v1.34.0

This release improves tool call handling and evaluation functionality with several technical fixes and optimizations.

Improvements

• Optimizes partial tool call streaming by sending only delta arguments instead of accumulated arguments
• Reduces evaluation summary display width for better terminal formatting
• Includes tool definition only on the first partial tool call to reduce redundancy

Bug Fixes

• Fixes schema conversion for OpenAI Responses API strict mode, resolving issues with gpt-4.1-nano
• Removes duplicate tool call data from tool call response events to reduce payload size

Technical Changes

• Updates evaluation system to not provide all API keys when using models gateway
• Removes redundant tool call information from response events while preserving tool call IDs for client reference

---

What's Changed

• docs: update CHANGELOG.md for v1.33.0 by @docker-read-write[bot] in #2159
• Fix (reduce) evals summary width by @gtardif in #2160
• Evals: don't provide all API keys when using models gateway by @gtardif in #2162
• Only send the delta on the partial tool call by @rumpl in #2105
• build(deps): bump google.golang.org/grpc from 1.79.2 to 1.79.3 in the go_modules group across 1 directory by @dependabot[bot] in #2164
• Remove the tool call from the tool call response event by @rumpl in #2163
• Fix schema conversion for OpenAI Responses API strict mode - Fixes tool calls with gpt-4.1-nano by @gtardif in #2168

Full Changelog: v1.33.0...v1.34.0

v1.33.0

This release improves file editing reliability, adds session exit keywords, and fixes several issues with sub-sessions and evaluation handling.

What's New

• Adds support for "exit", "quit", and ":q" keywords to quit sessions immediately
• Adds per-eval Docker image override via evals.image property in evaluation configurations
• Adds run instructions to creator agent prompt for proper agent execution guidance

Bug Fixes

• Fixes handling of double-serialized edits argument in edit_file tool when LLMs send JSON strings instead of arrays
• Fixes sub-session thinking state being incorrectly derived from parent session instead of child agent
• Fixes --sandbox flag when running in CLI plugin mode
• Fixes cross-model Gemini function calls by using dummy thought_signature
• Fixes event timestamps for user messages in SessionFromEvents to prevent duration calculation issues

Improvements

• Displays breakdown of failure types in evaluation summary for better debugging
• Declines elicitations in run --exec --json mode
• Validates path field consistently in edit file operations

Technical Changes

• Removes unused fileWriteTracker from creator package
• Simplifies UnmarshalJSON implementation for better path validation
• Updates evaluation image build cache to handle different images per working directory

---

What's Changed

• docs: update CHANGELOG.md for v1.32.5 by @docker-read-write[bot] in #2147
• Better rendering in tmux and ghostty by @dgageot in #2146
• Fix --sandbox when running cli plugin mode by @gtardif in #2151
• Display breakdown of types of failures in eval summary by @gtardif in #2150
• feat: support "exit" as a keyword to quit the session by @trungutt in #2152
• Add per-eval Docker image override via evals.image property by @dgageot in #2153
• Add run instructions to creator agent prompt by @dgageot in #2154
• Decline elicitations in run --exec --json mode by @dgageot in #2156
• Remove unused fileWriteTracker from creator package by @dgageot in #2157
• fix: use dummy thought_signature for cross-model Gemini function calls by @dgageot in #2155
• fix: sub-session thinking state derived from child agent, not parent session by @dgageot in #2149
• fix: handle double-serialized edits argument in edit_file tool by @trungutt in #2144
• fix: use event timestamps for user messages in SessionFromEvents by @dgageot in #2158

Full Changelog: v1.32.5...v1.33.0

v1.32.5

This release improves agent reliability and performance with better tool loop detection, enhanced MCP handling, and various bug fixes.

