# Changelog
All notable changes to this project will be documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [Unreleased]
## [0.1.5] - 2026-02-07
### Added
- **Perpetual Persona Swarms (Phase 0)**: Initial always-on cognition runtime in `codetether-agent`
- New cognition contracts for persona identity/policy/state, thought events, proposals, snapshots, and lineage graph
- In-memory perpetual loop with bounded buffers and guarded recursion limits (spawn depth + branching factor)
- New server APIs for cognition control and swarm lifecycle:
- `POST /v1/cognition/start`
- `POST /v1/cognition/stop`
- `GET /v1/cognition/status`
- `GET /v1/cognition/stream` (SSE)
- `GET /v1/cognition/snapshots/latest`
- `POST /v1/swarm/personas`
- `POST /v1/swarm/personas/{id}/spawn`
- `POST /v1/swarm/personas/{id}/reap`
- `GET /v1/swarm/lineage`
- Feature flags for runtime enablement, auto-start, loop interval, and cognition budgets
- Documentation added in `docs/perpetual_persona_swarms.md`
- **Amazon Bedrock Provider**:
- New provider module using Bedrock Converse API with bearer-token auth
- Registered provider aliases: `bedrock` and `aws-bedrock`
- Supports configurable AWS region from Vault provider secret metadata (`region`)
- **OpenAI-Compatible Model Catalog Defaults**:
- Added known-model fallback listings for `cerebras`, `novita`, and `minimax`
### Changed
- **A2A Worker Server Paths**:
- Updated worker registration, task polling/output, and heartbeat endpoints to `/v1/opencode/*`
### Fixed
- **Model Reference Translation**:
- Prevented `provider:model` conversion when model IDs already contain `/`
- Avoids corrupting model IDs that use `:` for versions (for example `amazon.nova-micro-v1:0`)
## [0.1.2] - 2026-02-05
### Added
- **TUI Swarm Mode UI**: Real-time visualization of parallel sub-agent execution
- `/swarm <task>` command to run tasks in parallel swarm mode
- Live progress display with subtask status, stages, and execution stats
- Toggle views with `Ctrl+S` or `/view` command
- **Session Management**: Resume conversations across TUI sessions
- `/sessions` - List recent saved sessions
- `/resume` - Load most recent session
- `/resume <id>` - Load specific session by ID
- `/new` - Start fresh session (auto-saves current)
- **Real-time Tool Streaming**: See tool calls as they execute
- Live status indicator shows current tool being run
- Tool calls and results displayed with distinct styling
- **TUI Enhancements**:
- Auto-scroll to bottom on message send/receive
- Theme configuration support
- Token usage display with cost tracking
- File edit confirmation prompts
- Keyboard navigation improvements
### Fixed
- **Token Cost Calculation**: Was ~1 trillion times too high, now correctly calculates costs
- **Vim Key Conflicts**: Added Alt/Ctrl modifiers so h,j,k,l,g,d,u can be typed normally
- **Swarm UI Updates**: Subtasks now show immediately after decomposition, not just at completion
- **F2 Key Binding**: Added `Ctrl+S` as reliable alternative (F2 doesn't work in many terminals)
### Changed
- **Default Model**: TUI now defaults to `zhipuai/glm-4.7` for better performance
- **Help Menu**: Condensed and added slash command documentation
## [0.2.0] - 2026-02-03
### Added
- **Library Support**: Added `lib.rs` for using codetether-agent as a library
- **Integration Tests**: Comprehensive A2A client integration tests with mock server
- **Design Documentation**: Architecture docs for agent-swarm integration, MCP metadata, session management
- **SwarmMetrics**: Execution metrics tracking for swarm operations
- **MCP Registry**: Tool discovery and registration via McpRegistry
### Changed
- **async-openai 0.32.4**: Updated to latest API with new chat completion types
- **A2A Client**: Made public with exposed `call_rpc()` for custom RPC calls
- **Tool Registry**: Now includes BatchTool and InvalidTool by default
- **Provider Logging**: Enhanced debug logging with API key validation
### Fixed
- **67 Compiler Warnings → 0**: All unused code now properly integrated via swarm dogfooding
- MCP server tool/resource metadata storage
- Session add_message and generate_title methods
- Provider API key usage in debug output
- Error constants (PARSE_ERROR, INVALID_REQUEST, etc.)
