Latest Features in Apache Maka: Runtime Revamp, Typed APIs, and CLI Tools

Apache Maka (Incubating) versions 0.2.0 and 0.1.11 introduce a typed request() API, consolidated RuntimeKernel, SessionTodo work items, active tool-result pruning, and a new /transcript TUI command with full CLI packaging.

Apache Maka is an incubating Apache project building a universal runtime for AI agents. The recent 0.2.0 (Unreleased) and 0.1.11 (2026-08-18) releases bring fundamental architectural changes to how agents execute tools, manage context, and expose interfaces to users. These latest features in Apache Maka streamline the execution layer while adding practical utilities for both developers and end users.

Core Runtime Revamp

The 0.2.0 release restructures the execution layer around a single RuntimeKernel, consolidating the previous RuntimeRunner, Flow, and Invocation abstractions into one authoritative module.

Typed request() API

The runtime now exposes a single, type-safe entry point in packages/runtime-host/src/request.ts. All legacy forwarding aliases were removed, leaving one public method that supports subscriptions, capabilities, listeners, and lifecycle hooks.

import { request } from '@maka/runtime-host';

const response = await request({
  model: 'qwen3.8-max',
  prompt: 'Summarize the Apache Maka architecture in 3 sentences.',
  maxTokens: 500,
});

console.log(response.output);

RuntimeKernel Architecture

Located in packages/runtime/src/runtimeKernel.ts, the RuntimeKernel now owns backend dispatch, terminal coalescing, stop/drain handling, and durable continuation admission. This replaces the fragmented execution shell of earlier versions, ensuring consistency across all agent turns.

SessionTodo and Unified Context

SessionTodo (packages/runtime/src/sessionTodo.ts) replaces the deprecated Task Ledger, becoming the single authority for in-session work items. The schema no longer contains the workflow_task_ledger_events table. Coupled with unified context management in packages/runtime/src/context.ts, the Runtime now owns all compaction, pruning, and token-usage accounting. Environment variables like MAKA_CONTEXT_* were removed, handing full control to the kernel.

Active Tool-Result Pruning

Enabled by default in packages/runtime/src/toolResultPruning.ts, this feature automatically archives tool results exceeding a configurable token threshold, replacing them with lightweight placeholders to prevent runaway context growth. Control the behavior via environment variable:


# Enable pruning (default)

export MAKA_CONTEXT_ACTIVE_TOOL_RESULT_PRUNE=on

# Disable pruning for debugging large outputs

export MAKA_CONTEXT_ACTIVE_TOOL_RESULT_PRUNE=off

Manual Model Registration

The connection catalog in packages/mcp/src/connectionCatalog.ts now supports per-connection model overrides, capacity fields, and atomic saving. This allows fine-grained control over model parameters without modifying global catalogs, feeding directly into the Model Adapter used by the Runtime Kernel.

User-Facing Enhancements

Version 0.1.11 focuses on accessibility, security isolation, and deployment options across desktop and command-line interfaces.

/transcript TUI Command

Long sessions no longer require terminal scrollback. The new command in apps/desktop/src/tui/commands/transcript.ts provides a dedicated pager interface supporting line, page, and first/last navigation. Inside the TUI, type:


/transcript

Navigation shortcuts:

  • j / k — Scroll line up/down
  • Ctrl-f / Ctrl-b — Page forward/backward
  • g / G — Jump to first/last line

Work Board Storage Foundations

New UI components in apps/desktop/src/workboard/ provide interfaces for creating, managing, and persisting work items scoped to a specific host, laying the groundwork for persistent task management.

Windows AppContainer Sandbox

Security isolation improved with brokered sandbox support for Windows in native/windows-sandbox/. This provides stronger boundaries for tool execution, isolating agent processes from the host system.

Qwen 3.8 Support

Built-in support for the Qwen 3.8 model with an 8,000-token plan is available in packages/mcp/src/models/qwen38.ts, accessible immediately through the model registry.

Packaged CLI

The new @maka/cli package (documented in packages/cli/README.md) distributes the maka command via npm, supporting maka run, maka eval, and maka cli:dev without requiring a full source build.

npm i -g @maka/cli
maka run "Summarize the latest features in Apache Maka"

Architecture Integration

The redesigned architecture centers on a single Runtime Host per state root. All front-ends—Desktop, TUI, CLI, bots, and Eval—dispatch work through the typed request() API to this host.

