#6390 [GUI] restore the local agent reply ceiling at large context sizes

open · @Held0fTheWelt · 0 commentaires · Sur GitHub

Setup & installModels & quantsWindows

Description

## Problem and result

A local agent profile caps its reserved reply budget at 32,768 tokens. For a 512K window that is smaller than one quarter of the window, so the host never applies the existing reply ceiling to a regular request. The same quarter-window heuristic can raise an explicit limit for a summary or title on a smaller window. The host now recognizes the profile default directly. Auxiliary requests and distinct explicit limits keep their own budgets.

## Related discussion

The larger Harness changes are tracked in #6342. This PR extracts the local request limit repair from fork commit `9034ca16`.

## Approach

Use the local model's configured output budget to recognize an ordinary agent request that qualifies for the existing `replyCap`. Leave the profile reserve, the cap calculation, and the model-side clamp in place. Requests labeled with a purpose stay on their existing requested limit.

## Validation

**Passed:**
- Windows, Node v24.15.0, commit `57edd5f66`. `node --test app/host/tests/agent-reply-ceiling.test.mjs app/host/tests/agent-context.test.mjs app/host/tests/agent-shape.test.mjs` passed 21 tests, with 9 Harness-dependent tests skipped.
- The scripted 512K fixture received the existing ceiling of 65,536 tokens, including when it omitted the explicit request limit. The 64K fixture preserved compaction, title and distinct explicit limits.
- `git diff --check upstream/main...HEAD` passed.

**Failed:**
- With only the regression tests applied to upstream `3142b66b9`, `node --test app/host/tests/agent-reply-ceiling.test.mjs` failed both tests. The 512K request remained at 32,768 and the 64K compaction request was raised from 19,660 to 32,768.

**Not run:**
- Live model execution at 512K and installer validation. The staged Harness was not configured for these runs.

## Limits

The regression verifies the request option passed to the model adapter. Actual available output still depends on the prompt and the model-side context limit.

Sur le site

Liens install, modèles, releases.