- Context summary logging in RLM repl
### Dogfooding
This release was refactored by the swarm system itself:
- **Model**: Kimi K2.5 via Moonshot
- **Duration**: ~25 minutes
- **Files Changed**: 54
- **Lines Added**: 5,507
- **Warnings Fixed**: 67 → 0
## [0.1.0] - 2026-02-03
### Added
#### Core Features
- **Ralph Loop**: Autonomous PRD-driven development loop
- Implements user stories from structured PRD files
- Quality gates (cargo check, clippy, test, build) per story
- Memory persistence via git history, progress.txt, and PRD updates
- CLI: `codetether ralph run -p prd.json -m "model" --max-iterations N`
- **Swarm Execution**: Parallel sub-agent execution for complex tasks
- Automatic task decomposition with multiple strategies (auto, domain, stage)
- Concurrent sub-agent execution with result aggregation
- CLI: `codetether swarm "task description" --strategy auto`
- **RLM (Recursive Language Model)**: Handle large contexts exceeding model windows
- REPL-like environment for context exploration (head, tail, grep, count)
- Sub-LM calls via `llm_query()` for semantic sub-questions
- Auto-detection of content types (code, logs, conversation, documents)
#### Providers
- **Moonshot Direct**: Native Moonshot AI integration (kimi-k2.5)
- **OpenRouter**: Access to 100+ models via OpenRouter API
- **StepFun**: Step 3.5 Flash support
- **OpenAI**: GPT-4, GPT-4o, o1, o3 models
- **Anthropic**: Claude 3.5 Sonnet, Claude 4 Opus
- **Google**: Gemini Pro, Gemini Flash
- **DeepSeek**: DeepSeek V3
#### Tools (24+ total)
- `ralph` - Autonomous PRD-driven agent loop
- `rlm` - Recursive Language Model for large contexts
- `prd` - Generate and manage PRD documents
- `lsp` - Language Server Protocol operations (definition, references, hover, completion)
- `batch` - Run multiple tool calls in parallel
- `task` - Background task execution
- `codesearch` - Semantic code search
- `websearch` - Web search integration
- `webfetch` - Fetch web pages with smart extraction
- `skill` - Execute learned skills
- `todo_read`/`todo_write` - Track task progress
- `question` - Ask clarifying questions
- `plan_enter`/`plan_exit` - Switch to planning mode
- `read_file`/`write_file` - File operations
- `edit`/`multiedit` - Search/replace edits
- `apply_patch` - Apply unified diff patches
- `grep` - Search file contents with regex
- `bash` - Execute shell commands
- `glob` - Find files by pattern
- `list_dir` - List directory contents
#### Infrastructure
- **A2A Protocol**: Worker, server, and client modes for agent-to-agent communication
- **MCP Support**: Model Context Protocol client/server
- **HashiCorp Vault**: Secure API key storage (no env var secrets)
- **Interactive TUI**: Terminal interface built with Ratatui
- **Session Management**: Persistent session history with git-aware storage
### Dogfooding Achievement
This release was **built using itself**. The Ralph loop autonomously implemented 20 user stories:
| Total Stories | 20 |
| Pass Rate | 100% |
| Time | 29.5 minutes |
| Cost | $3.75 |
| Model | Kimi K2.5 |
Stories implemented:
- LSP Client (10 stories): Transport, JSON-RPC, Initialize, didOpen, didChange, Completion, Hover, Definition, Shutdown, Configuration
- Missing Features (10 stories): External Directory, RLM Pool, Truncation, LSP Integration, References, and more
### Performance
| Binary Size | 12.5 MB | ~90 MB | 7.2x smaller |
| Startup Time | 13 ms | 25-50 ms | 2-4x faster |
| Memory (10 agents) | 55 MB | 280 MB | 5.1x less |
| Process Spawn | 1.5 ms | 5-10 ms | 3-7x faster |