v1.46.0

This release adds OAuth credential configuration for MCP servers, evaluation testing improvements, and numerous stability fixes.

What's New

• Adds support for explicit OAuth credentials configuration for remote MCP servers that don't support Dynamic Client Registration
• Adds --repeat flag to eval command for running evaluations multiple times
• Adds support for xhigh effort level in Anthropic adaptive thinking (Claude Opus 4.7+)
• Adds task_budget configuration field for Claude Opus 4.7 to cap total tokens across multi-step tasks
• Adds markdown rendering support in user_prompt dialog messages

Improvements

• Improves image attachment handling by inlining as base64 data URLs for cross-provider compatibility
• Improves robots.txt caching to store parsed data per host instead of boolean results
• Improves session database version detection with clear upgrade messages for newer databases

Bug Fixes

• Fixes --attach flag being silently ignored when used without a message argument
• Fixes data race in AddMessageUsageRecord by adding mutex lock
• Fixes data race in rule-based router by protecting lastSelectedID with mutex
• Fixes panic in extractSystemBlocks when system message is empty with CacheControl
• Fixes empty messages slice handling in SendUserMessage path
• Fixes symlink-based path traversal vulnerability in ACP filesystem toolset
• Fixes OAuth callback CSRF vulnerability by rejecting when expected state is not set
• Fixes MCP tryRestart to use context-aware select instead of time.Sleep
• Fixes assistant text being discarded when tool calls are present in Responses API conversion
• Fixes MCP OAuth token refresh by remembering the discovered auth server

Technical Changes

• Updates mutex handling for MCP Toolset.Instructions() method
• Updates Go dependencies including Anthropic SDK and various UI libraries

---

What's Changed

• docs: update CHANGELOG.md for v1.45.0 by @docker-read-write[bot] in #2427
• fix: reject OAuth callback when expected state has not been set (CSRF) by @dgageot in #2434
• fix: prevent symlink-based path traversal in ACP filesystem toolset by @dgageot in #2433
• fix: guard against empty messages slice in SendUserMessage path by @dgageot in #2432
• fix: prevent panic in extractSystemBlocks on empty system message wit… by @dgageot in #2431
• fix: hold mutex for instructions read in MCP Toolset.Instructions() by @dgageot in #2430
• fix: add mutex to protect lastSelectedID in rule-based router by @dgageot in #2429
• fix: add mutex lock to AddMessageUsageRecord to prevent data race by @dgageot in #2428
• Add --repeat flag to eval command for running evaluations multiple times by @dgageot in #2440
• fix: replace time.Sleep with context-aware select in MCP tryRestart by @dgageot in #2436
• fix: preserve assistant text when tool calls present in Responses API conversion by @dgageot in #2438
• fix: detect newer session database and show clear upgrade message by @dgageot in #2441
• Support explicit OAuth credentials for remote MCP servers by @dgageot in #2394
• fix: cache parsed robots.txt per host instead of boolean result by @dgageot in #2437
• bump direct Go dependencies by @dgageot in #2444
• Render markdown in user_prompt dialog messages by @dgageot in #2446
• Add a pokemon example by @dgageot in #2445
• Add an advanced coder example by @dgageot in #2447
• fix(mcp): reuse discovered auth server for token refresh by @pandego in #2448
• Fix --attach flag by @rumpl in #2449
• Support xhigh effort for Anthropic adaptive thinking (Opus 4.7+) by @dgageot in #2450
• feat(anthropic): add task_budget for Claude Opus 4.7 by @dgageot in #2453
• chore: update cagent-action to v1.4.1 by @derekmisler in #2454

Full Changelog: v1.45.0...v1.46.0

v1.45.0

This release improves template expression handling, adds circular navigation to completions, and fixes issues with skills and MCP toolset loading.

Bug Fixes

• Fixes evaluation of JavaScript template expressions to handle failures independently - when one expression fails, other valid expressions in the same template are still expanded
• Fixes skills loading functionality
• Fixes retry behavior for MCP toolset startup when server is unavailable
• Fixes MCP toolset creation to proceed even when command binary is unavailable

Improvements

• Adds circular navigation wrapping to completion component, allowing users to cycle through completion options

---

What's Changed

• docs: update CHANGELOG.md for v1.44.0 by @docker-read-write[bot] in #2403
• add circular navigation wrapping to completion component by @maxcleme in #2407
• fix: evaluate JS template expressions independently on failure by @dgageot in #2400
• fix: retry stdio MCP toolset when binary is unavailable at startup by @simon-agent-go-expert in #2413
• Fix skills loading by @dgageot in #2414

Full Changelog: v1.44.0...v1.45.0

v1.44.0

This release introduces TUI customization capabilities, session management improvements, and OAuth security enhancements, along with numerous bug fixes and stability improvements.

