v0.6.14

@workglow/webcontainer

Features

webcontainer

• list a huge directory's first entries without copying or sorting it
• lazySandboxPort and a coep option for bootWebContainer
• promote the sandbox core to the @workglow/webcontainer package

Bug Fixes

webcontainer

• drain a background program's output before reporting its failure
• one watcher for both entries, validate listSandboxFiles' root

@workglow/ai

Bug Fixes

ai

• check a round's submit_answer calls one at a time
• send no submit reminder a turn cannot answer
• report an answered agent turn as answered under an unmeasured budget
• fail maxInputTokens closed on uncounted rounds; estimate usage for Cactus and Chrome AI

huggingface-transformers

• offer only pipelines a run-fn serves, and infer the ones it does
• name a tool's image in the result a text-only model reads

openai

• stream a compatible server's raw reasoning on the reasoning port
• stream reasoning summaries on the tool-calling reasoning port

deepseek

• name a tool's image in its result instead of sending it

Refactors

ai

• split AgentTask's turn loop into budget, round, policy and tool-round modules
• a shared model-profile resolver; Anthropic and OpenAI read every per-model decision from it

workglow

Features

workglow

• re-export @workglow/web-search from the meta package

Chores

• update deps

Updated Dependencies

• tslog: ^5.4.0

@workglow/mcp

Build

• type-check the tests and agent-eval with bun check instead of tsc

@workglow/test

Features

workglow

• re-export @workglow/web-search from the meta package

huggingface-transformers

• send images to vision-language models
• constrain forced tool calls with a grammar

Bug Fixes

• hide an ambiguous URL authority; count Chrome's tool schemas
• read main's model profiles where the branch read raw ids
• refuse a truncated guard history; admit only the forced tool's call
• address code-review findings on the branch

ai

• send no submit reminder a turn cannot answer
• report an answered agent turn as answered under an unmeasured budget
• fail maxInputTokens closed on uncounted rounds; estimate usage for Cactus and Chrome AI

huggingface-transformers

• offer only pipelines a run-fn serves, and infer the ones it does
• refuse vision models that cannot chat before their weights load
• write history tool calls in the chat-template function form
• name a tool's image in the result a text-only model reads
• carry usage across the grammar fallback
• validate tool calls against the checkpoint prefix's tools
• harden constrained tool-call decoding

tasks

• redact credentials from URLs that do not parse, and from the fragment

openai

• stream a compatible server's raw reasoning on the reasoning port
• stream reasoning summaries on the tool-calling reasoning port

deepseek

• name a tool's image in its result instead of sending it

cactus

• carry each call and its labelled result into Needle's query

providers

• send the effort dial from the Gemini and Anthropic rewriter and summary run-fns
• send the effort dial from every text run-fn on Ollama, HF Inference and node-llama-cpp

tf-mediapipe,xai

• read a MediaPipe URL's task directory; retired Grok betas do not reason

anthropic,tf-mediapipe,xai

• review findings on the model profiles

anthropic

• read adaptive-thinking support through the id normaliser; make the table-agreement test non-tautological

Refactors

anthropic

• strip absent thinking only when known to be off
• one rule for when signed thinking is replayed

ollama,llamacpp-server,node-llama-cpp

• one model profile per provider

google-gemini,xai,deepseek,typesafeai

• one model profile per provider

huggingface-transformers,tf-mediapipe,chrome-ai

• one model profile per provider

openrouter,huggingface-inference

• one model profile per provider

ai

• a shared model-profile resolver; Anthropic and OpenAI read every per-model decision from it
• one model-profile table per provider for Anthropic and OpenAI

Tests

test-contract

• skip Gemini's depleted-credit 402 on CI
• wire six more providers into the tool-turn round trip
• pin the tool-turn round trip as a shared provider contract

ai-provider-api

• pin HF Inference conformance to Llama 3.1 8B on deepinfra
• pin HF Inference conformance to Llama 3.1 8B on deepinfra
• skip HF Inference conformance while tool calling is unroutable

huggingface-inference

• pin the live conformance model to Llama 3.1 8B on deepinfra

huggingface-transformers

• type the second tool in the constrained-call tests

@workglow/tasks

Bug Fixes

• hide an ambiguous URL authority; count Chrome's tool schemas
• address code-review findings on the branch

tasks

• redact credentials from URLs that do not parse, and from the fragment

@workglow/test-contract

Features

huggingface-transformers

• send images to vision-language models

Bug Fixes

deepseek

• name a tool's image in its result instead of sending it

openai

• stream reasoning summaries on the tool-calling reasoning port

Tests

test-contract

• skip Gemini's depleted-credit 402 on CI
• wire six more providers into the tool-turn round trip
• pin the tool-turn round trip as a shared provider contract

@workglow/openai

Bug Fixes

openai

• stream reasoning summaries on the tool-calling reasoning port

Refactors

ai

• a shared model-profile resolver; Anthropic and OpenAI read every per-model decision from it
• one model-profile table per provider for Anthropic and OpenAI

@workglow/llamacpp-server

Refactors

ollama,llamacpp-server,node-llama-cpp

• one model profile per provider

@workglow/ollama

Bug Fixes

providers

• send the effort dial from every text run-fn on Ollama, HF Inference and node-llama-cpp

