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

# Nimi — evidence analyst

> Current Strands registration, model, inputs, tools and authority.

## Role

Analyzes the supplied source manifest into evidence-grounded insights and clip candidates. Timed speech comes from Amazon Transcribe; Nimi reasons over the resulting text and supplied visual evidence.

## Contract

| Property                 | Value                                                               |
| ------------------------ | ------------------------------------------------------------------- |
| Registered specialist    | `nimi_analyst; nimi_research_analyst`                               |
| Default Bedrock model ID | `us.amazon.nova-2-lite-v1:0`                                        |
| Typed boundary           | `AnalystInput → SemanticSourceAnalysis → host-bound SourceAnalysis` |
| Returned state key       | `source_analysis`                                                   |

## Skills

The host preloads `nimi-analysis-skills` method guidance and approved references into the system prompt. No `load_skill` or `load_skill_resource` tool is registered. Host bootstrap records activation provenance. Skill guidance is method, never evidence.

## Tools and research

`nimi_analyst` has no tools. The host selects `nimi_research_analyst` only for a typed research request. Before inference the runtime retains exactly one permitted tool: `nimi_gateway_search` for public web research or `nimi_agent_search_agent` for the configured Bedrock Knowledge Base.

## Prompt method

Keep transcript, frame, performance, memory and research evidence distinct. Quotes must occur in their cited source segments. The host binds complete segment references and checks time ranges, visual evidence, confidence and source lineage.

## Evidence and provenance

Only the supplied typed input and authorized returned tool evidence can support factual claims. Preloaded skills, model memory and generated prose do not establish facts. The host retains request hashes and validated references without logging private prompts or chain-of-thought.

## Authority boundary

The specialist cannot approve, grant permissions, publish, schedule an external effect, access credentials, mutate workflow state or create verification receipts. Those operations belong to deterministic host code and persisted authorization.

## Handoff

The host validates semantic output, attaches authoritative source identities, computes the canonical digest and persists SourceAnalysis before Ryan receives it.

## Runtime enforcement

`LocalStrandsTeamRuntime` creates a fresh Strands `Agent` per invocation with `structured_output_model`, `AuthorityHooks` and automatic model retries disabled. The host selects the registration, projects bounded input and validates output before a stage mutation. Provider repair follows the bounded host policy. Authenticated AWS execution remains pending live verification.

## Implementation references

* Registration: `agent/harmonia_agent/agents.py`
* Prompt: `agent/harmonia_agent/nimi_prompt.py`
* Method/tool implementation: `agent/harmonia_agent/nimi_research.py`
* Runtime: `agent/harmonia_agent/team_runtime.py`
* Model assignment: `agent/harmonia_agent/role_models.py`