harness-integration-stinger

The integration playbook for harness-integration-worker-bee. Encodes how to wire one capability (a skill, an agent, a hook-driven behavior, an MCP-backed tool) across The Hive's four harnesses - Claude Code, Cursor, ChatGPT Codex, Claude Cowork - from the first decision (which component type, which wiring mechanism) through capability detection, graceful degradation, portability, and distribution. Includes a full worked case study of a real six-host integration (Hivemind, a prior product this stinger originally documented) as a concrete example of every decision point answered end to end.

Quick navigation

Task	Guide
Start here: what integration means, the four harnesses, the wiring-mechanism decision matrix	guides/00-decision-framework.md
Where does each component type (rules, commands, agents, skills, plugins) live, per harness	guides/01-component-placement.md
Wire hooks/lifecycle events, per harness, and know the real shared-event floor	guides/02-hook-lifecycle.md
Register an MCP server, per harness (JSON vs. TOML, cloud reachability)	guides/03-mcp-registration.md
Detect what a harness supports and degrade gracefully when it doesn't	guides/04-capability-detection-and-degradation.md
Keep a skill, rule, or tool contract portable (spec-six frontmatter, AGENTS.md, plugin manifests)	guides/05-portability-and-contracts.md
Ship through each harness's marketplace/install flow and its distribution gate	guides/06-distribution-and-audit.md
Worked example: a real six-host integration end to end	examples/case-study-hivemind-six-host-installer.md

Critical directives

These are the non-negotiables. Violating any of them is the most common cause of a broken harness adapter. See the relevant guide for code patterns.

1. Keep the tool and command contract identical across every host. hivemind_search/hivemind_read/hivemind_index (plus hivemind_goal_add/hivemind_kpi_add on OpenClaw) must have the same name, args, and return shape everywhere. A drift in one host silently breaks cross-harness recall. Source: research/external/2026-06-16-tool-contract.md.

2. Hooks must be fast and fail-open. Capture hooks run on the agent's critical path. Honor the per-event timeout (SessionStart 10s, PreToolUse 60s, capture 10-30s), dispatch heavy work async: true, and never let a hook crash block the host. Source: research/external/2026-06-16-hook-lifecycle.md.

3. Capability detection must be cheap and side-effect free. hivemind install auto-detects each assistant by probing for its home dir / binary (~/.claude/projects, ~/.codex, ~/.cursor, ~/.hermes, ~/.pi, OpenClaw). Detection runs on every install; it must not write files or spawn work. Source: research/external/2026-06-16-capability-detection.md.

4. Never hardcode bundle paths - resolve them per host. Claude Code forks node "${CLAUDE_PLUGIN_ROOT}/bundle/<entry>.js"; Cursor/Hermes use ~/.<host>/hivemind/bundle/. Use the host's own root variable so the marketplace plugin and local installs both resolve correctly. Source: research/external/2026-06-16-architecture-build.md.

5. The OpenClaw bundle must pass the ClawHub static scanner. ClawHub forbids bare spawn/execFileSync. Route subprocess access through createRequire-based indirection (see src/skillify/gate-runner.ts comments and scripts/audit-openclaw-bundle.mjs) or the bundle is rejected. Source: research/external/2026-06-16-openclaw-clawhub.md.

6. Register the MCP server only where the host supports it. Hermes wires the MCP server (src/mcp/server.ts) under mcp_servers.hivemind in ~/.hermes/config.yaml. Do not assume every host has an MCP transport - Claude Code and Cursor use hooks; pi/OpenClaw use native extensions. Source: research/external/2026-06-16-mcp-registration.md.

7. pi ships raw TypeScript; do not pre-compile it. harnesses/pi/extension-source/hivemind.ts is delivered as .ts and pi compiles it at load. Bundling or transpiling it in the installer breaks the load path. Source: research/external/2026-06-16-pi-extension.md.

8. Author portable skills against the six-field Agent Skills spec only. Outside Claude Code proper, only name, description, license, compatibility, metadata, allowed-tools are legal SKILL.md frontmatter. Any Claude-Code-only field (context: fork, disable-model-invocation, paths, hooks, etc.) throws a hard packaging error on claude.ai uploads, the Skills API, and Cowork's "Save skill" pipeline. Source: research/external/2026-08-14-agentskills-spec-six-fields.md; queen-bee-stinger distilled-research-articles.md, Claude Code §SUPPLEMENT.

9. Codex MCP config is TOML with an underscored mcp_servers key, not JSON mcpServers. A config block copy-pasted from Claude Code or Cursor silently fails on Codex - no parse error, the server just never registers. Always write the Codex TOML form as its own case. Source: research/external/2026-08-14-mcp-capability-negotiation.md; queen-bee-stinger distilled-research-articles.md, ChatGPT Codex §Plugins.

10. Cowork connectors must be reachable from the public internet, not localhost. Cowork's MCP connectors run through Anthropic's cloud, not the local network. A capability that only stands up a local/stdio MCP server will silently fail to connect from a Cowork session. Source: queen-bee-stinger distilled-research-articles.md, Claude Cowork §Plugins.

11. Know the real shared hook-event floor before designing a hook-driven capability. Claude Code exposes 26 hook events; Codex exposes 5. The verified shared floor across Claude Code and Codex is SessionStart, UserPromptSubmit, PreToolUse (Bash-only), PostToolUse (Bash-only native, Edit/Write approximated), Stop. Design for that floor first; treat richer per-harness events as an enhancement, not a baseline. Source: research/external/2026-08-14-hookbridge-loss-report-pattern.md.

Scope note

This stinger covers the integration surface for The Hive's four harnesses (Claude Code, Cursor, ChatGPT Codex, Claude Cowork): per-harness component placement, the wiring-mechanism decision (hooks, native extensions, MCP, AGENTS.md/instruction files), capability detection and graceful degradation, cross-harness portability, and distribution. It also retains, as a labeled case study, the original Hivemind six-host adapter work (Claude Code, Codex, Cursor, Hermes, pi, OpenClaw) this stinger was first built around. It does not cover vector-store schema/write-path internals, the embeddings runtime, retrieval ranking internals, or MCP wire-protocol details beyond registration - route those to the relevant stinger (vector-store-stinger, embeddings-runtime-stinger, retrieval-stinger, mcp-protocol-stinger).

Handoff map

• Vector-store schema and the capture/write-path internals behind an MCP-backed memory tool: route to vector-store-stinger.
• Embeddings model selection, batching, runtime cost: route to embeddings-runtime-stinger.
• MCP protocol internals (tool schemas, transport framing) beyond per-harness registration: route to mcp-protocol-stinger.
• Bundling pipeline topology and release CI: route to ci-release-stinger.
• Login/auth device flow and token vault security audit: route to security-stinger.

---

Part of The Hive, curated by Mario Aldayuz a.k.a @thenotoriousllama.