Skip to content

Changelog

Changelog

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

[Unreleased]

Documentation

  • Added uv installation instructions to README, getting-started, developer-guide, and contributing docs; uv is the recommended install method, pip retained as fallback.
  • Updated maxTokens default from 16000 to 8192 in docs/configuration.md to match code.
  • Added GitHub Copilot (OAuth) to README LLM Providers feature list.
  • Fixed ff agent → miqi agent in historical changelog entries.
  • Updated README and project docs to match current code paths and schema defaults:
  • corrected workspace-relative memory/session storage paths (<workspace>/memory, <workspace>/sessions)
  • documented current packaged gateway scope (Feishu wired today; other channel adapters remain extension modules)
  • refreshed config defaults for memory, sessions, self-improvement, heartbeat, cron, and shell execution
  • documented MCP lazy, description, headers, and environment-inheritance behavior
  • clarified that SQLite session storage, provider fallback chains, command approval, and smart routing exist as shipped modules/helpers but are not all enabled by default in the packaged CLI/gateway path

Added

  • Added paper-research skill (miqi/skills/paper-research/SKILL.md):
  • Full workflow: paper_search → paper_download → translate_pdf → summarize with references.
  • Covers scheduled briefing scenarios (cron + feishu delivery).
  • Includes web_search fallback for news and preprints not indexed in Semantic Scholar.
  • Added feishu-report skill (miqi/skills/feishu-report/SKILL.md):
  • Format decision table: plain text / card message / Feishu Doc / calendar event / task.
  • Covers send_message, send_card_message, create_document + write_document_markdown, create_calendar_event, create_task.
  • Includes user identity resolution via resolve_users_by_name and get_chat_members.
  • Added workspace-cleanup skill (miqi/skills/workspace-cleanup/SKILL.md):
  • Organizes ~/.miqi/workspace into structured artifacts/ subdirectories.
  • Archives files older than 30 days to archive/YYYY-MM/.
  • Defines sacred directories (memory/, skills/, sessions/, system .md files) that are never touched.

Changed

  • Updated cron skill (miqi/skills/cron/SKILL.md):
  • Corrected timezone fallback documentation: without tz, cron expressions are now evaluated in UTC (not server local time).
  • Updated at mode examples to always include timezone offset (e.g. +08:00) or explicit tz=.
  • Added China Standard Time (Asia/Shanghai) examples to time expression table.
  • Updated cron tool parameter descriptions (miqi/agent/tools/cron.py):
  • message: clarified it is the full task prompt executed at trigger time, not just a label.
  • cron_expr: added explicit warning that expressions default to UTC; must pass tz= for non-UTC users.
  • at: updated example to include timezone offset; documents tz= fallback for naive datetimes.
  • tz: extended to apply to both cron_expr and at modes.

Fixed

  • Fixed max_tokens default of 16000 exceeding DeepSeek API maximum (8192 output tokens): lowered AgentDefaults.max_tokens back to 8192. After migration from litellm (which auto-capped per model) to direct SDK calls (which do not), the previous 16000 default caused 400 BadRequest errors on DeepSeek (config/schema.py).
  • Fixed _match_provider() using bare api_key truthiness check, preventing is_local providers (vLLM, Ollama Local) from matching when they use api_base instead of api_key: replaced with _is_configured() helper that checks both api_key and api_base (config/schema.py).
  • Fixed build_provider() not passing provider_name and default_model to fallback-chain provider constructors, causing fallback providers to use wrong endpoints and model names (config/schema.py).
  • Fixed _make_provider() in onboard CLI requiring api_key for is_local providers (vLLM, Ollama Local) that only need api_base: added is_local exemption to the api_key requirement check (cli/commands.py).
  • Fixed AnthropicProvider passing empty-string api_key to SDK instead of None, preventing the SDK from falling back to ANTHROPIC_API_KEY environment variable (providers/anthropic_provider.py).
  • Fixed cron at mode silently using server local timezone for naive datetime strings: naive datetimes now interpreted as UTC when no tz is provided; tz= can be passed together with at= to override (miqi/agent/tools/cron.py).
  • Fixed tz parameter rejected when combined with at: removed erroneous validation that blocked tz + at combinations (miqi/agent/tools/cron.py).
  • Fixed cron cron mode using unpredictable server local timezone as fallback: now falls back to UTC for deterministic behavior across deployment environments (miqi/cron/service.py).
  • Fixed miqi agent mode (CLI) silently ignoring all cron jobs: on_job callback was never registered and cron.start() was never called; both now wired correctly in cli/agent_cmd.py.
  • Fixed miqi agent mode not propagating job_timeout from config to CronService (miqi/cli/agent_cmd.py).

