Changelog

0.9.0

Minor Changes

• 8a71557: Add CLI subcommands and MCP server

  CLI:

  • augure stop — graceful daemon shutdown via PID file
  • augure status — read-only overview (PID, config, memory, jobs, channels, Docker, skills)
  • augure doctor — validate config, Docker, LLM, and Telegram connectivity
  • augure memory list|show|edit — inspect and edit memory files
  • augure jobs list|add|remove|run — manage scheduled jobs
  • augure channels status — check channel connectivity
  • augure tools list|test — inspect and test registered tools
  • augure mcp — standalone MCP server (stdio transport)
  • augure start --daemon — run agent as background daemon with PID file
  • augure start --mcp — enable MCP HTTP server alongside the agent

  MCP Server:

  • Tool bridge: all registered Augure tools exposed as MCP tools (including browser and skill tools)
  • Resources: memory files (memory://) and scheduled jobs (jobs://)
  • Prompts: persona system exposed as MCP prompts
  • Transports: stdio (for Claude Desktop/Cursor) and HTTP (StreamableHTTP on configurable port)
  • Path traversal protection on memory resource reads

  Types:

  • Added McpConfig interface (enabled, port)
  • Added mcp? field to AppConfig

0.8.0

Minor Changes

• c19fca6: Add tiered approval for high-risk tools and file-based Docker bridge for code-mode tool calls.

  Tiered Approval:

  • New riskLevel field on NativeTool — tools marked "high" require explicit user approval
  • Channel-agnostic ApprovalGate with timeout auto-reject and fallback auto-approve
  • Telegram implementation using InlineKeyboard approve/reject buttons
  • sandbox_exec, opencode, and manage_skill marked as high-risk
  • Configurable via approval.enabled and approval.timeoutMs in augure.json5

  Docker Code-Mode Bridge:

  • Real tool execution from Docker containers via file-based bridge
  • Container harness writes .bridge-req-{id}.json, host polls and responds with .bridge-resp-{id}.json
  • 120s timeout prevents infinite poll inside container
  • Atomic temp+mv writes avoid partial reads

0.7.2

Patch Changes

• 2f42a7c: fix(scheduler): add structured logging and fix startup order preventing jobs from firing

  • CronScheduler now accepts a Logger via CronSchedulerOptions, replacing raw console.log with the standard Logger interface (info/debug/warn/error)
  • Added diagnostic warning when a one-shot job is added while the scheduler is not running
  • Added .catch() error handling on fire-and-forget promises in cron and one-shot callbacks
  • Fixed critical startup bug: await tg.start() (grammY long-polling) blocks forever, so scheduler.start(), heartbeat.start(), job trigger handlers, shutdown handlers, and update timers were never executed. All setup now runs before the blocking Telegram call.

0.7.1

Patch Changes

• d26c70d: Add browser automation via Stagehand. New @augure/browser package with BrowserSessionManager. Session-based browser NativeTool for LLM with act/extract/observe/screenshot actions. Supports local Playwright and Browserbase cloud providers.

0.7.0

Minor Changes

• 1f2df3d: Add GitHub tool with 17 actions (issues, PRs, repos, releases, search) via Octokit

0.6.1

Patch Changes

• 51f00d2: Add LLM call latency and per-message token/performance summary logs

0.6.0

Minor Changes

• 5d7e392: Add code mode support — LLM writes TypeScript calling typed APIs in a sandbox instead of individual tool calls

0.5.0

Minor Changes

• 406b451: ### Scheduler: one-shot jobs with runAt

  Jobs can now use an ISO 8601 runAt field instead of cron for one-shot scheduling. One-shot jobs fire once via setTimeout and are automatically removed after execution. Expired jobs are discarded on load.

  Sandbox: auto-build Docker image

  The sandbox Docker image is now built automatically at startup if it doesn't exist locally, removing the need for manual docker build setup.

  Core: startup auto-updates and job dispatch

  • Skill auto-update checks run at startup and periodically via SkillUpdater
  • CLI version check at startup with optional periodic Telegram notifications
  • Scheduled job triggers are dispatched to the agent and forwarded to Telegram
  • VersionChecker simplified: removed githubRepo, handles v-prefixed and pre-release versions

  Fixes

  • core: malformed tool call arguments from LLM no longer crash the agent loop (graceful fallback to {})
  • channels: GrammyError.error_code is now detected in retry logic alongside status
  • skills: SkillUpdater rejects downloaded skills with sandbox: false (defense-in-depth)

0.4.0

Minor Changes

• 7c430c9: Telegram channel overhaul and auto-update system

  Channels:

  • Add middleware pipeline (Koa-style) for outgoing messages: split, escape, retry
  • Refactor TelegramChannel with MarkdownV2 support, media handling (photos & documents), and plaintext fallback
  • Per-userId conversation isolation to prevent cross-talk
  • Graceful bot shutdown on SIGINT/SIGTERM
  • Custom rejectMessage for unauthorized users
  • Move FunctionSchema to @augure/types for proper layering; pass tools via LLM API instead of system prompt

  Auto-updates:

  • Add UpdatesConfig type for skill and CLI update settings
  • Add SkillUpdater for auto-updating hub skills with test-and-rollback
  • Add VersionChecker for CLI update notifications via npm registry

  CLI:

  • Add --env / -e flag for custom .env file path

  Docs:

  • Add channels overview and Telegram setup guide to Fumadocs site

0.3.0

Minor Changes

• 5be48b4: Bundle CLI with tsup for standalone npm distribution, add email tool (IMAP/SMTP), and add changelog + CLI reference to docs.

0.2.0

Minor Changes

• 00fae0e: Initial release: CLI with augure start and augure init commands, npm publishing pipeline with changesets, and GitHub Actions CI/CD.