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

The AI Toolchain — issue 010, August 29, 2026

THE AI TOOLCHAIN NO. 010
Tail
THE DAILY RELEASE FIREHOSE
PUBLISHED AUGUST 29, 2026 · EVERY WEEKDAY
EDITIONS tail grep head diff uniq

The daily firehose — everything the toolchain shipped today, already filtered.

// HOW THIS ISSUE IS MADE

We read every release from the 174 tools on our watchlist at the source — GitHub and GitLab release notes, vendor release pages and changelogs, project blogs and feeds, vendor press releases, and the source code behind the tag. Bug-fix-only releases and non-product newsroom noise are dropped; what's left is summarized down to the new capability, how to try it, and any screenshots or videos the release itself published. Every entry links to the sources it was built from.

VIEW
ISSUE VIEW full issue
Do you prefer this view? 
$ tct list   # 9 tools matched
AI & LLM Tooling
◆  AI Model & Data Infrastructure

Together AI

Sources Changelog → 1 RELEASE · 2026-08-29 CHANGELOG

Run, train, and serve open-source AI models on Together AI.

Together AI's CLI adds a full batch-inference command group, two new GLM-5.3 models join the catalog, and multi-project isolation reaches general availability alongside broader ACH payment access and fine-tuning quality upgrades.

└──▷ 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
`tg batches` CLI command groupNEW90

New tg batches command group manages the full batch inference job lifecycle: tg batches submit (upload a local JSONL file or an existing file ID), tg batches list, tg batches cancel, and tg batches download (streams to stdout or writes to disk with --output), plus aliases ls and get.

Submit a JSONL batch inference job from the CLI and download results once complete — useful for high-volume, cost-efficient offline inference.
$ tg batches submit requests.jsonl
tg batches list
tg batches download <batch-id> --output results/
— Named subcommands, flags, and a runnable example command.changelog-20260829-07a75e81
02
GLM-5.3 model family addedNEW70

Adds zai-org/GLM-5.3 (1,000,000-token context, FP4 quantization, function calling, structured outputs; $1.40 input / $4.40 output / $0.26 cached input per 1M tokens) and zai-org/GLM-5.3-Flash (1,000,000-token context, FP8 quantization, function calling, structured outputs; $0.15 input / $0.50 output / $0.03 cached input per 1M tokens).

— Full specs and pricing given but no call example.changelog-20260829-07a75e81
03
ACH bank transfers opened to all customersIMPROVED65

ACH bank transfers, previously enterprise-only, are now available to all customers; link a U.S. bank account via billing settings with instant verification and purchase credits directly.

— Clear before/after with a billing-settings starting point.changelog-20260829-07a75e81
04
Fine-tuning quality improved for several modelsIMPROVED65

Fine-tuning quality improved for Qwen/Qwen3.5-0.8B, Qwen/Qwen3.5-2B, Qwen/Qwen3.5-4B, Qwen/Qwen3.5-27B, Qwen/Qwen3.5-35B-A3B, Qwen/Qwen3.5-35B-A3B-Base, Qwen/Qwen3.5-122B-A10B, nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16, and nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16; no setting changes are required, just start a new job to pick it up.

— Names every affected model but no quality metrics.changelog-20260829-07a75e81
thinner coverage below
05
Multi-project isolation reaches GAIMPROVED55

Multi-project isolation is now generally available for all organizations, with no early-access limitations — clusters, fine-tuned models, endpoints, evaluations, files, and API keys are all fully scoped to projects.

— Lists scoped resource types but no navigation steps.changelog-20260829-07a75e81
06
Cancel expiration on API keysNEW45

Adds a Cancel expiration action, accessed via the three-dot menu, to remove a scheduled expiration from an existing project API key.

— UI action named but mechanism is minimal.changelog-20260829-07a75e81
Was this useful?

Groq

Sources Release page → 1 RELEASE · seen 2026-08-29 NOTES

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

Groq added two new open-weight reasoning models to its chat completions API, both offering extremely fast inference plus built-in browser search, code execution and structured output support.

└──▷ 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
gpt-oss-20b and gpt-oss-120b reasoning models addedNEW95

Added openai/gpt-oss-20b and openai/gpt-oss-120b models accessible via POST https://api.groq.com/openai/v1/chat/completions. The 20B MoE model runs at ~1000+ TPS and the 120B MoE model (128 experts) runs at ~500+ TPS; both share a 131K token context window, 32K max output tokens, and built-in browser search, code execution, and structured output support. Both carry built-in reasoning that surpasses OpenAI o4-mini on several benchmarks, including 98.7% on AIME 2025 (20B) and 90.0% on MMLU (120B).