What's New

• Adds framework-level tool loop detection to prevent degenerate agent loops when the same tool is called repeatedly
• Adds support for dynamic command expansion in skills using !\command`` syntax
• Adds support for running skills as isolated sub-agents via context: fork frontmatter
• Adds CLI flags (--hook-pre-tool-use, --hook-post-tool-use, etc.) to override agent hooks from command line
• Adds stop and notification hooks with session lifecycle integration

Improvements

• Reworks thinking budget system to be opt-in by default with adaptive thinking and effort levels
• Caches syntax highlighting results for code blocks to improve markdown rendering performance
• Optimizes MCP catalog loading with single fetch per run and ETag caching
• Derives meaningful names for external sub-agents instead of using generic 'root' name
• Optimizes filesystem tool performance by avoiding duplicate string allocations
• Speeds up history loading with ReadFile and strconv.Unquote optimizations

Bug Fixes

• Fixes context cancelling during RAG initialization and query operations
• Fixes frozen spinner during MCP tool loading
• Fixes model name display in TUI sidebar for all model types
• Fixes two data races in shell tool execution
• Fixes character handling issues in tmux integration
• Fixes binary download URLs in documentation to match release artifact naming
• Validates thinking_budget effort levels at parse time and rejects unknown values

Technical Changes

• Removes unused methods from codebase
• Hardens and simplifies MCP gateway code
• Adds logging for selected model in Agent.Model() for better observability
• Fixes pool_size reporting to reflect actual selection pool
• Reverts timeout changes for remote MCP initialization and tool calls

---

What's Changed

• Fix rag init context cancel by @gtardif in #2114
• docs: update CHANGELOG.md for v1.32.4 by @docker-read-write[bot] in #2112
• Bump dependencies by @dgageot in #2113
• Fix frozen spinner during MCP tool loading by @dgageot in #2115
• Support dynamic command expansion in skills (!command syntax) by @dgageot in #2116
• Fix model name display in TUI sidebar for all model types by @dgageot in #2118
• perf(markdown): cache syntax highlighting results for code blocks by @dgageot in #2119
• feat: framework-level tool loop detection by @derekmisler in #2123
• Fix issues on builtin filesystem tools by @dgageot in #2125
• Fix two data races in shell tool by @dgageot in #2127
• Fix a few characters for tmux by @dgageot in #2128
• Simplify MCP catalog loading: single fetch per run with ETag caching by @dgageot in #2124
• docs: fix binary download URLs to match release artifact naming by @dgageot in #2129
• More doc fixing with "agent serve mcp" by @gtardif in #2130
• Rework thinking budget: opt-in by default, adaptive thinking, effort levels by @dgageot in #2121
• Add timeouts to remote MCP initialization and tool calls by @dgageot in #2131
• Optimize start time by @dgageot in #2138
• gateway: harden and simplify MCP gateway code by @dgageot in #2133
• Derive meaningful names for external sub-agents instead of using 'root' by @dgageot in #2132
• Add --hook-* CLI flags to override agent hooks from the command line by @dgageot in #2135
• feat: support running skills as isolated sub-agents via context: fork by @dgageot in #2137
• Revert "Add timeouts to remote MCP initialization and tool calls" by @dgageot in #2141
• Add stop and notification hooks, wire up session lifecycle hooks by @dgageot in #2136
• Reject unknown thinking_budget effort levels at parse time by @dgageot in #2142
• Log selected model in Agent.Model() for alloy observability by @derekmisler in #2134

Full Changelog: v1.32.4...v1.32.5

v1.32.4

This release optimizes tool instructions, removes unused session metadata, and includes several bug fixes and improvements.

