dispatch-exploration

Dispatch Exploration

Send a small batch of research queries to specialist agents in parallel, apply the cross-verify consent gate when a returned finding triggers it, and return a compressed findings list.

Input

A list of query objects:

[
  {"question": "what does Ethan want re: relocation healthcare costs?", "target": "vault", "hint": "projects/relocation"},
  {"question": "current state of Spain non-lucrative visa for US remote workers", "target": "web"}
]

• target — "vault" or "web". Required.
• hint (vault only) — optional slug-prefix scope to focus the search.

The input shape is stable — callers do not need to opt into the cross-verify gate; it is always on for target: "web" results.

Process

1. Dispatch one agent per query, all in parallel (single message, multiple Agent tool calls):

   • target: "vault" → vault-researcher
   • target: "web" → web-researcher

2. Wait for all results.

3. Cross-verify intercept for each web result. See Cross-verify orchestration below. This is the only step where this skill may pause for user interaction; vault results are pass-through.

4. Build a compressed findings list, one line (or short block) per query:

   Vault result:

   <query> → <slug> — <relevance> — <excerpt>
   

   or <query> → no match if the agent returned none.

   Web result (default):

   <query> → <summary> — <top 1-2 citations> — cohort: <cohort_spread> — flags: <flag list>
   

   Pass through cohort_spread and flags when present. Omit the cohort: / flags: segments when the agent returned no flags and a single-cohort spread (e.g. an evergreen-fact lookup with a clean sample).

   Web result (pricing mode, single tier):

   <query> → new: median $Y, range $A–$B (N reputable sources) — flags: <flags>
   

   Web result (pricing mode, tiered after consent):

   <query>
     new:  median $Y, range $A–$B (N sources, manufacturer-direct + major-retailer)
     used: median $Z, range $C–$D (M sources, marketplace-used / refurbished / etc.)
     flags: <flags>