Dahawa Life Plan

An intelligent programming assistant who understands you
An intelligent assistant who understands you
An assistant who understands you

One is the bottom line reserved for yourself

Versions

v2.8.0 ~ v2.8.2 (2026-11-03 ~ 2026-11-05)
- Added agentMessageService, a unified message delivery system: unified ‘main’/‘agent:{agentId}’ addressing with Force/Try/Serial delivery modes plus waitMessages/peekMessages; fixed the loop scheduled task silently dropping prompts on busy
- Removed the deprecated TaskTool and TaskTeamTool; subagent task scheduling is now unified under AgentTool, which supports filtering the tool list by agentConfig.tools
- Web client enhancements
  - Subagents now support prompt input; fixed mailbox-injected messages (including main busy queued messages) being received by the model but not displayed on the page
  - Added pending prompt list (pending_prompts) and Task/TodoWrite task list (tasks_snapshot) displays above the input box, aligned with CLI ordering and styling
- Added a channel notification handling system: receives MCP server→client notifications, writes date-partitioned log files, and broadcasts dispatch via eventBus (shared routing between the main system and the codemode system)
- Fixed stdio subprocess leaks caused by MCP cache invalidation/reconnection
- Fixed PersistentShell external abortSignal being silently overridden by the internal signal; the timeout branch now correctly reports interrupted and ensures the process tree is cleaned up
- MCP main path now supports ${} dynamic variable interpolation (${env.NAME}, ${home}, ${cwd}, ${projectName}, etc.) with the same syntax as .codemode.json; interpolation happens in memory only and never writes back to config
v2.7.19 ~ v2.7.20 (2026-11-01 ~ 2026-11-02)
- Fixed the second tool confirmation getting stuck when confirmations arrive concurrently: the close path now guards with a timer, ignores duplicate closes, and avoids clearing the timer that displays the next confirmation
- Message Archiving placeholders switched to a bracket short-marker format (<system>[archived xxx: id xxx]</system>) to prevent the model from imitating them into files
- Global prompt prohibition rules now explicitly cover archiving placeholders: declared alongside markers as system-generated tokens, with the prohibition extended to files you write
v2.7.14 ~ v2.7.18 (2026-10-28 ~ 2026-10-29)
- ProxyToolLoad added an XML parameter description format: proxyToolSchemaFormat supports json/ts/xml; xml mode outputs a pure XML document with special character escaping
- Refactored ProxyToolLoad schema format building: extracted buildJsonMessage/buildXmlMessage/buildTsMessage into standalone functions
- MessageArchivingTool prompt slimmed down and translated into English
- Simplified TodoWriteTool prompts, removed forced task and todo reminders, and deleted the TodoWrite mandatory usage reminder from the system prompt
v2.7.10 ~ v2.7.13 (2026-10-24 ~ 2026-10-27)
- ProxyToolLoad supports switching parameter description formats via config: new settings proxyToolSchemaFormat (‘ts’|’json’) with a corresponding Config panel enum setting
- Config panel now supports Tab grouping: new General/Context tabs, switched with arrow keys
- codemode process lifecycle fixes: ESC cancel now takes effect through the mainLoop abortController; timeout/cancel/error paths converge on a unified terminate(reason) exit, eliminating subprocesses that ignore SIGTERM; describe removed the endsWith fallback in favor of exact matching, eliminating phantom method names
- Message Archiving enhancements
  - Addressing markers wrapped in <system> tags to prevent the model from repeating internal addressing tokens in user-facing replies
  - Added thinking block archiving (thinking / redacted_thinking); markers are injected at the end of the user message following the assistant (after tool_result, satisfying the deepseek tool_result-first compatibility constraint)
  - Marker insertion moved before the first reminder block, fixing the latent defect of reminders being included in the cache segment
  - Added __tool_input__ automatic input archiving: leaf string fields with more than 100 words (nested objects and array indices supported) are automatically archived; unarchive restores all archived input fields
  - Archiving result rendering slimmed: counts by tool result/input field/thinking block category, with skipped reasons deduplicated and merged
- Adjusted low-context-usage reminder wording to consistently guide toward archiving
- buildContextUsageMessage cooldown now uniformly returns only a plain usage sentence, always injecting the usage signal without inducing archiving
v2.7.6 ~ v2.7.9 (2026-10-17 ~ 2026-10-22)
- WebFetch added headless browser rendering for JS pages: playwright-core as an optional dependency, SPA shell detection, and dual settle detection (DOM silence + in-flight requests); new headlessBrowser switch in /config
- Bing Headless Search Provider: models without native search capability (OpenAI/DeepSeek and default) now search via www.bing.com with a three-tier fallback chain (headless rendering + in-page DOM extraction → static parsing → exception fallback)
- Table rendering upgrades
  - Layout algorithm migrated to the hawa-table component with water-filling column width allocation and display-width-aware wrapping, fixing overflow/misalignment of long table content
  - Fixed emoji display width misjudgment (miscellaneous technical symbols counted as width 2) and misalignment caused by counting variation selectors and zero-width joiners as width
  - Table headers now centered by default; added renderMarkdownTable to parse markdown pipe tables
