The AI Toolchain — issue 003, August 21, 2026

The computer use tool is promoted to general availability as computer_toolset_20260801, requiring no beta header. It adds batch actions (multiple actions per turn), zoom enabled by default, and per-member configuration via configs. Upgrading an existing integration to computer_toolset_20260801 changes the request shape and tool handling, so migration from computer_20251124 is required.

browser_toolset_20260801 is a new client toolset for driving an application-hosted browser viewport. It reads the page accessibility tree, elements, forms, and tabs, and adds element references, form input, tab management, download reporting, and opt-in file upload on top of screenshot-and-click control.

Adds background: transparent support (preview) for gpt-image-2 and gpt-image-2-2026-04-21 in the v1/images/generations, v1/images/edits, and v1/responses endpoints. Output format must be png or webp; jpeg does not support transparent backgrounds.

$ curl https://api.openai.com/v1/images/generations \
  -H 'Authorization: Bearer $OPENAI_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gpt-image-2",
    "prompt": "A sleek wireless headphone on a clean surface",
    "background": "transparent",
    "output_format": "png"
  }'

Releases a Prompt Caching dashboard on the OpenAI API platform showing cache hit rate over time, cache reads per write, and a breakdown of cache-read, cache-write, and uncached tokens; filterable by model and service tier.

A new Fully automatic option under Auto-remediation policy on the Repairs tab lets auto node repair run end-to-end without manual approval. Per-fault scoping via Repair actions lets operators choose which fault groups — Migrate to new host, Reprovision, VM reboot — run unattended under Fully automatic, so destructive repairs can stay gated while transient ones self-clear; a Wait policy set to Grace period can protect in-flight jobs during repair.

RunPod added a dedicated batch processing system for Serverless endpoints with a full REST API: POST /v2/{endpoint_id}/batch creates a batch (optionally seeded with an initial array of /run-shaped requests), POST /v2/{endpoint_id}/batch/{id}/requests appends up to 10 MiB of additional requests to a DRAFT batch across multiple calls, POST /v2/{endpoint_id}/batch/{id}/finalize locks the batch to FINALIZED and makes it eligible for execution, GET /v2/{endpoint_id}/batch/{id} returns requestTotal, requestInProgress, requestCompleted, and requestFailed counts for polling, GET /v2/{endpoint_id}/batch/{id}/requests returns paginated per-request status, output, error, startedAt, and completedAt fields via offset/limit query parameters with a hasMore flag, POST /v2/{endpoint_id}/batch/{id}/cancel cancels a batch (queued requests are cancelled and unbilled, in-progress requests finish normally), GET /v2/{endpoint_id}/batch lists all batches newest first, PUT /v2/{endpoint_id}/batch/{id} updates attributes like display name, and DELETE /v2/{endpoint_id}/batch/{id}/requests/{requestId} removes a single request from a DRAFT batch. Batches run on dedicated workers isolated from standard /run traffic, with limits of 10 active batches per endpoint, 5,000 requests per batch, and 50,000 queued requests per endpoint (up to 1,000,000 daily for enterprise); a Console Batch tab shows per-batch progress and per-request rows sorted failures-first, webhook and Console Inbox notifications fire once on terminal state (FAILED/CANCELLED) with batch ID, endpoint name, status, and counts, and batch jobs are billed at standard serverless request rates with flex worker discounts for enterprise.

$ # 1. Create batch with initial requests
curl -X POST https://api.runpod.io/v2/{endpoint_id}/batch \
  -H 'Authorization: Bearer {api_key}' \
  -H 'Content-Type: application/json' \
  -d '[{"input":{"text":"The quick brown fox"}},{"input":{"text":"Jumped over the lazy dog"}}]'

# 2. Finalize to start processing
curl -X POST https://api.runpod.io/v2/{endpoint_id}/batch/batch_01j9abc123/finalize \
  -H 'Authorization: Bearer {api_key}'

$ curl -X GET https://api.runpod.io/v2/{endpoint_id}/batch/batch_01j9abc123 \
  -H 'Authorization: Bearer {api_key}'
# Batch is done when requestCompleted + requestFailed == requestTotal

$ curl -X GET 'https://api.runpod.io/v2/{endpoint_id}/batch/batch_01j9abc123/requests?offset=0&limit=50' \
  -H 'Authorization: Bearer {api_key}'

The sandbox list SDK call gains an autoDestroyAt range filter, letting callers query sandboxes by their scheduled destruction time.

