Game Research Orchestrator

Workspace

Create at pipeline start:

WORKSPACE="/tmp/game-research-$(date +%s)"
mkdir -p "$WORKSPACE"/{discovery,extraction}

All subagents save structured JSON here. The orchestrator reads these files to coordinate between phases.

Data contract:

$WORKSPACE/
├── meta.json                # {timestamp, query, language, mode, query_type}
├── discovery/
│   ├── perplexity.json      # {overview, citations[], entities[], is_news_topic}
│   ├── brave_web.json       # raw Brave web API response
│   ├── brave_video.json     # raw Brave video API response
│   └── brave_news.json      # raw Brave news API response (if applicable)
├── source_matrix.json       # {articles[], videos[], community[], reviews[]}
├── refinement.json          # refinement search results (if applicable)
└── extraction/
    ├── articles.json        # {results[{url, title, content}]}
    └── community.json       # {results[{url, platform, content}]}

Full Report Mode — Orchestration Pipeline

Phase 0: Setup

1. Parse arguments: extract keyword and --brief flag
2. Run scripts/perplexity_search.py timestamp --tz=Asia/Seoul for the report's {date} field
   • Where scripts/ = <this skill's parent>/../web-research/scripts/ (resolve the path)
3. Create workspace directory
4. Detect search language: Korean if keyword contains Hangul (U+AC00-U+D7AF)
5. Classify query type for search expansion:

   Query pattern	Type	Search expansion
   Specific game title (e.g., "엘든 링", "Elden Ring")	title	Add "review", "metacritic", "gameplay"
   Genre/category (e.g., "소울라이크", "roguelike 2026")	genre	Add "best", "recommendations", "upcoming"
   Industry topic (e.g., "게임 구독 서비스", "UE6")	industry	Add "analysis", "comparison", "trend"
   Esports (e.g., "LCK 2026", "Valorant Champions")	esports	Add "results", "standings", "teams"
   Hardware (e.g., "스팀덱 OLED", "PS5 Pro")	hardware	Add "review", "specs", "comparison"

6. Detect gaming news keywords: 출시, 패치, 업데이트, DLC, 확장팩, 시즌, 발표, 논란, 서비스종료, release, launch, patch, update, nerf, buff, hotfix, early access, beta, etc.
7. Save meta.json to workspace (include query_type)

Phase 1: Parallel Discovery

Spawn two subagents simultaneously using the Agent tool. Read each agent template from agents/, construct the prompt with the variables below, and launch both in a single message.

Subagent A — Perplexity Discovery:

Read agents/discoverer-perplexity.md. Spawn with these variables:

• {query}: the research keyword
• {query_type}: classified query type from Phase 0
• {language}: detected search language
• {workspace}: workspace path
• {script_dir}: path to ../web-research/scripts/ (resolved absolute path)

Subagent B — Brave Discovery:

Read agents/discoverer-brave.md. Spawn with these variables:

• {query}: the research keyword
• {query_type}: classified query type
• {language}: detected search language
• {has_news_keywords}: whether news keywords were detected in Phase 0
• {workspace}: workspace path
• {script_dir}: path to ../web-research/scripts/ (resolved absolute path)

Wait for both subagents to complete before proceeding.

Phase 2: Classification & Refinement (inline)

This phase is lightweight enough to run inline in the orchestrator. No subagent needed.

1. Read discovery/perplexity.json and discovery/brave_web.json from workspace
2. Classify all URLs by pattern matching (refer to references/source-classification.md):
   • Video: YouTube, Twitch clips, etc.
   • Community: Reddit gaming subs, Steam community, 루리웹, 인벤, etc.
   • Review: Metacritic, OpenCritic, IGN, GameSpot, etc.
   • Article: everything else
3. Merge video sources: video URLs from web search + discovery/brave_video.json. Deduplicate.
4. If discovery/brave_news.json exists, merge news results into articles (leverage date metadata)
5. Deduplicate across all sources

Refinement decision: Read discovery/perplexity.json entities. If entities reveal subtopics not covered by initial results, run one refinement search inline:

Where scripts/ = resolved path to ../web-research/scripts/.

Save refinement results to refinement.json. Build and save source_matrix.json:

{
  "articles": [{"url": "...", "title": "..."}],
  "reviews": [{"url": "...", "title": "...", "platform": "..."}],
  "videos": [{"url": "...", "title": "...", "source": "...", "description": "..."}],
  "community": [{"url": "...", "platform": "..."}]
}

Phase 3: Parallel Extraction

Spawn extraction subagents simultaneously. Read agents/extractor.md for the template.

Subagent C — Article Extractor:

• {category}: "articles"
• {urls}: top 8 article URLs from source_matrix (comma-separated). Prioritize review sites and official sources.
• {query}: research topic (for relevance reranking)
• {workspace}: workspace path
• {script_dir}: resolved script directory path

Subagent D — Community Extractor (only if community URLs exist):

• {category}: "community"
• {urls}: top 5 community URLs from source_matrix (comma-separated)
• {query}: research topic
• {workspace}: workspace path
• {script_dir}: resolved script directory path

Video metadata is already available from Brave search results — no extraction subagent needed.

Wait for all extraction subagents to complete.

Post-extraction: Build extraction coverage summary. After extraction completes, build extraction_coverage to pass to the synthesizer:

{
  "extracted": ["url1", "url2"],
  "failed": ["url3"],
  "skipped": ["url4", "url5", "..."]
}

Phase 4: Synthesis

Spawn one synthesis subagent. Read agents/synthesizer.md for the template, and also read references/report-template.md to include in the prompt.

Provide the synthesizer with:

• {workspace}: workspace path
• {query}: research keyword
• {query_type}: classified query type
• {timestamp}: from Phase 0
• {script_dir}: resolved script directory path
• {report_template}: full contents of references/report-template.md
• {source_matrix}: contents of source_matrix.json
• {perplexity_overview}: the overview text from discovery/perplexity.json
• {extracted_articles}: contents of extraction/articles.json
• {extracted_community}: contents of extraction/community.json (if exists)
• {video_metadata}: video entries from source_matrix
• {refinement_data}: contents of refinement.json (if exists)
• {extraction_coverage}: the extraction coverage summary

The synthesizer decides whether to call perplexity_search.py reason based on:

The synthesizer outputs the complete report text.

Phase 5: Output & Save

1. Output the synthesizer's report directly in the chat response
2. Use AskUserQuestion to ask whether to save (in Korean):
   • Don't save → do nothing
   • Save to file → user provides path → Write the report verbatim

Brief Mode (--brief)

Simplified pipeline — no extraction or deep synthesis:

1. Phase 0: Setup (same as full mode)
2. Phase 1: Parallel Discovery (same — spawn both subagents)
3. Skip Phases 2-3
4. Quick Synthesis: Read discovery results. The orchestrator synthesizes directly (no subagent needed for brief mode — the data volume is small enough). Generate output following references/brief-template.md.
5. Output & Save: Same as Phase 5

Fallback Mode

Entry conditions (any one triggers):

• Phase 3 extraction subagents both return <= 2 successful extractions total
• Both Perplexity and Brave discovery subagents fail

Use scripts/tavily_search.py research "<query>" --model=mini as a single-call alternative. Note in the report that fallback mode was used.

Error Handling

• If a subagent fails or times out, proceed with results from the successful subagent(s)
• If a specific script call fails within a subagent, the subagent should note the failure and continue
• If no video results found, skip video section entirely
• A single subagent failure should never stop the entire pipeline — degrade gracefully