Context Usage Thresholds Not Working

[yiding-he]

Short description

execute_threshold_tokens seems not working

What happened?

I set the execute_threshold_tokens limit in the settings interface, but the context limit in Cache Diagnostics is inconsistent with my settings.

Diagnostics: magic-context-issue-20260724-102754.md

I am using OpenCode 1.18.4 with TUI.

Diagnostics

Plugin version

v0.32.4

OpenCode version

1.18.4

Platform

No response

Client

OpenCode TUI (CLI)

Log output (optional)

[ualtinok]

Short version: the setting is working — what you're seeing is a runtime clamp plus a display gap on our side.

execute_threshold_tokens is clamped to 80% × the model's context limit at runtime (this is in the field's description in the config UI). Your sidebar screenshot shows the mechanism: the active model resolved to a 128K context limit, so your 190,000 became min(190000, 0.8 × 128000) = 102,400 → the "80%" shown. The 190K value only takes effect on models with ≥237K windows. When the clamp fires we log a warning — check your log for execute_threshold_tokens clamped:.

One thing to verify on your side: 128K may be our fallback guess. For custom/proxy providers not in models.dev we default to 128,000 when the provider doesn't report a limit. If your kimi/LongCat/MiniMax deployments actually serve larger windows, set the context limit explicitly in your OpenCode provider config so the threshold math uses the real window.

On the Cache Diagnostics "inconsistency": that page displays the model's context window (the axis labels), never the execute threshold — the threshold is currently visible only in /ctx-status and the sidebar (84.4% / 80% = usage / effective threshold). We're taking two improvements from this report: an explicit "clamped" indicator in the status surfaces, and a threshold marker in the dashboard timeline.

Also worth noting from your timeline screenshot: frequent switching between a 1M-context model and 128K models forces a full history re-fold on every switch (each segment boundary in the chart). If that's your workflow, /ctx-wrapup before switching down to the small model will compact history under its threshold in one step.

[alfonso-magic-context]

v0.33.0 adds the clamp visibility discussed above: /ctx-status and the sidebar now show an explicit clamp note (e.g. 80% [token-mode, clamped: 190,000 > 80% of 128,000]) whenever execute_threshold_tokens exceeds 80% of the active model's window, instead of silently capping. If your kimi/LongCat/MiniMax windows are really larger than 128K, do also set explicit context limits in the provider config — see the earlier comment.