switch-dimension/molten-os-core/molten-search

Molten Search

Router only. Classify the query, name the route in one sentence, then hand off. Do not invent a search workflow of your own.

Operating Rules

1. Honor an explicit backend name over inference.
2. Pick one primary route. Add a second only when the user clearly needs two kinds of evidence (for example recent social sentiment and a cited web brief).
3. After routing, read the target `SKILL.md` and follow it. Do not summarize or improvise last30days, agent-reach, or parallel-web-search.
4. Local codebase lookup is not this skill. Use Grep, Glob, and Read instead.
5. Do not post, comment, like, or otherwise write to platforms. Agent-reach is fetch-only here.

Routes

Classification

Apply in this order. Stop at the first match.

1. Explicit override

User named a backend or close synonym:

• cursor, websearch, just google, quick search → cursor
• last30days, last 30 days, what people are saying, recent discussion → last30days
• agent-reach, agent reach, or a platform from its list (小红书/xhs, Twitter/X, B站/bilibili, Reddit, Facebook, Instagram, V2EX, LinkedIn, YouTube, GitHub code search, 小宇宙, 雪球) → agent-reach
• parallel, parallel-cli, parallel web search → parallel

2. Intent

last30days if the goal is recent community evidence: reaction, sentiment, discourse, "what are people saying", last month / last 30 days, social-platform pulse across Reddit/X/YouTube/TikTok/HN.

agent-reach if the goal is a specific platform or URL (including 调研/全网调研 and the platforms in that skill's description). Prefer this over last30days when they name one platform rather than a 30-day multi-source pulse.

cursor if the goal is cheap and fast: a fact, a definition, official docs, "what is X", "open this URL", a yes/no current check.

parallel for everything else that still needs the live web: compare tools, market/context brief, "research X", current info with citations, multi-source synthesis that is not a 30-day social pulse.

3. Default

If still ambiguous, use parallel.

Handoff

1. State the route: Routing to last30days (or cursor / agent-reach / parallel).
2. cursor: run WebSearch and/or WebFetch. Cite sources. Stop.
3. last30days / agent-reach / parallel: Resolve that skill's SKILL.md, read it immediately, then execute it as if the user had invoked it directly. Pass the user's query through unchanged unless the target skill requires a rewritten objective.

If parallel-cli is missing when the parallel route is chosen, resolve parallel-cli-setup and follow it, then continue.

Deep / exhaustive / comprehensive report requests are not parallel-web-search. Resolve and follow parallel-deep-research instead.

Resolve a backend skill

Look for <name>/SKILL.md in this order and use the first hit:

1. .agents/skills/<name>/
2. .cursor/skills/<name>/
3. .claude/skills/<name>/
4. skills/<name>/
5. ~/.agents/skills/<name>/
6. ~/.cursor/skills/<name>/
7. ~/.claude/skills/<name>/
8. ~/.codex/skills/<name>/

If the chosen backend is not installed, say so, keep the cursor route as a fallback for simple lookups, and tell the user how to install the missing skill (last30days, agent-reach, or parallel-web-search). Do not improvise that backend's workflow.

Examples