Consolidate 4-layer memory abstraction into unified Memory class:
- MemoryStore, QdrantMemoryStore, QdrantMemoryBridge, ContinuityManager
are now consolidated into a single src/memory/unified.ts
Single Qdrant collection with type field for discrimination:
- type: "memory" - facts, preferences, decisions
- type: "state" - personas orchestration state
- type: "conversation" - conversation continuity
- type: "session_chain" - session chain index
Key changes:
- New Memory class with save/search/get/delete for memories
- State persistence (saveState/loadState)
- Conversation continuity (startSession, processMessages, etc.)
- Cross-session memory injection for personas bootstrap
- Updated tiara.ts to use Memory instead of QdrantMemoryBridge
- Added bootstrap/personas.ts for lifecycle hooks
Legacy APIs preserved for backwards compatibility.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Create src/config/constants.ts with centralized values for:
- Qdrant URL and collection names
- Embedding model and dimensions
- Timeout values (drone, IPC, fact extraction, provider)
- Memory configuration limits
Update all consumers to import from constants instead of
using hardcoded values:
- src/personas/memory-bridge.ts
- src/personas/tiara.ts
- src/personas/drone-wait.ts
- src/personas/fact-extractor.ts
- src/memory/qdrant.ts
- src/memory/store.ts
- src/config/types.ts
- src/daemon/ipc-client.ts
All constants support env var override via QDRANT_URL etc.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Unify server port to use DEFAULT_API_PORT (3456) constant
- Make Qdrant URL configurable via QDRANT_URL env var in debug commands
- Fix plugin baseUrl to derive from Server.url() instead of hardcoded
- Update CLI help text to show correct default port
- Fix MDNS service name from opencode to agent-core
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
P0 Critical Fixes:
- Fix import ./tiara → ./claude-flow in plugin/builtin/index.ts
- Create AgentOrchestrator type bridge for council/tiara integration (src/tiara.ts)
- Fix hardcoded state ID in memory-bridge.ts (was causing cross-persona data loss)
P1 Core Functionality:
- Add persona isolation to memory bridge namespace (personas now have private memory)
- Load SKILL.md content into drone prompts (skills now available to drones)
- Inject persona-specific tools from AGENT_CONFIGS into drone prompts
P2 API Consistency:
- Standardize error response format in server.ts (MCP & gateway endpoints)
- WebSocket param validation already in place (verified)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- scripts/systemd/install.sh: fix help text CLI commands
- cli/cmd/pr.ts: fix spawned command and messages
- cli/error.ts: fix MCP and model error messages
- cli/cmd/uninstall.ts: also detect "# agent-core" in shell profiles
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add showThinking to SurfaceCapabilities type
- Set showThinking=true for CLI, Web, API (user toggleable)
- Set showThinking=false for WhatsApp, Telegram (locked to hide)
- Messaging platforms never show thinking/reasoning output
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Wrap CLI commands in Instance.provide for proper Config access
- Validate registry cache file against schema before use
- Prevents crash when Config.update() is called without Instance context
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add plugin discovery and management system:
- Registry index (plugins/index.json) with curated plugins
- Registry client with caching and fallback
- Plugin manager for install/remove operations
- CLI commands: plugin search, list, install, remove, info
Registry hosted as static JSON, uses existing npm/Bun infrastructure.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add text-to-speech capability to Telegram and WhatsApp gateways:
- TTSConfig interface with CLI command template system
- Template variables: {{Text}}, {{OutputPath}}, {{Voice}}
- sendVoice methods for synthesizing and sending audio
- Graceful fallback to text if TTS fails or not configured
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add three new canvas kinds for visual reasoning in WezTerm:
- diagram: flowcharts, architecture diagrams with box/diamond/oval nodes
- graph: nodes and edges with colors and weights (knowledge graphs)
- mindmap: hierarchical tree view for thinking/planning
Canvas features:
- ASCII art rendering with ANSI colors
- Works alongside nvim/LSP in WezTerm panes
- 67% width ratio (Claude:Canvas = 1:2)
- Reusable pane management
Updated tools with new kinds and config documentation.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Test pure functions: isExpiringSoon, isExpired
- Test file operations: get, all, set, remove
- Test status checking for multiple providers
- Test schema validation for OAuth, API, WellKnown types
- Test edge cases: missing files, invalid JSON, extra fields
44 tests covering auth storage and validation.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add showThinking to SurfaceCapabilities type
- Set showThinking=true for CLI, Web, API (user toggleable)
- Set showThinking=false for WhatsApp, Telegram (locked to hide)
- Messaging platforms never show thinking/reasoning output
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add showThinking to SurfaceCapabilities type
- Set showThinking=true for CLI, Web, API (user toggleable)
- Set showThinking=false for WhatsApp, Telegram (locked to hide)
- Messaging platforms never show thinking/reasoning output
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Rename PersonaConfig → OrchestrationPersona in personas/types.ts
to avoid confusion with agent/persona.ts:PersonaConfig
- Rename PERSONAS_CONFIGS → ORCHESTRATION_PERSONAS for clarity
- Remove duplicate inline configs in persona.ts, use shared constant
- getPersonaConfig now returns OrchestrationPersona type
Eliminates config duplication and clarifies type naming.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Change comments and commands in new code to use "agent-core"
- Rename OAUTH_DUMMY_KEY to agent-core-oauth-dummy-key
- Add AGENT_CORE_TEST_HOME env var (with OPENCODE_TEST_HOME compat)
- Keep external service URLs (opencode.ai, antigravity.opencode.ai)
- Keep OPENCODE_* flags for upstream compatibility
Part of technical debt cleanup (Phase 7).