- string-cli-width added truncateByWidth to truncate strings by terminal display width (correctly handling multi-byte characters and emoji); pending hints now uniformly use it
- Fixed npx tsx –test hanging forever: lspManager/fileFreshness switched to lazy loading, timers unref’d
- Fixed autoCompact summary extraction failure when thinking is enabled
- Context usage reminder added a high-water mark of 0.85: dynamic usage reminders are abandoned beyond it, leaving fallback to auto-compaction
- Archiving guidance now enforces a step interval: at least 5 requests between two archiving actions by default (adjustable in real time via HAWA_CODE_ARCHIVE_STEP_INTERVAL)
- Context optimization reminders split into a static guide and a dynamic usage line (gated by HAWA_CODE_CONTEXT_OPT_SPLIT)
- Simplified codemode.session API: kept only get/set, with value null as the delete sentinel
- Fixed CLI terminal forcibly breaking long words in thinking/text messages
- Moved LSP usage reminders from system reminders into tool descriptions and parameter descriptions
- Deduplicated archiving system reminder wording: static guide handles full explanations, dynamic usage keeps only the signal and minimal action suggestion
v2.7.0 ~ v2.7.5 (2026-10-13 ~ 2026-10-16)
- codemode Session v1: added codemode.session cross-process session storage capability
- codemode config merging extracted into mergeCodemodeConfigs, uniformly applying dynamic parameter interpolation to merged global and project-level .codemode.json results
- Added showWindowTokens setting: status bar now shows current window token count and context ratio
- getGitPrompt is now controlled by the HAWA_CODE_GIT_PROMPT environment variable (disabled by default)
- ProxyTool prompt optimization: loadOne responses add a Required arguments section and remove the Usage example section
- Added screenService as a unified clear-screen entry point, consolidating all REPL clear-screen call sites; resize now safely re-mounts Static to rewrite history at the new width
- Fixed tool confirmation dialog race conditions: preventing closed dialogs from resurrecting and zombie dialogs
- Fixed schemaToType index signature generation to use the actual additionalProperties type
v2.6.12 ~ v2.6.13 (2026-10-11)
- Removed the ps-tree dependency; added a self-developed processTree utility to unify process tree termination and liveness checks
- The update/config/mcp/plugin/approved-tools/doctor subcommands and –version/–help now skip eager loading of MCP tools
- Adjusted tool registration order so ProxyTool-related tools are registered first
- Simplified tsconfig paths wildcard fallbacks; bare path imports changed to relative imports
- Fixed react/jsx-runtime mapping pointing to .d.ts, which caused missing jsx exports in tsx runtime
- Removed the example prompt from the Agent tool
v2.6.10 ~ v2.6.11 (2026-10-09)
- Optimized startup time
- Codemode now excludes providers that failed to connect
- Unified codemode meta function signatures to object parameters
- Fixed PromptInput issues when entering large text
- Fixed the describe parameter in prompts
v2.6.2 ~ v2.6.9 (2026-10-06 ~ 2026-10-08)
- Startup optimizations
  - Moved the startup loading hint before parseArgs’s getCommands(), covering normal startup and –resume paths to avoid long periods without feedback
  - Added connection timeout protection for MCP HTTP/sse/streamable-http branches; the configured timeout now also applies to HTTP servers
- codemode enhancements
  - Unified tool name sanitization: prefixed tool names are converted to valid identifiers; only sanitized names are exposed to the sandbox and prompt, while original names are used when calling the MCP server
  - Unified snippet description syntax in describe to snippet., connecting the search → describe flow
  - Added memoize caching to getCodemodeTools to avoid repeated listTools calls on every execution
  - Improved Chinese search retrieval
  - Fixed inconsistency between codemode.search meta function outputSchema and its implementation
  - Added a CodeMode snapshot saving reminder to systemReminder
v2.6.0 ~ v2.6.1 (2026-10-03)
- Default configuration changes
  - Tool exposure mode lazyTool now defaults to proxy
  - Real-time memory extraction autoMemory.enabled and memory consolidation autoDream.enabled now default to false
- Removed the ThinkTool tool (including web-side render components and related references)
- Fixed a model configuration bug and added source name renaming
- ProxyToolCall self-invocation is now forbidden; the error message is unified to toolName cannot be ProxyToolCall
- codemode MCP descriptions are truncated by word count (English by words, Chinese by characters), with an ellipsis appended
- proxytoolload changed to single loading
- Unified codemode naming to camelCase
- Fixed a batch of compilation errors
v2.5.1 ~ v2.5.2 (2026-09-29 ~ 2026-09-30)
- Optimized ProxyTool prompts, emphasizing that tools must be loaded before calling
- proxy-tools now supports a maximum tool count; provider and methods layout optimized
- The maximum number of methods shown per provider in codemode is now configurable via /config (codemodeMethodsMax, 1-100, default 20)
- Markdown table improvements: separator lines added between data rows, green header highlight
- Simplified codemode usage instructions
- Removed unnecessary bun dependencies
v2.5.0 (2026-09-28)
- Split ProxyTool into two tools: ProxyToolLoad and ProxyToolCall
- Optimized useWindowSize non-TTY fallback size caching to avoid duplicate renders from invalid resize events
- Removed duplicate git command instructions from the PowerShellTool prompt
- Fixed a batch of compilation errors (agent.ts, apiTransformer.ts, NormalizedMessage, react key, etc.)
