GeezerVison

This release upgrades the Go toolchain to version 1.27.4, which includes security fixes.

Technical Changes

• Upgrades Go toolchain from 1.27.3 to 1.27.4, incorporating security fixes in cgo, the Go command, compiler, linker, and net/http

What's Changed

• docs: update CHANGELOG.md for v1.159.0 by @docker-read-write[bot] in #4900
• chore: upgrade Go to 1.27.4 by @aheritier in #4902

Full Changelog: v1.159.0...v1.160.0

Contributors

This release adds new providers and Anthropic fast mode support, completes built-in agent definitions, fixes several correctness and security issues, and includes extensive internal refactoring.

What's New

• Adds Obscurify.ai to the built-in provider catalog (OpenAI-compatible, default model deepseek/deepseek-v4-flash)
• Adds fast mode to the Anthropic provider via provider_opts.speed (fast | standard) for the Messages API
• Adds support for Claude Fable 5.2 and Mythos 5.2 on Bedrock
• Completes default and minimal built-in agents; the built-in default is now a coding agent with a minimal general-purpose alternative
• Emits a warning when a final reply is missing after tool use, instead of silently treating it as successful

Improvements

• Ends the Kitty graphics terminal probe on the DA1 reply, reducing startup delay for terminals that do not support Kitty graphics
• Loads prompt history off the startup path, cutting time to first TUI frame

Bug Fixes

• Fixes MCP tool results where separate text blocks were concatenated without separators, causing file lists and other multi-element results to run together
• Fixes Gemini handling of MALFORMED_FUNCTION_CALL / UNEXPECTED_TOOL_CALL to fail the stream with a non-retryable error instead of silently ending as a finished answer
• Fixes Gemini to omit thinkingConfig for non-reasoning models (fixes Gemini Flash Image)
• Fixes Gemini to keep invalid function calls non-retryable
• Fixes OpenAI-compatible stream adapter to treat content_filter finish reason as a refusal instead of passing it through unhandled
• Fixes config loading to accept agent configs and .env files saved with a UTF-8 BOM (affects files written by PowerShell 5 on Windows)
• Fixes .agentsignore parsing to remove the 64 KiB per-line scanner limit
• Fixes RAG package detection to stop dropping files with over-long lines
• Fixes transcript writes to be flushed before compaction is persisted, preventing ordering issues on Windows
• Fixes inherited Windows socket handles to be released with closesocket instead of CloseHandle
• Fixes docker agent serve acp to use the default agent when no agent argument is specified
• Fixes API read-only hints to apply only to GET requests; other HTTP methods are no longer incorrectly classified as read-only
• Fixes sandbox auto-kits to exclude workspace prompt symlinks, preventing host targets from being copied in
• Fixes message-merge probes to report aggregate prompt tokens correctly
• Fixes Anthropic catalogue pricing to be preserved for ANTHROPIC_BASE_URL endpoints in fast mode

Technical Changes

• Extensive internal refactoring across runtime, TUI, RAG, providers, shell, config, and other packages — removes redundant helpers, unused adapters, write-only state, and duplicate logic (net ~169 lines removed from production code)
• Refreshes the models.dev catalog snapshot (32 models added, 14 removed, 84 updated)
• Isolates test suites from the developer's home directory to prevent personal config from affecting test results
• Adds opt-in message-merging compatibility probes for provider testing