Introduces a stream idle watchdog that warns after 60 seconds of silence and cancels at 300 seconds, with automatic retries for throttling, 5xx errors, and connection drops. New config settings api.streamIdleSoftTimeout, api.streamIdleHardTimeout, and api.timeout let users tune the watchdog and a 60-minute streaming timeout for stalled or dropped connections.

Adds AI-generated session titles to the V3 session resume picker, surfaced via kiro-cli chat --resume-picker or /chat resume, so similar sessions are distinguishable at a glance.

$ kiro-cli chat --resume-picker

Adds the /handoff command to transfer a local session — conversation, attached images, and an optional follow-up prompt — to a Cline Cloud workspace that continues running after the app closes. Integrates Cloud into the existing Local / Remote environment menu, gated behind the Cloud sessions feature flag (shows 'Coming soon' when off), and adds interrupted-handoff recovery: if the app restarts, the network drops, or the branch moves mid-transfer, reopening the session resumes or cleanly retries the handoff and restores typed drafts and attachments.

Model selector in the composer and provider settings now lead with Recommended and Free tiers (labeled 'Subscribed' and 'Free' on ClinePass), showing display names with descriptions instead of an alphabetized list of raw model IDs. Image, voice, and other non-chat models are excluded from onboarding, model pickers, and ACP model listings, and the CLI's --model flag now rejects them outright.

Models that support image generation can now produce images during a task, rendered inline in the conversation. The CLI/TUI saves each generated file to a temporary path, HTML session exports embed images inline, and ACP clients receive them as image content; the SDK persists generated images in session history and exports.

Adds AMD, Arcee, Echo, Jalapeno, Kosmik, LLM Gateway, RunInfra, and SCNet as new model providers across desktop, CLI, and SDK, with updated model lists, pricing, and per-provider default models in the refreshed model catalog.

The provider scx is renamed to scx-ai in the model catalog; any configuration or code referencing scx by name will need to be updated.

Skill slash commands now load through the skills tool instead of being expanded into the user message, so instructions reach the model once instead of twice; the persisted transcript/history records the typed command (e.g. /my-skill ...) rather than the full SKILL.md body.

Adds additional_developer_instructions to managed requirements, exposed via configRequirements/read as additionalDeveloperInstructions, with a 10,000 estimated-token size limit, and support across compaction, resume, and agent forks.

Enables the multi-agent V1 protocol for Amazon Bedrock models, normalizing Bedrock model catalogs to advertise MultiAgentVersion::V1 to work around Bedrock's lack of support for response items required by multi-agent V2. Also switches Amazon Bedrock remote compaction to use compaction_trigger items sent through /v1/responses instead of the legacy dedicated compaction protocol, and explicitly enables remote compaction for OpenAI, Azure Responses, and Amazon Bedrock providers; the legacy dedicated compaction protocol is removed from provider capabilities, so remote compaction now defaults to unsupported and must be explicitly enabled per provider.

Verifies the Codex desktop app bundle signature using Apple's codesign --verify --deep --strict against OpenAI's bundle identifier com.openai.codex and Apple Team ID 2DC432GLL2 before launching or installing on macOS.

Expands codex doctor to diagnose endpoint protection, network/proxy failures, desktop app state, Windows sandbox provisioning, and update connectivity.

Adds a shellSnapshotV2 executor capability and an optional shell snapshot request to ExecParams, caching and restoring Unix shell state and profile exports for bash, zsh, and sh in an in-memory, attachment-scoped cache.

Adds codex queue command for sending messages to existing local or remote sessions.

$ codex queue <session-id> 'Run the full test suite and report failures'

Adds /cd, /pwd, and /cwd commands for managing the working directory inside TUI sessions.

Expands Vim editing in the TUI composer with character replacement and change motions including cw, c$, and cc.

Adds claude self-hosted-runner --defer-shutdown-max-min <minutes>: on SIGTERM the runner keeps serving already-attached sessions, parks any remaining sessions once the specified number of minutes elapses, then exits.

$ claude self-hosted-runner --defer-shutdown-max-min 5

Adds claude self-hosted-runner --proxy-authorization-command and --proxy-authorization-file so runners behind egress proxies that require a freshly issued Proxy-Authorization header on every connection can supply one dynamically.

$ claude self-hosted-runner --proxy-authorization-command 'vault read -field=token secret/proxy-auth'

Adds headersHelper to plugin marketplace URL entries and catalog entries: it runs a command to mint HTTP headers (e.g. a short-lived token) for catalog and same-origin archive fetches. A catalog entry's headersHelper runs only at claude plugin install or claude plugin update time, after its command is shown, behind a [y/N] prompt (or -y to skip).