v2.4.52 ~ v2.4.55 (2026-09-26 ~ 2026-09-27)
- Optimized ProxyTool prompts and JSON schema descriptions, clarifying the usage rules and parameter requirements for load and call
- Added explanations of incorrect invocation patterns to prompt examples (duplicate control parameters inside arguments, flattened parameter calls)
- Optimized the codemode describe output schema hint, emphasizing that the output structure should be inferred and persisted only after the first real call
- Optimized PowerShell tool use display
v2.4.51 (2026-09-25)
- Built-in default Skill support
  - Added repository root skills/ directory, copied to dist/skills during build and shipped with the package
  - SkillService added builtin scan path (lowest priority; user/project skills with the same name can override), watcher skips builtin
- Changed snippets storage path to .hcode/.codemode-snippets.json
- Optimized codemode prompt (emphasizing code invocation) and automatic output schema saving
- Optimized ProxyTool prompts, clarifying the load-then-call invocation flow
- Refactored loop architecture
v2.4.49 ~ v2.4.50 (2026-09-24)
- Renamed ProxyTool call parameter input to arguments; TS input type in load results uniformly renamed to arguments type
- Changed proxytool reminders to English
- Streamlined BashTool prompt: removed git commit/PR operation guides, directory verification steps, and unused imports
- Fixed unwrapMcpResult failing to parse JSON after concatenating multiple text contents
v2.4.45 ~ v2.4.48 (2026-09-23)
- ProxyTool parameter rename: action → proxyToolAction, name → toolName, avoiding nested name collisions with the target tool’s own action/name parameters that caused the model to omit arguments
- codemode enhancements
  - Custom Instructions support: .codemode.json MCPServer added instructions property, taking priority over the MCP Server’s own configuration
  - renderToolResultMessage adds single-line truncation: lines over 200 characters are truncated with a … marker
  - output schemas storage split by provider’s codemode config location: user-level stored in ~/.hcode/codemode-output-schemas.json, project-level stored in {projectDir}/.hcode/codemode-output-schemas.json; merged on read with project-level overriding user-level
  - Show loading hint when loading MCP tools at startup
- Fixed flickering and black-screen issues caused by BashTool/PowerShellTool outputting full progress during execution; display side truncated to the last 5 lines
- PowerShellTool terminal result display now reuses BashTool’s BashToolResultMessage, consistent with BashTool styling
v2.4.40 ~ v2.4.42 (2026-09-22)
- Streamlined product features
  - Removed release-notes, login(OAuth), and Sentry features
  - Spinner animation changed to braille dot matrix single-direction rotation; removed unreferenced dead code files
- Fixed memory files still being displayed when the memory feature is disabled
v2.4.33 ~ v2.4.39 (2026-09-18)
- Simplified TodoWriteTool prompt: removed all Examples sections, keeping only When/When Not/Task States core descriptions
v2.4.30 ~ v2.4.32 (2026-09-08 ~ 2026-09-17)
- ProxyTool loading mechanism upgrade
  - action renamed from describe to load; load became the default action, strengthening the load-before-call guidance
  - Prompt tags unified from to , consistent with tool list structure naming
  - load action displayed as LoadTool(target tool name) on CLI and Web, distinguished from call
  - Type signatures in load results no longer wrapped in typescript fences, output as plain text consistent with CodeMode describe
  - system prompt injects loading hint when ProxyTool is enabled: must load before call to obtain tool signatures
- rules compatible with ProxyTool proxy mode
  - path rules unwrap real tool names and arguments before matching; load action does not participate in matching
  - Rule expansion tag changed from to , inserted before the trailing reminder block when merging
  - Fixed dynamic rules cache hit issue
  - log.ts isMessageRecord recognizes path_rules type, matchedPathRules can be restored after resume
- Optimized lazy loading hint priority
v2.4.26 ~ v2.4.29 (2026-09-07)
- Code Mode snippet feature
  - Added codemode.snippet.<name>(input), supporting execution of saved snippets by name; newly saved snippets are available immediately at runtime
  - saveSnippet added optional inputSchema/outputSchema parameters with object validation
  - describe no longer returns source code for snippets; instead generates concrete type signatures based on schema (falls back to unknown without schema)
- Fixed error prompts for nonexistent tools
  - Only prompts to load via ToolLoader or ProxyTool when the tool exists in the full _tools list
  - Returns No such tool available when a tool can neither be loaded nor proxied
v2.4.24 ~ v2.4.25 (2026-09-06)
- Added appwrite provider configuration in .codemode.json
- Proxy mode exposure layer changed from ProxyTool+CodeMode only to ProxyTool + proxyWhiteList whitelist tools
- MCP client supports timeout configuration in .codemode.json; default connection timeout for both MCP paths unified to 20s
v2.4.22 ~ v2.4.23 (2026-09-04)
- Replaced JSON Schema with TypeScript type signatures as the description and validation basis for tool parameters
