Deep Research for OpenClaw

Install and wire a structured OpenClaw deep-research sub-agent with hybrid search, artifact-based runs, claim verification, report linting, and validated fin...

Security Scan

The skill's claimed purpose (wiring a deep-research OpenClaw sub-agent) matches its instructions and requirements; it is an instruction-only integration that asks you to clone a GitHub repo, copy files into your OpenClaw installation, and run the provided scripts — review the repo before installing and back up your OpenClaw config.

Purpose & Capability

The name/description describe installing a deep-research OpenClaw sub-agent. The declared requirements (openclaw, python, openclaw.json, a configured deep-researcher agent, optional Tavily API key) are coherent and expected for this purpose; there are no unrelated credentials or binaries requested.

Instruction Scope

SKILL.md instructions are focused: clone the public GitHub repository, copy the workspace-researcher files into your OpenClaw base, update openclaw.json, and run py_compile + a small init script and the openclaw agent CLI for a smoke test. The only noteworthy scope expansion is that you must write into and modify your OpenClaw config and execute scripts from the cloned repo — this is expected for installing a sub-agent but is something to inspect beforehand.

Install Mechanism

There is no packaged install spec; the workflow instructs cloning the repository from GitHub and copying files into your OpenClaw directory. Using a public GitHub repo is common and reasonable, but downloading and placing third-party scripts into your agent runtime can introduce remote code execution risk if you don't review the code first.

Credentials

No required environment variables or credentials are declared. The SKILL.md notes an optional TAVILY_API_KEY only if you choose the Tavily-backed path — this is proportional and clearly documented as optional.

Persistence & Privilege

The skill requires editing/adding the deep-researcher agent entry in openclaw.json and copying files into the OpenClaw installation, which creates persistent presence in your OpenClaw runtime. The skill does not set always: true and does not request platform-wide elevated privileges, but you should be aware it makes persistent changes to your agent configuration.

Assessment

This skill is internally consistent with its stated purpose, but treat it like any third-party code you install: 1) Inspect the GitHub repository before cloning (look at the scripts in openclaw/workspace-researcher/scripts and SOUL.md). 2) Back up openclaw.json and any OpenClaw directories you will modify. 3) Run the provided py_compile and review init_research_run.py output; consider running the init script in a sandboxed environment or VM first. 4) Only provide a Tavily API key if you intend to use that feature, and store it securely (e.g., in a vault or .env you control). 5) Verify the repository origin and maintainer reputation if this will run in production. Following these steps reduces risk from executing unreviewed code placed into your agent runtime.

What this skill is

This is an integration skill for installing and wiring the deep-researcher OpenClaw sub-agent from the public repository:

• deep-research-openclaw-agent

The repository contains:

• the workspace-researcher prompt pack;
• the local research helper scripts;
• the Main -> Deep Research orchestration contract;
• the report lint, validation, and finalization pipeline.

This skill is intended for OpenClaw users who want a reproducible deep-research workflow without assembling the runtime and contracts from scratch.

What it can do

• structured deep research through plan -> scout -> harvest -> verify -> synthesize;
• hybrid discovery with web_search, Tavily, and web_fetch;
• explicit source registry, claim ledger, and coverage tracking;
• report linting, validation, and final M2M JSON finalization;
• honest SUCCESS | PARTIAL | FAILURE delivery with explicit gaps and conflicts.

Requirements

• OpenClaw 2026.3.x or later
• Python available on the host
• a configured deep-researcher agent in OpenClaw
• Tavily API access if you want the Tavily-backed path

Install

1. Clone the repository:
   • git clone https://github.com/MilleniumGenAI/deep-research-openclaw-agent.git
2. Copy openclaw/workspace-researcher/ into your OpenClaw base directory, or point your agent config at that path directly.
3. Align the main-agent handoff with:
   • openclaw/main-deep-research-skill.md
4. Register or update the deep-researcher agent in openclaw.json.
5. If you want Tavily-backed scouting, ensure TAVILY_API_KEY is available in env or .env.

Validate

Run these checks before using the agent in real work:

python -m py_compile openclaw/workspace-researcher/scripts/*.py
python openclaw/workspace-researcher/scripts/init_research_run.py --workspace openclaw/workspace-researcher --topic "Smoke test" --language en --task-date 2026-03-10

Then run a first smoke task through OpenClaw once the agent is wired:

openclaw agent --agent deep-researcher --json --message "Perform deep research using your local SOUL.md contract. GOAL: confirm the runtime can initialize a fresh run and return PARTIAL if no external research is performed. SCOPE: in scope is only local init and artifact creation; out of scope is web research. SUCCESS CRITERIA: create fresh tmp artifacts and explain blocked evidence collection honestly. TASK DATE: 2026-03-10. DELIVERABLES: finalized M2M JSON. LANGUAGE: en. CONSTRAINTS: do not fabricate sources; return PARTIAL if evidence is insufficient."

Core references

• Root README: README.md
• Sub-agent contract: openclaw/workspace-researcher/SOUL.md
• Main handoff contract: openclaw/main-deep-research-skill.md
• Runtime scripts: openclaw/workspace-researcher/scripts/
• Agent config template: openclaw/agent-config.template.json
• Known limits: docs/known-limits.md

Notes

• This is an OpenClaw-only v1 package.
• ClawHub publishes skills under platform-wide MIT-0 terms.
• The runtime source of truth is openclaw/workspace-researcher/SOUL.md.
• Findings should be built only from traceable external sources, not from local artifacts.