What's Changed

• refactor(leantui): clone cached tool lines directly by @dgageot in #4868
• refactor(rag): reuse initialization file sets by @dgageot in #4867
• refactor(rag): stop collecting unused fusion ranks by @dgageot in #4865
• refactor(markdown): reuse incremental renderer reset by @dgageot in #4866
• refactor(leantui): inline single-use command filter by @dgageot in #4864
• refactor(cli): reuse model-list JSON encoder by @dgageot in #4863
• refactor(runtime): remove redundant JSON key sorting by @dgageot in #4862
• refactor(bedrock): remove lazy document forwarder by @dgageot in #4861
• refactor(effort): use standard first-match lookup by @dgageot in #4860
• refactor(tui): remove write-only inline edit content by @dgageot in #4859
• refactor(markdown): remove unused derived style state by @dgageot in #4856
• refactor(tui): reuse snapshot pluralization helper by @dgageot in #4857
• refactor(runtime): remove harness tool ID accessor by @dgageot in #4858
• refactor(tasks): remove JSON result forwarder by @dgageot in #4855
• build(deps): bump the actions group across 1 directory with 2 updates by @dependabot[bot] in #4827
• refactor(shell): remove shell detection forwarder by @dgageot in #4853
• refactor(mermaid): reuse sequence canvas for lifelines by @dgageot in #4854
• refactor(skills): consolidate skill lookup by @dgageot in #4851
• refactor(providers): remove stream adapter forwarders by @dgageot in #4852
• refactor(runtime): reuse prompt joiner for delegation notes by @dgageot in #4850
• refactor(codemode): use maps.Clone for schema copies by @dgageot in #4849
• refactor(codemode): remove unused comment indentation layer by @dgageot in #4848
• refactor(tui): remove redundant extractor state flags by @dgageot in #4846
• refactor(leantui): remove image forwarding layer by @dgageot in #4845
• refactor(fetch): remove HTML-to-text forwarding helper by @dgageot in #4844
• refactor(kit): remove unused auto-install check by @dgageot in #4834
• refactor(shellpath): collapse redundant Unix shell helper by @dgageot in #4835
• refactor(runtime): remove unused approval source mappers by @dgageot in #4833
• refactor(skills): remove unused git root helper by @dgageot in #4832
• refactor(httpclient): inline test-only ID source option by @dgageot in #4838
• refactor(path): delegate home shortening to shared helper by @dgageot in #4840
• refactor(tui): remove test-only validation adapters by @dgageot in #4841
• refactor(tui): remove write-only pointer state by @dgageot in #4843
• fix(runtime): flush transcript writes before persisting compaction by @dgageot in #4871
• refactor(anthropic): remove test-only content helper by @dgageot in #4837
• docs: update CHANGELOG.md for v1.158.0 by @docker-read-write[bot] in #4830
• fix(oaistream): end a content_filter turn as a refusal by @jaeminbird in #4825
• fix(acp): use default agent when none is specified by @dgageot in #4876
• Make the default built-in a coding agent by @dgageot in #4875
• refactor(runtime): reuse notifier pointer identity helper by @dgageot in #4847
• refactor(tui): inline agent registry index lookups by @dgageot in #4839
• refactor(rag): remove BM25 document-loading adapter by @dgageot in #4870
• refactor(config): share instruction-file decoder callback by @dgageot in #4869
• refactor(providerutil): remove redundant nil-map guards by @dgageot in #4842
• fix(config): accept agent configs and env files saved with a UTF-8 BOM by @jaeminbird in #4826
• chore: bump direct Go dependencies by @dgageot in #4878
• Bump golang.org/x/term to v0.48.0 by @dgageot in #4879
• fix(server): release inherited Windows sockets with closesocket by @dgageot in #4872
• fix(gemini): fail invalid function calls, end unfinished continuations by @jaeminbird in #4808
• feat: add Obscurify.ai to the built-in provider catalog by @tonydavis629 in #4751
• chore: refresh models.dev snapshot (+32 -14 ~84) by @github-actions[bot] in #4880
• docs: correct configuration and tool guidance by @dgageot in #4885
• feat: support Claude Fable and Mythos 5.2 by @aheritier in #4883
• test(providers): add opt-in message-merging compatibility probes by @onatozmenn in #4831
• refactor: remove accidental complexity in 33 small, behavior-preserving steps by @dgageot in #4887
• fix(mcp): keep separate text blocks of a tool result apart by @jaeminbird in #4802
• fix(rag): stop dropping files with over-long lines in package detection by @TDnorthgarden in https...

Contributors

This release adds optional config history and model catalog support, improves tool duration visibility in both TUIs, and fixes compaction overflow and theme hot-reload issues.

What's New

• Makes config history and embedded model catalog optional for Go hosts, reducing dependencies and binary data when those features are not needed
• Shows persistent tool duration footers in both TUIs, displaying live durations and retaining final "Completed in …" or "Failed after …" messages

Bug Fixes

• Fixes compaction overflow by retrying the summary request with less history when the context budget is exceeded
• Fixes embedded model snapshot not being included in image context
• Fixes theme watcher stopping across same-theme reloads, which could cause hot-reload edits to be dropped

Technical Changes

• Removes unused wrappers across multiple packages (chat, concurrent, markdown, TUI, promptfiles, genai, logging, session, httpclient, config, creator, skills)
• Moves test-only helpers into test files and adds coverage for configured backup limits
• Refreshes the embedded models.dev catalog snapshot (85 models added, 86 removed, ~258 updated)

What's Changed

• fix(compaction): retry the summary with less history when it overflows by @melmennaoui in #4817
• feat: show persistent tool duration footers in both TUIs by @rumpl in #4819
• docs: update CHANGELOG.md for v1.157.0 by @docker-read-write[bot] in #4812
• Make config history and embedded model catalog optional by @dgageot in #4814
• fix: flaky theme hot-reload, A2A agent card and stream idle-timeout tests by @aheritier in #4821
• refactor: remove unused wrappers and isolate test helpers by @dgageot in #4823
• chore: refresh models.dev snapshot (+85 -86 ~258) by @github-actions[bot] in #4829

Full Changelog: v1.157.2...v1.158.0

Contributors

Maintenance release with dependency updates.

