Summary

Integrate context-mode natively into OpenHuman so that all sub-agents benefit from automatic context window protection, session continuity across compaction, and sandbox code execution - without any manual configuration by the operator. This is the same capability already built into Claude.ai and adopted across Claude Code, Gemini CLI, VS Code Copilot, Cursor, and 11 other platforms.

Background

context-mode is an open-source MCP server (mksglu/context-mode) that solves four sides of the context window problem simultaneously. It hit #1 on Hacker News with 570+ points and is in active use across teams at Microsoft, Google, Meta, Amazon, NVIDIA, Stripe, GitHub, Canva, Notion, Supabase, Cursor, and Salesforce.

The four problems it solves

Benchmark savings

Session time extends from ~30 minutes to ~3 hours per run.

Platform support (context-mode is already on 15 platforms - OpenHuman is not one of them)

Problem

Every MCP tool call made by OpenHuman or a sub-agent dumps raw data into the context window. A Playwright snapshot costs 56 KB. Twenty GitHub issues cost 59 KB. After 30 minutes of a concurrent multi-agent run, 40% of the context window is gone. When the agent compacts to free space, it loses track of which files it was editing, what tasks are in progress, and what was last requested.

OpenHuman may not have sub-agent:

• Context sandboxing - no mechanism to keep raw tool output out of context
• Session continuity - no structured event tracking; compaction causes full state loss
• Sandbox code execution - no "think in code" primitive; agents process data inline, burning context from both sides
• BM25 knowledge retrieval - no indexed, searchable session knowledge base

At the scale OpenHuman operates - many concurrent agents, long-running tasks, deep tool call chains — this is a significant reliability and cost gap.

Solution

Scope: core (MCP tool registration + hook wiring) + config

Integration approach - native, not bolted-on:

Rather than asking operators to install context-mode manually as a separate MCP server, integrate it as a first-class native dependency of the OpenHuman agent runtime. When enabled, context-mode is:

1. Bundled with the OpenHuman install (npm install -g context-mode called by the installer)
2. Registered automatically as an MCP server available to all agents
3. Hooks wired at the agent runtime level (PreToolUse, PostToolUse, PreCompact, SessionStart) for each agent session
4. Routing enforced programmatically - agents are nudged to prefer sandbox tools over raw Bash/Read/WebFetch

This mirrors how Claude Code ships context-mode as a plugin marketplace integration — one install, fully automatic, no operator configuration required.

The 11 MCP tools exposed to every agent:

Session continuity for OpenHuman and sub-agents:

Every meaningful event during an agent session is captured and persisted in SQLite:

• File reads, edits, writes
• Git operations
• Task creation, update, completion
• Errors and error resolutions
• User decisions and corrections
• QC pass/fail/retry events

When the context window compacts, the PreCompact hook builds a priority-tiered snapshot (≤2 KB). The SessionStart hook restores it. The agent continues from the last prompt with full working state - no re-prompting, no state loss.

Configuration in config.yaml:

context_mode:
  enabled: true
  routing: hooks          # hooks (recommended) | instructions-only
  max_idle_timeout_ms: 900000
  per_agent: true         # each agent gets its own session DB

Fallback: context_mode.enabled: false disables the integration entirely. No existing behaviour changes.

Acceptance criteria

• Bundled install - context-mode is installed as part of the OpenHuman install process; no manual operator step required
• MCP server registered - all 11 context-mode MCP tools are available to Hermes and all sub-agents
• Hook wiring - PreToolUse, PostToolUse, PreCompact, and SessionStart hooks are wired at the agent runtime level
• Routing enforcement - agents are routed to sandbox tools (ctx_execute, ctx_batch_execute, ctx_fetch_and_index) over raw Bash/Read/WebFetch for large-output operations
• Session continuity - agent state (files, tasks, errors, decisions) survives context compaction and is restored automatically on SessionStart
• Per-agent session DB - each agent maintains its own SQLite session database; sessions do not share state
• ctx_stats observable - operators can run ctx stats in any agent session and see per-agent context savings
• ctx_doctor passes - ctx doctor reports all checks passing for the OpenHuman integration
• Config-driven - context_mode.enabled and routing mode are read from config.yaml
• No-op when disabled - context_mode.enabled: false disables the integration with no behaviour change
• Diff coverage ≥ 80% - the implementing PR meets the changed-lines coverage gate (Vitest + cargo-llvm-cov, enforced by .github/workflows/coverage.yml)

Related

• context-mode repository: https://github.com/mksglu/context-mode
• context-mode Hacker News Feat/gitbooks #1: https://news.ycombinator.com/item?id=47193064
• context-mode npm: https://www.npmjs.com/package/context-mode
• context-mode YouTube demo: https://www.youtube.com/watch?v=QUHrntlfPo4
• OpenHuman agent architecture: https://tinyhumans.gitbook.io/openhuman/
• Proposed: Global MCP tool registry (FR10)
• Proposed: Native MCP server exposure (FR8)
• Proposed: SearXNG native MCP tool (FR5)