research

name	research
version	1.0.0
description	Unified Research family. Use Research for citation-first investigation, bounded synthesis, decision support, and UI/UX or accessibility review.
invocation	blue research query="ZimaOS-Blue research architecture" --json
examples	["blue research query=\"ZimaOS-Blue research architecture\" --json","blue research mode=analyze topic=\"Product feedback analysis\" urls='[\"https://example.com/reviews\"]' --json","blue research mode=advisor question=\"Go vs Python for backend services\" category=language decision_mode=compare --json","blue research mode=ui_review url=https://example.com/pricing review_action=check_accessibility --json","blue research action=status job_id=job_123"]
capability_tags	["research","citations","web"]
interaction_mode	stateless
card_support	both

Research

This document describes the unified Research family.

Research is one public surface with four execution modes:

• mode=deep_research: citation-first, multi-source investigation
• mode=analyze: bounded synthesis over provided text, URLs, and/or search queries
• mode=advisor: decision memo for selection, replacement, migration, and best-practice questions
• mode=ui_review: UI/UX or accessibility review for a live URL or screenshot

How To Send

Use the canonical Research entry:

blue research query="ZimaOS-Blue research architecture" --json

Older prompts, wrappers, or internal routes may still refer to deep_research, analyze, or ui_reviewer. Treat those as compatibility aliases that now route to Research.

Routing Guide

Need	Recommended mode	What it is best at
Multi-source research with evidence, citations, comparisons, or timeline work	deep_research	Investigation with planning, retrieval, evidence merging, and confidence-backed answers
One synthesized report from known materials	analyze	Bounded synthesis over provided text, URLs, and search queries
Decision support for technology choices or migration tradeoffs	advisor	Recommendation memo with tradeoffs, risks, alternatives, and optional evidence
UI/UX, screenshot, or accessibility evaluation	ui_review	Visual and accessibility review for URLs or images
Local workspace files or README inspection	Not this family	Prefer exec / file tools instead

Common Parameters

• action (optional): run or status
• query (required for normal research runs unless another mode-specific input supplies the goal)
• mode (optional): auto, deep_research, analyze, advisor, or ui_review
• lang (optional): output language such as zh-CN or en-US
• report_style (optional): output style hint such as summary, timeline, or other mode-supported variants

Important:

• mode chooses the research family branch.
• research_depth chooses the depth only inside mode=deep_research.
• Do not overload mode with fast|standard|deep; use research_depth for that.

Mode: deep_research

Use this for citation-first investigation.

Example:

blue research query="Compare ZimaOS and CasaOS with citations and a timeline" mode=deep_research research_depth=deep --json

Key parameters:

• query (required)
• research_depth (optional): fast, standard, or deep
• strict_entity (optional): true|false
• time_windows (optional): array of timeline windows
• report_style (optional): typically summary or timeline

Auto recent-discussion routing:

• Queries like What are people saying in the last 30 days about ZimaOS?
• Queries like 最近30天大家怎么说 ZimaOS？

stay on the public Research entry. Blue now detects that intent and internally switches to a native recent multi-site retrieval profile across public web, Reddit, Hacker News, GitHub, Polymarket, X, TikTok, Instagram, and Bluesky. Known YouTube URLs are added as enrichment when they are already present in the request. There is no separate public last30days mode or command.

Behavior:

1. Plan sub-questions
2. Retrieve and compare evidence
3. Merge and deduplicate sources
4. Return answer, citations, confidence, and evidence count

Mode: analyze

Use this for a bounded report when the materials are already known or can be collected with a small number of URLs/searches.

Examples:

blue research mode=analyze topic="Product feedback analysis" urls='["https://example.com/reviews"]' search_queries='["product reviews 2026"]' --json
blue research mode=analyze topic="Survey insights" text="..." lang=en-US --json

Key parameters:

• topic (usually required for bounded synthesis)
• urls (optional, max bounded set)
• search_queries (optional, bounded supplemental discovery)
• text (optional)
• output_mode (optional): inline or report

Use analyze when:

• the user wants one synthesized report
• the source set is already mostly known
• citations are not the primary requirement

Mode: advisor

Use this for selection, replacement, migration, and best-practice decisions.

Examples:

blue research mode=advisor question="Go vs Python for backend services" category=language decision_mode=compare --json
blue research mode=advisor question="Should we replace LibreOffice with OnlyOffice?" category=library decision_mode=replace --json

Key parameters:

• question (required)
• category (optional): architecture, language, framework, library, process, migration, ops
• decision_mode (optional): recommend, compare, review, replace, best_practice
• candidates (optional)
• context / constraints (optional)
• grounding (optional): auto, none, web
• depth (optional): quick, standard, deep
• output (optional): decision_memo, scorecard, decision_pack
• scorecard_pack (optional): auto, solution_selection_v1, migration_v1, architecture_v1, process_v1
• scorecard_weights (optional): criterion-weight overrides

Deep advisor examples:

blue research mode=advisor question="Go vs Python vs Node" category=language decision_mode=compare output=scorecard --json
blue research mode=advisor question="Should we replace Python with Go?" category=migration decision_mode=replace depth=deep output=decision_pack --json

Mode: ui_review

Use this for visual quality or accessibility review of a webpage or screenshot.

Examples:

blue research mode=ui_review url=https://example.com/pricing --json
blue research mode=ui_review url=https://example.com/pricing review_action=check_accessibility --json
blue research mode=ui_review image=<base64_png_data> review_action=review_image --json

Key parameters:

• url for live-page review
• image for screenshot review
• review_action (optional): review_url, review_image, or check_accessibility
• device, channel, wait_ms, threshold, format, profile (optional review controls)

Important:

• Top-level action is reserved for run / status.
• Use review_action for the UI-review sub-action.

Status / Resume

Poll an existing research-family job with:

blue research action=status job_id=job_123

Output

Depending on mode, returns structured data such as:

• answer
• citations
• confidence
• evidence_count
• report metadata
• UI review findings and scores

Notes

• Think of this as one research family, not three unrelated tools.
• Prefer mode=deep_research for citation-first work.
• Prefer mode=analyze for one bounded synthesized report.
• Prefer mode=advisor for selection, replacement, and best-practice decisions.
• Prefer mode=ui_review for URL/screenshot quality review.
• If the task is only quick link discovery, prefer web_query.
• If the task is about local files already in the workspace, prefer exec / file tools.

Example Triggers

• "帮我做一个带引用的调研结论"
• "Compare these claims with citations and a timeline"
• "Summarize these links into one report"
• "Review the UI of this page for accessibility and layout issues"