Run a reasoning query against the 20B model to get fast, tool-augmented answers in CI or scripted workflows.
$ curl https://api.groq.com/openai/v1/chat/completions \
  -H "Authorization: Bearer $GROQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-oss-20b",
    "messages": [
      {"role": "user", "content": "Explain why fast inference is critical for reasoning models"}
    ]
  }'
Switch to the 120B model for higher-accuracy tasks such as health or multilingual analysis where benchmark quality outweighs raw throughput.
$ curl https://api.groq.com/openai/v1/chat/completions \
  -H "Authorization: Bearer $GROQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-oss-120b",
    "messages": [
      {"role": "user", "content": "Summarize the latest CVE advisories and assess patient-safety risk"}
    ]
  }'
— Names endpoint, model ids, specs, benchmarks, and runnable curl examples.snapshot-20260829
Was this useful?

OpenAI

Sources Release page → 1 RELEASE · seen 2026-08-29 NOTES

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

OpenAI added hard spend limits with automatic enforcement and proactive spend alerts for organizations and projects on the API platform.

└──▷ 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 alertsNEW63

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

— Names 429 enforcement and alerts but no config surface or exact setup steps.snapshot-20260829
Was this useful?
◆  AI Coding Agents

Replit Agent

Sources Release page → 1 RELEASE · seen 2026-08-29 NOTES

Replit Agent is an AI assistant that helps developers build, debug, and deploy code projects directly within the Replit IDE.

Replit Agent adds intelligent model routing with an Auto mode, restructures Enterprise Agent mode tiers, and introduces Growth Skills for customer acquisition alongside custom OAuth connector configuration for Enterprise Workspaces.

└──▷ 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.

01
Custom OAuth connectors for EnterpriseNEW85

Enterprise organizations can configure custom OAuth credentials — including Client ID, Client Secret, and scopes — for connectors like Gmail instead of using Replit's defaults, via the 'Use custom configurations' option. To set it up, go to Integrations > Connectors > Gmail in the Enterprise Workspace, select 'Use custom configurations', and enter the Client ID, Client Secret, and desired scopes so outbound traffic routes through the organization's verified identity.

Set up Gmail connector with your own OAuth app so outbound emails route through your organization's verified sender identity.
📍In the Enterprise Workspace, go to Integrations › Connectors › Gmail, select 'Use custom configurations', and enter your Client ID, Client Secret, and desired scopes.
— Names exact fields and UI navigation stepssnapshot-20260829
02
Growth Skills for customer acquisitionNEW75

Growth Skills is a new capability that runs Agent-powered go-to-market motions to find an app's first customers, with integrations for Apollo, Clay, SideShift, RevenueCat, Stripe, and PostHog. Users open Growth Skills in the Agent sidebar, select a partner integration, and run a motion against their deployed app.

Run a Growth Skills motion to find early customers for your app using a connected partner integration.
📍Open Growth Skills in the Agent sidebar, select a partner (e.g. Apollo or PostHog), and run a motion against your deployed app.
— Names six integrations and gives a starting workflowsnapshot-20260829
03
Auto mode for intelligent model routingNEW60

A new 'Auto' Agent mode automatically selects the model and effort level best suited to each part of a task, with the option to override to a specific model. Intelligent Model Routing is now the default Agent mode for Enterprise Workspaces, and account admins can control model availability per Workspace and toggle Auto mode on or off.

— Describes mechanism and admin controls but no exact UI pathsnapshot-20260829
thinner coverage below
04
Power and Max Agent mode tiers replace LiteBREAKING40

The Enterprise Agent mode menu now offers 'Power' and 'Max' tiers, replacing the former Lite option.

— Names the tier change but no migration guidancesnapshot-20260829
└──▷ BREAKING ON UPGRADE
  • !The Enterprise Agent mode menu no longer includes the Lite option; only Power and Max are available.
Was this useful?

Daytona

Sources Release page → 1 RELEASE · seen 2026-08-29 NOTES

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

Daytona 0.203.0 adds a CLI command for listing organization members and improves error handling consistency by propagating daemon error codes across all SDKs.

└──▷ 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
Organization members CLI commandNEW35

Adds an organization members command to the CLI for listing the members of an organization.

