feat: add ejentum-mcp skill and MCP catalog entry

[ejentum]

Summary

Adds three files for the Ejentum Reasoning Harness MCP server, following the existing notion pattern (skills/notion/SKILL.md + mcps/catalog/notion.json + marketplace entry):

1. skills/ejentum-mcp/SKILL.md — SKILL.md teaching the agent when to call each of the four harness tools, with MCP-first prose and a direct-HTTPS fallback.
2. mcps/catalog/ejentum.json — stdio kind, npx -y ejentum-mcp, EJENTUM_API_KEY envField.
3. marketplaces/openhands-extensions.json — entry in alphabetical position between docker and flarglebargle.

What ejentum-mcp does

Four agentic tools the agent calls during its reasoning loop:

• harness_reasoning(query) — analytical / diagnostic / planning / multi-step tasks (311 reasoning operations: abstraction, time, causality, simulation, spatial, metacognition).
• harness_code(query) — code generation / refactoring / review / debugging (128 software-engineering operations).
• harness_anti_deception(query) — prompts that pressure validation, manufactured agreement, authority appeals, fabricated commitments (139 operations: sycophancy, hallucination, deception, adversarial framing, judgment, executive control).
• harness_memory(query) — sharpening an observation already formed about cross-turn drift; filter-oriented, not write-oriented (101 perception operations).

Each call returns a structured cognitive scaffold with the same shape: a named failure pattern, an executable procedure, suppression vectors to block, and a falsification test. The agent reads the scaffold internally before producing its user-facing answer; the bracketed labels ([NEGATIVE GATE], [PROCEDURE], etc.) are agent-internal markers and should not be echoed to the user.

Rules followed

Per AGENTS.md:

• V1 SKILL.md frontmatter (name, description, triggers only; no type / version / agent).
• Progressive disclosure: concise summary first, full prose available when the skill is loaded.
• Specific and operational: per-tool example queries, an <IMPORTANT> env-var check block, direct HTTPS fallback with mode enum documented.
• No repo-local assumptions, no secrets, no marketing prose.
• Punctuation style: plain hyphens (-), no em dashes anywhere in the new files.
• One skill per directory; marketplace entry added per the synced-extensions check.

Test plan

• python -c "import json, io; json.load(io.open('marketplaces/openhands-extensions.json', encoding='utf-8'))" passes (plugin count 46 → 47)
• mcps/catalog/ejentum.json parses cleanly
• marketplaces/openhands-extensions.json diff is +13 lines (alphabetical insertion only; no reformat / no reorder of existing entries)
• Voice scrub: no pre-generation framing, no em dashes, no marketing language in any of the three new artifacts
• SKILL.md content matches the structural template from skills/notion/SKILL.md (MCP-first IMPORTANT block + direct-API fallback)
• Hands-on test from a maintainer (env: EJENTUM_API_KEY=zpka_test_xxx, then trigger one of the keywords in a conversation)

Notes

• Free and paid tiers for EJENTUM_API_KEY at https://ejentum.com/pricing
• npm: https://www.npmjs.com/package/ejentum-mcp (v0.1.x with OIDC trusted-publisher provenance)
• The same four harness tools are also reachable as twelve native framework packages on PyPI/npm (LangChain, LangGraph.js, CrewAI, Agno, PydanticAI, smolagents, Letta, AutoGen, Vercel AI SDK, Mastra, Genkit, LlamaIndex) for OpenHands users who do not configure MCP. The SKILL.md is opinionated about MCP-first, matching the notion / linear / slack patterns already in this registry.

cc @enyst for MCP catalog review per recent precedent on #246, #247, #248.