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add comprehensive cross-platform APIs for web/mobile integration:
Server endpoints:
- GET /sync?since=<ts> - Delta sync for offline clients
- GET /personas - List personas with capabilities and gateway status
- GET /themes - List available themes (built-in + persona-specific)
- GET /preferences/theme - Get current theme
- PATCH /preferences/theme - Set theme
Telegram gateway:
- Add voice note transcription support via configurable CLI
- Download voice/audio files from Telegram API
- Transcribe using {{MediaPath}} template command
Documentation:
- Update CROSS-PLATFORM.md with all new API endpoints
- Mark Phase 6 as complete in ROADMAP.md
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add SSE endpoints for real-time session sync:
- GET /events - global event stream for all sessions
- GET /session/:id/events - per-session event stream
- Events: session.created/updated/deleted, message.updated, todo.updated, status changes
- 30-second keepalive heartbeat
- Add session handoff API:
- POST /session/:id/handoff - prepare session for transfer to another platform
- Returns session context, last message, todos, and deep link URL
- Supports: mobile, web, cli, telegram, whatsapp targets
- Add unified notification endpoint:
- POST /notify - send notifications across platforms
- Supports Telegram and WhatsApp broadcast
- Includes deep linking to sessions
- Add docs/CROSS-PLATFORM.md with:
- Architecture overview
- API documentation
- Integration patterns for mobile/web
- Deep linking scheme (agentcore://)
- Offline support guidance
- Update ROADMAP.md marking Phase 6.2 as complete
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add Branches section to sidebar showing parent/siblings/children
- Clickable branch items to navigate between related sessions
- Add breadcrumb trail in header showing session ancestry
- Clickable ancestors to jump up the session tree
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add `theme` field to agent schema for per-persona themes
- Auto-switch theme when changing agents
- Configure persona themes: Zee (cobalt2), Stanley (matrix), Johny (rosepine)
- Add "Delegate to persona" command (<leader>d) to prepend @persona to prompt
- Create DialogDelegation component for persona selection
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Shows ✗ for expired and △ for expiring tokens next to provider names
in favorites, recent, and all providers sections.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add VoyageEmbeddingProvider for Voyage AI embeddings
- Add MemoryStore with high-level store/search/list API
- Configure Nebius cloud API for Qwen3-Embedding-8B (#1 on MTEB)
- Fix LocalEmbeddingProvider to support both /v1/embeddings and /embeddings
- Fix QdrantVectorStorage to set collection on existing collections
Model: Qwen3-Embedding-8B (70.58 MTEB score, 99% zero-shot, 4096 dims)
Provider: Nebius API (https://api.tokenfactory.nebius.com/v1)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Change serve command output from "agent-core server listening" to
"opencode server listening" for compatibility with @opencode-ai/sdk.
The SDK parses stdout looking for "opencode server listening" to
determine when the server is ready.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Document the shared configuration architecture between agent-core and zee:
- Shared primitives in src/config/shared.ts
- Architecture overview showing ownership patterns
- Key patterns for config, browser, memory, and sessions
- Migration path for zee to import shared types
- Current status and future work
This serves as a reference for the consolidation work and guides
future development on the shared configuration layer.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Create shared.ts with common configuration types that can be used
across agent-core and zee gateway:
- Session primitives: SessionScope, ReplyMode, TypingMode, ReplyToMode
- Policy types: DmPolicy, GroupPolicy
- Logging: LogLevel, ConsoleStyle, LoggingConfig
- Network: RetryConfig, BindMode
- Model/Provider: ModelApi, ModelDefinition, ProviderDefinition
- Processing: ThinkingLevel, QueueMode, QueueDropPolicy
Update types.ts to use shared primitives:
- GeneralSettings.logLevel now uses shared LogLevel
- WhatsAppConfig extended with dmPolicy, groupPolicy, allowFrom, retry
- TelegramConfig extended with dmPolicy, groupPolicy, allowFrom, retry
This is the first step toward config consolidation between agent-core
and zee gateway, enabling type-safe sharing of common configuration.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- bun run works more reliably for spawning child processes
- npx tsx has PATH issues when spawned as subprocess
- Enable all persona MCP servers by default
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements three MCP servers that expose persona capabilities to external
tools via the Model Context Protocol:
- personas-memory: Qdrant-backed semantic memory (store, search, list, delete, stats)
- personas-calendar: Google Calendar integration (events, create, update, delete, free slots, quick add)
- personas-portfolio: Financial tools bridging to Stanley CLI (portfolio, market data, SEC filings, research, backtest)
Each server uses @modelcontextprotocol/sdk with Zod schemas for type-safe
tool definitions and stdio transport for local process communication.