- Fixed issue where subagents could not obtain ProxyTool in proxy mode
- Optimized ProxyTool proxy invocation rendering logic, adding a <> prefix identifier
- Added CodeModeToolResult rendering component
v2.4.14 ~ v2.4.21 (2026-09-04)
- ProxyTool proxy tool enhancements
  - Merged lazyTool/proxyToolEnable into a three-state enum LazyToolMode
  - Fixed issue where directly invoking real tools in proxy mode bypassed ProxyTool isolation
  - Parameter validation failure prompts guide users to query the correct input structure via ProxyTool describe
  - Optimized proxy tool prompt, strengthening the describe-first guidance
v2.4.10 ~ v2.4.13 (2026-09-02 ~ 2026-09-03)
- Supports setting maximum token return count
- Optimized tool lazy loading order to increase cache utilization
- Optimized prompt cache points, reduced to only one
- Fixed opus-4-8 model thinking configuration
- Optimized codemode display
v2.4.3 ~ v2.4.6 (2026-09-01)
- Optimized codemode prompt wording
- Tool lazy loading no longer displays method hints
v2.4.1 ~ v2.4.2 (2026-08-31)
- codemode no longer displays stack traces for MCP errors; ZodError shows complete error message
- Added support for minimum, maximum, minLength, and maxLength properties in JSON Schema, with corresponding comments added to generated TypeScript types
- Added detailed input parameter comment feature; when HAWA_CODE_JSDOC is set to true, JSDoc comments include descriptions for each input field
v2.4.0 (2026-08-30)
- Code Mode 0.1 support
  - Added codemode tool, supporting natural language invocation of MCP server capabilities
  - Optimized codemode tool confirmation message prompts
  - Added codemode.search method documentation
  - Codemode Prompt displays parameters
  - Optimized codemode Prompt to display MCP server descriptions
- Added LS tool component on Web
  - Supports directory list display and expansion
- Web subagent disables user input when status is ended
v2.3.5 (2026-08-23)
- Web tool calls support notifications
- Web UI style optimizations
  - Changed web Prompt prefix to vertical bar
  - Optimized web message and tool result styles
  - Simplified PromptInput status hints and centered input box
  - Settings page scrollbar unified to custom dark theme style
  - Optimized Settings and Agent Viewer switching logic
v2.3.4 (2026-08-22)
- Web UI comprehensive upgrade
  - Web optimization 0.1
  - Optimized web layout
  - Frontend adopts dark theme
  - Fixed header style height issue
  - Changed web logo
  - Optimized frontend buttons
v2.3.3 (2026-08-22)
- Upgraded web UI display
v2.3.3 (2026-08-20)
- Anthropic API thinking block merge fix
  - Support thinking and redacted_thinking content blocks in mergeAnthropicChunks
  - Accumulate thinking_delta and signature_delta, preserving signature for multi-round passthrough
  - Fix related TypeScript type errors
  - Add api-anthropic merge test cases
v2.3.2 (2026-08-16)
- Web Prompt input optimization
  - Prevent IME Enter from sending messages directly
  - Expand input height
  - Center send button
- Fixed TextInput custom chunked paste detection
v2.3.1 (2026-08-10)
- Plugin marketplace list hides source; detail page shows source (except default)
v2.3.0 (2026-08-08)
- Split Plugin marketplace lifecycle into add/install/uninstall/remove
- Skill defaults user-invocable to false
- Fixed useTextInput Shift+Enter being reported as meta+return
v2.2.3 (2026-08-06)
- Translated system reminder wording to English
v2.2.2 (2026-08-05)
- MessageArchiving prompt simplification
  - Removed tool_use_id lookup guidance from the prompt; handled automatically by the tool
v2.2.1 (2026-08-04)
- Web Settings plugin management
  - Added plugin and marketplace management to the Web settings panel
v2.2.0 (2026-08-03)
- MessageArchiving enhancements
  - Skip archiving short tool_use input fields (≤100 characters) to avoid meaningless archiving
  - Pass historyMessages to tool calls for archival eligibility checks based on conversation history
  - Added settings.archiveToolUseInputs setting to control input field archiving
v2.1.5 (2026-07-30)
- WebServer model config validation enhancement
  - Return detailed field-level errors for invalid model configurations
v2.1.4 (2026-07-29)
- Plugin compatibility configuration
  - Changed Claude Code plugin compatibility to the settings.plugins.pluginClaudeCompat setting, enabled by default
v2.1.2 / v2.1.3 (2026-07-28)
- MCP HTTP server enhancement
  - Supports streamable-http as an MCP HTTP server type
v2.1.1 (2026-07-27 ~ 2026-07-28)
- Plugin plugin system enhancement
  - PluginManager marketplace detail page supports Update and Remove operations
- Terminal compatibility fix
  - Fixed inconsistent line endings on Windows, unified to \n
v2.1.0 (2026-07-23 ~ 2026-07-25)
- Plugin plugin system
  - Enabled plugin compatibility
  - Supported plugin 0.1, added PluginManager management interface
  - Added core modules such as MarketplaceService, PluginInstallService, PluginRegistry, PluginManifestService, PluginSecurity, PluginUpdateService
  - Supports fetching plugins from multiple sources including GitHub, Git, NPM, and local directories
  - Supports plugin MCP services; plugins can register custom MCP Servers
  - Added plugin-related tests (install, marketplaceValidator, schemas, PluginRegistry MCP, etc.)
