# Athos on Hermes (and open / self-hosted tool-calling models)

Hermes (Nous Research Hermes 2/3) is a **function-calling model**. It doesn't
load "skills" or run hooks — you give it a system prompt that (a) carries the
Athos coaching rules and (b) declares the Athos API as callable tools. Your
harness (vLLM, Ollama, LM Studio, TGI, or any OpenAI-compatible server) executes
the HTTP request when Hermes emits a tool call and feeds the result back. The
same pattern adapts to other open tool-calling models (Llama 3.x, Mistral, Qwen).

## 1. Load the coaching rules

```bash
curl -L <ATHOS_API_URL>/api/v1/skill/download -o athos-skill.tar.gz
tar xzf athos-skill.tar.gz            # -> athos/SKILL.md, athos/references/*.md
```

Put `athos/SKILL.md` (and any `references/*.md` you need) into the system prompt
or a retrieval store. That's the coaching logic, routing, and endpoint cheat sheet.

## 2. Auth — harness-side, never in the model

Keep the key in your harness/tool executor, not in the prompt. Every Athos call:

```
Authorization: Bearer <ATHOS_API_KEY>
Base URL: <ATHOS_API_URL>            # default https://app.athos.fitness
```

## 3. Hermes system prompt + tools

Hermes expects tools inside `<tools>` and returns calls as
`<tool_call>{...}</tool_call>`. Minimal working scaffold:

```
<|im_start|>system
You are Athos, an AI fitness & nutrition coach. Never invent training science or
nutrition numbers — always call a tool to read Athos state. get_program_week_current
is the source of truth for the current plan; keep planned vs logged distinct.
Return tool calls as <tool_call>{"name":..,"arguments":{..}}</tool_call>.
<tools>
{"type":"function","function":{"name":"get_program_week_current","description":"Canonical current week: workouts + meals. Source of truth.","parameters":{"type":"object","properties":{}}}}
{"type":"function","function":{"name":"get_onboarding_questions","description":"Fetch the onboarding interview questions.","parameters":{"type":"object","properties":{}}}}
{"type":"function","function":{"name":"submit_onboarding_answers","description":"Submit interview answers keyed by exact question id.","parameters":{"type":"object","properties":{"answers":{"type":"object","description":"map of question_id -> answer"}},"required":["answers"]}}}
{"type":"function","function":{"name":"get_onboarding_status","description":"Poll onboarding progress until strategy is ready.","parameters":{"type":"object","properties":{}}}}
{"type":"function","function":{"name":"confirm_strategy","description":"Confirm the proposed coaching strategy.","parameters":{"type":"object","properties":{}}}}
{"type":"function","function":{"name":"log_food_text","description":"Log a meal described in plain language.","parameters":{"type":"object","properties":{"description":{"type":"string"},"meal_type":{"type":"string"}},"required":["description"]}}}
{"type":"function","function":{"name":"log_activity","description":"Log an activity/workout.","parameters":{"type":"object","properties":{"description":{"type":"string"}},"required":["description"]}}}
{"type":"function","function":{"name":"modify_today","description":"Modify today's workout for pain/equipment/time.","parameters":{"type":"object","properties":{"instruction":{"type":"string"}},"required":["instruction"]}}}
</tools><|im_end|>
```

Map each tool to its endpoint in your executor (all with the Bearer header):

| tool | method + path |
|------|---------------|
| `get_program_week_current` | `GET  /api/v1/program/week/current` |
| `get_onboarding_questions` | `GET  /api/v1/onboarding/questions` |
| `submit_onboarding_answers` | `POST /api/v1/onboarding/submit-answers` |
| `get_onboarding_status` | `GET  /api/v1/onboarding/status` |
| `confirm_strategy` | `POST /api/v1/onboarding/confirm-strategy` |
| `log_food_text` | `POST /api/v1/food/entries/analyze-text` |
| `log_activity` | `POST /api/v1/activity/log` |
| `modify_today` | `POST /api/v1/program/today/modify` |

Loop: model emits `<tool_call>` → harness runs the HTTP request → return the JSON
as `<tool_response>{...}</tool_response>` → model continues. (Photo food logging
uses `POST /api/v1/food/analyze-image` if your harness can send images.)

## 4. No hooks → poll

Without a webhook transform you can't receive `onboarding.strategy_ready`,
`notification.due`, or `skill.updated` pushes. Poll instead:

- After `submit_onboarding_answers`, poll `get_onboarding_status` until ready,
  then `confirm_strategy`.
- Re-fetch `get_program_week_current` when the user re-engages.
- Re-download the package periodically to pick up skill updates.

If your deployment can expose an inbound URL, register it with
`POST /api/v1/webhooks` to reduce polling.

Full coaching rules, routing, and dedupe: `athos/SKILL.md` and `references/`.