This page answers which method skill an agent may load, which data tools it may call, where its
focused prompt lives, and which deterministic trace proves what actually ran.
Skill guidance is method, never evidence. Search is a tool, not a skill. Memory Bank is advisory context, never authority.
Runtime inventory
Prompt and runtime locations
Loading and grounding rules
Nimi, Ryan, Temi, Noni, and Dara load their named project-owned skill exactly once, then load only
approved references relevant to the request. Missing, duplicate, out-of-order, irrelevant, or
unapproved resource loads fail closed. Nova selects exactly one routing skill before a compatible
read tool. Maya has no skill or tool trace.
Search-capable agents never place native google_search beside filesystem or function tools in the
same model request. A dedicated ADK search agent returns a typed, request-bound result. Deterministic
code preserves and validates grounding chunks, supports, queries, titles, URLs, and search entry-point
metadata before search evidence IDs may be cited.
Evidence and authority
Valid evidence comes only from the typed invocation: operator context IDs, Nimi moment or angle IDs,
verified performance IDs, eligible Memory Bank fact IDs with Firestore provenance, approved internal
read-tool evidence IDs, or validated native-search evidence IDs. Skill files, prompts, model memory,
and source-coverage ledgers cannot prove customer, market, trend, performance, or source claims.
No agent may approve, transition workflow state, schedule an external effect, publish, access
credentials, create receipts, verify effects, or grant policy exceptions. See the
authority matrix and agent contracts. 