- Fixed conflict between auto mode thinking and tool_choice
v2.0.3 (2026-07-21)
- History storage optimization
  - History records now stored in history.jsonl
- Settings model configuration optimization
  - Fixed issue where comma input in the models field of model configuration was blocked
v2.0.2 (2026-07-17)
- Added /cd slash command
  - Supports quickly switching working directories
  - Supports generic argument hints (arg-hint)
- Configuration initialization refactoring
  - Unified configuration initialization logic into modelManager
v2.0.1 (2026-07-11 ~ 2026-07-16)
- Terminal framework upgraded to ink 7
  - Replaced custom paste detection with ink 7 usePaste
  - Fixed Esc key routing, terminal clear timeout, and functional state update issues
  - REPL terminal resize debouncing to reduce flicker and clear races
  - Slash command arrow/enter navigation restored
  - Spinner and periodic UI refresh migrated to ink 7 useAnimation
  - Unified window resize handling with shared useWindowSize hook
  - Removed unused sticker form and animated asterisk components
- Agent execution enhancement
  - Supports configuring agent depth limits
  - Waits for background subagent tasks to complete before continuing
- Agent Viewer enhancement
  - In-progress tool_use displays a spinner
v2.0.0 (2026-07-10)
- Settings form layout optimization
  - Replaced flex space-between with a grid layout using a fixed label column
  - Improved alignment of controls in General, Web Server, Auto Mode, Memory, and other settings to avoid excessive gaps
v1.42.3 (2026-07-09)
- Agent Viewer enhancement
  - Supports prompt input
  - Supports interrupting main agent requests
- Unified interactive tool rendering
  - Added toolUserResponseService, supporting unified rendering of interactive tools in terminal and Web
- Settings enhancement
  - Added planDirectory setting, supporting custom plan file storage directory
v1.42.2 (2026-07-04)
- Fixed compressed thinking failure, ensuring message content is correctly handled during compression to avoid errors caused by unexpected content formats
- Removed unused src/messages.ts and related references
v1.42.1 (2026-07-03)
- Fixed clearConversation function not triggering REPL re-render when clearing messages
v1.42.0 (2026-07-03)
- Web tool confirmation 2.0
  - Added ConfirmationService, decoupling tool permission confirmation flow from the UI layer
  - Unified tool confirmation support for CLI local UI and Web clients
  - Added ToolConfirmQueue component on Web, with modal-style tool call confirmation
  - WebSocket broadcasts tool confirmation requests and completion notifications
  - Other clients’ confirmation windows automatically close after one client responds
  - Pending confirmation requests can be restored after reconnecting or refreshing the page
- Settings migration
  - Memo and Message Archiving toggles migrated from environment variables to settings
  - Added memo and messageArchiving settings, disabled by default
  - Config panel supports enabling/disabling Memo and Message Archiving
  - Settings file hot-reload listener changed to unref to avoid blocking process exit
v1.41.4 (2026-07-02)
- Version bump
v1.41.3 (2026-07-02)
- Optimized context usage reminder
  - Fixed MessageArchiving and Memo tool loaded-state judgment logic
  - Prompts when tools are not loaded are more specific and actionable
  - Added reminder: even if context usage is low, historical messages with a lot of task-irrelevant content should be handled
- Web Settings panel optimization
  - Adjusted JSON editor height
  - Improved Model Config JSON editor styling
v1.41.2 (2026-07-02)
- Web model configuration Settings panel
  - Added /api/model-config API, supporting reading and updating config.json
  - Added ModelConfigSetting component, supporting form mode and JSON edit mode
  - Supports sources and routing configuration
  - Added model config validation logic and related tests
v1.41.1 (2026-07-01)
- Optimized Agent Viewer and supported Memo write detail display
  - AgentToolUse added prompt detail expand/collapse, showing model, type, resume, background execution, and other metadata
  - AgentToolResult long results support collapsing after expansion
  - MemoToolUse write operation added details button to view full written content
v1.41.0 (2026-07-01)
- Web Settings configuration panel
  - Added /api/settings REST API and configuration write validation
  - WebSocket real-time synchronization of configuration changes (settings_changed / settings_snapshot)
  - Frontend SettingsPanel component and basic editing controls
  - Configuration file hot reload broadcasts changes via eventBus
  - Added settingsValidation validation module and related tests
- Settings panel optimization
  - Adjusted Advanced configuration order to env/permissions/sandbox/sessionSummary
  - env configuration changed to KeyValue form mode, supporting add/delete/edit
  - Changed from real-time save to save-button batch submission mode
  - Improved env configuration validation rules and added tests
- /cost command UI panelization
  - Added token usage, context window ratio, and cache hit rate display
  - cost-tracker added cacheTokens and currentContextTokens tracking