Improvements

• Optimizes builtin tool instructions for conciseness by applying Claude 4 prompt engineering best practices
• Removes unused branch metadata and split_diff_view from sessions to clean up data storage

Bug Fixes

• Fixes emoji rendering issues in iTerm2
• Reverts keyboard enhancement changes that caused incorrect behavior in VSCode with AZERTY layout

Technical Changes

• Extracts compaction logic into dedicated pkg/compaction package for better code organization
• Updates skill configuration
• Improves evaluation system by validating LLM judge, disabling thinking for LLM as judge, and removing handoffs scoring
• Disallows unknown fields in configuration validation

---

What's Changed

• Bump dependencies by @dgageot in #2094
• Optimize builtin tool instructions for conciseness by @dgageot in #2091
• docs: update CHANGELOG.md for v1.32.3 by @docker-read-write[bot] in #2097
• Remove unused branch metadata and split_diff_view from sessions by @rumpl in #2078
• Revert "tui: improve tmux experience and simplify keyboard enhancements" by @gtardif in #2098
• Fix 2089 - emoji rendering in iTerm2 by @dgageot in #2099
• Extract compaction into a dedicated pkg/compaction package by @dgageot in #2101
• Improve evals by @dgageot in #2100

Full Changelog: v1.32.3...v1.32.4

v1.32.3

This release removes an experimental feature and improves error handling for rate-limited API requests.

Improvements

• Makes HTTP 429 (Too Many Requests) errors retryable when no fallback model is available, respecting the Retry-After header

Bug Fixes

• Gates 429 retry behavior behind WithRetryOnRateLimit() opt-in option to prevent unexpected retry behavior

Technical Changes

• Removes experimental feature from the codebase
• Adds optional gateway usage for LLM evaluation as a judge
• Refactors to use typed StatusError for retry metadata, with providers wrapping errors at Recv()

---

What's Changed

• Remove experimental feature by @dgageot in #2087
• docs: update CHANGELOG.md for v1.32.2 by @docker-read-write[bot] in #2090
• This can be retried by @dgageot in #2093
• [eval] Optionnally use the gateway for the llm as a judge by @dgageot in #2092
• fix: make HTTP 429 retryable when no fallback model, respect Retry-After header by @simon-agent-go-expert in #2096

New Contributors

• @simon-agent-go-expert made their first contribution in #2096

Full Changelog: v1.32.2...v1.32.3

v1.32.2

This release focuses on security improvements and bug fixes, including prevention of PATH hijacking vulnerabilities and fixes to environment file support.

Bug Fixes

• Fixes prevention of PATH hijacking and TOCTOU (Time-of-Check-Time-of-Use) vulnerabilities in shell/binary resolution (CWE-426)
• Fixes --env-file support for the gateway

Technical Changes

• Removes debug code from codebase
• Reverts user prompt options feature that was previously added

---

What's Changed

• docs: update CHANGELOG.md for v1.32.1 by @docker-read-write[bot] in #2084
• fix: prevent PATH hijacking and TOCTOU in shell/binary resolution by @dgageot in #2083
• Remove debug code by @dgageot in #2086
• Fix --env-file support for the gateway by @dgageot in #2085
• Revert "Add options-based selection to user_prompt tool" by @trungutt in #2088

Full Changelog: v1.32.1...v1.32.2

v1.32.1

This release fixes several issues with session handling, tool elicitation, and MCP environment variable validation.

Bug Fixes

• Fixes corrupted session history by filtering sub-agent streaming events from parent session persistence
• Fixes elicitation requests failing in sessions with ToolsApproved=true by decoupling elicitation channel from ToolsApproved flag
• Fixes MCP environment variable validation being skipped when any gateway preflight errors occur

Improvements

• Prevents sidebar from scrolling to top when clicking navigation links in documentation

Technical Changes

• Adds end-to-end test for tool result block validation
• Updates CHANGELOG.md for v1.32.0 release

---

What's Changed

• docs: update CHANGELOG.md for v1.32.0 by @docker-read-write[bot] in #2072
• Don't scroll sidebar to the top by @dgageot in #2076
• Fix corrupted session history by @dgageot in #2077
• Fix MCP env var check skipped when any gateway preflight errors by @dgageot in #2081
• fix: decouple elicitation channel from ToolsApproved flag by @dgageot in #2080

Full Changelog: v1.32.0...v1.32.1

v1.32.0

This release adds support for newer Gemini models, improves toolset documentation, and enhances user interaction capabilities.

