Issue 007 — The AI Toolchain

Groq

1 RELEASE · seen 2026-08-25 NOTES

Groq is a high-speed AI inference platform that runs large language models extremely fast with minimal latency.

Groq expanded its model lineup this window with two new OpenAI gpt-oss reasoning endpoints and added configurable reasoning effort levels for Qwen3 models.

└──▷ WHAT SHIPPED · 1 FEATUREmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Qwen3 model support with configurable reasoning effortNEW65

Adds support for Qwen3 models, with qwen/qwen3.8-27b supporting low, medium, and high reasoning effort levels, where high maps to the model's native xhigh mode. A default value uses the model's built-in default, and reasoning can also be disabled entirely.

— Names model and effort levels but no API field example shown.product docs

Was this useful?

OpenAI

1 RELEASE · seen 2026-08-25 NOTES

OpenAI provides APIs and tools for accessing advanced language models like GPT for building AI-powered applications.

OpenAI added hard monthly spend limits at the organization and project level, with API calls rejected once the cap is hit, plus alerting to warn teams beforehand.

└──▷ WHAT SHIPPED · 1 FEATUREmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Hard spend limits and spend alertsNEW65

Organizations and projects can now set hard monthly spend limits; once tracked monthly spend reaches the configured cap, API calls return HTTP 429 errors. Spend alerts notify teams before a limit is hit and traffic is interrupted.

— Names status code and behavior but no config UI/API detailssnapshot-20260825

Was this useful?

emisar

1 RELEASE · 2026-08-25 NOTES

An MCP that lets AI tools securely connect to your infrastructure, write IaaS code, debug issues, and assist during incidents - without risking production stability. Built for security teams to approve and infrastructure teams to experience like magic.

emisar v0.42.0 adds PKI-backed signed dispatch, new pack diff/verify/suggest tooling, full MCP Bridge command coverage with native Windows support, Nomad namespace support in packs, and a set of authorization, OAuth and SCIM hardening changes.

└──▷ WHAT SHIPPED · 11 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Pack diff, verify, and catalog-based suggestNEW80

emisar pack diff shows upgrade changes between pack versions, emisar pack verify proves pack setup is correctly configured, and pack suggest now accepts a registry catalog URL as a source.

— Names three exact CLI commands, each runnable todayv0.42.0

02

PKI-backed signed dispatchIMPROVED60

Signed dispatch now accepts certificates issued from your own PKI, with each certificate naming the runners it may target and runners verifying the chain against configured trust anchors.

— Explains verification mechanism but no exact command or config keyv0.42.0

03

Windows support for MCP BridgeNEW60

Windows support was added to the MCP Bridge, including VS Code integration, native ARM64 binaries, direct authentication, and safer client-config replacement with symlink-safe backups.

— Names several concrete Windows-specific additionsv0.42.0

thinner coverage below

04

MCP Bridge connection lifecycle managementIMPROVED50

The MCP Bridge now owns client connect and disconnect lifecycle, drains runbook output automatically, and returns actionable validation errors.

— Describes behaviour change but no config or flag to act onv0.42.0

05

Bridge commands for every MCP toolNEW45

Every MCP tool is now available as a bridge command, enabling scripts and clients without MCP support to invoke them directly.

— Clear capability but no named command syntax givenv0.42.0

06

Nomad namespace support in packsNEW45

Nomad namespaces are now supported in packs, with setup requirements surfaced in both the console and CLI.

— Names surfaces (console, CLI) but little mechanismv0.42.0

07

Transactional recheck of authority-granting decisionsIMPROVED45

MFA, SSO, email, invitation, device-grant, session-revocation, and SCIM decisions are now rechecked within the transaction that commits the resulting authority.

— Names all decision types but is a background security changev0.42.0

08

Revocation on OAuth refresh token replayIMPROVED45

Replaying a still-valid spent OAuth refresh token now revokes the backing connection and its active successors, requiring reconnection.

— Clear before/after behaviour but nothing for a reader to configurev0.42.0

09

SCIM delete and revival semanticsIMPROVED45

SCIM DELETE now retires the directory resource and its independent OIDC authority; an exact later POST revives only the directory resource.

— Precise behavioural rule but no endpoint or config namedv0.42.0

10

Break-glass override for approval quorumsNEW35

Owners and administrators gain an audited break-glass override for unreachable approval quorums.

— Names the capability but no mechanism or invocation pathv0.42.0

11

Cost and disclosure bounds on diagnosticsIMPROVED35

Metrics, logs, and process diagnostics now carry explicit cost and disclosure bounds; the Redis Sentinel down-state probe is structurally read-only.

— States the change but not what the bounds arev0.42.0

Was this useful?

HeyGen HyperFrames

2 RELEASES · 2026-08-24 → 2026-08-25 NOTES

Write HTML. Render video.

HyperFrames shipped catalog usage tracking for renders and a new drag-based editing interaction for automation lanes in Studio.

└──▷ WHAT SHIPPED · 2 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Catalog usage tracking via hyperframes addNEW65

The hyperframes add command now writes installed catalog items to hyperframes.json, enabling renders to report which catalog blocks the video actually used.

— Names command and config file but no further mechanism detail.v0.8.14

thinner coverage below

02

Drag-segment automation editing in StudioNEW55

Studio automation lanes now support dragging a segment between two automation points to move both endpoints together, with segments thickening on hover and preserving their shape within lane bounds.

— Describes interaction behavior but no named UI path or config.v0.8.13

Was this useful?

Microsoft ONNX Runtime

1 RELEASE · 2026-08-24 NOTES

ONNX Runtime: cross-platform, high performance ML inferencing and training accelerator

ONNX Runtime's WebGPU Plugin EP 0.3.0 expands operator and data-type coverage, adds quantized KV cache and GQA improvements, introduces device-free compile sessions and new performance optimizations, and removes the legacy dynamic WGSL-generation path.

└──▷ WHAT SHIPPED · 10 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Expanded integer data-type coverage across opsIMPROVED70

Expands integer data-type coverage: int64 for Add, Cast, Clip, Concat, Equal, Gather, Min, Max, ReduceSum, Reshape, Sub, Tile, and Where; uint8 for Cast, Expand, Gather, and Reshape; int32/uint32 for CumSum and Tile.

— Enumerates every op per type but no usage sample.plugin-ep-webgpu/v0.3.0

02

Quantized KV cache and GQA improvementsNEW65

Adds quantized KV cache support and expands GQA (Grouped Query Attention) with sliding-window cache, batched right-padded prompts, and FlashAttention graph capture for Gemma 4.

— Names mechanisms and target model but no config path.plugin-ep-webgpu/v0.3.0

03

New WebGPU operators and PagedAttention supportNEW60

Adds initial PagedAttention support, plus new operators: MRotaryEmbedding, GRU, DFT, PRelu, HardSwish, Trilu, Max, Min, and MatMulBnb4.

— Lists concrete new ops but no usage example.plugin-ep-webgpu/v0.3.0

04

2-bit GatherBlockQuantized and ONNX 1.22 supportNEW60

Adds 2-bit GatherBlockQuantized support and integrates ONNX 1.22 with opset 27.

— Names the op and version but no migration detail.plugin-ep-webgpu/v0.3.0

05

Removal of dynamic Duktape/Node WGSL generationBREAKING60

Removes the dynamic Duktape/Node WGSL-generation path in favor of statically generated shader sources, which may break build setups that relied on runtime WGSL generation.

— Names the removed path and warns of build impact.plugin-ep-webgpu/v0.3.0

thinner coverage below

06

Intel subgroup-matrix MatMul and FP16 Gemm kernelsNEW55

Adds Intel subgroup-matrix MatMul and FP16 Gemm kernels with batched-B and odd-N weight support and additional tiling optimizations.

— Describes kernel scope but not how to invoke it.plugin-ep-webgpu/v0.3.0

07

WebGPU JS API configuration optionsNEW50

Exposes WebGPU buffer-cache options through the JavaScript API, adds a robustness provider option, and makes the pending-dispatch limit configurable.

— Names options but no exact flag/key syntax given.plugin-ep-webgpu/v0.3.0

08

Deferred shader-compilation dispatchIMPROVED50

Adds deferred dispatch to parallelize shader compilation during cold start, reducing first-inference latency.

