SN Deep Research for Openclaw

1. Install or enable the SN Deep Research skill in the OpenClaw skills directory, then resolve the actual skill directory exposed by the runtime. OpenClaw uses the {baseDir} placeholder; replace it with the resolved absolute path.
2. Ensure the runtime provides file read/write, command execution, web search, and webpage-fetching capabilities. These are mandatory for reliable research.
3. Configure optional credentials in the repository root .env file or through the runtime environment. Do not place secrets in payloads, command-line arguments, reports, logs, or transcripts.
4. Create a unique report workspace before dispatching agents:

run=$(openssl rand -hex 2 2>/dev/null || printf '%04x' "$RANDOM")
report_dir="$PWD/deep-research-reports/$(date +%F)-{topic}-$run"
mkdir -p "$report_dir"/sub_reports "$report_dir"/board "$report_dir"/sections "$report_dir"/content_units
echo "$report_dir"

5. Write the original request and start progress tracking before research artifacts are generated. The shared helper is located under the plugin skills directory:

python3 {plugin_skills_dir}/sn-ppt-standard/scripts/progress_event.py \
  --deck-dir "$report_dir" \
  --workflow deep-research \
  --stage mode-selection \
  --status running \
  --artifact request.md \
  --label "Research workflow started"

6. Launch the required Research Workbench:

python3 {plugin_skills_dir}/sn-ppt-standard/scripts/launch_workbench.py \
  --deck-dir "$report_dir" \
  --product research \
  --progress-route / \
  --source-session-id "${HERMES_SESSION_KEY:-}" \
  --agent-managed 1 \
  --require-webui \
  --host 0.0.0.0

7. For Windows systems without python3, use python. In Git Bash or MSYS, prefix the command with MSYS_NO_PATHCONV=1 when passing the / route.
8. Render the final report after the selected workflow completes:

python3 {plugin_skills_dir}/sn-prepare-citations/scripts/prepare_citations.py \
  --report {report_dir}/stitched.md \
  --evidence {report_dir}/sub_reports/d*.evidence.json \
  --outline {report_dir}/outline.json \
  --no-l0 \
  --output {report_dir}/report.md

For quick and normal, use sections/s_full.md as the --report input and omit --outline. Optional credentials include SN_IMAGE_GEN_API_KEY, SN_API_KEY, platform cookies, TIKHUB_TOKEN, YOUTUBE_API_KEY, GitHub credentials, HF_TOKEN, SO_API_KEY, and academic API keys. Missing optional credentials reduce coverage or increase rate-limit risk but do not necessarily stop the workflow.

All artifacts live under one uniquely named absolute report directory:

Metadata and routing conventions include:

• Report directories follow YYYY-MM-DD-{topic}-{hex4} to prevent collisions between repeated runs.
• Research dimensions use stable IDs such as d1, d2, and d3; quick mode uses d1.
• Heavy-mode content units use IDs such as u1 and receive bounded evidence subsets.
• Every source record preserves its original URL and is selected according to a source category.
• language and format are request-level runtime parameters passed to every child agent; they are not stored as separate state files.
• mode determines the pipeline: quick uses research plus writer, normal uses plan plus parallel research plus writer, and heavy uses the complete audited pipeline.
• Evidence must record unresolved gaps honestly, including partial, no_data, or out_of_scope conditions.
• Citation rendering checks for orphan citations and unresolved claim identifiers such as [^dN.cM] before delivery.