lea_code 0.1.0

lea_code is a Dart command-line coding assistant powered by Genkit.

0.1.0

• New terminal UI modeled on Claude Code: a welcome box, a bordered input box with history, multi-line input, collapsed pastes and slash-command completion, a spinner with elapsed time, Markdown-styled streaming responses, ⏺ Tool(argument) blocks with short result summaries, a checkbox task list, and arrow-key menus for command approval and ask_user_question. Piped input falls back to plain lines.
• Added /clear, /tasks, /model, and /help. Esc now interrupts a request, and approving a command can allow it for the rest of the session; declining it stops the request so you can redirect Lea.
• Cancelling a request now reports the interruption and leaves the conversation unchanged; before, Genkit's aborted result was taken as a normal reply.
• Added the anthropic provider via genkit_anthropic, with claude-sonnet-5-5 as its default model and a 32000-token output limit (the plugin's default of 4096 is too small for whole-file writes).
• --model now defaults per provider; openai defaults to gpt-6-luna instead of the Gemini model name.
• Provider errors such as an invalid API key are now shown instead of producing an empty response.
• Replaced the analysis, execute, and verify agents with a single CodingAgent loop modeled on Claude Code. The full conversation, tool calls included, now carries across requests instead of only the verifier's messages.
• Added a built-in system prompt focused on code quality and verification. It is now sent as a real system message; the previous outputInstructions route did not reach the model as a system prompt.
• Added session context to the prompt (working directory, platform, date, model, git snapshot) and loading of LEA.md, CLAUDE.md, and AGENTS.md project instructions plus ~/.lea/LEA.md.
• Rewrote every tool description with usage rules for the model.
• edit now takes a single old_string/new_string, rejects ambiguous matches unless replace_all is set, and returns a snippet of the result.
• read defaults to 2000 lines, truncates very long lines, reports remaining lines, and rejects directories and binary files.
• glob now uses rg --files, so ** patterns work, results respect .gitignore, and they are sorted newest first.
• grep gained type, context lines, multiline, and a default head_limit.
• bash uses the user's shell, defaults to a 2 minute timeout, truncates long output, and no longer asks for approval on compound commands made only of read-only parts. Redirection into files, find -delete/-exec, and command substitution now require approval.
• Workspace checks now normalize .., so paths can no longer escape the workspace.
• Tool errors are returned to the model as Error: ... instead of aborting the request, and failures are shown in the terminal.
• Responses stream to the terminal, and Ctrl-C cancels the running request.
• powershell is only exposed on Windows, and the MCP tools only when an adapter is configured.
• Upgraded to genkit 1.0, genkit_google_genai 1.0, genkit_openai 1.0, and schemantic 1.0, and added path.

0.0.9

• Replaced the single GeneralAgent implementation with a staged agent workflow made up of AnalysisAgent, ExecuteAgent, and VerifyAgent.
• Added BaseAgent plus shared read-only and full-access tool mixins so each stage can expose only the tools it needs.
• Updated LeaCode to orchestrate analysis, execution, and verification passes for each request, including retrying execution up to three times when verification fails.
• Added verification prompt/status handling so verifier responses start with STATUS: PASS or STATUS: FAIL and only the final user-facing result is shown in the terminal.
• Turned lib/agents/general_agent.dart into a compatibility export for the staged agent classes.
• Bumped the package version from 0.0.7 to 0.0.8.

0.0.7

• Replaced the older file and web tool set with a larger runtime-managed tool suite: read, write, edit, glob, grep, sleep, powershell, ask_user_question, task_create, task_update, task_list, list_mcp_resources, and read_mcp_resource.
• Added ToolRuntime to track workspace state, remembered file reads, the current working directory, session tasks, approval requests, and interactive multiple-choice questions.
• Updated bash so it runs in the workspace, supports workdir and timeout_ms, formats structured command output, and requests approval for mutating, networked, or otherwise non-read-only commands.
• Replaced generated schema-backed tool input models with lightweight JSON schema helpers and removed the old generated files from lib/models/.
• Added tests for command safety rules and runtime file/task behavior.
• Bumped the package version from 0.0.6 to 0.0.7.

0.0.6

• Added OpenAI provider support via genkit_openai.
• Added --provider / -p to choose between google (default) and openai.
• Added --api_key / -k and --base_url / -b for provider configuration.
• Added --help / -h and --version / -v.
• Made --system_prompt optional and routed it through the agent as output instructions.
• Bumped the package version from 0.0.5 to 0.0.6.

0.0.5

• Refactored the CLI around new LeaCode and GeneralAgent classes, replacing LeaCodeEngine.
• Added a new file_edit tool backed by sed for in-place file edits.
• Added tool status messaging so tool start and completion updates are shown in the terminal.
• Added a --max_turns / -t CLI option to control the maximum number of model turns per request.
• Changed the installed executable from lea_code to lea.
• Bumped the package version from 0.0.3 to 0.0.5.

0.0.3

• Added a new web_fetch tool so the assistant can fetch web pages as text.
• Registered web_fetch in LeaCodeEngine so it is available during model generation.
• Bumped the package version from 0.0.2 to 0.0.3.

0.0.2

• Bumped the package version from 0.0.1 to 0.0.2.

0.0.1

• Added the initial lea_code interactive CLI entrypoint.
• Added Gemini-powered response generation through Genkit and genkit_google_genai.
• Added support for selecting a model with --model.
• Added support for passing a custom system prompt with --system_prompt.
• Added conversation management commands for resetting and exiting the session.
• Added local assistant tools for shell execution, file reads, file writes, string search, and file discovery.
• Added typed tool input models and generated schema files for tool registration.