— Explains mechanism and benefit but no toggle named.plugin-ep-webgpu/v0.3.0

09

Device-free compile-only sessionsNEW45

Adds device-free compile-only sessions for offline graph transformation without requiring a physical GPU.

— Describes capability but no API call shown.plugin-ep-webgpu/v0.3.0

10

Online softmax and tuned FlashAttention decodeIMPROVED45

Switches Softmax and non-flash Attention to an online softmax algorithm and adds tuned FlashAttention decode.

— Names the algorithm change but lacks concrete metrics.plugin-ep-webgpu/v0.3.0

└──▷ BREAKING ON UPGRADE

!The dynamic Duktape/Node WGSL-generation path has been removed and replaced with generated shader sources, which may break build setups that relied on runtime WGSL generation.

Was this useful?

Eigen Labs Darkbloom

7 RELEASES · 2026-07-29 → 2026-08-21 NOTES CODE

Private Inference Network on Idle Macs

Darkbloom shipped a default-on CBv2 prefill stack delivering +38% throughput on Apple Silicon, expanded OpenRouter alias routing into a dedicated catalog endpoint, added a production Qwen3.6-35B-A3B vision-language model with inline speculative decoding, and tightened provider privacy by closing plaintext egress paths and rebuilding remote media handling around an inline-encryption pipeline.

└──▷ WHAT SHIPPED · 12 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Default-on CBv2 prefill stack with escape hatchesIMPROVED100

CBv2 prefill stack is now enabled by default, delivering ~1,766 tok/s at 8K context on M4 Max (+38% vs v0.8.5 defaults) via expert-tile trust serving default, solo-prefill stripe (2048-token chunks), recurrent prompt narrowing, and packed prefill for Qwen3.6. Three environment variables opt out of pieces of the new behavior: DARKBLOOM_CBV2_SOLO_PREFILL_STRIPE=0 disables the solo-prefill stripe, which regresses TTFT ~12% under Low Power Mode on throttled or battery-backed providers; DARKBLOOM_CBV2_PREFILL_NARROWING=0 restores pre-v0.8.6 behavior where intermediate chunks return the full [1,512,248320] logits tensor (242.5 MiB/chunk) instead of a one-element handle; and DARKBLOOM_CBV2_MAX_PARTIAL_PREFILLS=1 caps prompt-work rows per scheduler step (FCFS), turning burst TTFTs into a staircase instead of forcing all requests to wait for the full makespan.

Disable the solo-prefill stripe on a battery-backed or Low Power Mode provider to avoid the known ~12% TTFT regression.

$ DARKBLOOM_CBV2_SOLO_PREFILL_STRIPE=0 darkbloom

Opt in to mean-TTFT prefill serialization to spread burst load as a staircase rather than forcing all concurrent requests to wait for the full prefill makespan.

$ DARKBLOOM_CBV2_MAX_PARTIAL_PREFILLS=1 darkbloom

— Named env vars, exact numbers, and runnable commands included.v0.8.6

02

Remote media fetching toggle and inline-encryption pipelineNEW95

The coordinator now accepts OpenAI-style remote http(s) image_url and video_url parts on POST /v1/chat/completions: it fetches each URL once, structurally validates it, and rewrites it as an inline base64 data: URI before coordinator-to-provider encryption, keeping provider Macs from ever fetching request-controlled URLs. Remote media URLs in sender-sealed requests are rejected with instructions to inline them, preserving end-to-end encryption; the behavior can be disabled live via EIGENINFERENCE_MEDIA_FETCH_ENABLED=false, without redeploy, restoring the prior HTTP 400 response.

Disable remote media fetching at runtime on a running coordinator without redeploying — useful when you want to enforce that clients must pre-inline all images.

$ EIGENINFERENCE_MEDIA_FETCH_ENABLED=false

— Endpoint, env var, and exact fallback behavior all named.v0.8.0

03

OpenRouter alias support in model catalogIMPROVED90

OpenRouter-only model aliases were first surfaced in GET /v1/models (v0.8.7), then extended so aliases can reference concrete catalog models as their source, not just standard rollout aliases, with alias source kind persisted to prevent silent retargeting on later takeovers. The coordinator now exposes them exclusively through a dedicated GET /v1/models/openrouter endpoint, omitting them from the main /v1/models list while preserving exact GET /v1/models/{id} retrieval — even at zero connected providers — and sharing source pricing, limits, features, readiness, capacity, and datacenter data in that feed.

Fetch the OpenRouter-specific alias catalog, which now includes aliases backed by concrete catalog models.

$ curl https://api.darkbloom.dev/v1/models/openrouter \
  -H 'Authorization: Bearer sk-db-...'

— Full endpoint evolution and mechanics documented across three releases.v0.8.9 v0.8.7 v0.8.6

04

MTP posture and acceptance metrics endpointNEW90

The local /metrics endpoint now exposes mtp_enabled, mtp_active, mtp_rounds_total, mtp_tokens_proposed_total, mtp_tokens_accepted_total, and mtp_inactive_reason{model,reason} counters in both --local and unified serving modes, making MTP acceptance observable without Datadog Logs.

Poll MTP posture and acceptance counters from a running local provider to verify speculative decoding is active and measure acceptance rates.

$ curl http://localhost:<port>/metrics | grep mtp

— Exact metric names and a runnable query provided.v0.8.5

05

Closed-vocabulary egress and darkbloom report commandBREAKING80

Provider-derived plaintext egress paths are closed: inference failures now cross the WebSocket boundary only as closed-vocabulary codes/reasons, and free-form telemetry plus automatic provider log reporting are retired in favor of the operator-initiated darkbloom report command, which supports --dry-run and authenticated upload.

Review what darkbloom report would upload before sending, preserving macOS unified-log privacy redaction.

$ darkbloom report --dry-run

— Names the exact command and flag, plus the security guarantee it replaces.v0.8.2

06

Qwen3.6-35B-A3B VLM with inline MTPNEW75

Adds production-path text, image, and tool inference for the combined Qwen3.6-35B-A3B VLM artifact via the EngineV2 provider path, with request-owned recurrent state, three-axis mRoPE state, causal vision attention, exact rollback, and source-matched target/assistant memory accounting. Inline MTP (depth-one, serial, exact-target-verified) ships alongside the model; video, prefix reuse, paged KV, compiled decode, packed prefill, and rectangular MTP remain fail-closed.

— Mechanism and scope detailed, but no runnable example given.v0.8.3

07

Adaptive persistent-history MTP for Qwen3.6IMPROVED65

An adaptive persistent-history capture-verify MTP stack for Qwen3.6 was added behind the mtp beta flag (default off), measured at 2.0077x throughput vs non-MTP in canary validation; it was then promoted onto the release pin with persistent head KV and adaptive width selection, still gated behind the same mtp beta flag.

— Beta flag and throughput figure given, but no usage command.v0.8.6 v0.8.5

08

Qwen3.6 E=256 expert-tile prefill routeNEW65

Adds a Qwen3.6 E=256 expert-tile prefill route with fused gate/up projection (SwitchGLU(fuseGateUp: true)) and an opt-in trust refinement, delivering up to +15.2% prefill throughput at 8k context on M4 Max.

— Names the mechanism and performance gain, no config command shown.v0.8.5

09

Config-controlled Gemma 4 26B optimizationsNEW65

The [gemma_optimizations] config block now controls Gemma 4 26B-A4B v0.8.2 behavior, gating layer-18 lazy prefill submission, coupled weighted-expert-unsort, and safe-R1 expert-QMM gate — both default-on.

— Config key and gated behaviors named, no example command.v0.8.2

thinner coverage below

10

In-memory inline video decodingIMPROVED50

Inline video (MP4/QuickTime) is kept off disk by decoding coordinator-inlined bytes through a bounded, memory-backed AVFoundation asset, and any legacy vlm-<UUID>.mp4 files are purged on launch.

— Mechanism and file pattern named, no user-facing action.v0.8.2

11

Self-serve provider payout bank updatesNEW45

Providers can now update their own Stripe Express payout bank account via a new self-serve endpoint, removing the need to contact support for bank account changes.