Full Changelog: v1.157.1...v1.157.2

Maintenance release with dependency updates.

Full Changelog: v1.157.0...v1.157.1

This release delivers several bug fixes across runtime token handling, model error recognition, and provider-specific behavior, along with a UI fix for text selection and a new build tag for embedders.

What's New

• Adds docker_agent_no_hook_templates build tag, allowing embedders to drop text/template from the runtime to reduce binary size (relevant for WebAssembly targets)
• Adds worktree: true opt-in for background agents to isolate concurrent tasks in separate git worktrees with independent checkouts, tools, and hooks
• Validates structured output using google/jsonschema-go

Bug Fixes

• Fixes tool calls cut off by the output-token limit being executed with truncated arguments instead of being retracted
• Fixes llama.cpp context overflow errors not being recognized as token overflow, preventing auto-compact from triggering on Docker Model Runner
• Fixes Gemini adapter not reporting MAX_TOKENS as a length stop or SAFETY stops as refusals, which could cause refused tool calls to execute and loop
• Fixes sort -ofile and tree -ofile (option operands) being incorrectly classified as safe commands
• Fixes server-side fallback token usage counting only the final attempt's tokens instead of all attempts, allowing token budgets to be exceeded
• Fixes Google SDK endpoint precedence so that GOOGLE_GEMINI_BASE_URL and GOOGLE_VERTEX_BASE_URL environment variables are not overridden by injected catalog endpoints
• Fixes OPENAI_BASE_URL not being considered when determining whether hosted (native) tool search should be disabled
• Fixes double-click text selection to use Unicode word boundaries and grapheme-aware hit testing, highlighting, and clipboard extraction
• Fixes runtime not preserving the raw thinking state of a length turn that retracts nothing

What's Changed

• docs: update CHANGELOG.md for v1.156.2 by @docker-read-write[bot] in #4797
• docs: auto-update for merged PRs (2026-11-03) by @aheritier in #4798
• build(deps): bump the actions group across 2 directories with 7 updates by @dependabot[bot] in #4788
• fix(runtime): retract a tool call cut off by the token limit by @PerryLink in #4784
• fix(modelerrors): recognize llama.cpp context overflow by @jaeminbird in #4785
• fix(gemini): report MAX_TOKENS as length and safety stops as refusal by @jaeminbird in #4786
• fix(openai): exclude OPENAI_BASE_URL from hosted tool search by @oyobbeb in #4793
• fix(google): preserve SDK endpoint precedence by @TDnorthgarden in #4790
• fix(runtime): count all server fallback tokens toward budgets by @TDnorthgarden in #4799
• fix(safety): do not classify sort and tree with an option operand as safe by @arxhr007 in #4787
• test(hooks): skip stdin JSON test when jq is missing by @onatozmenn in #4800
• Fix double-click text selection by @dgageot in #4801
• docs: auto-update for merged PRs (2026-11-05) by @aheritier in #4809
• Let embedders drop text/template from the runtime by @trungutt in #4811

New Contributors

• @TDnorthgarden made their first contribution in #4790
• @arxhr007 made their first contribution in #4787
• @onatozmenn made their first contribution in #4800

Full Changelog: v1.156.2...v1.157.0

Contributors

Maintenance release with dependency updates.

Full Changelog: v1.156.1...v1.156.2

This release introduces a TypeScript SDK that generates and publishes wire types directly from the Go source, making docker-agent the source of truth for TypeScript API integrations.

What's New

• Adds a TypeScript SDK (@docker/docker-agent-ts-sdk) that generates TypeScript types for the API wire types, sessions, and events from the Go source code
• Adds generated Zod schemas for built-in tool arguments as part of the TypeScript SDK
• Adds CI checks to build the TypeScript SDK package and verify its contents

What's Changed

• Generate and publish a TypeScript SDK for the API wire types by @trungutt in #4795

Full Changelog: v1.156.0...v1.156.1

Contributors

This release adds new MCP transport options, a CloudGPU provider, and improved elicitation forms for Claude Code, alongside a range of bug fixes for MCP environment handling, streaming finish reasons, and tool schema correctness.