Removed

  • Removed clawhub skill (miqi/skills/clawhub/) — not applicable to this deployment.

[Unreleased — previous]

Added

  • Added Feishu group chat @mention filtering (channels/feishu.py):
  • New config channels.feishu.requireMentionInGroups (default true).
  • In group chats, messages are only forwarded to the agent when the bot is @mentioned.
  • @mention placeholder (@_user_N) is automatically stripped from the message text.
  • Private chats (p2p) remain unaffected and always forwarded.
  • Added MCP tool heartbeat progress reporting (agent/tools/mcp.py, agent/loop.py):
  • New config tools.mcpServers.<name>.progressIntervalSeconds (default 15).
  • During long-running MCP tool calls, periodic status messages are sent to the user (e.g. "⏳ raspa_run_simulation 正在执行中... (已用时 1m30s)").
  • Heartbeat task is automatically cancelled when the tool finishes.
  • Existing MCP SDK progress callbacks remain functional alongside heartbeat.
  • Added task queue tracker with user notifications (agent/loop.py):
  • New TaskTracker class tracks active and pending tasks.
  • New config channels.sendQueueNotifications (default true).
  • When agent is busy, new messages are queued and senders receive position notifications (e.g. "✅ 收到！当前正在处理 Alice 的任务，您排在第 2 位，请稍候。").
  • When a queued task starts processing, sender receives "🚀 开始处理您的任务...".
  • CLI and system messages bypass the queue for immediate processing.
  • Added sender_name field to InboundMessage (bus/events.py) for display-friendly queue notifications.
  • Added sender_name parameter to BaseChannel._handle_message() (channels/base.py).
  • Added tests for new features:
  • tests/test_task_tracker.py: TaskTracker unit tests, MCP heartbeat integration test, InboundMessage sender_name tests.
  • tests/test_feishu_mention_filter.py: Feishu group @mention filtering tests, config schema tests.
  • Added modular CLI command files:
  • miqi/cli/onboard.py
  • miqi/cli/agent_cmd.py
  • miqi/cli/gateway_cmd.py
  • miqi/cli/management.py
  • Added core regression tests:
  • tests/test_agent_loop_core.py
  • tests/test_cron_service_core.py
  • Added project documentation:
  • docs/API.md
  • docs/DEVELOPER_GUIDE.md
  • docs/ARCHITECTURE.md
  • CONTRIBUTING.md
  • Added Feishu business tools in miqi/agent/tools/feishu.py:
  • feishu_doc for cloud doc creation and optional plain-text write.
  • feishu_calendar for calendar event creation and attendee assignment.
  • feishu_task for task creation with group-member assignment.
  • feishu_drive for Drive folder creation (create_folder / ensure_folder_path) and workspace file uploads.
  • feishu_handoff as a generic collaboration handoff layer that orchestrates delivery steps.
  • Added Feishu mention metadata extraction (sender_open_id, mentions) in miqi/channels/feishu.py.
  • Added Feishu tool usage guidance in miqi/templates/TOOLS.md.
  • Added paper_download tool in miqi/agent/tools/papers.py to download PDFs into workspace.
  • Added documentation updates for Feishu collaboration + paper delivery workflow:
  • README feature/capability updates
  • API docs for paper_download, feishu_drive, and feishu_handoff
  • Feishu backend permission checklist in docs/API.md

