Skip to content

CHANGELOG.md

Latest commit

 

History

History
88 lines (70 loc) · 3.32 KB

CHANGELOG.md

File metadata and controls

88 lines (70 loc) · 3.32 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and uses version tags such as v0.0.1. Every published version must have:

  • A dedicated version entry in this file
  • A matching Git tag
  • A GitHub Release created from that tag

[Unreleased]

Added

  • Interactive chat command exposed in the CLI
  • Cross-platform shell and file search behavior
  • Automated tests for tools, storage, and Ollama client
  • lefthook configuration for local quality gates
  • GitHub Actions release workflow triggered by version tags
  • Goal mode with persistent GoalRun and WorkItem tracking
  • Lightweight planner and multi-step MCP execution loop
  • Repository-aware tools: git_status, git_diff, glob_files, and run_validation
  • Safe apply_patch tool for contextual file edits
  • Session summaries for goal-driven runs
  • Heuristic repository context selection for goal preparation using goal text, recent session messages, summaries, and git changes
  • Goal/session visibility for selected context files in REPL status and session show
  • MCP execution safeguards for per-iteration tool-call limits and repeated tool or validation failure limits
  • Ollama-first multi-provider support with OpenAI-compatible remote providers for OpenAI, OpenRouter, Groq, and DeepSeek
  • Encrypted provider profile storage and a provider CLI command for saving and activating remote backends
  • Fixed home-directory persistence under ~/.mini-agent for the database, config file, MCP server file, and encryption key
  • Provider management commands for show, remove, and rotate-key
  • Advanced per-profile remote options for temperature, timeout_seconds, extra_headers, and options_json
  • Optional sandbox modes for workspace-constrained tool execution

Changed

  • Chat sessions now support /goal, /status, and /stop
  • Goal execution now supports bounded iteration and clean stop requests
  • Session history now exposes persisted goal runs
  • Goal execution now attempts bounded repair passes after failed validation
  • Planner prompts now include risk level and suggested starting context for normal and repair flows
  • Goal preparation now persists selected context files with each GoalRun
  • Chat and run flows now resolve the active LLM backend through a provider factory instead of a hardcoded Ollama client
  • Goal execution now reduces repeated SQLite polling and redundant in-flight progress writes during long-running runs
  • Chat, run, and manual tool execution now share the same sandbox policy wiring

Fixed

  • Shell execution safety now blocks a broader set of destructive commands by default
  • Goal repair detection now catches broader validation failure wording such as generic fail output
  • Goal cancellation now propagates into the Ollama HTTP request via context.Context
  • Goal completion is now surfaced without waiting for another REPL input line
  • Remote provider requests now honor per-profile headers, timeout, and extra options payload settings
  • File writes and patch operations can now be confined to a sandbox root, and sandboxed shell execution blocks write-risk commands

Version Entry Template

## [v0.0.1] - 2026-07-02

### Added

- Describe new features here

### Changed

- Describe behavior changes here

### Fixed

- Describe bug fixes here