects without name IMPROVED 55

TypeScript SDK (v3.31.0) scorer functions in the test runner can now return a score object without a name field, using 'score' as the key when name is omitted.

— Exact field behavior and version named, easy to apply snapshot-20260910

08 07

Dashboard sections for charts NEW 55

Group a dashboard's charts into named sections that can be reordered, duplicated, and collapsed, with collapsed state persisted per browser; available on Pro and Enterprise plans.

— Describes behavior and plan availability, no exact UI path snapshot-20260910

09 08

Patterns AI trace investigator NEW 50

Patterns is an AI-driven scheduled investigation of traces that surfaces recurring problems, cost trends, and cohort-level failure modes; each finding is saved with supporting traces and a suggested fix.

— Explains mechanism but no config, endpoint, or navigation path snapshot-20260910

10 00

Pydantic AI span metadata parameters IMPROVED 50

Adds temperature, max_tokens, and similar invocation parameters to Pydantic AI LLM span metadata in the Python SDK (v0.37.0) when model settings are provided.

— Names fields and version but no usage example snapshot-20260910

11 09

Shareable dashboard chart links NEW 40

Adds a shareable link to individual dashboard charts that opens the chart in fullscreen mode.

— Clear UI action but thin on mechanism product docs

12 00

toolCallId in Vercel AI SDK v6 spans IMPROVED 40

Vercel AI SDK v6 tool call spans now include toolCallId in span metadata (TypeScript SDK v3.31.0).

— Names the field and SDK version, thin on further detail snapshot-20260910

└──▷ BREAKING ON UPGRADE

!Task and Classifier implementations must now be thread-safe due to concurrent eval case execution.
!Task, Scorer, and Classifier implementations in the Java SDK (v0.3.23) must now be thread-safe, as eval cases run concurrently.

Was this useful?

AI-Infra-Guard

1 RELEASE · 2026-09-10 NOTES Rank

AI-Infra-Guard is an AI red teaming platform that scans AI infrastructure for CVE vulnerabilities, MCP/Agent skill risks, jailbreak weaknesses, and API relay security issues.

AI-Infra-Guard's v4.6.1 release expands the API Checker's model fingerprinting coverage to include Gemini, Gemma, and GLM models.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Gemini, Gemma, and GLM fingerprints in API Checker IMPROVED 45

The API Checker now identifies Gemini, Gemma, and GLM AI models via fingerprints, and the GLM-5.3-Flash fingerprint baseline was updated.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names models added but no mechanism or usage detail v4.6.1

Was this useful?

ai-safe2-framework

1 RELEASE · 2026-09-10 NOTES Rank

Ai-Safe2-Framework provides governance, risk, and compliance controls for securing agentic AI systems and non-human identities.

safe2 CLI 0.2.0 adds offline Challenge evidence bundle generation and verification, a hardened Skill Trust Gate command that replaces a retired script, and stricter AISM ingestion validation, alongside a small adapter flag and expanded CI coverage.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 6 FEATURES 5 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

Skill Trust Gate CLI command and detection rules BREAKING 95

Adds safe2 gate skill PATH --strict to replace the retired standalone skill-gate script — the old script now exits with an error, so all uses must be replaced with safe2 gate skill PATH --strict.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ The new command enforces expanded file-inspection scope across scripts, extensionless files, and unfamiliar extensions, and ships new Skill Trust Gate detection rules covering injection directives, credential paths, dynamic execution, and other suspicious patterns, with coverage counts reporting what was inspected.

— Names exact command, flag, and migration path in full. 2026-09-09_safe2_CLI_0.2.0

02 02

Offline Challenge 001 evidence bundle generation NEW 90

Adds safe2 challenge quickstart 001 --output-dir <dir> to generate an offline Challenge 001 evidence bundle — including Decision Cards, inert scenario results, and optional Ed25519 signatures — without model credentials or paid APIs.

— Exact command and bundle contents given. 2026-09-09_safe2_CLI_0.2.0

03 03

Challenge bundle verification command NEW 90

Adds safe2 challenge verify-bundle <dir> to verify a Challenge bundle's internal consistency, checking file inventory, hashes, source binding, and artifact integrity; accepts --expected-sha256 to pin the manifest against a separately obtained trusted fingerprint.

— Exact command, flag, and verification checks named. 2026-09-09_safe2_CLI_0.2.0

04 04

Stricter AISM ingestion validation BREAKING 60

AISM ingestion now rejects ambiguous JSON, unsafe paths, invalid Challenge runs, and existing output files — use a new output filename when re-running.