— No exact command syntax or flags givensnapshot-20260829
02
Daemon error codes propagated to SDKsIMPROVED33

Propagates missing daemon error codes to all SDKs for consistent error handling across clients.

— No specific error codes or SDKs namedsnapshot-20260829
Was this useful?

Earendil Works Pi

Sources Release notes → 1 RELEASE · 2026-08-28 NOTES

AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI

Pi's v0.84.4 window centered on terminal and TUI control: manual overrides for terminal capability detection, an RPC clear_queue command, extension lifecycle events for UI prompts, thinking-level and model-picker keyboard controls, fullscreen copy-on-select behavior, and an experimental DeepSeek vision model.

└──▷ 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
Terminal capability detection overridesNEW92

Adds env vars PI_HYPERLINKS=1|0|auto, PI_IMAGE_PROTOCOL=kitty|iterm2|none|auto, and PI_TRUE_COLOR=1|0|auto, plus JSON-only config keys terminal.hyperlinks, terminal.images ('kitty'|'iterm2'|false|'auto'), and terminal.trueColor, letting users manually override Pi's auto-detected OSC 8 hyperlink support, inline image protocol, and truecolor support when running behind a terminal proxy or multiplexer.

Force kitty inline image protocol when Pi's auto-detection fails inside a tmux session.
$ PI_IMAGE_PROTOCOL=kitty pi
Permanently disable OSC 8 hyperlinks and set truecolor for a terminal path that corrupts those sequences.
json 
{
  "terminal": {
    "hyperlinks": false,
    "trueColor": true
  }
}
Force kitty image protocol when auto-detection picks the wrong backend in your terminal emulator.
json 
{
  "terminal.images": "kitty"
}
— Names exact env vars, config keys, and values with runnable examples.v0.84.4
02
Per-model and per-session thinking level controlsNEW90

Adds the modelThinkingLevels config key to store per-model startup thinking levels keyed by "provider/modelId" (e.g. "anthropic/claude-sonnet-4-20250514": "high"), the /thinking command to choose a thinking level for the current session, Ctrl+S inside the thinking-level picker to save it as the startup default, and Shift+Tab to cycle through thinking levels.

Pin a specific model to high thinking level at startup so every new session uses extended reasoning without manual selection.
json 
{
  "modelThinkingLevels": {
    "anthropic/claude-sonnet-4-20250514": "high"
  }
}
— Names config key, command, shortcuts with a worked config example.product docs
03
Extension UI prompt lifecycle eventsNEW80

Adds notification-only ui_prompt_start and ui_prompt_end extension lifecycle events that fire around ctx.ui.select(), ctx.ui.confirm(), ctx.ui.input(), and ctx.ui.editor() calls, letting host/status integrations show a 'waiting for user' state instead of just 'running'. Each event carries event.reason ('ui_prompt'), event.kind ('select'|'confirm'|'input'|'editor'|'custom'), and event.title; nested or overlapping prompts are coalesced into a single outer waiting span.

— Explains mechanism, event fields, and coalescing behavior in detail.v0.84.4
04
RPC clear_queue commandNEW78

Adds a clear_queue RPC command ({"type": "clear_queue"}) that removes queued steering and follow-up messages from the RPC queue and returns their text in the steering and followUp response fields, enabling interactive Esc-style cancel-and-restore workflows.

Cancel in-flight queued guidance (Esc behavior) and recover the text so the user can edit and resend it.
$ {"type": "clear_queue"}
— Names exact RPC command, fields, with example payload.v0.84.4
05
Fullscreen copy-on-select and Ctrl+X behaviorNEW70

Adds the fullscreenCopyOnSelect config key controlling whether selected text is automatically copied in fullscreen mode; when disabled, Ctrl+X copies the active fullscreen selection, falling back to the selected message or the last assistant message depending on context.

Disable automatic selection copying in fullscreen so Ctrl+X controls exactly what gets copied.
json 
{ "fullscreenCopyOnSelect": false }
— Names config key and shortcut with example, but thin on scope.v0.84.4
06
Model picker keyboard shortcuts and startup defaultsNEW65

Adds Ctrl+L to open the model picker and choose a model for the current session, Ctrl+S inside the model picker (and via the /model command) to save the selected model as the startup default, and Ctrl+P / Shift+Ctrl+P to cycle forward and backward through scoped models.

— Names shortcuts but lacks a dedicated config key for defaults.product docs
thinner coverage below
07
DeepSeek vision model supportNEW45

Adds the experimental deepseek-v4-flash-vision-exp vision-capable model through Pi's built-in DeepSeek provider.

