saggamer/orchestrator-v1.5

Orchestrator v1.5

Orchestrator v1.5 is an everyday agentic model system for research, browser workflows, coding coordination and connected-app tasks in Kira Superapp. Its task-routed expert adapters and bounded Diagonal orchestration bring action proposals, tool observations and follow-up execution into one local workflow. The published Kira evaluations below describe measured model-plus-controller results.

Kira Superapp ecosystem

Kira Superapp brings AI chat, browser research, coding, connected apps, MCP tools, voice and artifact workflows into a local workspace. The application supplies the tools, execution loop, memory and permission handling around compatible models. Each model's specific runtime and integration requirements are described in this card.

Explore Kira Superapp · GitHub and installation · Diagonal Browser source

Orchestrator v1.5 is Kira Superapp's local, Apple Silicon research-preview model bundle for everyday assistance, browser work, and multi-step agentic tasks. It combines a Gemma 4 E4B instruction base with task-routed MLX adapters and Kira's bounded execution controller. Kira handles tool execution and returns observations to the model before it continues.

What it does well

• Carries browser research into follow-up actions: reads linked sources, completes observed forms, saves once, and checks the visible result.
• Handles everyday browser tasks including page reading, navigation, search, form entry, scrolling, comparison, and screenshot capture through Diagonal Browser.
• Uses Kira Superapp's Diagonal mode to evaluate independent action proposals against observed evidence within a finite step budget.
• Continues conversations with Kira's local chat history and retrieved preferences, while treating remembered text as context rather than a fresh instruction.
• Routes app, coding-agent, browser, and presentation requests to their corresponding Kira control paths.
• Discovers tools from an approved Kira MCP app connection and uses advertised read-only tools with schema-checked arguments.

Measured Kira results

Project evaluation	Result
Everyday Diagonal Browser suite	16/16 passed
Long Diagonal workflows	6/6 passed
Research followed by verified browser execution	2/2 passed
Isolated Notes-like MCP read task over stdio	1/1 passed

These scores are from Kira Superapp's local task suites. They measure the complete model-plus-controller workflow. In the research-to-execution suite, the three-source recommendation task finished in 53.05 seconds with eight browser actions; the public-web documentation-to-form task finished in 112.73 seconds with seven actions.

Run with Kira Superapp

Use the Kira Superapp runtime on Apple Silicon. Select Orchestrator v1.5 and choose low, medium, high, or Diagonal thinking effort. The browser and app actions are performed by Kira's tool handlers, so the adapters are intended for the matching Kira runtime rather than standalone text generation.

Keep the downloaded bundle together on a local drive. The runtime_overlay folder contains the matching Kira controller and browser files for this preview; review and apply it to a Kira Superapp checkout. Install requirements-orchestrator-v15.txt, then point Kira to the bundle manifest:

export KIRA_V15_MANIFEST_PATH="/path/to/orchestrator-v1.5/manifest.json"
.venv-orchestrator-v15/bin/python interface.py

The bundle includes a local MLX conversion of the Gemma 4 E4B instruction model as its base. The release manifest records exact base and adapter hashes and task routing. The base model and third-party components retain their original attribution and license notices.

Runtime update — October 26, 2026

This update keeps the published base and adapter weights unchanged. It adds native PDF/Word routing to Orchestrator v1.5, corrects negative web/image intent, repairs document output formatting, and connects the public thinking selector to its execution settings. Browser and MCP skills and model-selected tool planning are included.

Live local checks verified one generated one-page PDF and action proposals at all four effort levels. Low used one proposal; Medium added a decision memo; High used two proposals and a memo; Diagonal used three branches plus independent evaluation. These are functional checks, not a new general benchmark. The focused regression suite previously passed 172 tests.

Apply the updated runtime_overlay to the matching Kira Superapp checkout, install the project dependencies (including fpdf2 and PyMuPDF for document generation/verification), and restart Kira. Browser profiles, generated screenshots and Python environments are excluded from the current release.