— Clear before/after behaviour with migration note. 2026-09-09_safe2_CLI_0.2.0

thinner coverage below

05 05

SkillSpector adapter --executable flag NEW 55

Adds --executable flag to the SkillSpector adapter for configuring a separate provider environment.

— Flag named but mechanism thinly described. 2026-09-09_safe2_CLI_0.2.0

06 00

Expanded CI Python version coverage IMPROVED 30

Extends CI coverage to Python 3.11–3.14 (previously only Python 3.11 was validated).

— Simple internal CI change, minimal reader action. 2026-09-09_safe2_CLI_0.2.0

└──▷ BREAKING ON UPGRADE

!The retired standalone skill-gate script now exits with an error; replace all uses with safe2 gate skill PATH --strict.
!AISM ingestion now rejects ambiguous JSON, unsafe paths, invalid Challenge runs, and existing output files — use a new output filename when re-running.

Was this useful?

Doberman-Core

1 RELEASE · 2026-09-09 NOTES Rank

Doberman-Core is a runtime security tool that blocks unsafe or unintended AI agent inputs, outputs, and tool calls before execution.

Doberman-Core v0.18.7 adds phone-based Approve/Deny challenges over ntfy, tightens detection against command-wrapper bypasses, and closes a fail-open gap in Claude Code hooks, alongside more auditable decision logs and tighter elevation revocation.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 6 FEATURES 3 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

Phone-based Approve/Deny challenges via ntfy NEW 90

doberman phone setup pushes 2FA and elevation challenges to a phone via ntfy with Approve and Deny buttons; on 2FA tiers the tap replaces the TOTP code.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ --server points setup at a self-hosted ntfy instance instead of the public one.

Enroll phone approvals against a self-hosted ntfy server so challenge commands never reach the public ntfy instance.

$ doberman phone setup --server https://ntfy.example.com

— Names command, flag and mechanism with a runnable example. v0.18.7

02 02

Timeout pin on Claude Code hook entries BREAKING 90

Claude Code hook entries now pin a timeout that outlasts Doberman's challenge ceiling, preventing a timed-out hook from letting a call through before Doberman denies. Existing hook entries lack this pin and will continue to fail open on timeout until you re-run doberman install-hooks (add --host cursor for Cursor); doberman doctor flags un-pinned entries.

Refresh hook entries after upgrading to pick up the timeout pin that prevents fail-open on challenge timeout.

$ doberman install-hooks --host cursor

— Exact migration command and flag given, with mechanism and impact. v0.18.7

03 00

Wrapper-bypass detection for destructive commands IMPROVED 70

Destructive or egress commands wrapped in builtin, eval, strace, flock, unshare, or taskset now classify the same as their unwrapped form instead of passing the scanner unnoticed.

— Names all six wrappers and the before/after behaviour; no user action needed. v0.18.7

thinner coverage below

04 03

JSONL decision log output NEW 55

doberman log --jsonl output records which code path resolved an authentication and whether a person actually approved it in every decision-log row.

— Names the flag but gives no usage example. v0.18.7

05 00

Re-verification of approved deletes before execution IMPROVED 45

Deletes approved through a host hook are re-verified before execution, showing the blast radius and denying if the file set changed since approval.

— Mechanism described but no named surface or user action. v0.18.7

06 00

Elevation revocation on post-approval BLOCK IMPROVED 45

A post-approval BLOCK now revokes the elevation it just granted instead of leaving it live until its TTL expires.

— Behaviour change explained but no configurable surface or example. v0.18.7

└──▷ BREAKING ON UPGRADE

!Existing Claude Code hook entries lack the new timeout pin and will continue to fail open on timeout until you re-run doberman install-hooks (add --host cursor for Cursor); doberman doctor flags un-pinned entries.

Was this useful?

garak

1 RELEASE · 2026-09-09 NOTES Rank

Garak scans language models and applications for vulnerabilities including prompt injection and data leakage.

garak v0.17.0 adds EU AI Act risk-category mapping to probe results, extends the Ollama generator with authentication and custom client support, and drops Python 3.10 in favor of 3.13.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ WHAT SHIPPED · 3 FEATURES 2 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 00

Ollama generator auth and parameter forwarding IMPROVED 55

The Ollama generator now supports authentication and custom client kwargs, and forwards generation parameters to Ollama via the options dict.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names the options dict but no exact flags or config keys v0.17.0

02 01

EU AI Act risk-category tags for probes NEW 45

Adds EU AI Act risk-category reference tags so probe results can be surfaced and grouped by EU AI Act risk categories.