— Names Stripe Express integration but not the endpoint path.v0.8.2

12

Human-approved agent production accessNEW20

Human-approved agents can now operate in production environments.

— One-line description with no mechanism or surface named.v0.8.5

Was this useful?

Augment Code

1 RELEASE · seen 2026-08-25 NOTES

Augment Code is an AI-powered code completion and generation tool that helps developers write code faster with intelligent suggestions.

Augment Code's Auggie CLI 0.36.0 adds control over model reasoning effort and removes the secret get --reveal flag.

└──▷ WHAT SHIPPED · 2 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

thinner coverage below

01

Removal of secret get --reveal flagBREAKING45

The secret get --reveal flag has been removed from the CLI.

— Names exact flag but no migration guidancesnapshot-20260825

02

Reasoning effort flag in Auggie CLINEW40

Auggie CLI gains a new flag that lets developers control the model's reasoning effort level from the command line.

— Flag exists but exact name not givensnapshot-20260825

└──▷ BREAKING ON UPGRADE

!The secret get --reveal flag has been removed.

Was this useful?

Daytona

2 RELEASES · seen 2026-08-25 NOTES

Daytona is an open-source development environment platform that enables developers to spin up standardized, reproducible coding environments instantly.

Daytona expanded Sandbox object metadata for infrastructure control and improved CLI and SDK consistency with a new organization members command and uniform daemon error codes.

└──▷ WHAT SHIPPED · 3 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

New Sandbox object fields for class and poolingNEW63

The SDK now exposes sandboxClass, warmPoolId, gpuType, desiredState, daemonVersion, and otelEndpointOverride fields directly on the Sandbox object.

— Names all six fields but no usage detailSandbox class and warm pool fields on Sandbox

thinner coverage below

02

Organization members CLI commandNEW35

A new CLI command lists organization members.

— Names the capability but not the exact command syntaxsnapshot-20260825

03

Consistent daemon error codes across SDKsIMPROVED35

Daemon error codes are now propagated consistently to all SDKs, enabling uniform error handling across SDK consumers.

— Explains benefit but no specific codes or SDKs listedsnapshot-20260825

Was this useful?

OpenAI Codex CLI

1 RELEASE · 2026-08-25 NOTES CODE

Lightweight coding agent that runs in your terminal

Codex CLI's alpha release adds a new AWS Bedrock authentication mode, a Guardian v2 scope restricted to computer-use tools, and several TUI refinements around thread naming, event persistence, and Vim editing.

└──▷ WHAT SHIPPED · 6 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Computer-use-only scope for Guardian v2NEW85

Adds features.guardianv2.review_scope.computer_use_only config option to restrict async classification and fast approvals exclusively to browser and computer-use REPL tools, giving this scope precedence over sandboxed_exec_commands when both are enabled.

— Exact config key and precedence rule given, ready to setrust-v0.150.0-alpha.9

02

Managed AWS access keys for Amazon BedrockNEW80

Adds BedrockAccessKeys as a distinct auth mode, enabling the experimental amazonBedrockAccessKeys app-server login flow with SigV4-signed Bedrock requests; credentials are persisted in the configured auth store and reported as bedrockAccessKeys in login status.

— Names auth mode, login flow, and status field but no CLI command shownrust-v0.150.0-alpha.9

thinner coverage below

03

Thread naming, /rename, and event persistence in TUINEW50

Generates descriptive TUI thread titles and suggests conversation-based titles via /rename; also persists realtime events in the thread timeline and adds persisted thread artifact models.

— Names the `/rename` command; persistence mechanism left vaguerust-v0.150.0-alpha.9

04

Credential alias brokering for child processesIMPROVED45

Brokers credential aliases in child environments by discovering, replacing, and restoring credentials inherited from the parent even when their canonical provider variable is filtered out of the child environment.

— Mechanism explained but no config key or command namedrust-v0.150.0-alpha.9

05

Dot-repeat in Vim modeNEW35

Adds dot-repeat support to Vim mode in the TUI.

— Clear editor feature but no further detail on scoperust-v0.150.0-alpha.9

06

Hooks for interrupted turnsNEW25

Adds hooks for interrupted turns.

— Bare mention with no hook name or trigger detailrust-v0.150.0-alpha.9

Was this useful?

Cline

2 RELEASES · 2026-08-24 → 2026-08-25 NOTES CODE

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

Cline's desktop app consolidated its extension surfaces into a single Customize hub and overhauled the Models page with OAuth sign-in, while the SDK added two new model providers and capped the hub's event log to keep disk usage bounded.

└──▷ WHAT SHIPPED · 9 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Unified Customize hub for extensionsBREAKING70

Plugins, MCP, Skills, Rules, Hooks, and Tools are merged into a single Customize hub with tabbed sections, live counts, and an inline Browse catalog; installing from the catalog immediately surfaces the item above the Browse section. The separate Marketplace page is removed, so any workflow relying on it must move to the Customize hub.

— Names all six merged areas and the removed page, but no config surfacedesktop-v0.0.17

02

Models page redesign with OAuth sign-inIMPROVED70

Providers are grouped into Connected, Popular, and All categories with explicit Connect / Disconnect / Sign out actions. OAuth providers now offer browser sign-in instead of an API key field, with a collapsed manual-key fallback where supported.

— Describes exact UI states and the OAuth mechanism changedesktop-v0.0.17

03

Hub event log capped at 64 MiBIMPROVED70

Caps the hub's durable event log at 64 MiB on disk; oldest events are dropped first and the SQLite file is vacuumed to reclaim space. Pruning now also runs after every 16 MiB appended, instead of waiting for the previous hourly sweep.

— Concrete limits and mechanism given, though not user-facing actionsdk/sdk/v0.0.79

thinner coverage below

04

Dedicated Voice settings pageNEW55

Voice input moves to its own Settings → Voice page that lists only transcription-capable connected providers and preselects a default model; the composer microphone button now appears only after a voice model is configured.

— Clear navigation path and behaviour, no deeper mechanism givendesktop-v0.0.17

05

Sidebar, session, and schedule navigation overhaulIMPROVED55

Session search now opens in a dialog and searches full history rather than only sessions already loaded in the sidebar. Sidebar sessions are always grouped by project, with pinned sessions leading each group (the Favorite action is renamed Pin) and scheduled sessions marked by an inline clock icon. Dedicated New, Schedule, and Customize rows are added below the logo — New starts a fresh task and places the cursor directly in the composer — and the Schedule page also gains suggested schedule templates.

— Several distinct UI tweaks named but no single actionable entry pointdesktop-v0.0.17

06

Platform-native line endings for new filesIMPROVED40

New files are now created with the platform-native line endings — CRLF on Windows, LF on macOS/Linux — instead of a fixed convention.

— Specific behaviour named but limited scope and no configdesktop-v0.0.17

07

Default model change for Aki.io and NanoGPTBREAKING40

The resolved default model for the Aki.io and NanoGPT providers changes in this release — users of either provider without a pinned model will get a different default after upgrading.

— Names affected providers and the fix (pin a model), no further detailsdk/sdk/v0.0.79

08

Removed agent todo tool and Agenda panelBREAKING35

The agent todo tool and the Agenda panel are removed from the desktop app; scheduled tasks themselves are unaffected by the removal.

— States what was removed but not why or migration pathdesktop-v0.0.17

09

AgentRouter and Opper model providersNEW35

Adds two new model providers — AgentRouter and Opper — to the SDK's model catalog.

— Names the providers but gives no further detailsdk/sdk/v0.0.79

└──▷ BREAKING ON UPGRADE

!The Favorite action on sidebar sessions is renamed to Pin; any workflow or documentation referencing 'Favorite' must be updated.
!The separate Marketplace page is removed; extension discovery now lives inline within the Customize hub.
!The agent todo tool and Agenda panel are removed from the desktop app.
!The resolved default model for Aki.io and NanoGPT changes in this release — if you use either provider without pinning a model you will get a different default after upgrading.

Was this useful?

Diagram Design

changes since 2026-08-11 CODE

38 editorial diagram types for Claude Code, Codex, and Pi. Self-contained HTML + SVG.

Diagram Design adds draw.io import, native Pi support, named client profiles, new chart types, and automatic plugin updates.