What's New

• Adds options-based selection to user_prompt tool, allowing the agent to present users with labeled choices instead of free-form input
• Documents {ORIGINAL_INSTRUCTIONS} placeholder for enriching toolset instructions rather than replacing them

Bug Fixes

• Fixes support for Gemini 3.x versioned models (e.g., gemini-3.1-pro-preview) to ensure proper model recognition and thinking configuration
• Fixes gateway handling when using docker agent without a command
• Fixes broken links in documentation

Technical Changes

• Adds check for broken links in CI
• Updates .gitignore to exclude cagent-* binaries from being committed

---

What's Changed

• docs: update CHANGELOG.md for v1.31.0 by @docker-read-write[bot] in #2063
• doc: document {ORIGINAL_INSTRUCTIONS} placeholder for toolset instructions by @dgageot in #2062
• fix: support Gemini 3.x versioned models (e.g., gemini-3.1-pro-preview) by @dgageot in #2054
• Fix gateway handling with docker agent without command by @dgageot in #2064
• Fix broken links by @dgageot in #2067
• gitignore cagent-* binaries by @derekmisler in #2069
• Check for broken links by @dgageot in #2068
• Add options-based selection to user_prompt tool by @trungutt in #2071

Full Changelog: v1.31.0...v1.32.0

v1.31.0

This release enhances the cost dialog with detailed session statistics and improves todo tool reliability for better task completion tracking.

What's New

• Adds total token count, session duration, and message count to cost dialog
• Adds reasoning tokens display for supported models (e.g. o1)
• Adds average cost per 1K tokens and per message metrics to cost analysis
• Adds cost percentage breakdown per model and per message
• Adds cache hit rate and per-entry cached token count display

Improvements

• Improves todo tool reliability by reminding LLM of incomplete items and including full state in all responses

Bug Fixes

• Fixes Sonnet model name
• Fixes various edge-case bugs in cost dialog formatting

Technical Changes

• Adds cache to building hub image in CI
• Optimizes CI by building and testing Go on the same runner to avoid duplicate compilation
• Freezes config to v6
• Deduplicates tool documentation into individual pages
• Adds docs-serve task for local Jekyll preview via Docker

---

What's Changed

• docs: update CHANGELOG.md for v1.30.1 by @docker-read-write[bot] in #2050
• Add cache to building hub image in CI by @gtardif in #2037
• cost dialog: enrich with session stats, per-model percentages, and formatting fixes by @dgageot in #2046
• Fix sonnet model name by @dgageot in #2052
• fix: improve todo completion reliability by @trungutt in #2048
• Freeze config v6 by @dgageot in #2059
• Improve the toolsets documentation by @dgageot in #2056

Full Changelog: v1.30.1...v1.31.0

v1.30.1

This release improves command history handling, adds sound notifications, and includes various bug fixes and performance optimizations.

What's New

• Adds sound notifications for long-running tasks and errors (opt-in feature, disabled by default)
• Adds LSP multiplexer to support multiple LSP toolsets simultaneously
• Adds per-toolset model routing via model field on toolsets configuration
• Adds click-to-copy functionality for working directory in TUI sidebar
• Makes background_agents a standalone toolset that can be enabled independently

Improvements

• Improves tmux experience with better keyboard enhancements and focus handling
• Optimizes BM25 scoring strategy for better performance
• Reduces redundant work during evaluation runs
• Fixes animated spinners inside terminal multiplexers
• Repaints terminal on focus to fix broken display after tab switch in Docker Desktop

Bug Fixes

• Fixes loading very long lines in command history that previously caused crashes
• Fixes LSP server being killed by context cancellation and restart failures
• Fixes session-pinned agent usage in RunStream instead of shared currentAgent
• Fixes sidebar context percentage flickering during sub-agent transfers
• Fixes concurrent map writes by moving registerDefaultTools to constructor
• Returns clear error when OPENAI_API_KEY is missing for speech-to-text