What's New

• Adds support for extending and customizing TUI with additional commands through new Immediate flag and Parser struct
• Adds session delete functionality to session browser
• Adds click-to-select support for agents in the sidebar
• Adds /fork slash command to duplicate current session into a new tab
• Adds mid-turn message steering for running agent sessions with new /steer and /followup API endpoints
• Adds OAuth token storage in OS keychain with silent refresh token support
• Adds debug OAuth commands: list, remove, and login
• Adds support for shell expansions (~, env vars) in config paths
• Adds total session count display in session browser dialog title

Improvements

• Improves TUI rendering to match sandbox template
• Makes Ctrl+W context-aware to preserve word deletion in editor when focused
• Makes /exit close only the current tab when multiple tabs are open

Bug Fixes

• Fixes crash when opening empty websocket frames in OpenAI provider
• Fixes Gemini thinking tokens not included in output token count for cost calculation
• Fixes tool calls getting stuck as running when moved out of active reasoning block
• Fixes missing type in schema and orphaned function calls in Responses API
• Fixes spurious blank line appearing in every assistant message
• Fixes layout shift when hovering over assistant messages to reveal copy button
• Fixes concurrent RunSession calls causing tool_use/tool_result mismatch
• Fixes panic in code mode when tool handler is nil
• Fixes suggestion ghost text remaining when completion dialog closes on backspace
• Fixes skill frontmatter parsing when description contains a colon
• Fixes sidebar agent click zones mapping all lines to first agent
• Fixes OAuth token security vulnerabilities and infinite recursion issues
• Fixes auto-detect tool install failures being treated as fatal
• Fixes background agent context being cancelled with parent message lifecycle

Technical Changes

• Stores OAuth tokens in OS keychain with graceful fallback to in-memory storage
• Serializes concurrent RunSession calls to prevent race conditions
• Sanitizes message history to ensure all tool calls have results
• Adds regression tests for SSE comment lines from OpenRouter
• Uses in-memory store in keyring tests to avoid macOS keychain permission dialog
• Separates steer and follow-up into distinct queues with lock/confirm semantics
• Adds documentation for OpenAPI toolset
• Optimizes PR CI build process

---

What's Changed

• docs: update CHANGELOG.md for v1.43.0 by @docker-read-write[bot] in #2347
• Better sandbox by @dgageot in #2348
• support session delete to session browser by @shaj13 in #2352
• fix: include Gemini thinking tokens in output token count for cost calculation by @dgageot in #2353
• fix(openai): ignore empty websocket frames by @pandego in #2350
• Store OAuth tokens in OS keychain and add silent refresh token support by @dgageot in #2355
• bump direct Go dependencies by @dgageot in #2358
• Allow to extend and customize TUI with additional commands by @shaj13 in #2346
• Debug oauth by @dgageot in #2365
• fix: handle missing type in schema and orphaned function calls in Responses API by @dgageot in #2362
• Fix tool call stuck as running when moved out of active reasoning block by @dgageot in #2360
• Add regression tests for SSE comment lines from OpenRouter by @dgageot in #2359
• feat: click on agent in sidebar to switch to it by @dgageot in #2356
• fix(tui): remove spurious blank line from every assistant message by @steilerDev in #2369
• fix: use in-memory store in keyring tests to avoid macOS keychain permission dialog by @dgageot in #2367
• docs: add documentation for OpenAPI toolset by @dgageot in #2371
• fix(tui): reserve stable top row for copy icon to prevent layout shift by @dgageot in #2374
• fix: serialize concurrent RunSession calls to prevent tool_use/tool_result mismatch by @dgageot in #2375
• Faster PR CI by @rumpl in #2378
• Sanitize message history by @rumpl in #2377
• fix(toolinstall): soft-fail auto-detect installs by @pandego in #2386
• fix: prevent panic in code mode when tool handler is nil by @dgageot in #2388
• optional title and app name by @shaj13 in #2366
• fix: /exit closes only the current tab when multiple tabs are open by @dgageot in #2387
• Add support for shell expansions (~, env vars) in config paths by @dgageot in #2389
• Bump direct Go dependencies by @dgageot in #2398
• fix: OAuth token security and bug fixes by @dgageot in #2395
• fix: decouple background agent context from parent message lifecycle by @dgageot in #2392
• Show total session count in session browser dialog title by @dgageot in #2391
• fix: make Ctrl+W context-aware to preserve word deletion in editor by @dgageot in #2390
• Add /fork slash command to duplicate current session into a new tab by @shaj13 in #2385
• Fix: allow tab auto-complete without executing by @joshbarrington in #2325
• Fix skill frontmatter parsing when description contains a colon by @dgageot in #2401
• fix: clear suggestion ghost text when completion dialog closes on backspace by @dgageot in #2399
• Add mid-turn message steering for running agent sessions by @trungutt in #2363
• Fix sidebar agent click zones mapping all lines to first agent by @dgageot in #2402