└──▷ GET THIS VERSION

$ git clone --branch commits-2026-08-11 https://github.com/cathrynlavery/diagram-design.git
# already have the repo? check out this version:
$ git checkout commits-2026-08-11

›Adds the /diagram-design:import workflow to redraw draw.io files (raw, compressed, PNG-embedded, and SVG-embedded) at a chosen format, size, and detail level into the project design system.
›Adds native Pi package support for compatibility with the Pi platform.
›Adds named client profiles (PR #61).
›Adds automatic plugin updates via native marketplaces with a version gate.
›Adds treemap diagram type for part-of-whole visualization by area.

+11 moreshow less

›Adds dumbbell as a variant of the Bar diagram type.
›Adds slopegraph as a variant of the Line diagram type for showing change between two states.
›Ships ten additional editorial diagram types in a single release (commit 4691a2f).
›Adds verify-ridgeline.py script to verify that a ridgeline chart's drawn ridges match their declared bin values, enforcing eight geometric invariants including single amplitude, even pitch, shared bins, and zero closure.
›Adds lint-render.py script to lint diagram examples as rendered in headless Chromium, catching clipped SVG content, collapsed SVGs, page overflow, and runtime errors that source-level linting cannot detect.
›Adds adversarial test suite scripts/test-verify-sankey.py for the Sankey verifier, covering both polarities (fires when it should, stays quiet when it shouldn't).
›Adds a pre-draw checkpoint, docs-sync gate, packaged self-check, and ADRs for the skill workflow.
›Adds animated example example-queue-animated.html for Semantic Pattern #1 (Fan-in Queue / Bottleneck).
›Adds animated example example-paved-road-animated.html for Semantic Pattern #5 (Secure Paved Road).
›Deploys a live diagram gallery to GitHub Pages.
›Adds a multi-OS CI testing matrix covering Linux, Windows, and macOS with automated visual artifact packaging and upload on linter failures, plus a GitHub Step Summary table for matrix run results.

Was this useful?

Cotool

1 RELEASE · 2026-08-24 NOTES

Cotool is an AI-powered security tool that analyzes code for vulnerabilities and provides automated remediation recommendations for developers.

Cotool v0.65.0 expands its integration surface with new Microsoft Sentinel, Control D, and GitHub Actions connectors, alongside improvements to structured outputs and detection accuracy.

└──▷ WHAT SHIPPED · 5 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

thinner coverage below

01

Microsoft Sentinel integrationNEW55

Adds a Microsoft Sentinel integration for querying logs, triaging incidents, managing analytics rules, and powering detection authoring and environment mapping.

— Names platform and functions but no setup or command detail.v0.65.0

02

Control D integration for DNS investigationNEW45

Adds a Control D integration for investigating DNS activity and managing custom filtering rules across sub-organizations.

— Names integration and scope, no configuration steps given.v0.65.0

03

GitHub Actions workflow toolsNEW45

Adds GitHub Actions workflow tools for creating and updating pull requests, protected by explicit permissions.

— Names permission model but no exact flag or command.v0.65.0

04

Structured outputs with nested views and schema editingIMPROVED45

Enhances structured outputs with table-based nested data views and drag-and-drop schema editing that preserves existing fields.

— Describes UI behaviour but no navigation path or config key.v0.65.0

05

Detection coverage linking to verified rulesIMPROVED30

Improves detection coverage accuracy by linking external alerts to verified rules from connected security platforms.

— Vague on which platforms or mechanism used.v0.65.0

Was this useful?

llama.cpp

1 RELEASE · 2026-08-25 NOTES CODE

LLM inference in C/C++

llama.cpp's v0.3.0 release adds multi-GPU tensor-split inference for DeepSeek 4, a new dots3-note multimodal model, and an upgraded ggml backend, alongside several smaller model, UI and performance additions.

└──▷ WHAT SHIPPED · 9 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Tensor-split mode for DeepSeek 4 multi-GPUNEW75

Adds the -sm tensor flag to enable tensor-split mode for DeepSeek 4 multi-GPU inference, letting the model be split across multiple GPUs.

Run DeepSeek 4 across multiple GPUs using tensor-split mode for better VRAM distribution.

$ llama-cli -m deepseek4.gguf -sm tensor

— Named flag with a runnable example commandv0.3.0

02

Slot debug diff window env var in llama-serverNEW70

Adds the LLAMA_SERVER_SLOTS_N_DIFF environment variable to widen the slot debug diff window in llama-server, for diagnosing slot assignment issues.

Widen the slot debug diff window when diagnosing slot assignment issues in llama-server.

$ LLAMA_SERVER_SLOTS_N_DIFF=20 llama-server -m model.gguf

— Named env var with a runnable example commandv0.3.0

03

ggml upgraded to v0.22.0IMPROVED65

Upgrades ggml to v0.22.0, bringing meta-backend tensor split, per-op Metal kernels with parallel compilation, and non-in-place ggml_clamp.

— Names version and internal changes but no direct usage stepsv0.3.0

thinner coverage below

04

Mamba2 throughput improvementIMPROVED55

Improves mamba2 throughput by flattening in/out projections to dispatch GEMM instead of GEMV.

— Explains mechanism but gives no benchmark numbersv0.3.0

05

dots3-note multimodal model supportNEW50

Adds the dots3-note multimodal model along with a new DSA-ISWA KV cache type, including vision and audio support.

— Names model and cache type but no usage steps givenv0.3.0

06

New model architecture support: MTP and DSparkNEW40

Adds MTP (multi-token prediction) support for the GLM-4.5-Air model and DSpark architecture support for the bailingmoe3 model.

— Names two model/architecture additions with no further detailv0.3.0

07

`-no-cnv` CLI option removedBREAKING40

The -no-cnv CLI option has been removed.

— States removal but offers no replacement or migration pathv0.3.0

08

WebP image decoding in multimodal subsystemNEW35

Adds WebP image decoding support in the multimodal subsystem via ffmpeg.

— Names format and dependency but no usage stepsv0.3.0

09

Tabbed chat navigation in web UINEW30

Adds tabbed navigation for chat conversations in the built-in web UI.

— Describes UI addition without exact navigation pathv0.3.0

└──▷ BREAKING ON UPGRADE

!The -no-cnv CLI option has been removed.

Was this useful?

llmfit

3 RELEASES · 2026-08-03 → 2026-08-17 NOTES CODE

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

llmfit expanded its model catalog and RamaLama discovery, and gave the fit command a new provider filter alongside several TUI polish updates.

└──▷ WHAT SHIPPED · 5 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Provider filter for fit commandNEW90

Adds a --providers flag to llmfit fit, accepting a comma-separated, case-insensitive list of provider names (e.g. unsloth,bartowski) to narrow fit results to specific canonical or GGUF providers.

Filter fit results to models from specific GGUF providers to find what unsloth or bartowski quantizations will run on your hardware.

$ llmfit fit --providers unsloth,bartowski -n 10

Pipe provider-filtered fit results as JSON into jq for scripted or agent consumption.

$ llmfit fit --json --providers unsloth,bartowski --perfect

— Exact flag with usage syntax and runnable example commandsv1.1.7

thinner coverage below

02

Qwen3.8 model family added to catalogNEW55

Adds the Qwen3.8 model family to the built-in catalog, making Qwen3.8 variants scoreable and recommendable against local hardware.

See how the new Qwen3.8 models rank against your hardware, filtered to good-or-better fits.

$ llmfit recommend --json --limit 5 | jq '.models[] | select(.name | test("Qwen3"))'

— Runnable example ranks new models, but detail stays briefv1.1.10

03

TUI fit view usability improvementsIMPROVED55

The TUI now marks list rows matched through a GGUF source so it's visually clear which results come from GGUF-backed providers, and the Fit box spells out active advanced range filters (e.g. Too Tight ≤2B, mem≥50%) in warning color instead of cryptic R/M markers, with the empty-results hint naming the active filter explicitly.

— Concrete UI labels named but only a navigation path, no commandv1.1.7

04

RamaLama model discovery via MCP and local storeNEW50

llmfit discovers RamaLama runtime models via MCP, and separately detects RamaLama store models via ramalama ls when no server is running, so locally stored models appear in fit results without requiring an active runtime.

