podman/docs/gemini.md

Gemini Integration Spec

PodMan uses Gemini for two distinct jobs: vision (understanding screens) and voice (speaking nudges).

---

1. Vision — Screen Understanding

Model: gemini-2.0-flash (fast, cheap, strong multimodal)

Trigger: every 30s per active engineer, when Hermes receives a POST /ingest frame

Input: base64-encoded JPEG, max 1280×720, ~50–80KB after compression

Prompt:

You are analyzing a software engineer's screen during a coding session.
Extract the following JSON. If you cannot determine a field with confidence above 0.7, set it to null.

{
  "currentFile": "string | null",         // active file visible in editor tab or title bar
  "inferredTask": "string | null",        // 1 sentence: what the engineer appears to be doing
  "terminalVisible": true | false,        // is a terminal or CLI panel visible
  "recentTerminalOutput": "string | null", // last meaningful line of terminal output if visible
  "confidence": 0.0–1.0                  // your overall confidence in this extraction
}

Respond with valid JSON only. No explanation. No markdown.

Confidence gate: if confidence < 0.6, Hermes discards the frame — no state update, no event detection triggered.

Rate limit: 1 call per engineer per 30s. With 3 engineers = 6 calls/min ≈ $0.002/min at Flash pricing.

Demo setup requirement: editors must have large font (18pt+), single window, file name clearly visible in tab. This is the primary reliability lever.

---

2. Event Detection — Coordination Awareness

Model: gemini-2.0-flash (text only, fast)

Trigger: after every successful state write to MongoDB, Hermes runs event detection over all active engineer contexts.

Input: JSON snapshot of all engineers' current states + ownership map

Prompt:

You are a team coordination agent. Below is the current state of each engineer on the team.

Engineer states:
{{engineerStates}}

Ownership map (who owns which files):
{{ownershipMap}}

Detect if any of these coordination events are occurring:
- DEPENDENCY_READY: an engineer who was blocked or waiting now has what they need because another engineer completed relevant work
- BLOCKER_DETECTED: an engineer appears stuck (same file, error in terminal, no progress) and another teammate could help
- DUPLICATE_WORK: two or more engineers are working on the same file simultaneously

If an event is detected, respond with:
{
  "event": "DEPENDENCY_READY" | "BLOCKER_DETECTED" | "DUPLICATE_WORK" | null,
  "involvedEngineers": ["engineerId", ...],
  "file": "string | null",
  "reason": "1 sentence explanation"
}

If no event, respond with { "event": null }.
Respond with valid JSON only.

---

3. Nudge Generation — Voice Message

Model: gemini-2.0-flash (text only)

Trigger: when event detection returns a non-null event

Input: event type + engineer names + file + reason

Prompt:

You are PodMan, a friendly AI teammate. Generate a short spoken message (1–2 sentences max) to notify the team about this coordination event.

Event: {{eventType}}
Engineers involved: {{engineerNames}}
File: {{file}}
Context: {{reason}}

Rules:
- Use first names only
- Be direct and specific
- Do not use filler words
- Sound natural when spoken aloud
- Do not start with "Hey" or "Attention"

Respond with the message text only.

Example output:

"Carol — Alice just got the auth endpoint running. You're clear to integrate."

---

4. Voice Output — Gemini Live 2.5 via LiveKit

Model: gemini-live-2.5-flash (confirm exact ID from LiveKit Agents docs)

Integration: LiveKit Agents framework — Hermes runs as a LiveKit Agent with Gemini Live 2.5 as the voice provider

Flow:

1. Nudge message text generated (step 3)
2. Hermes passes text to Gemini Live via LiveKit Agents
3. Gemini Live streams audio back in real-time
4. LiveKit publishes audio into the pod room
5. All participants hear it through their audio output

Why Gemini Live (not plain TTS):

• Streams audio directly — no intermediate WAV file conversion
• Latency ~300–500ms from text to first audio packet
• Natural-sounding voice
• Strong prize story: Gemini Live 2.5 is the headline model

---

Cooldown

Per-pod cooldown of 3 minutes between nudges. Prevents spam if multiple events fire simultaneously. Implemented in Hermes, not in Gemini.