MCP headersHelper in a project .mcp.json, and inline MCP servers in project or --add-dir agent files, now require the folder's trust dialog to have been accepted (also enforced under claude -p). MCP headersHelper from a project .mcp.json, plugin, or agent file now runs without inherited credential environment variables; user-, managed-, and claude.ai-scope helpers instead run from the Claude config directory.

Cross-session messaging now reports 'refused' back to the sender when the target session has crossSessionInbound: "refuse" set, instead of silently succeeding, and notifies the sending session when the target inbox drops messages due to a rate limit or full queue instead of messages vanishing silently.

Adds a keybindingFlavor setting; set to "readline" to make Ctrl+W delete back to the previous whitespace (Bash-style), while the default "classic" behavior is unchanged.

Changes claude mcp list and claude mcp get to show disabled servers as ⊘ Disabled instead of connecting to them for a health check.

Adds a --resume flag to /review, review run, and CI retries so an interrupted review can continue rather than restart, as long as the PR head commit has not moved since the last run.

$ qwen review run --resume

Adds the stable qwen3.8-max model to the Token Plan model list, selectable via /model alongside the existing preview version.

$ /model qwen3.8-max

Adds authenticated HTTPS Git extension installs with configurable credential persistence, enabling secure cloning of private repositories via the daemon, plus batch APIs to update extension activation states for up to 100 extensions at once, either globally or for specific trusted workspaces.

Web Shell now supports inserting file attachments into active turns via the composer or @ selection, with preview and queue management; file uploads are unified to offer reference-or-upload options with persistent storage and duplicate name handling; and Web Shell adopts Goal v3 controls, letting goals be managed independently of chat messages via a compact composer row.

The /review skill now supports posting comments and approvals to Aone Code via the a1 CLI when using the --comment flag.

ACP child processes now track and report V8 old-generation heap metrics, including peak committed memory and major-GC counts, in daemon status.

Adds the --mmproj-device CLI flag (short form -mmdev) and the MTMD_BACKEND_DEVICE environment variable to select which compute device runs the multimodal projector (mmproj), enabling setups like offloading mmproj to an iGPU while the main model runs on a dGPU. Passing none disables GPU offload for the projector entirely.

$ llama-server -m model.gguf --mmproj mmproj.gguf --mmproj-device CUDA1

$ MTMD_BACKEND_DEVICE=CUDA1 llama-cli -m model.gguf --mmproj mmproj.gguf

The Metal backend now runs a preprocessing pass for GGML_OP_FLASH_ATTN_EXT that dequantizes quantized KV caches (Q4_0, Q4_1, Q5_0, Q5_1, Q8_0) into contiguous F16 scratch buffers via a new kernel_flash_attn_ext_dequant_to_f16 kernel, instead of dequantizing in-kernel, improving accuracy and performance on Apple Silicon. It skips redundant V dequantization when V is a view of K (as in MLA-based models), and new test-backend-ops cases cover the MLA shape, pad path (kv=113), and large context (kv=16384).

$ llama-perplexity -m model-q8_0.gguf -fa 1 -ctk q8_0 -ctv q8_0

Adds DSpark speculative decoding support for LFM2 models, including partial rollback for LFM2 recurrent state, enabling draft-model-accelerated inference for the LFM2.5 series via flags like --spec-draft-n-max and --spec-draft-n-min.

$ llama-server -m LFM2.5-1.2B-Instruct-F16.gguf -md LFM2.5-1.2B-Instruct-DSpark-5L-draft-f16.gguf -ngl 99 -ngld 99 -fa on --spec-draft-n-max 7 --spec-draft-n-min 0

Adds per-hardware and per-quantization-type batch crossover thresholds for the CUDA mul_mat_vec_q to MMQ (int8 tensor-core) dispatch path, with tuned switch points for Blackwell (RTX 5090, DGX Spark) and Ada (RTX 4090) GPUs, delivering +23-41% decode throughput at batch size 8 for K-quants (Q2_K, Q3_K, Q4_K, Q5_K, Q6_K) on dense models with no low-batch regression.

The /v1/models endpoint is now private when server authentication is enabled, blocking unauthenticated access to model file paths that may contain PII such as home directory names. Unauthenticated clients that previously queried /v1/models without a key will now receive an auth error.

