docker/docker-agent v1.94.0

on GitHub

This release significantly expands model provider support with 12 new built-in providers, adds mouse support and new commands to the TUI, introduces a SafetyPolicy primitive, and includes numerous bug fixes and performance improvements.

What's New

• Adds Baseten as a built-in model provider (provider: baseten)
• Adds OVHcloud AI Endpoints as a built-in model provider (provider: ovhcloud)
• Adds Groq as a built-in model provider (provider: groq)
• Adds DeepSeek as a built-in model provider (provider: deepseek)
• Adds Cerebras as a built-in model provider (provider: cerebras)
• Adds Fireworks AI as a built-in model provider (provider: fireworks)
• Adds Together AI as a built-in model provider (provider: together)
• Adds Hugging Face Inference Providers as a built-in model provider (provider: huggingface)
• Adds Moonshot AI (Kimi) as a built-in model provider (provider: moonshot)
• Adds Vercel AI Gateway as a built-in model provider (provider: vercel)
• Adds Cloudflare Workers AI and Cloudflare AI Gateway as built-in model providers
• Adds mouse support to the agent picker: hover highlights a card, single click selects it, double-click immediately starts the session, and scroll wheel navigates the YAML details panel
• Adds metadata tags field to agent config and surfaces tags as coloured chips in the agent picker
• Adds /effort slash command to select the model's reasoning level directly (available in both full and lean TUI)
• Adds SafetyPolicy primitive (unsafe/safer/strict) to sessions, forwarded to hooks via hooks.Input.SafetyPolicy
• Enforces lifecycle.startup_timeout for MCP/LSP toolset startup, which was previously parsed but never acted on
• Adapts safer_shell hook classification to the session SafetyPolicy
• Collapses file picker keyboard shortcuts to a single line on wide dialogs
• Strengthens default memory toolset instructions

Improvements

• Memoizes successful OCI config reads per process, halving warm startup latency for OCI agent refs
• Suppresses spurious empty-response warning for benign post-tool stops in forked-skill sub-sessions
• Agent picker dialog is larger and truncates long lines instead of wrapping; cards are 70 columns wide

Bug Fixes

• Fixes coalescing of system messages for self-hosted and OpenAI-compatible endpoints that reject multiple system messages (e.g. vLLM with Qwen)
• Fixes OVHcloud default model to Qwen3.5-397B-A17B
• Fixes macOS Option-key characters not triggering file picker alt+h/alt+i visibility toggles
• Removes duplicate "Initializing MCP servers…" spinner from the sidebar that fired on every turn even for agents with no MCP servers
• Adds 5-second timeout to graceful-shutdown calls that previously used WithoutCancel without a deadline, preventing indefinite blocking
• Fixes concurrent Connect races and goroutine leak on startup timeout for MCP/LSP toolsets
• Fixes auto-deny of tool asks in non-interactive sessions instead of blocking indefinitely
• Fixes tool-call confirmations being dropped without a Resume under --yolo in JSON mode
• Fixes Anthropic provider falling back to token thinking for effort levels on models without adaptive thinking support (e.g. Haiku 4.5)
• Fixes tool call render briefly disappearing or shrinking while JSON arguments are still streaming in the lean TUI
• Clamps max_tokens to the context window for OpenAI-compatible providers to prevent "context window exceeded" errors on self-hosted vLLM
• Fixes cost and context limit display in the lean TUI

Technical Changes

• Derives provider API-key env vars forwarded to eval containers from the provider registry instead of a hard-coded list
• Refactors provider client boilerplate into shared helpers across openai, anthropic, gemini, bedrock, and dmr clients
• Consolidates OpenAI-compatible alias provider tests into a single shared test
• Extracts shared calleeObject helper in lint cops to remove duplication
• Adds OTelTracerName lint cop to enforce scoped OpenTelemetry tracer names
• Replaces 44 time.Sleep calls in the test suite with deterministic synchronization primitives