Refactors

ollama,llamacpp-server,node-llama-cpp

• one model profile per provider

@workglow/node-llama-cpp

Bug Fixes

huggingface-transformers

• name a tool's image in the result a text-only model reads

providers

• send the effort dial from every text run-fn on Ollama, HF Inference and node-llama-cpp

Refactors

ollama,llamacpp-server,node-llama-cpp

• one model profile per provider

@workglow/anthropic

Bug Fixes

• read main's model profiles where the branch read raw ids

providers

• send the effort dial from the Gemini and Anthropic rewriter and summary run-fns

anthropic,tf-mediapipe,xai

• review findings on the model profiles

anthropic

• read adaptive-thinking support through the id normaliser; make the table-agreement test non-tautological

Refactors

anthropic

• strip absent thinking only when known to be off
• one rule for when signed thinking is replayed

ai

• a shared model-profile resolver; Anthropic and OpenAI read every per-model decision from it
• one model-profile table per provider for Anthropic and OpenAI

Performance

anthropic

• keep rewriter and summary streams off the SDK's generic stream type

@workglow/google-gemini

Bug Fixes

providers

• send the effort dial from the Gemini and Anthropic rewriter and summary run-fns

Refactors

google-gemini,xai,deepseek,typesafeai

• one model profile per provider

@workglow/xai

Bug Fixes

tf-mediapipe,xai

• read a MediaPipe URL's task directory; retired Grok betas do not reason

anthropic,tf-mediapipe,xai

• review findings on the model profiles

Refactors

google-gemini,xai,deepseek,typesafeai

• one model profile per provider

@workglow/deepseek

Bug Fixes

deepseek

• name a tool's image in its result instead of sending it

Refactors

google-gemini,xai,deepseek,typesafeai

• one model profile per provider

@workglow/typesafeai

Refactors

google-gemini,xai,deepseek,typesafeai

• one model profile per provider

@workglow/huggingface-transformers

Features

huggingface-transformers

• send images to vision-language models
• constrain forced tool calls with a grammar

Bug Fixes

• read main's model profiles where the branch read raw ids
• refuse a truncated guard history; admit only the forced tool's call
• address code-review findings on the branch

huggingface-transformers

• offer only pipelines a run-fn serves, and infer the ones it does
• refuse vision models that cannot chat before their weights load
• write history tool calls in the chat-template function form
• name a tool's image in the result a text-only model reads
• carry usage across the grammar fallback
• validate tool calls against the checkpoint prefix's tools
• harden constrained tool-call decoding
• type the tool-call logits processor

Refactors

huggingface-transformers,tf-mediapipe,chrome-ai

• one model profile per provider

@workglow/tf-mediapipe

Bug Fixes

tf-mediapipe,xai

• read a MediaPipe URL's task directory; retired Grok betas do not reason

anthropic,tf-mediapipe,xai

• review findings on the model profiles

Refactors

huggingface-transformers,tf-mediapipe,chrome-ai

• one model profile per provider

@workglow/chrome-ai

Bug Fixes

• hide an ambiguous URL authority; count Chrome's tool schemas

ai

• fail maxInputTokens closed on uncounted rounds; estimate usage for Cactus and Chrome AI

Refactors

huggingface-transformers,tf-mediapipe,chrome-ai

• one model profile per provider

@workglow/openrouter

Refactors

openrouter,huggingface-inference

• one model profile per provider

@workglow/huggingface-inference

Bug Fixes

providers

• send the effort dial from every text run-fn on Ollama, HF Inference and node-llama-cpp

Refactors

openrouter,huggingface-inference

• one model profile per provider

@workglow/cactus

Bug Fixes

cactus

• carry each call and its labelled result into Needle's query

ai

• fail maxInputTokens closed on uncounted rounds; estimate usage for Cactus and Chrome AI

@workglow/webcontainer-agent

Features

webcontainer

• promote the sandbox core to the @workglow/webcontainer package

examples

• add webcontainer-agent, an AgentTask coding agent sandboxed in a WebContainer

Bug Fixes

webcontainer-agent

• stop_servers also stops background starts still in flight
• serialize background server startups
• exact truncation flag, race-free server slots, tolerant tree refresh
• bound the page's file-tree walk like list_files
• report previews from shared tools, honour cancel during spawn
• keep preview servers alive across follow-up turns

Chores

• update deps

webcontainer

• record the typecheck budget, drop the example's test script

@workglow/eval

Chores

• update deps

Updated Dependencies

• hyparquet: ^1.31.5

@workglow/cli

Features

cli

• register WebSearchTask with no search provider

Chores

• update deps

Updated Dependencies

• preact: ^11.0.2

@workglow/web

Chores

• update deps

Updated Dependencies

• @codemirror/view: 6.43.15
• @lezer/common: 1.5.4
• react-dom: ^19.3.1
• react-resizable-panels: ^4.15.0
• jsdom: ^30.1.3
• vite: ^8.3.5

@workglow/agent-eval

Bug Fixes

ci

• green the nightly type-drift guard and Bun parity runs

Build

• type-check the tests and agent-eval with bun check instead of tsc