Skip to content

CHANGELOG.md

Latest commit

 

History

History
50 lines (39 loc) · 3.49 KB

CHANGELOG.md

File metadata and controls

50 lines (39 loc) · 3.49 KB

Changelog

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

[0.2.19] - 2026-03-21

Added

  • OpenAI Skill: AIR SDK is now available as an OpenAI hosted shell skill. Upload via /v1/skills API, use with gpt-5.4 or gpt-5.4-mini. Requires network_policy allowlist for api.agentinternetruntime.com.
  • Content-type pre-router (S1): extract_url now detects RSS, Atom, and JSON API endpoints via HEAD request and parses them directly, bypassing WebKit. Fed RSS feeds extract at 85% confidence with 20+ items. GitHub Atom feeds return full release data.
  • JSON-API nested field extraction: jsonObjectToContentItem() now resolves nested fields (e.g., commit.message, commit.author.date) via dot-path traversal and one-level auto-traverse. Fixes GitHub API extraction returning only author logins.
  • Universal pattern abstraction (S2): After macro synthesis, abstract patterns (e.g., "search_form") are extracted and stored in sdk_universal_patterns. Execute_capability falls back to pattern_matched tier (0.6 confidence) for unindexed domains.
  • Mandatory diagnostics in MCP responses (S3): extract_url and browse_capabilities now include [AIR-DIAG] lines with method, confidence, items, cached status, credits, and tier distribution.
  • Browser observation capture (S4): report_outcome accepts browserObservations (pageStructure, contentQualityVsBrowser, extractionFallbackUsed). Domains flagged as browser_much_better get navigation_policy: 'requires_js_rendering' on capability_index.
  • SDK quality-gated billing: SDK keys no longer consume execution quota for meta-only/zero-value extractions. Previously, all SDK extract calls charged 1 execution regardless of quality.
  • Claude Desktop support in install-skill: install-skill now auto-detects Claude Desktop (macOS, Windows, Linux) alongside Claude Code, Cursor, Windsurf, and OpenClaw.
  • Absolute binary path in MCP configs: install-skill writes the full global binary path (e.g., /opt/homebrew/bin/air-sdk) instead of bare air-sdk, avoiding npx version caching issues.

Fixed

  • Updated MCP tool descriptions (S5): extract_url now mentions RSS/Atom/JSON API support and browser fallback guidance. browse_capabilities mentions universal patterns.
  • browserQuality and fallbackMethod fields now stored in sdk_synthesis_queue action_pattern (previously silently dropped by report.ts).

[0.2.12] - 2026-03-20

Fixed

  • report_outcome schema: success and selectorMatched now accept both boolean and string types (LLM agents often send "true" instead of true)
  • Cross-request correlation: execute_capability generates a requestId that links to report_outcome for richer telemetry
  • Heuristic execution hints for description_only capabilities — keeps agents in the AIR workflow even without verified macros

Added

  • Benchmarks: real API latency measurements across 8 domains, cost/speed comparison tables
  • Benchmark script (benchmark/run.ts) for independent verification

[0.1.1] - 2026-03-19

Fixed

  • Unified MCP server with extract_url tool
  • Telemetry field mapping for cloud ingestion
  • Version reporting consistency

[0.1.0] - 2026-03-18

Added

  • Initial release
  • Playwright, Puppeteer, and Browser Use adapters
  • Smart selector resolution with fallback cascading
  • Privacy-first telemetry with PII filtering
  • Capability discovery and macro execution
  • MCP server for Claude Code and Cursor
  • CLI init for zero-config setup
  • Machine Payments Protocol (MPP) support