  - llmodel updates cache and context token statistics on each API call
- Web tools and message optimization
  - Memo tool Web component implementation, including usage and result display components
  - Filtered internal messages displayed on Web (INNER_MESSAGE, SYSTEM_NOTIFICATION, system-reminder)
  - Fixed archive function Web display
- Lazy-loaded tool prompt optimization
  - Fixed incorrect prompt when lazy-loaded tool does not exist
  - Optimized prompt copy for unloaded tools
v1.40.1 (2026-06-24)
- Unified lazy-loaded tool reminders to use tools parameter, optimized LSP prompt copy
  - getToolLazy and getLSPServer methods now use the current agent’s actual tool list instead of global context data
  - Optimized unloaded tool prompt copy: replaced Unloaded tools with Available but not loaded tools
  - LSP reminders now determine whether LSP tools are loaded via the tools parameter, removing direct dependency on CONTEXT_LOADED_TOOLS
v1.40.0 (2026-06-24)
- Subagent tool lazy loading
  - Full support for subagent tool lazy loading, with loaded tool state isolated per agent
  - ToolLoader / ToolUnLoader isolate loaded state by agentId, so subagents do not pollute the main agent’s tool list
  - Fork-mode subagents pass the complete original tool list (_tools), ensuring lazy loading can access all available tools
  - Tool.prompt added optional context parameter, allowing ToolLoader to perceive the current agent’s tool scope
  - In lazy loading mode, ToolLoader / ToolUnLoader are always available (fixTools dynamically supplements based on enableLazyTool)
  - Unloaded tool error prompts also apply to subagents
v1.39.2 (2026-06-23)
- Optimized context usage reminder
  - system reminder context usage prompt is now shown at all usage levels, not only when ≥50%
  - Optimized MessageArchiving and Memo tool load/loaded prompt copy to be more concise and clear
  - Multiple tool suggestions (archive + memo) are presented as a list for clearer structure
v1.39.1 (2026-06-22)
- Fixed webServer frontend process is not defined error caused by referencing @constants/product.js
v1.39.0 (2026-06-21)
- Message archiving and context optimization
  - Message archiving 1.0 officially launched
  - Supports MessageArchiving lazy loading and optimized guidance reminders
  - Decoupled Message Archiving and Memo context optimization reminders
  - Centralized Message Archiving enablement judgment in SettingManager
  - Archive tool results support optional summary field for summarizing key information
  - Skips archiving for tools with little return content (e.g., Cwd, Think, TodoWrite, MessageArchiving, etc.) to avoid meaningless replacement
  - Applies messageArchivingService.applyToMessages before automatic compression to ensure archived content is also replaced
  - Centralized AUTO_COMPACT_THRESHOLD_RATIO in constants/product.ts
  - system reminder context usage prompt supplemented with automatic compression trigger threshold (92%)
- Memo tool enhancement
  - Simplified prompt and extracted constants, schema added basic validation
  - Added input validation to prohibit saving content containing …(argument truncated) placeholders
  - Changed to truncation mode: placeholders replaced with truncated text, keeping first 20 characters and adding a hint
  - Removed duplicate summary hint from successful Memo write results
  - Added SettingManager.isMemoEnable, controlled by HAWA_CODE_MEMO environment variable
  - system prompt added explanation of Memo content placeholders
  - Added content parameter to filter content during queries to reduce context usage
  - After truncating old Memo write content, insert a synthesized explanation message to reduce model hallucination
- Configuration and interaction
  - Added autoMode toggle option in settings panel
  - /model command added Open Config button
  - Slash commands support controlling tool lazy loading
  - autoModel enabled by default
  - Fixed bug where autoDream still executed after being disabled
- Tools and permissions
  - Restored LS directory listing tool
  - Backend added /api/config endpoint returning userType
  - Frontend added useConfig hook to get userType and filter tools based on userType
  - Added description fields to BashTool and PowerShellTool input schemas
  - Added missing agent id parameter
- Other fixes
  - Fixed websocket module import issue, compatible with CommonJS and ES modules
v1.38.12 (2026-06-16)
- Fixed error when reading empty files
v1.38.11 (2026-06-15)
- Fixed agentviewer bug
- Architecture optimization: separated tool execution from agent execution
v1.38.10 (2026-06-14)
- Fixed AgentTabs main tab display issue, keeping it consistent with backend agentId
- Fixed bug where main agent could not be awakened after subagent ended
- Added agentType, agentName, and description parameters to hawaContext.onQueryStart to facilitate tracking sub-agent lifecycle
v1.38.9 (2026-06-12)
- Hidden tool unload messages
- webviewer 2 added usage and result components for multiple tools: NotebookEdit, WebSearch, WebFetch, TaskStop, TodoWrite, Think, AskUserQuestion, EnterPlanMode, ExitPlanMode, Enter WorktreeTool, ExitWorktreeTool, SlashCommand, Skill, CronCreate, CronDelete, CronList, TeamCreate, TeamDelete, SendMessage