New Contributors

• @pandego made their first contribution in #2350
• @steilerDev made their first contribution in #2369
• @joshbarrington made their first contribution in #2325

Full Changelog: v1.43.0...v1.44.0

v1.43.0

This release adds non-interactive mode capabilities, improves TUI interactions with mouse support, and includes several bug fixes for RAG tools and streaming responses.

What's New

• Adds auto-stop for max iterations in non-interactive mode to prevent hanging when tools are approved
• Adds non-interactive mode flag to distinguish from tools approval scenarios
• Adds mouse drag-to-move support for TUI dialogs, allowing repositioning by clicking and dragging the title area
• Adds custom session ID support through WithID option instead of relying on UUID generation
• Adds support for custom providers in RAG embedding and reranking models
• Adds underline styling for URLs on mouse hover

Improvements

• Evolves providers config to support any provider type with shared model defaults
• Improves mise build output to show go build command and resulting binary
• Exempts background-agent polling from loop-termination detection to prevent false positives

Bug Fixes

• Fixes agent accent color application for working spinners in sidebar
• Fixes duplicate RAG tool names and nil pointer panic in file watcher
• Fixes toolset startup triggering from emitToolsChanged callback to avoid spurious timeout warnings
• Fixes missing Models map in RAG ManagersBuildConfig for model alias resolution
• Fixes nil pointer dereference in BM25Strategy.watchLoop during session teardown
• Fixes extraction of reasoning_content from DMR streaming responses
• Fixes scrollbar rendering in web terminals by replacing problematic characters

Technical Changes

• Adds nocgo build support for rag/treesitter
• Updates error message display when only one model is available

---

What's Changed

• docs: update CHANGELOG.md for v1.42.0 by @docker-read-write[bot] in #2317
• fix: use agent accent color for working spinners in sidebar by @jiahuipaung in #2315
• Exempt background-agent polling from loop-termination detection by @krissetto in #2319
• fix: pass Models map to RAG ManagersBuildConfig for model alias resolution by @dgageot in #2324
• fix: avoid triggering toolset startup from emitToolsChanged callback by @dgageot in #2323
• fix: resolve duplicate RAG tool names and nil pointer panic in file watcher by @dgageot in #2322
• rag/treesitter: add nocgo build support by @shaj13 in #2329
• Bump go dependencies by @dgageot in #2330
• session: add WithID option for custom session IDs by @shaj13 in #2331
• feat(runtime): add auto-stop for max iterations in non-interactive mode by @tdabasinskas in #2208
• Fix nil pointer dereference in BM25Strategy.watchLoop during session teardown by @dgageot in #2334
• Fix scrollbar rendering in web terminals by @dgageot in #2340
• Improve mise build output to show go build command and resulting binary by @dgageot in #2338
• fix: extract reasoning_content from DMR streaming responses by @dgageot in #2335
• feat: add mouse drag-to-move support for TUI dialogs by @dgageot in #2339
• Evolve providers to support any provider type with shared model defaults by @dgageot in #2343
• feat: support custom providers in RAG embedding and reranking models by @dgageot in #2344
• Underline URLs on mouse hover by @dgageot in #2316
• Nicer message by @rumpl in #2345

New Contributors

• @jiahuipaung made their first contribution in #2315
• @shaj13 made their first contribution in #2329

Full Changelog: v1.42.0...v1.43.0

v1.42.0

This release improves evaluation output with structured JSON results and fixes several Windows compatibility issues.

