Skip to content

CHANGELOG.md

Latest commit

 

History

History
143 lines (116 loc) · 13.6 KB

CHANGELOG.md

File metadata and controls

143 lines (116 loc) · 13.6 KB
Raw

Changelog

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

[0.6.3] - 2026-06-13

Changed

  • Codex System Prompt Survives Compaction (memory_mode="codex_thread"): The agent system prompt is now passed as thread-level instructions on every thread_start and thread_resume, instead of being text-injected into the first user turn. Codex keeps these in the compaction-preserved prefix (outside the compactable conversation body), so the persona no longer degrades as the thread grows, and it is re-applied on every turn (also covering thread recovery after a lost thread).
  • New instructions_mode option (client_config_override): "developer" (default) layers the system prompt on top of Codex's native coding-agent/tool harness; "base" replaces that harness for full persona control. Both were verified to keep runtime MCP tool calls working and to honor the persona; "developer" is the default as the non-destructive choice.

Compatibility

  • SDKs that do not accept the instructions kwarg on thread_start/thread_resume fall back to the previous first-turn text injection automatically (a warning is logged).

[0.6.2] - 2026-06-13

Added

  • Agent Convenience Class: New from agentify import Agent one-liner with batteries-included defaults. Only model is required; an in-process store and a default conversation address are created automatically when memory/memory_address are omitted, and extra keyword arguments are forwarded to AgentConfig. The explicit BaseAgent/AgentConfig/MemoryService API is unchanged.
  • Top-Level Exports: Agent and InMemoryStore are now exported from agentify, so the common case no longer needs deep imports.
  • Richer Tool Schemas: The @tool decorator now emits items for typed lists (list[str]), and enum for Literal[...] and Enum parameters, improving tool-calling accuracy.

Fixed

  • Concurrent Streaming Isolation: Streaming tool-call/reasoning state is no longer stored on the agent instance (_last_stream_*); it flows through a per-call sink, so concurrent streaming arun() calls on a shared agent instance no longer clobber each other. The legacy self._last_stream_* attributes remain populated for direct callers.

Changed

  • Logging Is No Longer Configured at Import Time: agentify.memory.service attaches a NullHandler on import; the human-readable colored stream output is wired up lazily only when a MemoryService is created with log_enabled=True, so importing Agentify no longer hijacks the host application's logging configuration.

[0.6.1] - 2026-06-13

Added

  • System Prompt Seeding for Codex Thread Memory: In memory_mode="codex_thread", the agent's system prompt is now sent once on the first turn of each new Codex thread (prepend_system_prompt / extract_system_prompt in codex_inputs.py), so the agent identity reaches Codex even though only the latest user message is sent per turn. Subsequent turns on the persistent thread do not re-send it.
  • CodexThreadBackend.drop_session(session_id): Forget the Codex thread mapped to a session (and remove it from the persisted thread_map_path), so a memory reset starts a fresh thread instead of resuming the old one. The Codex thread itself remains on disk under ~/.codex/.

Fixed

  • Concurrent Thread-Map Saves No Longer Clobber: _save_thread_map now merges with the on-disk map before writing, so two backends sharing one thread_map_path (e.g. a CLI and a Telegram bridge) don't overwrite each other's session → thread mappings.

[0.6.0] - 2026-06-13

Added

  • Per-Session Runtime MCP Bridge: Tools, timeouts, and memory-bound executors are now registered on the runtime MCP bridge per session (register_session), so concurrent turns for different Agentify sessions on the same Codex backend are fully isolated. The legacy single-session bridge API keeps working.
  • Durable Codex Thread Mapping: New client_config_override={"thread_map_path": ...} persists the session → Codex thread ID mapping as a JSON file, so memory_mode="codex_thread" sessions survive process restarts (Codex threads themselves are persisted by the CLI under ~/.codex/).
  • Codex Thread Recovery: If a mapped Codex thread no longer exists, Agentify logs a warning and transparently starts a new thread for that session; transient resume errors are still raised so context is never silently discarded.
  • Native Thread Introspection: CodexThreadBackend.get_thread_id(session_id) and await CodexThreadBackend.read_session_history(session_id) expose the native Codex thread state (ThreadReadResponse).
  • Typed Codex Errors: New actionable exception hierarchy (CodexAuthError, CodexCLINotFoundError, CodexModelNotSupportedError, CodexUsageLimitError, CodexMCPStartupError, CodexToolNotFoundError, CodexEmptyTurnError, CodexStreamTimeoutError), all carrying retry metadata honored by BaseAgent.
  • MCP Startup Diagnostics: When a Codex turn completes without reconstructible text, the error now reports whether the runtime MCP bridge ever received a connection and (best effort) which MCP servers/tools Codex can see via mcpServerStatus/list.
  • Real-Codex Test Suite & Benchmark: Env-gated E2E tests against real ChatGPT OAuth (AGENTIFY_CODEX_E2E=1, plus AGENTIFY_CODEX_E2E_SLOW=1 for a >30s tool regression test) and scripts/benchmark_codex_memory_modes.py to compare memory modes.

