Cut Gemini cost: default to flash-lite + optional reply drafting

[krMaynard]

Summary

Input-token cost on the every-5-minutes email pipeline was dominated by the premium gemini-3.5-flash model. Two changes drastically reduce it:

• Default GEMINI_MODEL → gemini-2.0-flash-lite (~10× cheaper input tokens than gemini-3.5-flash) for both analyze_email and the standup. The env var still overrides for higher accuracy when needed. This is the single biggest lever — it multiplies down the cost of the fixed system prompt, the injected user context, and every email body.
• New ENABLE_DRAFT_REPLIES flag (default true, mirrors the existing ENABLE_ACTION_ITEMS). When set to false, the large REPLY section is dropped from the analyze_email prompt entirely, so the model never classifies replies or drafts one — cutting both input and output tokens on every email — and no Gmail drafts are created. Calendar events, action items, and context updates are unaffected.

To go maximally cheap, set both:

gcloud run services update assistant \
  --region us-central1 \
  --update-env-vars GEMINI_MODEL=gemini-2.0-flash-lite,ENABLE_DRAFT_REPLIES=false

Changes

• assistant/event_extractor.py: GEMINI_MODEL default → gemini-2.0-flash-lite; build_system_prompt() and analyze_email() gain a draft_replies param that omits the REPLY section.
• main.py: reads ENABLE_DRAFT_REPLIES, passes it through, skips draft creation (and the get_user_email call) when off; standup default model updated to match.
• tests/test_prompt_building.py: deterministic unit tests (no API key) verifying each flag drops its section and sections renumber contiguously.
• README.md, .env.example: document the model default change and the new flag.

Test plan

• pytest tests/ — 24 passed, 8 skipped (Gemini eval tests skip without a key locally)
• Deploy with ENABLE_DRAFT_REPLIES=false and confirm in Cloud Logging that analyze_email prompt tokens drop and no drafts are created
• Confirm calendar events still get created

🤖 Generated with Claude Code

https://claude.ai/code/session_01CSnZNqUEh8uDyMCPDmzBTa

---

Generated by Claude Code

[gemini-code-assist]

Code Review

This pull request switches the default model to gemini-2.0-flash-lite and introduces a new ENABLE_DRAFT_REPLIES configuration flag to conditionally disable reply drafting and reduce API costs. Unit tests are also added to verify the prompt-building logic. The reviewer pointed out that when reply drafting is disabled, the underlying Pydantic schema EmailAnalysis still requires the needs_reply field, which may cause validation errors or hallucinated outputs from the model. It is recommended to set a default value of False for needs_reply in the schema.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

When draft_replies is set to False, the REPLY section is omitted from the system prompt. However, the EmailAnalysis Pydantic model (used as the response_schema for Gemini) defines needs_reply as a required boolean field:

class EmailAnalysis(BaseModel):
    ...
    needs_reply: bool

Because it is required, the Gemini model is forced to output needs_reply in the JSON response, even though the system prompt contains no instructions or context on how to evaluate it. This can lead to unpredictable model behavior, validation errors, or arbitrary/hallucinated values.

To resolve this, please update the EmailAnalysis class (around line 122) to provide a default value for needs_reply:

    needs_reply: bool = False