[Feature Request] Support dynamic/configurable prompt overhead in cost estimation for reasoning & large-system-prompt models #1538

Background & Current State

In internal/agent/estimate.go and internal/scan/estimate.go, the pre-run cost projection uses hard-coded estimation constants:

• promptOverheadTokens = 2000
• avgMainRoundsPerFile = 7
• avgOutputTokensPerRound = 700

As stated in the code comments, these heuristics are meant to provide users with an order-of-magnitude projection before a large review run.

Motivation

With the growing adoption of reasoning-oriented models (such as GPT-5, o-series, and Claude with extended thinking capabilities) or custom setups with extensive custom rules and long system prompts, the fixed 2000 token overhead and 700 output token heuristic can substantially underestimate the pre-run cost projection for large diff change sets.

Proposed Improvement

Consider making the estimation heuristics slightly more adaptive or configurable:

1. Model-aware defaults: Apply higher baseline overhead estimates when a reasoning/extended-context model is resolved.
2. Template-derived scaffolding: Optionally derive the base overhead from the actual rendered template lengths rather than a static 2000 constant.
3. User override in config: Allow power users to optionally configure estimation_overhead_tokens in config.yaml if they use heavily customized rule sets.

Code References

• internal/agent/estimate.go:25-37
• internal/scan/estimate.go