— Names discovery mechanisms but gives no example commandv1.1.7 v1.1.10

05

Benchmark counts in TUI hardware pickerIMPROVED30

Displays per-preset benchmark counts in the TUI hardware picker, so you can see at a glance how many community measurements back each hardware preset.

— Describes the addition but offers no usage detailv1.1.8

Was this useful?

PydanticAI

1 RELEASE · 2026-08-25 NOTES

How Python does AI: agents, realtime voice, image generation, embeddings. Every model, every interface, typed end to end.

PydanticAI v2.34.0 adds support for a new model provider and a skill to help teams migrate off LangChain.

└──▷ WHAT SHIPPED · 2 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

thinner coverage below

01

GLM-5.3 support via ZhipuModelNEW45

Adds support for GLM-5.3 through a new ZhipuModel class.

— Names the class but no usage detailsv2.34.0

02

LangChain migration skillNEW35

Adds a LangChain migration skill to help teams port existing LangChain agents to PydanticAI.

— States purpose but no mechanism or usage stepsv2.34.0

Was this useful?

OpenWorker

1 RELEASE · 2026-08-24 CODE

OpenWorker v0.2.0 ships a major permission engine overhaul with new approval outcomes and a shadow evaluation mode, adds Memory and Skills, introduces an OpenAI Responses provider with custom endpoint support alongside new BytePlus/Ark providers, and adds persona sharing with built-in security coworker bundles.

└──▷ WHAT SHIPPED · 15 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Append-only board event store with hash-chainingNEW80

Introduces an append-only board event store with hash-chaining per space for tamper-evidence, verifiable via verify_chain, and supports taint provenance marking.

— Names `verify_chain` and `taint` but no example invocationv0.2.0

02

OpenAI Responses provider with custom endpoint supportNEW80

Adds the OpenAI Responses provider, enabling reasoning and tool use for native OpenAI models; leaving the Endpoint field blank for OpenAI now automatically routes the session to the Responses provider. Custom Responses-compatible endpoints are also supported by entering a URL in the Endpoint field.

Trigger a session that routes to the new Responses provider with reasoning enabled by leaving the endpoint blank for a native OpenAI model.

📍In the provider settings screen, select OpenAI and leave the Endpoint field blank — the session is automatically routed to the Responses provider, enabling reasoning and tools.

Point OpenWorker at a custom Responses-compatible endpoint (e.g. a self-hosted or third-party proxy) instead of the default OpenAI URL.

📍In the provider settings screen, select OpenAI and enter your custom Responses endpoint URL in the Endpoint field.

— Exact UI steps for routing and custom endpoints givenv0.2.0

03

Persona sharing and security coworker bundlesNEW80

Adds persona sharing v1: export and import persona bundles with versioning and consent metadata via 'Export bundle' and 'Import bundle' in the Coworkers panel. Ships built-in security coworker persona bundles in the sidecar, plus a self-contained findings report page for security bundles.

Export a persona bundle to share a coworker configuration — including skills, MCP connectors, and consent metadata — with another user or team.

📍In the Coworkers panel, open the persona you want to share, then choose Export bundle. The resulting file includes version and consent fields and can be imported on another instance via Import bundle.

— Exact UI navigation for export/import given, bundles namedv0.2.0

04

Permission engine approval outcomes and command handlingNEW70

The permission engine gains ALWAYS_DOMAIN and READONLY_SESSION approval outcomes, giving session-wide grants for classifier-approved read-only shell commands; a session read-only command grant surfaced on the consent card; compound command splitting so each part of a multi-part shell command is checked independently; and server-side validation of approval answers with auditing of autonomy changes.

— Named enum outcomes and mechanism described, but no config surface shownv0.2.0

05

Board API endpoints with per-actor tokensNEW70

Adds /v1/board/ API endpoints with per-actor board tokens (identity + access) designed to be handed to external harnesses and other machines.

— Named endpoint and token model, but no example call shownv0.2.0

06

Shadow evaluation mode for permission reviewsNEW65

A new shadow evaluation mode lets the auto-approve reviewer record its decisions while a human still approves each action, showing a deny card with a one-shot 'Allow anyway' option when the reviewer would have denied.

— Clear mechanism described, no config or API name givenv0.2.0

07

Mode.AUTO renamed to Mode.BYPASS_APPROVALSBREAKING65

The Mode.AUTO enum value is renamed to Mode.BYPASS_APPROVALS; any code or tests referencing Mode.AUTO will break.

— Exact enum rename given as clear migration notev0.2.0

thinner coverage below

08

BytePlus and Volcengine Ark provider supportNEW55

Registers BytePlus and Volcengine Ark as new providers, including their model catalogs, credential verification via the Responses provider, and branding.

— Named providers and mechanism, no setup steps shownv0.2.0

09

Artifact viewer with sandboxed HTMLNEW55

Introduces the Artifact viewer with an airtight sandbox for agent-generated HTML and an 'Open in browser' option.

— Names sandbox and open-in-browser option, no technical detail on sandboxingv0.2.0

10

Intel Mac build supportNEW55

Pins Intel Mac (darwin-x86_64) builds via a macos-15-intel GitHub Actions runner, adding DMG artifacts and auto-update manifest entries for Intel Macs.

— Names runner and artifacts but no user-facing actionv0.2.0

11

Per-coworker skills and connector permissionsNEW45

Adds Skills support: per-persona skills wired through a manifest into sessions. Also makes connectors a per-coworker allowlist.

— Names manifest wiring and allowlist but no config keys shownv0.2.0

12

Upgraded ask_user with rich optionsIMPROVED45

Upgrades ask_user with rich options, grouped questions, and option previews.

— Names the ask_user function but no usage examplev0.2.0

13

Reviewer metering and statsNEW40

Adds reviewer metering with a badge, a mode-menu summary, and durable stats tracking the auto-approve reviewer's activity.

— Only names UI elements, no mechanism or numbers givenv0.2.0

14

Coworker picker and folder selection at sendNEW40

Adds a coworker picker with setup chips above the composer and folder selection at send.

— Describes UI elements only, no navigation path givenv0.2.0

15

Memory V1 for remembered facts and instructionsNEW35

Introduces Memory V1: remembered facts and user instructions are surfaced together in a single screen.

— Thin description, no config surface or example givenv0.2.0

└──▷ BREAKING ON UPGRADE

!The Mode.AUTO enum value is renamed to Mode.BYPASS_APPROVALS; any code or tests referencing Mode.AUTO will break.

Was this useful?

deepset Haystack

1 RELEASE · 2026-08-24 NOTES

Open-source AI orchestration framework for building context-engineered, production-ready LLM applications. Design modular pipelines and agent workflows with explicit control over retrieval, routing, memory, and generation.

Haystack v3.1.0 adds context compaction hooks and token counters to keep long-running agents within context limits, introduces AgentTool for wrapping agents as delegatable tools, and ships a wide set of breaking changes spanning deserialization safety, pipeline snapshot shape, evaluator scoring, and retriever/document-store behavior.

└──▷ WHAT SHIPPED · 14 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

Context compaction hooks for agentsNEW95

Adds CompactionHook (from haystack.hooks.compaction) with context_window, compact_at, and compact_to parameters, wired into Agent via hooks={'before_llm': [hook]} to shorten conversation history before LLM calls. SlidingWindowCompactor drops oldest full turns then individual steps, replacing removed content with an omission note. ToolResultPruningCompactor takes min_keep_steps and min_tokens parameters and replaces older/large tool results with placeholders while preserving the most recent tool-calling steps.

Prevent context-window overflows in a long-running agent by sliding off old turns automatically.

python

from haystack.hooks.compaction import CompactionHook, SlidingWindowCompactor
from haystack.components.agents import Agent
from haystack.components.generators.chat import OpenAIResponsesChatGenerator

hook = CompactionHook(
    compactor=SlidingWindowCompactor(),
    context_window=128_000,
    compact_at=0.8,
    compact_to=0.5,
)
agent = Agent(
    chat_generator=OpenAIResponsesChatGenerator(model="gpt-4o"),
    tools=[web_search],
    hooks={"before_llm": [hook]},
)
result = agent.run([ChatMessage.from_user("Summarize recent AI news.")])