— Names model id but no usage or benchmark detail given.v0.84.4
08
Transcript usage notices for compaction and branch summariesIMPROVED45

Adds transcript usage notices for compaction and branch summaries when cache miss notices are enabled.

— Describes condition and scope but no exact config name.v0.84.4
09
detectSupportedImageMimeTypeFromFile library exportNEW35

Adds detectSupportedImageMimeTypeFromFile() to Pi's public library exports.

— Bare function name with no usage context provided.v0.84.4
Was this useful?
Other / Uncategorized
◆  AI OBSERVABILITY

LangChain LangSmith

Sources Release page → 1 RELEASE · 2026-08-10 NOTES

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

LangSmith's July window centers on new debugging and testing surfaces — OpenTelemetry resource metadata on traces, an API to validate thread evaluators before saving, and bulk split editing in experiment views — alongside zstd-compressed exports, trace-root scoped sharing, clearer evaluator/conflict error handling, and removal of legacy dataset comparison helpers from the public API and SDKs.

└──▷ 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
Validate thread evaluators via APINEW90

Test a multi-turn thread evaluator against a real conversation thread before saving it, using POST /runs/rules/validate with test_thread_id, session_id, and evaluator fields, to verify it grades correctly.

Test a multi-turn thread evaluator against a real conversation thread before saving it, to verify it grades correctly.
$ 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-id>", "session_id": "<session-id>", "evaluator": { ... }}'
— Named endpoint and payload fields with runnable curl examplesnapshot-20260829
02
zstd compression for bulk exportsIMPROVED80

Bulk exports now default to zstd compression; self-hosted deployments can keep the previous gzip behavior via FF_BULK_EXPORT_DEFAULT_COMPRESSION=gzip in .env.

On self-hosted deployments, keep gzip compression for bulk exports instead of the new zstd default.
$ FF_BULK_EXPORT_DEFAULT_COMPRESSION=gzip
— Exact config key and file named for overridesnapshot-20260829
thinner coverage below
03
Bulk split editing in experiment viewsNEW55

In experiment views, select multiple rows — or all matching the current filters — to add, replace, or remove dataset splits in one action, or copy the selected examples to another dataset.

— UI action described but no flag or command givensnapshot-20260829
04
Removal of legacy dataset comparison helpersBREAKING55

Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs, though existing HTTP routes continue to work for LangSmith UI clients.

— Scope of removal named but no replacement path givensnapshot-20260829
05
Trace-root scoped sharing for runsIMPROVED50

Sharing, viewing, or unsharing any run in a trace now operates on the trace root, making every run in a shared trace publicly viewable and opening public run links within the shared trace context.

— Behavior change explained but no UI path or commandsnapshot-20260829
06
Clearer 409 Conflict messages for duplicate runsIMPROVED50

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

— Named status code but no endpoint or example givensnapshot-20260829
07
Error recording for failed code evaluatorsIMPROVED35

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.

— Behavior described but no surface or command namedsnapshot-20260829
└──▷ BREAKING ON UPGRADE
  • !Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs (existing HTTP routes continue to work for LangSmith UI clients).
Was this useful?

Braintrust

Sources Release page → 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's biggest changes this window are a full Dashboards rework of monitoring views, public-preview MCP write tools for end-to-end pipeline automation, and a bt CLI overhaul that consolidates coding-agent tracing and auth commands; alongside these, the JS, Python and Java SDKs gained dozens of new auto-instrumentation targets, span metrics, and provider integrations including Azure AI Gateway, new Gateway open-source models, and an AWS Lambda trace extension.

└──▷ WHAT SHIPPED · 21 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
SDK instrumentation package migrations for OpenAI Agents and LangChainDEPRECATED83

Introduces the braintrust/instrumentation package with OpenAIAgentsTraceProcessor and auto-instrumentation via node --import braintrust/hook.mjs, replacing @braintrust/openai-agents, which is deprecated and has received its final release. @braintrust/langchain-js is likewise deprecated and has received its final release; users must migrate to BraintrustLangChainCallbackHandler.

— Names replacement package, class, and import path preciselyproduct docs
02
Dataset initialization by ID in Python and TypeScript SDKsNEW78

Adds a dataset_id parameter to init_dataset() in Python SDK v0.35.0 to initialize a dataset by ID, taking precedence over project, project_id, and name; adds the equivalent datasetId parameter to initDataset() in TypeScript SDK v3.29.0 to look up a dataset by ID without specifying a project.

