v26.8.6

🔭 Vision & Analysis

• Multi-Modal analyze_local_image Upgrade (analyzeImage.ts): Migrated from hardcoded Gemini SDK to the user-configured LLM provider via executeWithRetry. Now supports any OpenAI-compatible vision provider (OpenAI, Anthropic, Gemini, xAI, etc.) using the standard image_url message format. Added automatic content-type detection (screenshot, chart, diagram, document, photo) to inject specialized analysis instructions per content type.
• Native PDF Analysis Skill (analyzePdf.ts): Introduced analyze_pdf skill wrapping the existing extract_pymupdf.py playbook script. Supports three modes — text (raw extraction), structured (pages + metadata), and summary (LLM-generated summary of extracted content). Gracefully handles encrypted and scanned PDFs, returning page count and word count metadata.
• Chart & Diagram Analyzer (analyzeChart.ts, vision.py): Added analyze_chart skill backed by a new POST /vision/analyze-chart ML Engine endpoint. Returns structured data: title, axes, data points, trend direction, and natural-language key insight. Uses the configured LLM's vision capabilities — no hardcoded provider.
• Visual Verification Loop (verifyVisual.ts, vision.py): Added verify_visual_output skill backed by POST /vision/verify-screenshot. Agent can autonomously screenshot → describe expected state → verify → iterate. Returns { matches, confidence, issues[], suggestions[] }. Pairs naturally with the existing computer_use skill for full autonomous UI testing.
• ML Engine Vision Router (packages/ml-engine/routers/vision.py): New FastAPI router registered at /vision/* providing chart analysis and visual verification endpoints. Uses configured LLM provider via config.py — no hardcoded API keys.

🔧 Software Engineering Upgrade

• Autonomous Test Runner (runTestsAndFix.ts): New run_tests_and_fix skill that executes the project's test command, parses output (vitest, jest, pytest, cargo test), and returns a structured report: passing count, failing count, per-failure details (file, line, error). Designed as an iterative tool — LLM reads results, applies fixes via edit_local_file, then calls again. State tracked via ~/.nyxora/test_loop_state.json. Max iterations configurable (default 5).
• Migration Plan Orchestrator (planMigration.ts): New plan_migration skill that scans the codebase for affected files, assesses migration scope, and generates a structured step-by-step migration plan with recommended execution order, module batching, and test gates. Returns actionable plan text — execution is handled by the LLM via existing tools.
• Software Engineering SOP Prompts (3 new cognitive skills):
  • large-scale-refactor.md: Pre-refactor checklist, batch-by-module strategy, rollback triggers, progress tracking with todo_write.
  • autonomous-testing.md: Test-fix loop strategy, flaky test detection, regression prevention, escalation rules.
  • migration-orchestration.md: Scope assessment, incremental migration, circular dependency handling, checkpoint + rollback strategy.
• Cognitive Manager SOP Mappings (cognitiveManager.ts): Registered keyword triggers for all 3 new SOPs — migrate/migration/migrasi, run tests/fix tests/test loop, large scale/bulk change/refactoring — enabling automatic SOP injection when relevant tasks are requested.

🎯 Long-Horizon Task System

• Persistent Goal Manager (goalManager.ts): Introduced goalManager — a JSON-backed persistent store at ~/.nyxora/goals.json for multi-day autonomous tasks. Supports full lifecycle: createGoal, advanceStep, updateCheckpoint, pauseGoal, resumeGoal, markComplete, markFailed. Auto-pauses goals after 3 consecutive failures. Unlike cronManager (recurring schedules), goalManager handles one-time multi-step tasks that span days.
• Background Goal Worker (goalWorker.ts): startGoalWorker() polls active goals every 5 minutes. For each due goal, resumes from last checkpoint, calls processUserInput with accumulated context, saves progress, then advances the step counter. Pushes Telegram notifications on step completion, milestone, and failure. Auto-starts 30 seconds after daemon boot to allow services to initialize.
• Goal Management Skills (goalSkills.ts): Four new LLM-callable tools: create_long_term_goal (register a multi-step async goal), get_goal_status (formatted overview of all goals), pause_goal / resume_goal (lifecycle control).
• Active Goals System Prompt Injection (promptBuilder.ts): Active goals are now injected into every system prompt turn via getGoalSummary(), keeping the LLM aware of ongoing long-running work without requiring the user to re-explain context.
• Goal Worker Daemon Integration (launcher.ts): startGoalWorker() is now called at daemon startup, ensuring long-horizon tasks survive restarts and automatically resume from their last checkpoint.

Bug Fixes

• Streaming Tool Interceptor (toolInterceptor.ts): Fixed 4 critical bugs — raw JSON leak on stream end, buffer overflow inside tool block, tool_params type safety, and empty payload guard.
• Telegram Tool Status Flood (telegram.ts): Fixed onProgress sending a new Telegram message per tool call when Computer Use tool runs multiple sequential actions (click, key, screenshot, etc.). Reverted to intended multi-bubble behavior while fixing the root cause.
• Computer Use XML Parameter Leak (osAgent.ts): getToolLabel() was passing raw firstArgValue for computer tool which contained multiline XML-like parameter tags (<parameter=coordinate> [100, 100]). Fixed by: (1) reading args.action key directly for computer tool, (2) adding sanitizeLabel() helper that strips all <parameter=...> XML tags and collapses whitespace across all tool labels.
• Self-Healer 5 Bugs (selfHealer.ts, agentskills.ts): Fixed invalid max_tokens field, broken ?t= query string cache-bust (replaced with proper require.cache deletion), greedy first-match code block regex (now takes last/largest match), over-strict export sanity check (added module.exports support), and wasteful LLM calls on un-healable runtime errors (added isHealableError() filter).

Installation & Distribution

• One-Line Installer Script (docs/public/install.sh): Linux & macOS zero-warning installer via curl.
• Windows Installer (docs/public/install.ps1): PowerShell installer via winget/Chocolatey.
• 3-Option Installation Documentation: npm install -g, curl installer, and developer source install.

Full Changelog: v26.8.5...v26.8.6

Full Changelog: v26.8.5...v26.8.6