— Full parameter list plus a runnable wiring example.v3.1.0

02

Token counting utilities moduleNEW90

New haystack.token_counters module ships three classes, each exposing a .count(messages) method: ApproximateTokenCounter (dependency-free, configurable via chars_per_token), TiktokenCounter (local estimation via encoding parameter, requires pip install tiktoken), and OpenAITokenCounter (calls OpenAI's counting API for exact model-specific counts).

Estimate token usage before sending a request to decide whether compaction is needed, without any extra dependencies.

python

from haystack.token_counters import ApproximateTokenCounter, TiktokenCounter
from haystack.dataclasses import ChatMessage

messages = [ChatMessage.from_user("Explain quantum entanglement.")]

# No extra install required
approx_count = ApproximateTokenCounter(chars_per_token=4.0).count(messages)

# Closer estimate for OpenAI models — requires: pip install tiktoken
tiktoken_count = TiktokenCounter(encoding="o200k_base").count(messages)

print(f"Approximate: {approx_count}, Tiktoken: {tiktoken_count}")

— Named classes, install step, and runnable example.v3.1.0

03

AgentTool for multi-agent delegationNEW85

AgentTool (from haystack.tools) wraps any Agent as a Tool so an orchestrating agent can delegate to it; it exposes name and description parameters and surfaces only the wrapped agent's final reply to the caller.

Build a multi-agent system where a coordinator delegates web research to a specialist agent.

python

from haystack.components.agents import Agent
from haystack.components.generators.chat import OpenAIResponsesChatGenerator
from haystack.dataclasses import ChatMessage
from haystack.tools import AgentTool

researcher = Agent(
    chat_generator=OpenAIResponsesChatGenerator(model="gpt-4o-mini"),
    system_prompt="You are a research specialist. Investigate the task and report your findings.",
    tools=[web_search],
)

research_specialist = AgentTool(
    agent=researcher,
    name="research",
    description="Research a question on the web and report the findings",
)

coordinator = Agent(
    chat_generator=OpenAIResponsesChatGenerator(model="gpt-4o"),
    tools=[research_specialist],
    system_prompt="You coordinate specialists. Delegate research questions, then answer the user.",
)

result = coordinator.run([ChatMessage.from_user("What are the latest LLM benchmarks?")])
print(result["last_message"].text)

— Named class with a full coordinator/specialist example.v3.1.0

04

HAYSTACK_UNSAFE_DESERIALIZATION env varBREAKING85

New HAYSTACK_UNSAFE_DESERIALIZATION environment variable (truthy values 1 or true) bypasses all deserialization safety checks process-wide for Pipeline.load, Pipeline.loads, Pipeline.from_dict, Tool.from_dict, State.from_dict, and the ConditionalRouter/OutputAdapter Jinja sandbox; the value is read once and frozen for the process lifetime. As a result, OutputAdapter and ConditionalRouter components serialized with unsafe: true now raise DeserializationError on load unless Pipeline.load(..., unsafe=True) (or Pipeline.loads/Pipeline.from_dict with unsafe=True) is used.

— Exact env var, values, and affected load methods named.v3.1.0

05

Pipeline snapshot input shape changeBREAKING80

Adds an inputs_format field to PipelineSnapshot.pipeline_state to distinguish the new per-sender input shape {component: {socket: [{sender: ..., value: ...}]}} from the legacy flattened shape. PipelineSnapshot.pipeline_state.inputs and BreakpointException.inputs now use the new shape, so code reading inputs['my_component']['my_socket'] must become inputs['my_component']['my_socket'][0]['value'].

— Exact before/after access pattern given for migration.v3.1.0

06

Agent exit_reason output and reserved schema keyBREAKING75

Agent.run() now returns exit_reason, one of 'text', the name of the tool that satisfied an exit condition, or 'max_agent_steps', also accessible in hooks via state.get('exit_reason'). exit_reason is now a reserved key in Agent.state_schema: defining a custom schema key with that name raises ValueError at initialization, and Agent.state_schema now excludes internally managed keys (messages, step_count, token_usage, exit_reason) — use the new agent.resolved_state_schema attribute to see the full effective runtime schema.

— Names field, values, and migration path via resolved_state_schema.v3.1.0

07

Agent and Toolset warm_up() called on every runBREAKING65

The Agent now calls warm_up() on hooks before every run, not only the first, so hooks with expensive setup must guard against repeated calls (e.g. if self._client is not None: return). The internal _is_warmed_up flag that prevented repeated warm_up() calls on Toolset is also removed, so every call reaches warm_up() directly and custom Tool/Toolset subclasses need their own guard.

— Gives the exact guard pattern needed to migrate.v3.1.0

08

PDF annotation link parsing in convertersNEW60

Adds a link_format parameter to PyPDFToDocument and PDFMinerToDocument that parses PDF annotation links and appends them to page content, matching existing DOCXToDocument behavior.

— Names the exact param and affected components, no usage example.v3.1.0

09

SentenceWindowRetriever window_size=0 now errorsBREAKING60

Passing window_size=0 to SentenceWindowRetriever.run or .run_async now raises ValueError instead of silently falling back to the constructor value; pass None or omit the argument to use the constructor default.

— Exact old/new behavior and the fix to use.v3.1.0

thinner coverage below

10

Agent hook tracing spanNEW55

Adds a content-free haystack.agent.hook tracing span emitted for every Agent hook invocation, recording hook point, hook name, hook type, compaction strategy, estimated context size, compaction trigger status, token target, and whether the compactor returned a replacement.

— Names span and recorded fields but no action for readers.v3.1.0

11

DocumentMAPEvaluator scoring changeBREAKING55

DocumentMAPEvaluator scores may change: average precision now uses all unique valid ground-truth values as the denominator and credits each value at most once; existing evaluation baselines must be recalculated.

— Explains formula change but no example of new vs old scores.v3.1.0

12

InMemoryDocumentStore metadata value matching changeBREAKING55

InMemoryDocumentStore.get_metadata_field_unique_values and its async counterpart now match search_term against the metadata field value (case-insensitive substring) instead of the document content; callers relying on content-matching must filter documents themselves.

— Names the exact methods and behavior shift.v3.1.0

13

Agent.clone() methodNEW50

Agent.clone() returns a new Agent with the same configuration, accepting keyword arguments to override init parameters, e.g. agent.clone(system_prompt='Answer in German.').

— Exact call shown but no separate runnable example.v3.1.0

14

Resource-release methods on retriever componentsNEW50

Adds close() and close_async() methods to AutoMergingRetriever, CacheChecker, DocumentWriter, FilterRetriever, and SentenceWindowRetriever.

— Names all five components but no usage detail.v3.1.0

└──▷ BREAKING ON UPGRADE

!OutputAdapter and ConditionalRouter components serialized with unsafe: true now raise DeserializationError on load unless Pipeline.load(..., unsafe=True) (or Pipeline.loads / Pipeline.from_dict with unsafe=True) is used.
!exit_reason is now a reserved key in Agent.state_schema; defining a custom state_schema key named exit_reason raises ValueError at Agent initialization.
!Agent.state_schema now contains only the user-provided schema, excluding internally managed keys (messages, step_count, token_usage, exit_reason); use agent.resolved_state_schema to get the full effective schema.
!PipelineSnapshot.pipeline_state.inputs and BreakpointException.inputs changed shape from {component: {socket: value}} to {component: {socket: [{sender: ..., value: ...}]}}; reading inputs['my_component']['my_socket'] must become inputs['my_component']['my_socket'][0]['value'].
!DocumentMAPEvaluator scores may change: average precision now uses all unique valid ground-truth values as the denominator and credits each value at most once; existing evaluation baselines must be recalculated.
!Passing window_size=0 to SentenceWindowRetriever.run or SentenceWindowRetriever.run_async now raises ValueError instead of silently falling back to the constructor value; pass None or omit the argument to use the constructor default.
!InMemoryDocumentStore.get_metadata_field_unique_values and its async counterpart now match search_term against the metadata field value (case-insensitive substring) instead of the document content; callers relying on content-matching must filter documents themselves.
!The Agent now calls warm_up() on hooks before every run (not only the first); hooks with expensive setup in warm_up() must guard against repeated calls (e.g., if self._client is not None: return).
!The internal _is_warmed_up flag that prevented repeated warm_up() calls on Toolset is removed; every call now reaches warm_up() directly, so custom Tool or Toolset subclasses with expensive setup in warm_up() must add their own guard.