Initialize a Braintrust dataset directly by its ID in Python, without needing to know the project name.
python 
import braintrust

ds = braintrust.init_dataset(dataset_id="<dataset-id>")
— Named params, SDK versions, and runnable code examplesnapshot-20260829
03
New tracing integrations for coding agents and frameworksNEW76

Adds a LangChain4j tracing integration for debugging prompts, evaluating models and monitoring production usage; a DeepSeek Harness integration (@braintrust/deepseek-harness, installed via dsh plugin --profile <profile> add @braintrust/deepseek-harness) that traces user turns, LLM steps, tool calls and nested child sessions; an OpenCode integration via the trace-opencode plugin with access to Braintrust data through built-in tools or MCP; and a pi coding-session integration via the pi-extension package that traces turns, model calls, tool executions and compactions. Also publishes OpenAI Agents JS Integration v0.1.6 and OpenTelemetry JS Integration v1.0.0 as new TypeScript SDK packages.

— Names packages and plugins but omits install commands for mostproduct docs
04
Dashboards replace monitoring views in ObserveNEW75

Monitoring views are now full Dashboards: each dashboard gets a dedicated page, a project-scoped list with search and starring, and clone/duplicate/rename/delete via the Dashboard actions menu, with chart, filter and grouping changes auto-saved. Dashboards can be duplicated across projects, built from a clone of the built-in dashboard, started empty, or duplicated from an existing one, and support time series, top list and big number chart types with export, duplicate and move-between-dashboards actions. New knowledge-base guides cover calculating per-trace averages with custom measures and creating a blank dashboard via API.

— Names UI actions and chart types but no API detailssnapshot-20260829
05
New auto-instrumentation targets across Python and TypeScript SDKsNEW73

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; Python SDK v0.32.0 adds Pipecat auto-instrumentation for real-time voice AI pipelines, covering LLM turns, STT, TTS, and tool calls. TypeScript SDK v3.28.0 adds Voyage AI auto-instrumentation for embeddings, multimodal embeddings, reranking, and contextualized embeddings. TypeScript SDK v3.27.0 adds Ollama and @cloudflare/think instrumentation, Anthropic beta sessions tracing via anthropic.beta.sessions.turn and anthropic.beta.sessions.thread.turn, Flue v2 support, and a vitest-evals span input override via meta.eval.input.

— Names many targets and functions across two SDKssnapshot-20260829
06
Java instrumentation adds OpenAI Responses and AWS Bedrock Converse supportNEW72

Java instrumentation v0.3.22 adds coverage for the OpenAI Responses API via OpenAiResponsesChatModel and OpenAiResponsesStreamingChatModel (LangChain4j 1.14.0+), and captures x-request-id and request-id vendor IDs as span attributes. It also adds AWS Bedrock Converse span capture for streaming tool-use, non-text content blocks, normalized reasoning blocks, prompt caching metrics, and tool definitions.

— Names classes and mechanisms but no code snippet givenproduct docs
07
Gateway token budget policiesNEW69

Adds gateway token budget policies scoped to a project, user, or API key to cap spend and token usage on every gateway request, documented at deploy/gateway-token-budgets. Policies are created under Deploy > Braintrust Gateway > Token Budgets, where you set the scope and configure token and spend limits.

Cap gateway token spend for a specific project to prevent runaway costs in production.
📍Go to Deploy › Braintrust Gateway › Token Budgets, create a new budget policy, set the scope to 'Project', select your project, and configure the token and spend limits.
— Doc path and UI steps given, no API endpointproduct docs
thinner coverage below
08
MCP write tools for pipeline automationNEW57

The Braintrust MCP server now exposes write tools (public preview) that let coding agents build dashboards, create alerts and scheduled jobs, run evals, edit dataset rows, and create or update prompts, scorers and classifiers end to end, using the permissions of the authenticated Braintrust account.

— Lists capabilities but no endpoint or tool names givensnapshot-20260829
09
Dataset rows can reference a group of up to 64 tracesIMPROVED54

A dataset row can now reference a group of up to 64 traces instead of a single one, with each trace rendered inline and unavailable traces flagged.

— Concrete limit of 64 traces, no UI path givensnapshot-20260829
10
Inline scorer rationale on Evaluators tabIMPROVED48

Applied scorers and classifiers on the Evaluators tab now show their recorded rationale inline, without requiring the user to open the producing span.

