Universal Search Orchestrator

Problem

search-orchestrator targets common failure modes in agent search: queries becoming too generic, multiple tools consuming the same prompt, results that seem topical but cannot be traced back to the original question, and research stopping after one shallow round. Instead of wrapping web_search only, it chains tool discovery, subquery construction, parallel execution, convergence checks, and output auditing into a repeatable pipeline, making search results more evidence-like than a pile of snippets.

Orchestration and key constraints

• Discover before invoking: discover.js scans installed search skills, extracts ready tools and call templates, and falls back to built-in tools such as web_search and web_fetch when no external skill is available.
• Complexity routing: L0, L1, and L2 levels control depth. Single-fact questions use light verification; comparisons and news require iterative checks; deep research forces multiple rounds until convergence or the hard limit.
• Anchor control: the workflow extracts 2–4 exclusive anchor terms from the original question. Every round must preserve the anchors and expand only subdimensions. It filters topic drift using anchor presence, anchor-removal, and traceability tests.
• Parallel execution and agent co-execution: scripts can run shell-based search tools, while the agent must also call built-in web_search in parallel, fetch key pages after the first URL list, and use browser when fetch is blocked.
• Output audit: each claim must answer the original question; core facts should be supported by at least two independent sources. Generic background, filler, and untraceable content are removed.

Boundaries

It fits multi-source verification and iterative research, but it depends on usable search skills or built-in tools. Cross-language anchor matching, long compound-word segmentation, and factual-conflict detection still have limits, so sources and claims still need manual review.