— Describes the capability but no concrete usage path given v0.17.0

03 02

Python version support updated BREAKING 35

Python 3.10 is no longer supported; Python 3.13 is now supported.

— Clear version change but minimal explanation of impact v0.17.0

└──▷ BREAKING ON UPGRADE

!Python 3.10 is no longer supported; Python 3.13 is now supported.

Was this useful?

Anthropic

1 RELEASE · seen 2026-09-10 NOTES Rank

Anthropic extended per-message effort changes (beta) to Google Cloud for Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Anthropic provides Claude AI models and APIs for building applications that generate, analyze, and automate work.

Anthropic extended per-message effort changes (beta) to Google Cloud for Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Per-message effort changes on Google Cloud NEW 70

The per-message effort changes beta, accessed via the mid-conversation-output-config-2026-07-01 beta header, is now available on Google Cloud for Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

— Names header and models but not mechanism of effort changes. snapshot-20260910

Was this useful?

DeepSeek Models

1 RELEASE · 2026-09-10 NOTES Rank

DeepSeek launched V4.1-Flash, a new smallest-in-family model with native multimodal visual understanding and lower pricing, while retiring older V4 Flash variants in favor of automatic routing to the new model.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

The DeepSeek API uses an API format compatible with OpenAI/Anthropic. By modifying the configuration, you can use the OpenAI/Anthropic SDK or softwares compatible with the OpenAI/Anthropic API to access the DeepSeek API.

DeepSeek launched V4.1-Flash, a new smallest-in-family model with native multimodal visual understanding and lower pricing, while retiring older V4 Flash variants in favor of automatic routing to the new model.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

Legacy V4 Flash model names rerouted to V4.1 Flash BREAKING 70

After 12:00 Beijing Time on September 14, 2026, all requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at the V4.1 Flash price.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ The model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are now temporarily routed to V4.1 Flash, and the original V4 Flash and V4 Flash Vision Exp models have been retired.

— Exact model names and cutover date given, no migration steps Date: 2026-09-10

02 02

DeepSeek-V4.1-Flash model launch NEW 67

DeepSeek released DeepSeek-V4.1-Flash, callable via the model name deepseek-flash on the DeepSeek API. It is the smallest model in a new architecture family designed for a higher capability ceiling, faster inference, and higher throughput, and it includes native multimodal visual understanding built in. API pricing for DeepSeek-V4.1-Flash was also reduced relative to prior generation models.

— Names model, call name, and pricing change but no numbers Date: 2026-09-10

└──▷ BREAKING ON UPGRADE

!After 12:00 Beijing Time on September 14, 2026, all requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at the V4.1 Flash price.
!The model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are now temporarily routed to V4.1 Flash; the original V4 Flash and V4 Flash Vision Exp models have been retired.

Was this useful?

Cohere Models

1 RELEASE · 2026-09-09 NOTES Rank

Cohere launched North Small Translate, a dedicated machine-translation model available through the API, as open weights, and for private enterprise deployment.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

Cohere develops Command, Aya, Embed, Rerank, transcription, and document-understanding models.

Cohere launched North Small Translate, a dedicated machine-translation model available through the API, as open weights, and for private enterprise deployment.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

└──▷ 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

North Small Translate machine-translation model NEW 90

north-small-translate-1-0 is a 218B-parameter (25B active) mixture-of-experts model purpose-built for machine translation across more than 50 languages and locale variants.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠ It is available via the Chat V2 API on the free tier with a 16K context length, as FP8 open weights on Hugging Face under CC BY-NC 4.0 for non-commercial use, and for private enterprise deployment through Model Vault under a commercial license with a suggested hardware configuration of two H100 GPUs or one B200 GPU.

— Names model, params, API, tiers, weights license, and hardware specs September 9, 2026

Was this useful?

The AI Toolchain — issue 022, September 10, 2026

Approve a risky agent action from your phone, and stop hook failures from silently allowing it

Provision Pinecone projects, service accounts and API keys from code instead of the console

Run a 780B MoE and new ASR and speech models from the same library you already use

Run eval and debugging loops without maintaining the runner yourself

Ship an MCP gateway stack from a declarative file instead of wiring clients by hand

Put a human approval step in front of dangerous agent actions

Merge and clean agent memory stores across machines and CI

On your toolchain