Harness extensibility research¶

Research date: September 6, 2026. This is a source/documentation comparison, not a performance benchmark or a claim that every upstream feature was runtime-tested. Upstream default branches and experimental contracts can change.

This page records research and design direction, not a feature-parity promise. For implemented behavior and current limitations, use the ngn guide and release availability.

Findings¶

Project	Context and compaction	Extensions	Lesson for nagents
OpenCode	Per-request message/system transforms; compaction hook can append context or replace the summary prompt	TypeScript plugins, tool/permission hooks, custom tools, agent profiles, on-demand skills, MCP, commands	Separate observer events from hooks that can change execution; expose more than a compaction prompt
pi	Non-destructive context transforms; custom compaction checkpoint with retained-history boundary	TypeScript SDK/extensions, tool replacement, custom tools and terminal components, Python-accessible JSONL RPC	Keep agent core, coding harness, and terminal independent; native Python should not require a JS bridge
OpenClaw	Replaceable context engine owns ingestion, assembly, and optionally compaction	TypeScript plugins, typed gates/modifiers/observers, skills, per-agent workspaces, local or gateway TUI	Make context ownership explicit; permission checks and run lifecycle need defined semantics
Grok Build	Automatic compaction and preservation instructions; pre/post compaction hooks are passive	Shell/HTTP hooks, MCP, skills/plugins, subagent definitions, headless streaming	Hooks are not necessarily algorithm replacement; document authority and failure behavior
Community grok-cli	Token reserves, recent-history retention, configurable summary instructions	TypeScript/Bun application with lifecycle hooks, MCP, skills and subagents	Internal exported functions do not constitute a stable compaction-plugin API
Grok Bot	Persistent cloud-computer agents; no custom context engine verified	Skills, routines, connectors/MCP, approvals	Persistent agents sharing a computer are not separate security boundaries
LangChain middleware	Python before-model and wrapped model/tool calls, request overrides	Typed middleware on LangGraph	Borrow useful function-level contracts without importing graph orchestration

"Grokbot" is ambiguous. Grok Bot, official Grok Build, and the community grok-cli are different projects. Clawdbot is an earlier name of OpenClaw, not an additional current framework. The comparison does not assume which name was intended.

Important distinctions¶

Skills are instructions and resources loaded on demand, not executable hooks or a permission system.
Changing a summary prompt is not the same as replacing a compaction algorithm.
A model context is a projection of the transcript. Request-time filtering must not silently destroy the stored conversation.
Permission gates must fail closed and approve the arguments that actually execute, after any plugin transformation.
Trusted Python plugins have the privileges of the host process. Path checks on built-in tools do not sandbox a plugin, a shell, or an MCP server.
Cancelling a UI worker is not sufficient to terminate an underlying thread or subprocess. Stopping execution and repairing conversation state are separate.
A crash after a side effect but before its result is stored cannot be repaired by blindly retrying the tool. Record an unknown outcome and inspect it instead.

Direction for ngn¶

The library exposes ordered Python hooks for incoming messages, model requests, tool calls/results, event observation, and cleanup. A separate compaction strategy can select its own trigger and replace the active message context, not just the summary prompt. Old transcript rows remain available in storage.

The coding harness supplies workspace discovery, instructions/skills, guarded file tools, managed shell execution, approval requests, and session selection. The TUI and headless CLI use that same harness. Neither contains the LLM loop.

Executable project configuration is opt-in. Even a project-controlled API endpoint could expose a user's provider credential, so trust applies to more than .py files. Configuration and diagnostics must make this boundary visible.

The initial implementation deliberately does not attempt feature parity with these products: no marketplace, remote daemon, agent swarm, worktree orchestrator, LSP client, or arbitrary terminal-component plugin API is required to validate Python-extensible harness behavior.

Primary sources¶

The Grok Bot design articles distinguish state-driven avatars from orchestration: idle, thinking, working, waiting, blocked, and done are visual lifecycle states. Actual collaboration is asynchronous messaging between agents with separate role context. Some notch/cursor-companion animations are explicitly exploratory prototypes, not shipped interaction modes. For ngn, a future team view should render real task events rather than simulate peer collaboration the backend does not support. Color themes and collaboration views are separate concerns.

Chrome was started for browser research. Google returned an automated-traffic challenge, and the supplied SearXNG hostname did not resolve in this environment. Official documentation and upstream source were reachable directly and used for the findings above.