Adds conversion and inference support for GraniteSWAForCausalLM and GraniteMoeSWAForCausalLM architectures, enabling IBM Granite models with interleaved Sliding Window Attention (SWA) and Attention Sinks. Also adds per-layer RoPE/NoPE determination via llama_hparams::has_rope, allowing models with mixed attention layers to load and run correctly.

Native MTP (speculative decoding) for Qwen3.8 and dots3-note now uses an aligned draft-head context cache plus generalized skip-replay so rejected drafts no longer replay through the main model at any depth. Warm in-app decode on Qwen3.8-27B rises from ~22 to 38–48 t/s, and dots3-note reaches ~41 t/s.

DFlash2 now reuses session prefixes via end-of-turn cache checkpoints, draft hidden-state gap splice, and prompt-boundary snapshots, dropping warm multiturn TTFT by roughly 20x and sustaining 60+ t/s on Qwen3.8 (68.7 t/s measured); the DFlash runtime now ships in the bundle.

Prefill admission rejections now include a wired-limit advisory naming the exact sysctl, the macOS ~84%-of-RAM default, and the reset-on-reboot behavior.

The RAM preflight check that refused large model loads has been removed; memory estimates now advise rather than block, enabling a 101GB dots3-note bundle to load on a 128GB Mac.

Introduces a classifier: knn router type that does similarity-weighted voting over a persisted JSONL corpus stored under <data path>/router-corpus, governed by a configurable knn.similarity_threshold; prompts below threshold return an undecidable fallback instead of a forced guess. The corpus is managed via POST /api/router/{name}/corpus, GET /api/router/{name}/corpus/stats, and DELETE /api/router/{name}/corpus, all admin-gated and also exposed as MCP tools.

$ curl -X POST http://localhost:8080/api/router/my-router/corpus \
  -H 'Authorization: Bearer <token>' \
  -H 'Content-Type: application/json' \
  -d '{"text": "What is my account balance?", "label": "finance"}'

Every HTTP route now requires credentials unless explicitly listed in a public registry, closing bypasses for previously unauthenticated aliases such as /moderations, /models, /backends, and /mcp/chat/completions. With database auth or legacy API keys configured, /version and generated audio, image, video, and 3D asset URLs also now require credentials by default; narrow public paths can be restored via ApplicationConfig.PathWithoutAuth.

Adds MiniMax-H3 video+audio generation to the vllm-cpp backend (ABI v12): a second engine handle loads the H3 checkpoint set, with parameters.model naming the DiT and the text encoder and two VAEs named under options:. GenerateVideo renders jointly into an MP4 with a real AAC audio track.

A new --web-rtc-udp-port CLI flag and LOCALAI_WEBRTC_UDP_PORT environment variable let all Realtime WebRTC calls reuse a single shared Pion ICE UDP mux, simplifying container and firewall setup by needing only one open UDP port for every call.

$ docker run -ti --name local-ai -p 8080:8080 -p 3478:3478/udp -e LOCALAI_WEBRTC_UDP_PORT=3478 localai/localai:latest

An opt-in pii.reverse_in_response config turns masked PII values into request-scoped deterministic pseudonyms (e.g. EMAIL_001, EMAIL_002) that are restored when the backend echoes them back, including across SSE tokens split over multiple writes.

An opt-in per-model compression config compresses older complete conversation turns through a configured LocalAI model before inference, while preserving system prompts, the newest messages, and whole tool-call/result units intact. The resulting compression ratio and duration are returned as response metadata and metrics.

vllm-cpp CUDA architecture coverage on amd64 grows from 120a;121a to 80;86;89;90a;100a;103a;120a;121a, and on arm64 from 121a to 87;90a;100a;110;121a, adding support for A100, L4, 4090, H100/H200, B200, Jetson Orin, and Jetson Thor.

/app/models now owns the Explore and Installed views, /app/backends owns the Catalog and Installed views, and /app/manage redirects to the new pages while preserving legacy query state.

Adds Qwen3-TTS support to the llama-cpp backend across CUDA, ROCm, SYCL, Vulkan, Metal, and L4T using upstream GGUF conversion, with gallery entries qwen3-tts-llamacpp and qwen3-tts-llamacpp-q4 provided.

Adds Stripe Link wallet blocks for wallet-based payments and Stripe subscription webhook trigger blocks for automating workflows on Stripe subscription events.

Adds a user-facing memory settings page for managing what each expert remembers, isolated expert memory so each expert's memory store is kept separate from others, and an expert memory admin viewer for inspecting and managing per-expert memory as an administrator.

