docker/docker-agent v1.142.0

on GitHub

This release adds several new features including background agent synchronization, native Anthropic compaction, expanded Gemini capabilities, OpenAI provider options, and a WebAssembly shared runtime, alongside numerous bug fixes for token accounting, streaming usage, and cost tracking.

What's New

• Adds wait_background_agents tool that performs an all-settled join on a caller-supplied list of task IDs, waiting up to a configurable timeout and returning an ordered result array
• Adds native Anthropic conversation compaction with replay guards and prefix mismatch controls, opt-in progress updates, strict tool argument enforcement, cache diagnostics, and request-context secret redaction
• Adds cache_diagnostics, preserve_reasoning, and native_tool_search provider opts for OpenAI's Responses API
• Adds native Gemini text embeddings support, enabling RAG use-cases on Gemini via the dedicated embeddings API
• Adds support for Gemini service tiers via provider_opts.service_tier, including flex tier with extended idle timeout handling
• Adds support for Gemini URL Context via provider_opts.url_context: true, allowing the model to fetch and reason over public URLs at inference time
• Adds Docker Model Runner model metadata and context limit discovery via the /engines/_configure endpoint, enabling auto-compaction and context gauge support for DMR models
• Replaces the WebAssembly bespoke event loop with the shared embeddedchat-backed agent runtime, adding portable tools, in-memory storage, egress proxy, session-scoped MCP OAuth token stores, and opt-in cloud provider builds

Improvements

• Defers TUI frame composition while the terminal is blurred, avoiding unnecessary rendering work during unfocused sessions

Bug Fixes

• Fixes deferred tail buffer not being cleared when a different assistant message becomes the active owner, preventing old content from bleeding into new message output
• Fixes Gemini native tool-call IDs being stripped instead of preserved, and fixes tool-response name matching to prevent malformed conversation history
• Fixes Gemini native tool-call ProviderID not being preserved in the WASM accumulator
• Fixes session compaction not drawing from the session's cost or token budgets, allowing spend beyond configured ceilings
• Fixes TUI animations continuing to run in unfocused and detached (tmux) sessions
• Fixes empty, refused, or reasoning-only assistant responses being skipped even when the provider reported billable usage
• Fixes OpenAI WebSocket pool allowing a single idle connection to be shared across concurrent streaming requests, which could corrupt both responses
• Fixes concurrent map and slice access on session model-state fields (AgentModelOverrides, CustomModelsUsed) causing data races
• Fixes unsafe concurrent Close and Next/retry calls on WebSocket and Anthropic streams
• Fixes runtime streams being abandoned before fully drained, which could release streaming locks or turn tokens prematurely
• Fixes cache tokens not being counted toward max_tokens token budgets; previously only input+output tokens were counted, allowing cache-heavy calls to bypass configured limits
• Fixes Gemini tool-use input tokens (ToolUsePromptTokenCount) being dropped, understating token usage and cost
• Fixes Anthropic streaming input token counts and cache read/write metrics being dropped because they were read from message_delta instead of the accumulated usage object
• Fixes per-call telemetry costs being reported incorrectly; previously cumulative session cost was emitted before the current call was charged
• Fixes context limits not being bounded before narrowing to int in the runtime
• Fixes deferral state not being preserved across session reloads for tools
• Fixes Bedrock SDK bearer authentication scheme configuration
• Fixes concurrent indexing usage accounting in RAG

Technical Changes

• Refactors external reference name and config import suffix parsing to use strings.CutLast
• Simplifies test sleep-and-wait pairs using Go 1.27's synctest.Sleep
• Simplifies composite literals in runtime and config packages using Go 1.27 promoted fields
• Adds Lint/ExclusiveStreamLease analyzer to detect shared WebSocket stream leases
• Adds Lint/SessionStateAccessors cop to require session model-state accessors
• Adds Lint/StreamCloseSafety cop to detect unsafe stream close access
• Adds lint detection for abandoned runtime streams
• Consolidates and deduplicates documentation across custom commands, thinking/task budgets, provider credentials, MCP configuration, server CLI flags, generated media, provider inheritance examples, and tool concepts

What's Changed

• refactor: parse reference names and import suffixes with strings.CutLast by @dgageot in #4325
• test: simplify sleep-and-wait pairs with synctest.Sleep by @dgageot in #4323
• refactor: simplify literals with Go 1.27 promoted fields by @dgageot in #4324
• docs: update CHANGELOG.md for v1.141.0 by @docker-read-write[bot] in #4326
• chore(deps): bump openai-go/v3 to v3.61.0 by @dgageot in #4327
• fix: preserve final TUI responses across deferred message transitions by @dgageot in #4328
• feat: add wait_background_agents join tool by @dgageot in #4330
• docs: consolidate custom command documentation by @dgageot in #4331
• docs: consolidate thinking and task budget references by @dgageot in #4332
• docs: centralize provider credential reference by @dgageot in #4333
• docs: separate MCP and shared tool configuration references by @dgageot in #4334
• fix(gemini): preserve native tool-call IDs and fix tool-response matching by @dgageot in #4339
• docs: centralize server CLI flag tables by @dgageot in #4335
• docs: consolidate generated media documentation by @dgageot in #4336
• docs: reuse provider inheritance examples by @dgageot in #4337
• docs: link tool concepts to the canonical catalog by @dgageot in #4338
• fix: charge compaction to shared session budgets by @dgageot in #4343
• feat(dmr): discover model metadata and context limits by @dgageot in #4340
• feat(openai): add cache_diagnostics, preserve_reasoning, and native_tool_search provider opts by @dgageot in #4355
• fix: report per-call costs in telemetry by @dgageot in #4341
• feat: support native Gemini text embeddings by @dgageot in #4350
• feat(anthropic): native compaction, progress updates, strict tools, cache diagnostics, and secret redaction by @dgageot in #4354
• feat: support Gemini service tiers by @dgageot in #4348
• fix(app): guard session model state with an accessor cop by @dgageot in #4347
• fix(tui): pause animations in unfocused and detached sessions by @dgageot in #4344
• fix(runtime): retain costs for empty assistant responses by @dgageot in #4345
• feat: support Gemini URL Context by @dgageot in #4349
• perf(tui): defer frame composition while blurred by @dgageot in #4342
• fix(openai): lease WebSocket connections exclusively by @dgageot in #4346
• fix: account for Gemini tool-use input tokens by @dgageot in #4359
• fix(runtime): count cache tokens toward token budgets by @dgageot in #4358
• fix: preserve Anthropic streaming input and cache usage by @dgageot in #4360
• test(tui): fix ordering race in TestThemeWatcher_WatchUserTheme by @aheritier in #4305
• fix: drain runtime streams before releasing turn ownership by @dgageot in #4353
• chore(deps): bump anyio from 4.9.0 to 4.14.2 in /examples/dhi/dhi_mcp_server in the pip group across 1 directory by @dependabot[bot] in #4361
• docs: auto-update for merged PRs (2026-09-19) by @aheritier in #4362
• feat(wasm): shared runtime, portable tools, in-memory storage, egress proxy, and cloud provider builds by @dgageot in #4357
• docs: auto-update for merged PRs (2026-09-21) by @aheritier in #4367

Full Changelog: v1.141.0...v1.142.0