What's Changed

• docs: update CHANGELOG.md for v1.93.0 by @docker-read-write[bot] in #3339
• docs: sync documentation with recent main merges by @aheritier in #3340
• feat: add Baseten provider support by @dgageot in #3341
• feat: add OVHcloud AI Endpoints provider support by @aheritier in #3343
• chore: bump github.com/anthropics/anthropic-sdk-go to v1.55.0 by @dgageot in #3345
• feat: add Groq as a supported model provider by @Sayt-0 in #3358
• fix(openai): coalesce system messages for self-hosted and open-model endpoints by @Sayt-0 in #3357
• feat: add DeepSeek as a supported model provider by @Sayt-0 in #3361
• docs: group and collapse left navigation sections by @aheritier in #3360
• docs: list providers alphabetically by name in nav by @aheritier in #3364
• fix(docs): restore inline code contrast in light mode by @aheritier in #3365
• feat: collapse file picker shortcuts to one line on wide dialogs by @dgageot in #3366
• fix(tui): remove duplicate MCP-init spinner from sidebar by @dgageot in #3367
• fix: add 5s timeout to graceful-shutdown calls that used WithoutCancel by @dgageot in #3370
• feat: add Cerebras as a supported model provider by @Sayt-0 in #3368
• fix(tui): accept macOS Option-key chars for file picker toggles by @Sayt-0 in #3333
• feat: add Fireworks AI as a supported model provider by @Sayt-0 in #3369
• test: consolidate OpenAI-compatible alias provider tests by @dgageot in #3374
• refactor(config): derive provider API-key env vars from config by @dgageot in #3375
• feat: enforce lifecycle.startup_timeout for MCP/LSP toolset startup by @dgageot in #3373
• feat: add Hugging Face Inference Providers as a supported model provider by @Sayt-0 in #3379
• docs: tighten comment guidance in AGENTS.md by @dgageot in #3380
• fix: unblock headless tool-Ask paths (dispatcher + JSON runner) by @melmennaoui in #3377
• feat(session): add SafetyPolicy primitive and hook pass-through by @melmennaoui in #3378
• feat: add Together AI as a supported model provider by @Sayt-0 in #3376
• fix(anthropic): use token thinking for effort levels on models without adaptive support by @Sayt-0 in #3381
• fix(tui): keep tool call render stable while args stream by @rumpl in #3382
• feat(hooks/safer_shell): adapt classification to session SafetyPolicy by @melmennaoui in #3383
• feat: add Moonshot AI (Kimi) as a supported model provider by @Sayt-0 in #3384
• feat: add Vercel AI Gateway as a supported model provider by @Sayt-0 in #3385
• feat: add Cloudflare Workers AI and AI Gateway as model providers by @Sayt-0 in #3389
• test: make models-list tests hermetic and parallelizable by @dgageot in #3388
• feat(lint): add OTelTracerName cop to enforce scoped tracer names by @dgageot in #3386
• feat: add tags field to agent config metadata by @dgageot in #3390
• feat: add mouse support to the agent picker (hover, double-click, wheel) by @dgageot in #3391
• fix(runtime): suppress spurious empty-response warning for benign post-tool stops by @dgageot in #3392
• feat(memory): strengthen default toolset instructions by @dgageot in #3395
• fix: clamp max_tokens to the context window for OpenAI-compatible providers by @Sayt-0 in #3393
• test: make the test suite faster and sleep-free by @dgageot in #3399
• perf(config): memoize successful OCI reads to halve warm startup latency by @dgageot in #3397
• fix: show lean tui cost and context limit by @rumpl in #3400
• refactor: extract shared calleeObject helper in lint cops by @dgageot in #3396
• feat: show metadata tags as coloured chips in the agent picker by @dgageot in #3394
• refactor(provider): deduplicate provider client boilerplate by @dgageot in #3398
• feat(tui): add /effort command to select the model's reasoning level by @Sayt-0 in #3403

Full Changelog: v1.93.0...v1.94.0