- Optimized normal MCP tool display
v1.38.8 / v1.38.7 / v1.38.6 (2026-06-12)
- Three tags point to the same commit
- Hawa Code - Agent Viewer centered display
v1.38.5 (2026-06-12)
- hands tool supports cwd tool
v1.38.4 (2026-06-12)
- hands tool server supports Session reuse of current mcpserver
- Optimized command message display
v1.38.3 (2026-06-12)
- Fixed error where retries could not be interrupted
- Supports proxy setting in config, allowing users to enable or disable proxy functionality
- Fixed edit tool display issue in webviewer
- Optimized hands remote identifier
- Optimized hands tool rendering in webviewer
- Optimized overall webviewer rendering
- Removed cost display
v1.38.2 (2026-06-08)
- Fixed port conflict when starting multiple hcode webserver instances
v1.38.1 (2026-06-08)
- Fixed webviewer static file path issue
v1.38.0 (2026-06-08)
- Supports webviewer 1.0
- Web viewer supports table display
- Optimized webview frontend page rendering of tool use and tool result correspondence
- Switched to incremental messages for better performance
- TaskCreate, TaskGet, TaskUpdate tools no longer render on Web, consistent with CLI
- Fixed duplicate display of default and main agent
- Optimized webserver startup log output
- Optimized webserver display in config
1.37.4 (2026-06-07)
- Optimized hands tool local rendering
1.37.0 (2026-06-03)
- Supports Goal slash command
- Stop Hook added last_assistant_message parameter
1.36.0 (2026-05-24)
- Supports built-in Hands tool, directly start hands command without installing mcp
1.35.5 (2026-05-21)
- Fixed Edit tool line misalignment bug during file comparison
- Supports Http request powerful proxy
- Supports custom use agent, supports authorization, x-api-key
- Supports enabling hawa code built-in computer use via environment variable
1.35.0 (2026-05-16)
- Supports setting openai image detail via environment variable
1.32.0 (2026-05-08)
- hawa code fully supports windows, skill, slash commands, and user instructions all support powershell
- Fixed image paste bug
1.31.4 (2026-05-08)
- Optimized hawa code update performance on windows
1.31.3 (2026-05-08)
- Optimized Shell execution performance
1.31.2 (2026-05-07)
- Optimized tool lazy loading performance
- Fixed openai response request bug
1.31.0 (2026-05-07)
- Supports automatic tool unloading in tool lazy loading mode
- Supports thinking depth switching
- Compatible with openai response api search
- Fixed openai chat completions thinking protocol conversion bug
- Optimized multi-model token compression estimation calculation
- Supports multi-model load balancing
1.28.0
- Supports powershell tool
- Supports memory recall during query
- Agent Team idle messages are suppressed
- Memory slash command supports manual memory clearing
1.27.0
- autoDream supports team memory
- autoMemory supports automatically recalling up to 5 related memories before Query
- Supports /memory slash command to manage memory function on/off and memory files
- Supports auto permission mode, where the model autonomously judges and authorizes tool execution
- Supports /auto slash command to quickly enable auto mode
- Supports switching auto mode via terminal shortcut
- Optimized error log output
1.26.0
- Supports session memory automatic extraction, autoMemory, currently disabled by default
- Supports session memory automatic integration, autoDream, currently disabled by default
1.25.3
- Supports session summary compression mode
- Fixed resume slash command duplicate message bug
1.24.0
- Supports git worktree tool
- subagent supports git worktree
- Fixed agent team tab switching loading position anomaly error
1.23.0
- Fixed Agent Team Tab logo not displaying
- Optimized terminal rendering performance
1.22.8
- Fixed automatic compression bug
- Fixed bash path issue
1.22.5
- Fixed tree-sitter-bash testing issue
1.22.3
- Optimized compressed messages
- Fixed TreeSitterNode path bug
1.22.2
- Added permission release for acceptEdits mode, allowing all files without file restrictions
- Fixed permission path matching issue
1.22.1
- Message logs support linked list association
- Fixed error log printing bug
- Fixed crash when accessing Kimi process bug
1.22.0
- Optimized message log storage performance
- Fixed terminal message display disorder bug
1.21.0
- Fixed some tool configuration permission anomaly bug
1.20.0
- Supports autoAllowBashIfSandboxed parameter, if enabled sandbox Bash no longer validates
- Permission configuration supports ask configuration, can force tool execution to require user confirmation
- In bypassPermissions mode, writing to .git, .hcode, .vscode, and .idea directories still prompts confirmation, but writing to .hcode/commands, .hcode/agents, and .hcode/skills is exempted
- Supports disableBypassPermissionsMode parameter to disable bypassPermissions mode
- Uses tree-sitter-bash to interpret shell, optimizing parsing performance
- Fixed shell prefix matching bug
- Supports shell saving as multiple subcommands instead of a single named match
1.19.0
- Supports sandbox, can isolate network, disk files, users can specify allowed domains, disk file methods, etc.