Config entries added (disabled by default):
- "personas-memory": npx tsx src/mcp/servers/memory.ts
- "personas-calendar": npx tsx src/mcp/servers/calendar.ts
- "personas-portfolio": npx tsx src/mcp/servers/portfolio.ts
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Mark existing TUI improvements as complete:
- Favorites system in model selector
- Recently used models tracking
- Cost display (Free badge)
- Persistent agent state via ContinuityManager
- Better compaction summaries via continuity.ts
Remaining Phase 4 items:
- Provider status indicators
- Agent-specific themes/colors
- Agent handoff
- Conversation branching/export
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Mark Phase 3.1 Memory System as COMPLETE:
- Qdrant vector storage with full CRUD
- MemoryStore high-level API
- BGE-M3 embeddings via vLLM
- ContinuityManager for session persistence
- Cross-persona memory sharing with namespaces
Mark Phase 3.2 Orchestration as COMPLETE:
- Orchestrator with full worker lifecycle
- WezTerm pane management
- Tiara/SPARC integration
- DroneWaiter for async completion
- Fact extraction hooks
Remaining: MCP servers for external tool exposure
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add Auth.refreshToken() to refresh a specific provider's token
- Add Auth.refreshAllExpiring() to refresh all tokens expiring within 10 min
- Add Auth.status() to get token validity info for all providers
- Add Auth.isExpiringSoon() and Auth.isExpired() helpers
- Call refreshAllExpiring() during provider initialization
Supported providers: anthropic, openai
Note: google/antigravity uses plugin-handled custom flow
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Disable google-vertex and google-vertex-anthropic (redundant with Antigravity)
- Remove non-working antigravity provider config
- Use built-in antigravity-* models from provider.ts
- Add zee, stanley, johny agent definitions
- Update compaction to use google/antigravity-gemini-3-flash
- Update johny.md to use google/antigravity-claude-opus-4-5-thinking
- Remove top_p from agent files (handled by ProviderTransform)
- Add docs/PROVIDERS.md with provider setup documentation
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add topP: 0.95 to antigravity Claude thinking model definitions
- Add topP check in transform.ts for claude+thinking models
- Wrap plugin fetch to remove x-goog-api-key header (conflicts with OAuth)
- Add createEffect to properly initialize agent selection on load
- Update CLAUDE.md with binary update process documentation
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add favorite property to model store and favorite() method required
by dialog-model.tsx after upstream merge.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The google-antigravity provider was redundant - the opencode-antigravity-auth
plugin already handles OAuth, request transformation, and API calls through
the standard 'google' provider.
Testing confirmed google/gemini-2.5-flash works via the plugin:
bun run ./src/index.ts run --model google/gemini-2.5-flash "Say hello"
Changes:
- Remove google-antigravity custom loader and provider definition
- Remove antigravity.ts token refresh helper (unused)
- Keep auth OAuth passthrough for extra fields like projectId, email
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Fix API message format: use `parts` array instead of `content`
- Add `agent` parameter to route messages to correct persona
- Add `x-opencode-directory` header for proper session handling
- Same fixes as WhatsApp gateway
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Fix model selection not persisting across restarts by saving/loading
the model field in model.json (was only saving recent and variant)
- Call save() on every model change, not just when adding to recents
- Add google-antigravity provider for free Gemini access via OAuth
- Create antigravity.ts helper for token refresh via antigravity proxy
- Provider inherits all Google models but marks them as free (cost: 0)
- Auto-refresh tokens with 5-minute buffer before expiry
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- WhatsApp is exclusively Zee's channel using GLM 4.7 via Z.ai (cerebras)
- Remove persona detection patterns (Stanley/Johny use Telegram instead)
- Fix API message format: use `parts` array instead of `content`
- Add `agent` parameter to route messages to correct persona
- Add `x-opencode-directory` header for proper session handling
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Terminal QR codes can be distorted by font size/line height. Save QR as
PNG file that can be opened with an image viewer for reliable scanning.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Artur Do Lago