Changed

  • Codex Memory Mode Guidance: Documentation now explains how Codex thread memory works (storage under ~/.codex/, persistence, auto-compaction) and recommends memory_mode="codex_thread" for interactive assistants (~1.5–1.7x faster per turn in benchmarks; the gap grows with history length).
  • Agentify Memory Preamble: Hardened the prompt preamble used by memory_mode="agentify" so the model no longer leaks the memory instructions into its replies.
  • Stable Session Keys: Native backends now receive MemoryAddress.key_str() as the session key instead of an unstable fallback.

Fixed

  • Transient Codex Errors No Longer Abort Turns: Codex error events are parsed from their real payload (ErrorNotification.error / willRetry); errors Codex retries internally are recorded as warnings instead of failing turns that complete successfully.
  • Slow Tools Over MCP: The runtime MCP proxy socket timeout is now derived from tool_timeout (--call-timeout), so tools slower than 30 seconds return their structured result instead of killing the proxy connection. Validated against real Codex with a 35s tool.
  • MCP Tool Config Never Silently Dropped: The thread_start compatibility fallback no longer discards MCP tool configuration on older SDKs; it raises an actionable upgrade error instead.
  • Error Classification: codex_error_info root models are unwrapped; unauthorized (OAuth missing/expired → run codex login) and contextWindowExceeded are now classified; a missing codex binary raises an installation hint instead of a raw FileNotFoundError.

[0.5.0] - 2026-06-02

Added

  • Codex Provider Parity: provider="codex" now behaves much closer to standard Agentify providers while using native Codex threads internally.
  • Agentify-managed Codex Memory: Codex uses Agentify memory stores (SQLite, in-memory, Redis, Elasticsearch) as the source of truth by default with memory_mode="agentify". Native Codex thread memory remains available via memory_mode="codex_thread".
  • Runtime MCP Tool Adapter: Normal BaseAgent(tools=[...]) tools are automatically exposed to Codex through a runtime MCP bridge. Users no longer need to manually wrap tools for common Codex usage.
  • Persisted MCP Tool History: Codex MCP tool calls are logged and stored in Agentify memory as assistant tool intents plus tool result messages with metadata.source="codex_mcp".
  • Codex Multimodal Input: image_path=... is converted to Codex SDK image input when supported, while preserving the Agentify multimodal memory message.
  • Codex Structured Output: Added support for Codex output_schema through model_kwargs={"output_schema": ...} and OpenAI-style response_format={"type": "json_schema", ...}.
  • Codex Streaming Events: stream=True now emits text chunks reconstructed from Codex thread.turn(...).stream() events.
  • Codex Tool Iteration Limits: Runtime MCP tool calls now respect AgentConfig.max_tool_iter inside a Codex turn.
  • Provider Lifecycle Cleanup: Added BaseAgent.close(), BaseAgent.aclose(), and CodexThreadBackend.close() to release provider resources such as runtime MCP bridges.
  • Codex Diagnostics Script: Added scripts/manual_codex_feature_diagnostics.py for manual validation of structured output, image input, streaming, and MCP tool limits.
  • Typed Package Marker: Added agentify/py.typed to match package-data configuration.