Was this useful?

LangChain LangSmith

1 RELEASE · 2026-08-10 NOTES

LangSmith is a platform for debugging, testing, and monitoring LLM applications built with LangChain.

LangSmith's biggest change is IAM role-based authentication for S3 export destinations, alongside new per-project/user trace limits, bulk dataset-split actions in experiment views, evaluator reliability fixes and a new validation endpoint, OTel resource-attribute support for traces, a default switch to zstd compression for bulk exports, and removal of legacy dataset comparison helpers from the public API.

└──▷ WHAT SHIPPED · 8 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

IAM role assumption for S3 export destinationsNEW93

Adds an aws_role_arn field to the PATCH body for S3 export destinations, letting LangSmith assume an AWS IAM role instead of storing static credentials (GCP SaaS deployments only). Switching between static credentials and IAM role assumption no longer requires recreating the destination — supplying aws_role_arn clears stored credentials, and supplying static credentials clears the stored role ARN — and every switch is validated with a test write to the destination bucket before saving, returning an error if the new configuration lacks sufficient write permissions.

Switch an existing S3 export destination from static credentials to IAM role assumption so no long-lived keys are stored in LangSmith.

$ curl -X PATCH https://api.smith.langchain.com/api/v1/export-destinations/<destination_id> \
  -H 'Content-Type: application/json' \
  -H 'X-API-Key: <your_api_key>' \
  -d '{"aws_role_arn": "arn:aws:iam::123456789012:role/LangSmithBulkExportRole"}'

— Names exact field, endpoint, and includes runnable PATCH commandproduct docs

02

Custom code evaluator reliability and testingIMPROVED80

Custom code evaluators that time out or fail on a run now record an error on that run instead of silently leaving it without feedback. Code evaluator upload now accepts Python entrypoints annotated with PEP 604 union return types (e.g. -> dict | None). Evaluators can also be tested against a real conversation before saving, via POST /runs/rules/validate with a test_thread_id, session_id, and the evaluator payload.

Test a multi-turn thread evaluator against a real conversation before saving it, to catch mapping or logic errors early.

$ curl -X POST 'https://<your-langsmith-host>/runs/rules/validate' \
  -H 'Content-Type: application/json' \
  -H 'X-API-Key: <your-api-key>' \
  -d '{"test_thread_id": "<thread-uuid>", "session_id": "<session-uuid>", "evaluator": { ... }}'

— Names endpoint and payload fields with a runnable validation callsnapshot-20260825

03

OTel resource attribute support for tracesNEW75

Deployment metadata (e.g. user ID, service name) can now be attached to every trace without modifying span emission logic, by setting OTEL_RESOURCE_ATTRIBUTES (e.g. service.name=my-agent,user.id=u_12345); the attributes can then be filtered on as otel.resource.* fields in LangSmith.

Attach deployment metadata (e.g. user ID, service name) to every trace without modifying span emission logic, then filter on otel.resource.* fields in LangSmith.

$ export OTEL_RESOURCE_ATTRIBUTES='service.name=my-agent,user.id=u_12345'

— Exact env var and filter field names given with runnable examplesnapshot-20260825

04

Bulk export default compression switched to zstdIMPROVED65

Bulk exports now default to zstd compression; self-hosted deployments that need to keep the previous behavior can set FF_BULK_EXPORT_DEFAULT_COMPRESSION=gzip to retain legacy gzip compression.

Keep the legacy gzip compression for bulk exports on a self-hosted deployment instead of the new zstd default.

$ export FF_BULK_EXPORT_DEFAULT_COMPRESSION=gzip

— Names the flag and old/new format but not full rollout scopesnapshot-20260825

thinner coverage below

05

Bulk row actions in experiment viewsNEW50

Experiment views now support selecting multiple rows — or all rows matching the current filters — to add, replace, or remove dataset splits in one action, or to copy the selected examples to another dataset.

— Describes the actions but no UI path or endpoint givensnapshot-20260825

06

Legacy dataset comparison helpers removed from public APIBREAKING45

Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs; the existing HTTP routes continue to work only for LangSmith's own UI clients.

— States removal scope but not which helpers or migration pathsnapshot-20260825

07

Clearer 409 Conflict messages for duplicate run payloadsIMPROVED40

LangSmith now returns clearer 409 Conflict messages when duplicate run create or update payloads are submitted, indicating whether the duplicate was a create or update request.

— Names the status code but no reproduction stepssnapshot-20260825

08

Per-project and per-user monthly trace limitsNEW35

New traces exceeding a configured per-project or per-user monthly limit are now rejected, while patches and feedback for already-accepted traces continue to flow through.

— No config surface or limit values namedsnapshot-20260825

└──▷ BREAKING ON UPGRADE

!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?

Langfuse

1 RELEASE · 2026-08-25 NOTES

Open source AI engineering platform: LLM evals, observability, metrics, prompt management, playground, datasets. Integrates with OpenTelemetry, LangChain, OpenAI SDK, LiteLLM, and more. YC W23

Langfuse v4.18.0 sharpens its evaluation pipeline with payload-aware filtering and refreshed scoring templates, while adding auditability for human-in-the-loop tool approvals and better visibility into OTLP ingestion traffic.

└──▷ WHAT SHIPPED · 4 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

thinner coverage below

01

Payload-backed evaluation rule filtersNEW53

Evaluation rules can now filter on trace/generation payload fields, letting eval rules target specific fields within the trace or generation payload rather than only top-level attributes.

— Names the mechanism but no config key or exact field listv4.18.0

02

Tool approval source in agent event streamNEW42

Records tool approval source in the in-app agent event stream, giving auditability of human-in-the-loop decisions when tools require approval.

— Describes purpose but not the UI path or data shownv4.18.0

03

Refreshed built-in evaluation templatesIMPROVED40

Refreshes built-in evaluation templates for all-caps detection, exact-match, and keyword-overlap scoring.

— Names three templates but not what changed in eachv4.18.0

04

OTLP ingestion request size loggingIMPROVED35

Logs encoded request body sizes for OTLP ingestion to aid observability of ingest traffic.

— States what is logged but not where to view itv4.18.0

Was this useful?

Braintrust

1 RELEASE · 2026-08-01 NOTES

Braintrust is an open-source evals framework for testing and monitoring AI applications with custom test cases and metrics.

Braintrust gave its MCP server full write access so coding agents can create and update prompts, evals and monitors, added two new built-in open-source models to the Gateway, and shipped a wave of SDK auto-instrumentation for new AI providers across Python, TypeScript and Go — alongside three breaking changes to trace and span formats.

└──▷ WHAT SHIPPED · 8 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

TypeScript SDK provider and framework instrumentationNEW75

v3.28.0 adds Voyage AI auto-instrumentation for embeddings, multimodal embeddings, reranking, and contextualized embeddings, and scorer functions in Eval() now receive the id and tags of each case. v3.27.0 adds Ollama and @cloudflare/think instrumentation, Anthropic beta sessions tracing (anthropic.beta.sessions.turn and anthropic.beta.sessions.thread.turn), Flue v2 support, and a vitest-evals span input override via meta.eval.input. v3.26.0 adds auto-instrumentation for Cloudflare Agents, Cloudflare AI Chat, and Hugging Face Transformers.js, plus system prompt capture for Strands Agents SDK spans.

— Names every integration and field but no example usage.snapshot-20260825

02

Go SDK model and tool instrumentationNEW75

v0.11.1 adds Firebase Genkit WithProvider and WithModel options on NewMiddleware for explicit model attribution, plus traced tool wrappers DefineTool, DefineToolWithInputSchema, and DefineMultipartTool, with auto-instrumentation now automatically replacing the un-traced genkit.DefineTool, genkit.DefineToolWithInputSchema, and genkit.DefineMultipartTool calls. v0.11.0 adds Anthropic span capture for prompt_cache_creation_5m_tokens and prompt_cache_creation_1h_tokens, Bedrock Runtime spans capturing audio and video content blocks with full instrumentation of InvokeModelWithResponseStream for Anthropic Claude models, and Google ADK spans that include reasoning and cached token metrics.