What's New

• Adds URL click detection for terminals with mouse tracking support
• Includes structured results, run configuration, and summary in evaluation JSON output
• Includes judge reasons for passed relevance criteria in evaluation results

Bug Fixes

• Fixes Windows OS detection typo in session environment (corrects "window" to "windows")
• Replaces removed claude-3-7-sonnet-latest alias with explicit model ID in examples
• Uses platform-aware shell detection for Windows compatibility in skill expansion, script_shell, post-edit hooks, and bang commands

Technical Changes

• Pre-populates criterion names in CheckRelevance results
• Fixes lint issues including gci formatting and testifylint float comparisons

---

What's Changed

• docs: update CHANGELOG.md for v1.41.0 by @docker-read-write[bot] in #2307
• eval: include structured results, run config, and summary in JSON output by @hamza-jeddad in #2309
• fix: correct Windows OS detection typo in session environment by @trungutt in #2312
• fix: replace removed claude-3-7-sonnet-latest alias in examples by @trungutt in #2313
• fix: use platform-aware shell for skill expansion, script_shell, post-edit hooks, and bang command by @trungutt in #2314
• tui/messages: Add URL click detection for terminals with mouse tracking by @vvoland in #2308

Full Changelog: v1.41.0...v1.42.0

v1.41.0

This release introduces a new models discovery command, contextual help system, and several TUI improvements including persistent warnings and simplified lean mode.

What's New

• Adds docker agent models command to list available models for the --model flag
• Adds contextual help dialog accessible via Ctrl+H (or F1/Ctrl+?) showing all keyboard shortcuts
• Adds --lean flag for simplified TUI mode with minimal interface (just message stream and editor)
• Adds copy button on hover for assistant messages to copy content to clipboard
• Adds Vertex AI Model Garden support for non-Gemini models (Claude, Llama) hosted on Google Cloud

Improvements

• Makes TUI warnings persist until manually dismissed instead of auto-dismissing after 3 seconds
• Preserves recent messages during session compaction to maintain conversational context
• Shows elapsed time and warning for long-running tool calls in the TUI
• Adds desktop_uuid in telemetry alongside user_uuid for better tracking

Bug Fixes

• Fixes markdown rendering in callout notes by adding markdown="1" attribute
• Fixes panic on closed channel by making chanSend non-blocking
• Fixes recursive run_skill loop in context:fork skill sub-sessions
• Fixes docker run --sandbox functionality
• Fixes eval tool_call_response to use correct event field names
• Fixes guard against nil tool_definition in buildTranscript

Technical Changes

• Replaces kin-openapi with pb33f/libopenapi for OpenAPI parsing
• Removes trailing headers handling for rate limit headers
• Tracks command errors with success=false and error details in telemetry
• Ports build system to mise
• Updates Go module dependencies

---

What's Changed

• Add desktop_uuid in telemetry (next to user_uuid) by @gtardif in #2279
• docs: update CHANGELOG.md for v1.40.0 by @docker-read-write[bot] in #2281
• Preserve recent messages during session compaction by @rumpl in #2259
• Make TUI warnings persist until manually dismissed by @dgageot in #2252
• Fix markdown rendering in documentation callout notes by @dgageot in #2285
• go mod tidy by @dgageot in #2292
• docs: add Vertex AI Model Garden section to Google provider docs by @dgageot in #2290
• Add copy button on hover for assistant messages by @dgageot in #2288
• Add Vertex AI Model Garden support for non-Gemini models by @dgageot in #2287
• fix: make chanSend non-blocking to prevent panic on closed channel by @dgageot in #2286
• tui: show elapsed time and warning for long-running tool calls by @dgageot in #2291
• Bump direct Go module dependencies by @dgageot in #2284
• Fix TUI stuck in Working state after failed sub-agent transfer_task by @dgageot in #2294
• Add --lean flag for simplified TUI mode by @dgageot in #2253
• Port to mise by @dgageot in #2293
• Track command errors with success=false and error details by @gtardif in #2283
• Remove trailing headers handling for rate limit headers by @dgageot in #2298
• fix: prevent recursive run_skill loop in context:fork skill sub-sessions by @dgageot in #2289
• Replace kin-openapi with pb33f/libopenapi for OpenAPI parsing by @dgageot in #2299
• fix: eval tool_call_response uses correct event field names by @hamza-jeddad in #2302
• Fix docker run --sandbox by @dgageot in #2301
• use DD proxy when available, also from WSL by @gtardif in #2306
• Add contextual help dialog (Ctrl+H) by @masegraye in #2305
• feat: add docker agent models command by @masegraye in #2304

New Contributors

• @hamza-jeddad made their first contribution in #2302
• @masegraye made their first contribution in #2305

Full Changelog: v1.40.0...v1.41.0

v1.40.0

This release improves AI assistant capabilities with better response tracking and Google integration, plus fixes a critical exit hang issue.