- Supports sandbox slash command
1.18.0
- Skill dynamic parameter replacement, $ARGUMENTS, $ARGUMENTS[N], $N, ${HCODE_SESSION_ID}, ${HCODE_SKILL_DIR}
- Skill supports disable-model-invocation, user-invocable parameter configuration
- Skill supports !command dynamic shell execution
1.17.0
- Skill supports context fork parameter for subagent execution, supports agent parameter
1.16.0
- Plan mode supports local file storage, stored at ~/.hcode/plan
- Added plan command to enter plan mode
- Agent Team supports mode parameter
- Supports Read, Edit tool permission configuration
1.15.0
- Supports scheduled task scheduling
1.14.0
- Supports automatic memory, Hawa code automatically extracts memories from conversations, memories are cross-session
1.13.5
- Fixed bug where Agent Team was enabled but Tasks were not enabled
1.13.4
- Optimized Agent Team list display
1.13.2
- Supports terminal display of Agent Team member messages
1.13.1
- Enabling Agent Team automatically enables Tasks function
- Tasks list added owner display
1.13.0
- Supports Agent Team
1.12.2
- Supports SessionEnd Hook
1.12.1
- Supports SessionStart Hook
- Fixed TODO restore bug
1.12.0
- Task planning mode upgraded from TodoWrite to Tasks mode, supporting task local file storage and cross-session communication
- Fixed terminal title display anomaly
- Optimized tool name usage
1.11.3
- Fixed Subagent deadlock issue
- Optimized log output performance
1.11.2
- Added PostToolUse Hook, supports triggering custom operations after successful tool execution
- Added PostToolUseFailure Hook, supports triggering custom operations when tool execution fails
- Optimized Hook Matcher matching performance
1.11.1
- Supports tool inputScheme lazy loading, optimizing startup time
- Optimized Task Tool progress display, solving terminal flicker issue
- Supports disabling Prompt prompt function
- Fixed message duplicate publishing bug
- Fixed system reminder message duplicate bug
- Optimized tool execution performance
1.11.0
- Supports PreToolUse Hook
1.10.2
- Supports using shortcut keys to switch tasks to background execution
- Supports controlling background task running reminders
- Optimized Markdown display
- Supports displaying output during shell execution
- Fixed inaccurate background shell task count display issue
1.10.0
- Added TaskStop tool, supports aborting background subagent tasks and shell task execution
- Removed KillShell tool
- After background shell tasks complete, supports restarting conversation
- Supports displaying background shell task count
- Fixed background subagent conversation restart deadlock issue
1.9.3
- Supports SubagentStart hook
1.9.2
- Supports Stop, SubagentStop hook
1.9.0
- Supports user-level rules, project-level rules, and on-demand loading by path
- Supports startup commands
- Optimized message reminders
1.8.0
- Optimized terminal MCP auto-authorization configuration
1.7.0
- Remote MCP supports auto-authorization
- Optimized usage effect in githubcodespace
- MCP Tool permissions.allow supports wildcard setting
- Optimized compatibility with non-standard Tool input JSON and erroneous JSON formats in Qwen-code-plus
1.6.0
- Supports MCP Tool lazy loading
- Fixed Task Tool verbose mode crash bug
- Optimized network request retry performance
1.5.0
- Supports settings.json configuration and hot reload
1.4.0
- AskUserQuestion tool supports multiple questions and custom input
- Optimized Grep Tool
1.3.0
- Supports prompt suggestion statements at the end of conversations
1.2.0
- Supports submitting prompt statements during conversations
1.1.0
- Supports subagent background execution
- Supports subagent resume execution
- Optimized LSP tool file path validation
1.0.16
- Optimized tool lazy loading performance
- Optimized LSP tool terminal display
1.0.15 Restored TODO LIST when restoring session