Changed

  • Refactored miqi/cli/commands.py into an entry/compatibility layer that registers split command modules.
  • Refactored arXiv XML parsing in miqi/agent/tools/papers.py by extracting shared _parse_arxiv_entry logic.
  • Replaced remaining built-in print() calls with loguru.logger warnings in miqi/config/loader.py.
  • Split memory implementation into package modules with MemoryStore facade in miqi/agent/memory/store.py.
  • Normalized provider api_base handling in LiteLLMProvider to auto-fill missing default base paths (for example /v1, /api/v1) when users provide host-only URLs, while preserving explicit custom paths.
  • Updated AgentLoop tool wiring in miqi/agent/loop.py to:
  • auto-register Feishu business tools when Feishu credentials are configured.
  • propagate message context (channel, chat_id, message_id, sender_id, metadata) into Feishu tools for group assignment resolution.
  • include the generic Feishu handoff tool in runtime registration/context propagation.
  • Added paywall-aware PDF download behavior:
  • detects common login/paywall HTML responses and returns structured errors instead of saving invalid .pdf files.
  • Updated Feishu Drive upload implementation to support multipart upload (分片上传) for files > 20 MB:
  • _upload_file_multipart(): full 3-step flow — upload_prepare → upload_part × block_num → upload_finish.
  • Auto-routing in _upload_file(): ≤ 20 MB → upload_all, > 20 MB → multipart.
  • Added _FEISHU_UPLOAD_ALL_MAX_BYTES = 20 MB constant (per Feishu official docs 2024-10-23 hard limit).
  • Added feishu_drive action grant_permission: calls POST /open-apis/drive/v1/permissions/:token/members to grant doc/file access to a user or group chat.
  • Added _grant_permission() helper to FeishuToolBase (shared across doc/drive tools).
  • Added grant_chat_access parameter to feishu_doc: when true, auto-grants current group chat view access after doc creation.
  • Added docs/API.md corrections based on Feishu Open Platform official docs (2026-02-26 verification):
  • Permission Management vs Events & Callback setup guide reorganized.
  • Two-layer permission model explanation (scopes ≠ resource access).
  • Drive: full multipart upload flow documented; Adler-32 checksum requirement noted.
  • Drive: permission-member/create API fully documented including group-chat limitations.
  • Calendar: corrected HTTP method from GET to POST for calendar.v4.calendar.primary.
  • Task: corrected API version — v2 is the latest (POST /open-apis/task/v2/tasks, updated 2025-06-04).
  • Updated "Current Implementation Limits" to reflect new capabilities.