Changed

  • Codex Tool Architecture: Moved provider-specific Codex MCP adaptation out of BaseAgent into agentify.llm.tool_adapters.
  • Codex Backend Structure: Split Codex input building and retry-aware error classification into dedicated modules (codex_inputs.py, codex_errors.py) for maintainability.
  • Package Extras: agentify-core[all] now includes the optional Codex dependency.
  • Project Metadata: Updated package license metadata to modern SPDX-style license = "MIT".
  • Documentation: README, PyPI README, API reference, and core concepts now document native Codex usage, login, runtime MCP tools, structured output, multimodal input, streaming, and lifecycle cleanup.

Fixed

  • Codex Retry Behavior: Non-retryable Codex errors such as usage limits, unsupported models, MCP startup failures, and unknown MCP tools now stop retries immediately.
  • Codex Tool Observability: MCP tool usage is now visible in callbacks/logs and persisted memory history.
  • Codex Packaging Hygiene: Removed reliance on private Codex SDK input imports and verified package build contents.

[0.4.1] - 2026-05-30

Added

  • Tool Hooks System: New tool_pre_hooks and tool_post_hooks in BaseAgent for executing custom logic before and after tool execution.
  • SpawnAgentTool Enhancements: Support for passing tools, pre_hooks, post_hooks, tool_pre_hooks, and tool_post_hooks to spawned sub-agents.
  • MCP Tool Name Validation: New _safe_function_name() function to safely convert MCP tool names into valid Python identifiers.
  • Experimental Codex Native Provider with MCP-backed Tools: Agentify now supports Codex as a native experimental provider using ChatGPT OAuth and Codex threads. Unlike OpenAI providers, Codex does not use tool_calls; Agentify exposes tools through an MCP stdio server and reconstructs Codex responses from thread event streams.

Changed

  • MCP Adapter: convert_mcp_tools_to_agentify() is now synchronous (no longer async) with improved error handling and null-coalescing for inputSchema.
  • Shell Safe Tool: Removed restrictive allowlist; now accepts any shell command with configurable timeout (increased from 60s to 1800s max).
  • System Message Integrity: Enhanced _aensure_system_initialized() to verify system message position and eliminate duplicates in conversation history.
  • Type Hints Modernization: Updated type aliases to use modern TypeAlias syntax and PEP 604 union operator (|) instead of Union.
  • LLM Client Factory: Added support for new provider "llama" alongside existing providers.

Fixed

  • Event Loop Detection: Simplified has_running_loop() to use asyncio.get_running_loop(), eliminating DeprecationWarning on Python 3.10+.
  • MCP Tests: Fixed async/await mismatch in test cases for convert_mcp_tools_to_agentify().
  • Tool Execution Rollback: New _arollback_last_tool_turn() method to recover from interrupted tool execution sequences.

[0.4.0] - 2026-04-18

Added

  • Local Provider Support: Official support for local LLM servers (LM Studio, Ollama, etc.) via the "local" provider.
    • Automatically configured for LM Studio's default port (http://localhost:1234/v1).
    • Supports custom tools, streaming, and vision on local models.
    • New configuration environment variables: LOCAL_API_BASE and LOCAL_API_KEY.
  • Dual execution bridge via run() (sync) and arun() (async) across agents and multi-agent runtimes.
  • agentify/core/sync_bridge.py with loop-safety checks, sync coroutine execution, and async-to-sync streaming bridge.
  • Delegation recovery controls in AgentConfig:
    • delegation_recovery_enabled
    • delegation_recovery_mode
    • delegation_max_retries
    • delegation_retry_backoff_ms
  • tool_timeout in AgentConfig to control tool execution timeout independently from model timeout.
  • agentify/core/multimodal.py to centralize image encoding and multimodal content building.

Changed

  • Lazy Client Loading: Improved performance by making the synchronous client instantiation lazy in BaseAgent. The client is only created when first accessed, reducing overhead for async-first workflows.
  • Refactored BaseAgent to use arun() as the execution source of truth; run() now bridges to async runtime.
  • Increased default timeout from 60 to 300 seconds for long-running reasoning scenarios.
  • Set default tool_timeout to 300 seconds.
  • Improved delegated tool-call recovery for consistency errors to reduce user-facing failures in concurrent multi-agent flows.
  • Reduced verbosity/noise for recoverable consistency errors in callback logging.
  • Updated docs and examples to consistently document dual API usage (run()/arun()).

[0.3.1] - 2026-03-28

Removed

  • Removed legacy internal sync execution paths in BaseAgent that were no longer part of the active runtime.