What's New

• Adds Google Search, Google Maps, and code execution capabilities for Gemini models
• Surfaces finish_reason information on assistant messages and token usage events to track why the AI stopped generating responses

Bug Fixes

• Fixes process hang when using /exit command due to bubbletea renderer deadlock

Technical Changes

• Adds tests reproducing bubbletea renderer deadlock on exit
• Adds safety-net exit mechanism for bubbletea renderer deadlock prevention

---

What's Changed

• docs: update CHANGELOG.md for v1.39.0 by @docker-read-write[bot] in #2265
• Fix process hang on /exit due to bubbletea renderer deadlock by @aheritier in #2269
• Fix url by @dgageot in #2277
• Google grounding by @dgageot in #2276
• Surface finish_reason on assistant messages and token usage events by @trungutt in #2254

Full Changelog: v1.39.0...v1.40.0

v1.39.0

This release adds new color themes for the terminal interface and includes internal version management updates.

What's New

• Adds Calm Roots theme with warm white accents, sage green info messages, and charcoal background
• Adds Neon Pink theme with vibrant pink tones and high-contrast white accents for readability

Technical Changes

• Freezes v7 version
• Updates CHANGELOG.md for v1.38.0

---

What's Changed

• docs: update CHANGELOG.md for v1.38.0 by @docker-read-write[bot] in #2256
• Add Calm Roots and Neon Pink themes by @ambigois in #2260
• Freeze v7 by @dgageot in #2264

Full Changelog: v1.38.0...v1.39.0

v1.38.0

This release improves OAuth configuration and fixes tool caching issues with remote MCP server reconnections.

Improvements

• Changes OAuth client name to "docker-agent" for better identification
• Reworks compaction logic to prevent infinite loops when context overflow errors occur repeatedly

Bug Fixes

• Fixes tool cache not refreshing after remote MCP server reconnects, ensuring updated tools are available after server restarts

Technical Changes

• Updates CHANGELOG.md for v1.37.0 release documentation

---

What's Changed

• docs: update CHANGELOG.md for v1.37.0 by @docker-read-write[bot] in #2243
• Change the oauth client name to docker-agent by @rumpl in #2245
• fix: refresh tool and prompt caches after remote MCP server reconnect by @dgageot in #2246
• Refactor compaction by @dgageot in #2242

Full Changelog: v1.37.0...v1.38.0

v1.37.0

This release adds support for forwarding sampling parameters to provider APIs, introduces global user-level permissions, and includes several bug fixes and improvements.

What's New

• Adds support for forwarding sampling provider options (top_k, repetition_penalty, etc.) to provider APIs
• Adds global-level permissions from user config that apply across all sessions and agents
• Adds a welcome message to the interface
• Adds custom linter to enforce config version import chain

Improvements

• Refactors RAG from agent-level config to standard toolset type for consistency with other toolsets
• Restores RAG indexing event forwarding to TUI after toolset refactor
• Simplifies RAG event forwarding and cleans up RAGTool

Bug Fixes

• Fixes Bedrock interleaved_thinking defaults to true and adds logging for provider_opts mismatches
• Fixes issue where CacheControl markers were preserved during message compaction, exceeding Anthropic's limit
• Fixes tool loop detector by resetting it after degenerate loop error
• Fixes desktop proxy socket name on WSL where http-proxy socket is not allowed for users

Technical Changes

• Documents max_old_tool_call_tokens and max_consecutive_tool_calls in agent config reference
• Documents global permissions from user config in permissions reference and guides
• Pins GitHub actions for improved security
• Updates cagent-action to latest version with better permissions

---

What's Changed

• docs: update CHANGELOG.md for v1.36.1 by @docker-read-write[bot] in #2227
• Add custom linter to enforce config version import chain by @dgageot in #2225
• feat: forward sampling provider_opts (top_k, repetition_penalty) to provider APIs by @dgageot in #2226
• Refactor RAG from agent-level config to standard toolset type by @dgageot in #2210
• docs: add max_old_tool_call_tokens and max_consecutive_tool_calls to agent config reference by @dgageot in #2229
• Pin GitHub actions by @dgageot in #2231
• update cagent-action to latest (with better permissions) by @derekmisler in #2233
• Bump direct Go module dependencies by @dgageot in #2238
• fix: strip CacheControl from messages during compaction by @dgageot in #2236
• Reset tool loop detector after degenerate loop error by @dgageot in #2237
• Add global-level permissions from user config by @trungutt in #2230
• docs: document global permissions from user config by @trungutt in #2241
• Fix desktop proxy socket name on WSL by @gtardif in #2240

Full Changelog: v1.36.1...v1.37.0