long_decode_max_output_tokens (agent/gemini_long_decode.py)

Utilities for managing Gemini long‑decode token budgets and continuation handling.

# Source: https://github.com/NousResearch/hermes-agent | License: MIT | Copyright 2025 Nous Research
# Provided via justclaw.it.com; keep this notice with the code.
"""Gemini long decode: one response spread over several generateContent hops.

Gemini 4 caps each request (32,768 tokens unary, 65,536 streamed) but lets ``maxOutputTokens`` reach
1,048,576 minus the input. A hop that hits the per-request cap with budget left ends with
``finishReason: CONTINUATION`` plus a ``continuationToken``; the client resends the SAME request with
that token (never the partial output) until any other finish reason. Each hop re-reads the prompt plus
everything generated so far, so per-hop ``promptTokenCount`` grows: usage is summed for billing (what
the google-genai SDK reports and Google bills) while the f

... (truncated -- full source via MCP)