> ## Documentation Index
> Fetch the complete documentation index at: https://docs.useharmonia.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Nimi — evidence analyst

> Nimi's source-analysis contract, evidence method, and authority boundary.

## Role

Nimi turns one typed source package into grounded moments and defensible angles for Ryan. Nimi
analyzes supplied transcript segments, bounded frame evidence, verified performance observations,
and eligible Memory Bank facts without defining strategy or writing copy.

## Contract

| Property        | Value                                                                                                                        |
| --------------- | ---------------------------------------------------------------------------------------------------------------------------- |
| ADK name        | `nimi_analyst`                                                                                                               |
| Default model   | `gemini-3.5-flash`                                                                                                           |
| Input           | `AnalystInput`                                                                                                               |
| Output          | `SourceAnalysis`                                                                                                             |
| Runtime skills  | `nimi-analysis-skills` with seven allow-listed method references                                                             |
| Skill interface | ADK filesystem `SkillToolset` exposing `load_skill` and `load_skill_resource`                                                |
| Research tools  | Native `google_search` in an isolated public-search agent; optional `VertexAiSearchTool` in an isolated private-search agent |

## Skills

`nimi-analysis-skills` provides seven allow-listed analysis methods. Nimi loads it exactly once and
only loads references relevant to the active evidence problem. Skill guidance is method, never evidence.

## Tools and research

Public research runs only through `nimi_google_search_agent`; configured private indexed retrieval
runs through `nimi_agent_search_agent`. Both require an exact application-supplied `researchRequest`
and remain isolated from the filesystem skill tools.

## Prompt method

The prompt requires the skill exactly once and only the references relevant to the analytical
problem. The skill establishes an evidence hierarchy: transcript segments and supplied frames are
primary source evidence; verified performance and eligible memory are advisory and retain distinct
provenance. Exact quotes must occur in cited segments, timestamps must remain in bounds, and visible
facts require supplied frame IDs. Evidence kinds cannot be mixed.

Skill guidance shapes method but is never evidence. Weak evidence becomes an explicit assumption with lower confidence. Nimi cannot invent trends,
speaker identity, customer outcomes, or unseen visuals.

## Evidence and provenance

Transcript segments and supplied frames are primary evidence. Verified performance, eligible memory,
and validated research retain distinct IDs and provenance. Exact quotes must occur in cited segments;
timestamps and frames remain in supplied bounds. Unknown or mixed-kind references fail closed.

## Authority boundary

Nimi cannot define objectives, positioning, pillars, campaigns, channels, cadence, KPIs, briefs,
final copy, CTAs, approvals, schedules, publishing actions, receipts, or verification claims.

## Handoff

Deterministic code persists the complete `SourceAnalysis` and canonical digest. Ryan receives that
exact typed object, not a prose summary or a new source-analysis opportunity.

## Runtime enforcement

Deterministic validation checks the independent skill-loader and research-tool traces, the exact
request binding, native grounding metadata, transcript support, media bounds, reference existence, evidence-kind compatibility,
confidence, authority overreach, canonical digesting, persistence, and the exact Ryan handoff.

## Implementation references

* Prompt: `agent/harmonia_agent/nimi_prompt.py`
* Contracts: `AnalystInput` and `SourceAnalysis` in `agent_models.py`
* Skill: `agent/harmonia_agent/skills/nimi-analysis-skills/SKILL.md`
* Tests: `agent/tests/test_nimi_contracts.py`, `test_nimi_validation.py`, and `test_nimi_skills.py`