— Names every function, option, and field with no runnable example.snapshot-20260825

03

MCP server gains write toolsNEW60

The Braintrust MCP server now exposes write tools so coding agents can create and update prompts, scorers, classifiers, Topics pipeline config, monitor views, alerts, scheduled jobs, evals, and dataset rows — previously the server only supported reading them.

— Names all writable resource types but gives no example call.snapshot-20260825

04

SQL sandbox query sidebarNEW60

The SQL sandbox organizes queries in a collapsible sidebar with search by name, drag-to-reorder, command-bar navigation, and a per-query menu to rename, duplicate, or delete; a later update adds a 'Copy share link' action that generates a shareable URL.

— Names UI capabilities but no exact navigation path given.snapshot-20260825

thinner coverage below

05

Permission group management from member dialogNEW55

Member permission groups are now editable directly from the member dialog in Settings > Members, showing direct group memberships, inherited memberships, and available groups to add — without opening each group individually.

Audit or update which permission groups a team member belongs to without hunting through each group.

📍Go to Settings › Members, open a member's dialog — the panel shows direct group memberships, inherited memberships, and groups available to add.

— Exact UI path given, no API or config surface.snapshot-20260825

06

Dataset rows can reference multiple tracesNEW50

Dataset rows can now reference a group of up to 64 traces instead of a single trace, rendering each inline and flagging any that are no longer available.

— Gives the exact 64-trace limit but no config or command.snapshot-20260825

07

Span-name log filteringNEW50

Clicking a span's name in the trace panel header filters the logs list to that span name, instead of requiring a manually built filter expression.

Quickly scope logs to a single span without manually constructing a filter expression.

📍In the trace panel, click the span name in the header — the logs list automatically filters to that span name.

— Exact click action described, but no named filter syntax.product docs

08

Searchable filter value pickerIMPROVED25

The filter tab's value picker is now searchable and accepts values not present in the list by typing them directly.

— Thin one-line description with no further mechanism.product docs

└──▷ BREAKING ON UPGRADE

!Python SDK v0.32.0: LiveKit Agents audio attachments on agent_speaking spans are now disabled by default. Set BRAINTRUST_CAPTURE_AGENT_AUDIO_ATTACHMENTS=true to restore the previous behavior.
!Go SDK v0.11.1 (Google GenAI): Provider metadata changed from 'gemini' to 'google'. Trace queries that filter on the previous provider value must be updated.
!Go SDK v0.11.1 (Eino): ChatModel span output is now an OpenAI-compatible choices array ([{"index": 0, "finish_reason": "...", "message": {...}}]) instead of a flat message map. Embedding input is now {"inputs": [{"content": "..."}]} and output is {"count": N}, removing embedding_length and renaming embeddings_count. Provider metadata is now lowercase (e.g. 'openai' instead of 'OpenAI'). Trace queries relying on the previous formats must be updated.

Was this useful?

Weaviate

1 RELEASE · 2026-08-25 NOTES

Weaviate is an open-source vector database that stores both objects and vectors, allowing for the combination of vector search with structured filtering with the fault tolerance and scalability of a cloud-native database.

Weaviate v1.39.1 rounds out its REST Search API with four new endpoints, lets backups include role definitions, and adds a gRPC transport option for GCS backups alongside several schema and index reliability fixes.

└──▷ WHAT SHIPPED · 7 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

RUNTIME_REINDEX_ENABLED kill-switch env varNEW65

Introduces the RUNTIME_REINDEX_ENABLED environment variable as a kill-switch, off by default.

— Exact env var name and default given, no further mechanismv1.39.1

02

REST Search API endpoints for BM25, hybrid, near-object, aggregateNEW60

Weaviate completes its REST Search API rollout with four new endpoints: BM25 keyword search (e.g. POST /v1/search/bm25), hybrid search, near-object search, and aggregate counts — labelled REST Search API 1/4 through 4/4.

— Names endpoints but only one exact path givenv1.39.1

thinner coverage below

03

Backup role inclusion via includeRolesNEW55

Adds an includeRoles parameter to the backup/restore process so role definitions can be included in backups.

— Names the parameter but not full request shapev1.39.1

04

Auto-schema default named vectorIMPROVED40

Auto-schema now creates a default named vector instead of a legacy vector when inferring schema.

— Describes the change but not migration impactv1.39.1

05

Resumable vector-index drop operationsIMPROVED40

Interrupted vector-index drop operations now resume from the recorded pending set, improving reliability of drop-vector-index across restarts.

— Explains mechanism but no named API surfacev1.39.1

06

gRPC transport for backup-gcs backendNEW35

Adds an opt-in gRPC transport option for the backup-gcs backend.

— Names the backend but no config detailv1.39.1

07

MUVERA usage calculations for multi-vector indexesNEW23

Adds MUVERA-specific usage calculations for multi-vector index accounting.

— Bare mention with no mechanismv1.39.1

Was this useful?

Pinecone

1 RELEASE · 2026-08-01 NOTES

Pinecone is a managed vector database service for storing and querying high-dimensional embeddings at scale.

Pinecone launched Nexus, a BYOC RAG product with grounded KnowQL queries and per-query tracing, alongside new S3/GCS/Azure Blob storage integrations, API versioning, a CLI targeting subcommand, and new Terraform resources for account and org management.

└──▷ WHAT SHIPPED · 7 FEATURESmost 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.

depth0–40what it does and how it works or what changed
specificity0–30names real surfaces — APIs, flags, formats, limits, numbers
actionability0–30enough to go use it — a named endpoint, flag, or config key tops this; a UI path is a starting point

Code and config rank first by construction: a documented endpoint, flag or config key scores at the top of specificity and actionability, so it sorts above a feature described only in prose. Under 60 the notes go thin — everything below the “thinner coverage” line is thinner documentation, not smaller work. Hover any meter for that feature's three sub-scores.

01

API versioning via X-Pinecone-Api-Version headerNEW80

New X-Pinecone-Api-Version header controls which stable API version your requests use, with documented support windows, letting requests be pinned to avoid unintended behaviour from version upgrades.

Pin API requests to a specific stable Pinecone API version to avoid unintended behaviour from version upgrades.

$ curl -X GET 'https://api.pinecone.io/indexes' \
  -H 'Api-Key: <PINECONE_API_KEY>' \
  -H 'X-Pinecone-Api-Version: 2024-07'

— Exact header name plus runnable curl exampleproduct docs

02

Nexus BYOC RAG with KnowQL and query tracingNEW65

New Pinecone Nexus product supports BYOC (Bring Your Own Compute) deployment, KnowQL queries that return grounded multi-document answers with citations, and per-query tracing showing reasoning steps, retrieval tool calls, token usage, latency, cost breakdown, and cache savings.

— Mechanism and scope described but no API/CLI surface givenproduct docs

03

Bulk import/export via S3, GCS, Azure BlobNEW65

New storage integrations enable bulk import of vector data into indexes and export of audit logs from Amazon S3 (via IAM role), Google Cloud Storage (via service account key), and Azure Blob Storage (via service principal).

— Names three providers and auth methods but no config keysproduct docs

04

pc target CLI context commandNEW60

New pc target CLI subcommand sets the target organization and project context for all commands, including CI/CD pipelines.

— Names the subcommand but no flags or example runproduct docs

thinner coverage below

05

Standard plan trial with $300 creditsNEW55

New Standard plan trial provides $300 in credits over 21 days, including bulk import, backup and restore, RBAC, and higher scale limits.

— Concrete numbers but no activation steps givenproduct docs

06

Terraform resources for org and access managementNEW55

Adds Terraform resources for managing service accounts, role bindings, organization invites, and removing organization members in Pinecone.

— Names resource types but no resource block names or examplesnapshot-20260825

07

API errors reference documentationNEW25

New API errors reference documents all HTTP 2xx, 4xx, and 5xx status codes returned by the Pinecone API.

— Bare mention of a documentation resource onlyproduct docs

Was this useful?

The AI Toolchain — issue 007, August 25, 2026