— Names exact tab; no further mechanism describedproduct docs
11
Permission group management from Settings > MembersIMPROVED48

Members' permission groups can now be managed directly from Settings > Members, showing direct groups, inherited groups, and available groups to add.

— Exact UI path given, thin on mechanismsnapshot-20260829
12
Auto-instrumentation applied to braintrust evalIMPROVED46

TypeScript SDK v3.29.0 makes braintrust eval apply auto-instrumentation automatically, so supported AI clients can emit spans before eval files load.

— Names command and behavior but no config detailsnapshot-20260829
13
generateImage calls traced as LLM spans with attachmentsNEW45

JS SDK v3.29.0 traces Vercel AI SDK generateImage calls as LLM spans, storing the generated images as Braintrust attachments.

— Names function and storage mechanism, but purely automaticsnapshot-20260829
14
Standalone billing address sectionNEW45

Adds a standalone Billing address section with an Edit address action to update the billing address without replacing the card on file.

— Names UI section and action; thin beyond thatproduct docs
15
Annotated version history for prompts, parameters, and scorersIMPROVED42

Prompt, parameter, and scorer version history now supports annotated saves with a description note pinned to each version, and shows each version alongside the one it replaced.

— Explains behavior but no UI path givensnapshot-20260829
16
Thinking and reasoning tokens recorded as span metricsIMPROVED41

The JS SDK now records thinking and reasoning tokens as span metrics; TypeScript SDK v3.29.0 specifically records Anthropic thinking tokens and LangChain reasoning tokens as span metrics.

— Automatic metric; no config or usage detail givensnapshot-20260829
17
Case id and tags passed to scorer functionsIMPROVED40

TypeScript SDK v3.28.0 passes the id and tags of each case to scorer functions in Eval().

— Bare field names with no usage examplesnapshot-20260829
18
apiKey parameter for loadPrompt()NEW38

TypeScript SDK v3.29.0 adds an apiKey parameter to loadPrompt().

— Bare parameter name with no usage contextsnapshot-20260829
19
Loop active observability featureNEW35

Introduces Loop, an active observability feature for investigating project data, editing Braintrust objects, and detecting recurring patterns via interactive and scheduled analysis.

— Vague description; no named surfaces or commandsproduct docs
20
New knowledge-base guide and feature lifecycle docsNEW27

Adds a knowledge-base guide for setting up Claude Code with OTel tracing, and a Feature lifecycle page documenting what private preview, public preview, and general availability mean for Braintrust features.

— Purely documentation; no product capability describedproduct docs
21
Sandboxes public previewNEW10

Sandboxes are now available in public preview.

— Bare mention with no detail on scope or mechanismproduct docs
└──▷ BREAKING ON UPGRADE
  • !bt auth logout is replaced by bt logout; bt auth profiles is replaced by bt profiles; bt auth refresh is replaced by bt login --refresh; the --fresh flag on bt login is renamed to --force.
  • !Coding-agent plugins (claude, codex, opencode, pi) must migrate to bt trace setup <agent> for tracing configuration — previous plugin-specific API key, project, tracing, and config.json/environment-variable settings are broken by this version.
  • !@braintrust/openai-agents is deprecated and has received its final release; users must migrate to braintrust/instrumentation using OpenAIAgentsTraceProcessor or node --import braintrust/hook.mjs.
  • !@braintrust/langchain-js is deprecated and has received its final release; users must migrate to BraintrustLangChainCallbackHandler.
  • !For trace-codex, config.json now takes precedence over environment variables.
  • !bt now handles authentication and trace routing for Claude Code, Codex, OpenCode, and pi plugins; older plugin-specific API key, project, tracing, and config-file settings must be migrated per the bt CLI migration guide.
  • !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 previous behavior.
Was this useful?
◆  VECTOR DB RAG

Pinecone

Sources Release page → 1 RELEASE · 2026-08-01 NOTES

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

Pinecone added Terraform support for managing organization-level access controls, including service accounts, role bindings, and invites.

└──▷ 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.

thinner coverage below
01
Terraform resources for organization access managementNEW57

Pinecone added Terraform resources for managing service accounts, role bindings, and organization invites, as well as for removing organization members.

— Names four Terraform resource types but no exact resource names or syntax.snapshot-20260829
Was this useful?
// END OF TRANSMISSION

The AI Toolchain · curated daily · August 29, 2026
Releases & features from the open-source and commercial AI tools that practitioners actually run.

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

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

    browse all tools →