Technical Changes

• Splits monolithic runtime.go into focused files by concern
• Refactors code to use slices and maps stdlib functions instead of manual implementations
• Enables modernize and perfsprint linters with all findings resolved
• Migrates tool output to structured JSON schemas for todo tools
• Replaces json.MarshalIndent with json.Marshal in builtin tools
• Uses errors.AsType consistently instead of errors.As with pre-declared variables

---

What's Changed

• Don't ignore GITHUB_TOKEN by @dgageot in #2002
• history: Fix loading very long lines by @vvoland in #1940
• docs: update CHANGELOG.md for v1.30.0 by @docker-read-write[bot] in #2003
• Fix broken links to pages subsections by @gtardif in #2005
• feat: add sound notifications for task completion and errors by @Mostamhd in #1870
• Add LSP multiplexer to support multiple LSP toolsets by @dgageot in #1970
• fix(#2012): Return clear error when OPENAI_API_KEY is missing for speech-to-text by @aheritier in #2013
• Bump direct Go dependencies by @dgageot in #2011
• Fix LSP server killed by context cancellation and restart failures by @dgageot in #2008
• Replace duplicated mockEnvProvider test types with shared environment providers by @dgageot in #2014
• fix: use session-pinned agent in RunStream instead of shared currentAgent by @dgageot in #2009
• docs: fix hallucinated CLI flags, commands, and config formats by @dgageot in #2020
• refactor: use slices and maps stdlib functions instead of manual implementations by @dgageot in #2021
• refactor(anthropic): deduplicate sequencing, media-type, and test helpers by @dgageot in #2019
• refactor: split runtime.go and extract pkg/modelerrors by @dgageot in #2010
• tui: improve tmux experience and simplify keyboard enhancements by @dgageot in #2017
• codemode: fix Start() fail-fast and use tools.As for wrapper unwrapping by @dgageot in #2007
• Simplify rulebased router: remove redundant types and score aggregation by @dgageot in #2016
• Unify streamAdapter/betaStreamAdapter retry logic into generic retryableStream by @dgageot in #2018
• Fix task deploy-local by @dgageot in #2024
• fix: default sound notifications to off (opt-in) by @dgageot in #2025
• feat: add per-toolset model routing via model field on toolsets by @dgageot in #2015
• refactor: use errors.AsType consistently instead of errors.As with pre-declared variables by @dgageot in #2028
• tui: repaint terminal on focus to fix broken display after tab switch by @dgageot in #2026
• Enable modernize and perfsprint linters, fix all findings by @dgageot in #2027
• update Slack link in readme by @derekmisler in #2032
• Replace json.MarshalIndent with json.Marshal in builtin tools by @dgageot in #2031
• refactor(dmr): split client.go into focused files by concern by @dgageot in #2029
• Fix last brew install cagent mention by @gtardif in #2034
• tui: fix animated spinners inside terminal multiplexers by @dgageot in #2035
• refactor(runtime): split monolithic runtime.go into focused files by @dgageot in #2030
• feat: click to copy working directory in TUI sidebar by @dgageot in #2036
• Use ResultSuccess/ResultError helpers in tasks and user_prompt tools by @dgageot in #2040
• fix: move registerDefaultTools to constructor to prevent concurrent map writes by @dgageot in #2041
• Fix sidebar context % flickering during sub-agent transfers by @dgageot in #2042
• eval: reduce redundant work during evaluation runs by @dgageot in #2047
• perf: optimize BM25 scoring strategy by @dgageot in #2043
• refactor: remove duplication in model resolution, thinking budget, and message construction by @dgageot in #2038
• feat: make background_agents a standalone toolset by @dgageot in #2033
• todo: migrate tool output to structured JSON schemas by @dgageot in #2045

Full Changelog: v1.30.0...v1.30.1