What's New

• Adds remote.standalone_sse option to MCP toolsets, enabling the streamable-HTTP standalone SSE stream via YAML configuration
• Adds CloudGPU as a built-in provider in the embedded catalog, discoverable via CLOUDGPU_API_KEY with deepseek-v4.1-flash as the default model
• Adds multi-select and titled choices support in elicitation forms (checkboxes for array fields, option titles and descriptions for oneOf/anyOf choices)
• Puts Claude Code's questions (AskUserQuestion) and plan approvals (ExitPlanMode) to the user as elicitation forms instead of plain text

Improvements

• Centralizes providers into an embedded catalog, replacing hard-coded provider and alias tables

Breaking

• Evaluation containers no longer receive credentials or other environment variables automatically; explicitly forward each variable with -e NAME or -e NAME=VALUE.

Bug Fixes

• Fixes ${env.X} expansion in MCP toolset env preflight to match runtime expansion behavior
• Fixes MCP supervisors failing to start after an explicit stop; allows reconnection without public API changes
• Fixes text content being dropped when a provider sends the last content delta in the same chunk as a terminal finish reason
• Fixes MCP toolset env values being silently overridden by the host environment; toolset env now takes precedence
• Fixes model_context_window_exceeded stop reason on Anthropic and Bedrock being treated as a normal stop instead of truncation
• Fixes typeless enum/const schemas being incorrectly typed as object; type is now derived from the enum/const values
• Fixes OPENAI_BASE_URL not being treated as a custom endpoint, causing incorrect behavior with local servers
• Fixes held-back stop/length finish reason being lost or reported as stop when usage arrives in a separate final chunk
• Fixes concurrent session compaction requests; concurrent manual or overflow compactions are now rejected with compaction already in progress
• Fixes tool results in evaluation judge transcripts being silently truncated to 500 bytes; full tool responses are now preserved
• Fixes MCP stop and resume to be transport-safe

Technical Changes

• Refreshes the embedded models.dev catalog snapshot
• Updates CloudGPU documentation example to use kimi-k3 model

What's Changed

• fix: allow MCP supervisors to start after an explicit stop by @rumpl in #4716
• refactor: centralize providers in an embedded catalog by @rumpl in #4755
• docs: update CHANGELOG.md for v1.154.0 by @docker-read-write[bot] in #4758
• docs: add model provider contribution guide by @aheritier in #4761
• chore: update Go dependencies by @dgageot in #4762
• fix: reject concurrent session compaction requests by @rumpl in #4769
• fix(runtime): keep text sent with a terminal finish reason by @locker95 in #4718
• fix(mcp): give the toolset's env precedence over the host environment by @PerryLink in #4744
• fix(codemode): keep runtime-handled tools out of the script toolset by @PerryLink in #4754
• fix(anthropic,bedrock): treat model_context_window_exceeded as truncation by @jaeminbird in #4759
• feat(mcp): add remote.standalone_sse to open the streamable-HTTP standalone stream by @steilerDev in #4741
• fix(evaluation): preserve full tool results in judge transcripts by @melmennaoui in #4771
• feat: add CloudGPU provider (centralized catalog) by @mrhy7176 in #4763
• chore: bump Go dependencies by @dgageot in #4778
• docs(cloudgpu): use kimi-k3 in examples by @mrhy7176 in #4781
• feat(harness): put Claude Code's questions and plan approvals to the user by @Sayt-0 in #4767
• fix(tools): type typeless enum/const schemas from their values by @jaeminbird in #4760
• fix(openai): treat OPENAI_BASE_URL as a custom endpoint by @oyobbeb in #4773
• fix(oaistream): keep a held-back stop/length finish reason by @jaeminbird in #4774
• fix(config): expand ${env.X} in the MCP toolset env preflight by @PerryLink in #4712
• docs: auto-update for merged PRs (2026-10-30) by @aheritier in #4772
• chore: refresh models.dev snapshot (+116 -124 ~284) by @github-actions[bot] in #4789

New Contributors

• @jaeminbird made their first contribution in #4759
• @mrhy7176 made their first contribution in #4763
• @oyobbeb made their first contribution in #4773

Full Changelog: v1.154.0...v1.155.0

Contributors

This release adds support for the Lunaroute AI provider and fixes compatibility with recent sbx sandbox releases.

What's New

• Adds Lunaroute as an OpenAI-compatible provider, including credential-based selection and model discovery

Bug Fixes

• Fixes compatibility with recent sbx sandbox releases, accepting both array and legacy envelope listings from sbx ls --json and updating the cloud flag for sbx 0.48–0.50.0

What's Changed

• feat: add Lunaroute provider support by @aheritier in #4750
• fix: support recent sbx sandbox releases by @dgageot in #4753
• docs: update CHANGELOG.md for v1.153.0 by @docker-read-write[bot] in #4752

Full Changelog: v1.153.0...v1.154.0

Contributors