# PR#49: the README explains why output limit and timeout depend on the model

> **WARNING:** This document describes only the change applied in this PR.
> It may already be outdated once the next PR is merged.
> Historic PR documentation is not maintained along with new PRs; treat it as a snapshot.

## The Problem

On 2026-10-08 the review of PR#47 got no answer from MiniMax M3:
the model used up `max_output_tokens = 32000` with reasoning and returned empty content with `finish_reason: length`.
The operator raised the limit on mih09 to 48000 and the review budget to 0.25 USD,
since a higher output limit lowers the input limit within the same budget.
config/werkrevisor.toml in the repository still holds the old values.

The README lists `max_output_tokens` and `timeout_seconds` without saying that their useful values depend on the model.
An operator who changes the model therefore keeps limits that fit the old one,
and finds out only when reviews fail or time out.

## Non-Goals

- The code stays unchanged: werkrevisor gets no built-in values per model, since only the README documents them for now.
  Built-in profiles per model may follow when more hosts use different models.
- The built-in defaults stay unchanged; only the example config config/werkrevisor.toml gets new values.

## The Solution

- The README section `[model]` explains why the output limit, the timeout and the reasoning effort depend on the model,
  and how the output limit and the budget depend on each other.
- A table in the README lists the values measured per model on 2026-10-06 and 2026-10-07.
- config/werkrevisor.toml gets the values deployed on mih09: `max_output_tokens = 48000` and `max_usd_per_review = 0.25`.

## Additional Changes

- The header of config/werkrevisor.toml and config/werkprobe.toml said "The PR never changes this file";
  it now says that a PR which changes the file is still reviewed with the deployed copy.

## Follow-up PRs

- werkrevisor shall not repeat a hint of the PR document check that the author already resolved,
  as long as the section of the hint stays unchanged.
  In this PR, the same false hint on the Non-Goals came back after each commit.

## Prerequisite PRs

- PR#47: werkrevisor 0.23.0, merged.