The RuntimeKernel manages:

  • SessionManager — Session-level identity and lifecycle
  • AgentRun — Per-turn execution, tool runtime, and event logging
  • Tool Runtime — Validation, permission checks, watchdog handling, and artifact recording
  • Runtime Event Log — Append-only source of truth for every model message and tool call

SessionTodo feeds the Agent Graph with versioned work items, while active tool-result pruning keeps the Runtime Event Log lean, preserving full history in the durable off-load store (packages/runtime/src/readImage.ts).

Code Examples

Configuring the Typed Request API

import { request } from '@maka/runtime-host';

// Query with the new unified interface
const result = await request({
  model: 'qwen3.8-max',
  prompt: 'Explain the RuntimeKernel consolidation',
  capabilities: ['file-system', 'network']
});

Creating SessionTodo Work Items

import { createTodo } from '@maka/runtime';

await createTodo({
  title: 'Investigate token-usage anchors',
  description: 'Verify that token_usage now records model+connection.',
  dueDate: new Date(Date.now() + 24 * 60 * 60 * 1000)
});

Use the TUI command and keyboard shortcuts to browse session history without scrollback buffers, reading directly from the Runtime Event Log via the new Read API.

Summary

  • RuntimeKernel consolidates execution logic previously split across RuntimeRunner, Flow, and Invocation into a single module in packages/runtime/src/runtimeKernel.ts
  • Typed request() API provides the sole public entry point to the Runtime Host, removing legacy forwarding aliases and standardizing the interface in packages/runtime-host/src/request.ts
  • SessionTodo replaces the Task Ledger with a cleaner schema and single authority for in-session work items, eliminating the workflow_task_ledger_events table
  • Active tool-result pruning (default-on) archives large outputs to control token usage, configurable via MAKA_CONTEXT_ACTIVE_TOOL_RESULT_PRUNE
  • /transcript TUI command enables scroll-free navigation of long sessions using line and page shortcuts in apps/desktop/src/tui/commands/transcript.ts
  • Windows AppContainer sandbox adds brokered isolation for tool execution on Windows in native/windows-sandbox/
  • @maka/cli package distributes the runtime via npm with maka run, eval, and cli:dev commands

Frequently Asked Questions

What is the RuntimeKernel in Apache Maka?

The RuntimeKernel is a consolidated execution engine introduced in version 0.2.0 that replaces the previous RuntimeRunner, Flow, and Invocation abstractions. Implemented in packages/runtime/src/runtimeKernel.ts, it owns backend dispatch, terminal coalescing, stop/drain handling, and durable continuation admission, providing a single authority for all runtime operations according to the Apache Maka source code.

How do I disable active tool-result pruning?

Set the environment variable MAKA_CONTEXT_ACTIVE_TOOL_RESULT_PRUNE to off before starting the runtime. This prevents automatic archival of large tool outputs, keeping full results in memory for debugging. The default value is on, which archives results exceeding the token threshold to a durable off-load store managed by the Runtime.

What replaced the Task Ledger in recent Apache Maka versions?

SessionTodo replaces the deprecated Task Ledger as of version 0.2.0. Located in packages/runtime/src/sessionTodo.ts, it serves as the single authority for in-session work items, and the underlying schema no longer contains the workflow_task_ledger_events table previously used for task tracking.

How does the new /transcript command improve session navigation?

The /transcript command, available in the TUI via apps/desktop/src/tui/commands/transcript.ts, eliminates the need for terminal scrollback by providing a dedicated pager interface. Users navigate using j/k for line movement, Ctrl-f/Ctrl-b for paging, and g/G to jump to the start or end of the session log.

Have a question about this repo?

These articles cover the highlights, but your codebase questions are specific. Give your agent direct access to the source. Share this with your agent to get started:

Share the following with your agent to get started:
curl -s "https://instagit.com/install.md"

Works with
Claude Codex Cursor VS Code OpenClaw Any MCP Client

Maintain an open-source project? Get it listed too →