Fixed

  • Fixed Feishu Drive upload checksum algorithm: replaced MD5 (_md5_file) with Adler-32 (_adler32_file via zlib.adler32). The Feishu upload_all API requires Adler-32 decimal string; MD5 caused API error 1062008.
  • Fixed Moonshot/Kimi requests failing with 404 (/chat/completions) when apiBase was configured without /v1.
  • Fixed heartbeat false trigger: empty HEARTBEAT.md no longer fires a heartbeat notification (heartbeat/service.py).
  • Fixed reflection prompt (_REFLECT_PROMPT) being visible to users — demoted from role: "user" to role: "system" so it is excluded from session history and never sent to the user (agent/loop.py).
  • Fixed long-task max-iteration message: Chinese-language "processing limit" notification now correctly delivered to users (agent/loop.py).
  • Fixed progress messages suppressed under default config: milestone and MCP heartbeat notifications now set _tool_hint=False so they pass through when send_progress=True (default) (agent/loop.py).
  • Fixed queue position notification: elapsed time now computed from start_time (when task started processing) instead of enqueue_time (when task arrived) (agent/loop.py).
  • Fixed record_tool_feedback never called: tool execution results now feed into LessonStore for self-improvement learning (agent/loop.py).
  • Fixed stale HISTORY.md references: removed from context.py, AGENTS.md, skills/memory/SKILL.md, commands.py — aligned with RAM-first memory architecture.
  • Fixed LLM API errors silently passed as normal replies: finish_reason == "error" now returns a user-friendly error message instead of raw error content (agent/loop.py).
  • Fixed _save_turn filter for reflect prompt: updated to match new role: "system" and made tool result truncation configurable via session_tool_result_max_chars (agent/loop.py, config/schema.py).
  • Fixed cron.status() called before cron.start() in gateway startup: moved into async run() after service initialization (cli/gateway_cmd.py).
  • Fixed subagent missing paper tools: PaperSearchTool, PaperGetTool, PaperDownloadTool now registered in subagent with paper_config propagated from main agent (agent/subagent.py).
  • Fixed cron every-type schedule losing continuity on restart: _compute_next_run now uses last_run_at_ms + interval when available, with is not None guard for zero values (cron/service.py).
  • Fixed subagent defaults misaligned with main agent: temperature 0.7→0.1, max_tokens 4096→8192 (agent/subagent.py, agent/loop.py).
  • Fixed MessageBus queues unbounded: added maxsize=1000 default to prevent memory leak under load (bus/queue.py).
  • Fixed skills frontmatter YAML boolean parsing: now returns Python True/False instead of truthy strings, preventing always: false skills from being force-loaded (agent/skills.py).
  • Fixed queue notification metadata using _progress key: changed to distinct _queue_notification key with corresponding dispatch filter in ChannelManager, preventing double-filtering by send_progress config (agent/loop.py, channels/manager.py).
  • Fixed save_config() writing config file with default permissions (644): now calls chmod 0o600 after write to protect API keys from other OS users (config/loader.py).
  • Fixed AgentLoop.stop() leaving consolidation tasks orphaned on shutdown: stop() now cancels all in-flight _consolidation_tasks before clearing the running flag (agent/loop.py).
  • Fixed tool execution having no timeout: ToolRegistry.execute() now wraps each tool call with asyncio.wait_for(timeout=120s) to prevent one hung tool from blocking all agent processing (agent/tools/registry.py).
  • Fixed MCP reconnect retrying on every message with no delay: _connect_mcp() now uses exponential backoff (2 s → 4 s → … → 60 s cap) via _mcp_retry_after / _mcp_backoff_secs fields (agent/loop.py).
  • Fixed LiteLLM retry loop using linear backoff (0.5 * attempt): replaced with exponential backoff and random jitter (2^(attempt-1) * 0.5–1.0 s, capped at 30 s) (providers/litellm_provider.py).
  • Fixed json_repair.loads() silently fixing malformed LLM tool arguments with no visibility: now logs a WARNING message including the tool name and the original malformed JSON when repair is triggered (providers/litellm_provider.py).
  • Fixed cron _execute_job() having no execution timeout: now wrapped with asyncio.wait_for(timeout=600s) to prevent a stuck job from blocking the entire cron scheduler (cron/service.py).
  • Fixed MessageBus.publish_inbound() blocking forever when inbound queue is full: replaced await put() with drop-oldest strategy (logs warning, evicts oldest message); publish_outbound() now uses a 10 s wait_for timeout and logs an error if outbound consumers stall (bus/queue.py).
  • Fixed channel send failures silently dropping messages: _dispatch_outbound() now retries up to 3 times with linear backoff (0.5 s, 1.0 s) before logging an error and discarding (channels/manager.py).
  • Fixed ToolRegistry 120 s outer timeout overriding per-MCP toolTimeout: Tool base class gains execution_timeout property (default None); MCPToolWrapper overrides it to toolTimeout + 5 s; ToolRegistry.execute() prefers per-tool timeout when set (agent/tools/base.py, agent/tools/mcp.py, agent/tools/registry.py).
  • Fixed unbounded parallel session consolidations: AgentLoop now holds asyncio.Semaphore(5) used by every consolidation task to cap concurrency and prevent memory spikes (agent/loop.py).
  • Fixed critically underestimated MCP tool timeouts in configure_mcps.sh: raspa2 60 s → 21 600 s (6 h), zeopp 300 s → 600 s, mofstructure/miqrophi 120 s → 600 s — RASPA GCMC/MD simulations routinely run 4-5+ hours (scripts/configure_mcps.sh).
  • Fixed pdf2zh MCP timeout severely underestimated at 800 s: default raised to 3 600 s (1 h) in both miqi config pdf2zh and configure_mcps.sh — translating a 50+ page paper can easily exceed 30 min depending on LLM response time (cli/config_cmd.py, scripts/configure_mcps.sh).
  • Fixed max_tool_iterations default of 40 stopping complex scientific workflows mid-task: raised to 100. A RASPA GCMC + result parsing workflow can easily consume 30-50+ steps; 40 caused "⚠️ 已达到最大执行步数" errors. Configurable via agents.defaults.maxToolIterations in config.json (config/schema.py).
  • Fixed max_tokens default of 8192 being too small for DeepSeek-R1 and extended Claude 4 responses: raised to 16 000 in AgentDefaults. LiteLLM itself has no cap; the 8192 ceiling was miqi's own config default (config/schema.py).
  • Fixed DeepSeek-R1 (deepseek-reasoner) multi-turn conversations breaking when used via the official DeepSeek API: _sanitize_messages now preserves reasoning_content in assistant message history when the provider spec sets supports_reasoning_history=True; DeepSeek spec updated accordingly. OpenRouter/gateway paths are unaffected — reasoning is still stripped there (providers/registry.py, providers/litellm_provider.py).
  • Fixed agent returning a silent empty response when a reasoning model (e.g. DeepSeek-R1 via OpenRouter) exhausts max_tokens during thinking and emits no visible answer text: _strip_think returning None now logs a warning and surfaces "⚠️ 模型完成了推理但未输出最终回复（可能是 max_tokens 设置过低）" to the user instead of sending nothing (agent/loop.py).
  • Fixed cron _execute_job() hardcoded 600 s timeout killing long-running scientific jobs: CronService now accepts job_timeout parameter (default 86 400 s / 24 h) read from config.cron.job_timeout_seconds; gateway plumbs the value from config (cron/service.py, config/schema.py, cli/gateway_cmd.py).

Tests

  • Added provider routing regression tests for API base normalization behavior in tests/test_provider_routing.py.