Adds All Quiet incident management blocks for integrating AutoGPT workflows with All Quiet alerting, and a DataForB2B provider block for B2B data enrichment within workflows.

Adds 'Raise your own expert v1', a flow for creating custom experts with attachments, budget, and skills configuration.

Adds a day-one kickoff flow that activates automatically after hiring an expert, writing-style capture during the hire flow so a new expert can match a user's communication style, and a launch roster with real workflow bundles for bootstrapping expert assignments.

Removes the enable_user_memories, search_session_history, num_history_sessions, and num_past_session_runs parameters; any code passing these will break on upgrade.

Human-in-the-loop configuration now requires human_review=HumanReview(...), replacing the previous flat HITL kwargs, which are removed and will break existing code on upgrade.

The reasoning=True shortcut is removed; callers must now pass an explicit reasoning_model argument.

The /runs/rules/validate endpoint now supports testing a multi-turn thread evaluator against a real conversation, using test_thread_id and session_id parameters, before saving the evaluator rule.

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

Bulk export compression now defaults to zstd instead of gzip on cloud deployments; self-hosted environments retain the gzip default via the FF_BULK_EXPORT_DEFAULT_COMPRESSION environment variable.

$ export FF_BULK_EXPORT_DEFAULT_COMPRESSION=gzip

Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs; code using those SDK methods must migrate to POST /v2/datasets/{dataset_id}/experiment-runs. Existing HTTP routes continue to work for LangSmith UI clients.

Setting OTEL_RESOURCE_ATTRIBUTES (e.g. user.id, deployment.environment) now surfaces those values on LangSmith traces under otel.resource.* fields without modifying span emission.

$ export OTEL_RESOURCE_ATTRIBUTES="user.id=usr_123,deployment.environment=production"

The batched-run ingestion log now emits run_verbs as a list of run_id and verbs objects instead of a map keyed by run UUID; structured-log aggregators parsing the old map format need to be updated.

The trace detail view now shows input/output previews at the top level, automatically extracted from the first user message and final assistant response, displayed above the span tree. This is especially useful for multi-step traces involving tool calls.

The span details panel is redesigned with a tabbed interface separating metadata, inputs/outputs, and entity references, plus hover cards that let users preview a span quickly without opening the full panel.

Go SDK v0.11.x changes trace span shapes across several provider integrations. For Google GenAI, provider metadata changed from 'gemini' to 'google'. For 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, and provider metadata is now lowercase (e.g. 'openai' instead of 'OpenAI'). For Anthropic, span metadata no longer includes endpoint, the output field is now a single message object instead of an array, and non-streaming spans no longer emit time_to_first_token. For Bedrock, span metadata renames stop_sequences to stop, removes additional_model_request_fields, and aligns image, document, and tool block shapes with Bedrock's native wire format. Trace queries relying on the previous field names or formats need updating.

Go SDK v0.11.1 adds WithProvider and WithModel options on NewMiddleware for explicit model attribution in Firebase Genkit, plus traced tool wrappers DefineTool, DefineToolWithInputSchema, and DefineMultipartTool; auto-instrumentation now replaces genkit.DefineTool, genkit.DefineToolWithInputSchema, and genkit.DefineMultipartTool calls with their traced equivalents.

The Braintrust MCP server now exposes write tools so coding agents can create and update prompts, scorers, classifiers, monitor views, alerts, scheduled jobs, dataset rows, and Topics pipeline configuration, not just read them. Write tools operate under the authenticated account's permissions and several can replace or remove objects, so Braintrust recommends configuring the client to require confirmation before use.

Adds kimi-k3 and deepseek-v4-flash-0731 as built-in open-source models served directly by Braintrust with no AI-provider setup required. They can be selected under the Braintrust provider in playgrounds, prompts, and scorers, or requested by name through the Braintrust Gateway; usage draws from monthly model credits shared with Topics.

Python SDK v0.33.0 adds Vercel AI SDK for Python auto-instrumentation, enabled by default in auto_instrument(), and Cursor SDK Python instrumentation for tracing agent runs, model turns, and tool calls.

import braintrust

braintrust.auto_instrument()  # Vercel AI SDK for Python instrumented by default in v0.33.0+

Traces that share a metadata.conversation_id now surface as related traces automatically for end-to-end multi-turn conversation review without requiring grouping configuration. The 'Group by' and 'Cluster by' controls have moved into the row-type selector in the toolbar.

Adds trace-group references to dataset rows so a multi-turn session or related log set becomes one example, with up to 64 traces per row rendered inline.