[Bug] Compaction ignores "auto: false" config and OPENCODE_DISABLE_AUTOCOMPACT env vars #32385

Description

Description:

In version 1.17.7, OpenCode is automatically triggering context compaction even when it is explicitly disabled via both the configuration file and environment variables.

I am managing my context externally using a DCP setup, so I need compaction entirely disabled. Currently, the compaction cycle fires unpredictably, taking over the session despite all configuration flags explicitly telling it not to.

Regardless of how the internal token math evaluates the current context limits, the strict disable flags are being bypassed entirely.

Steps to Reproduce:

1. Set the following environment variables (e.g., in ~/.bashrc):

   export OPENCODE_DISABLE_AUTOCOMPACT=true
   export OPENCODE_DISABLE_PRUNE=true

2. Configure your OpenCode settings with the following block:

   "compaction": {
     "auto": false,
     "prune": false
   }

3. Run a session using a model with a large context window.
4. Observe that OpenCode eventually ignores these overrides and triggers the "Compacting" process anyway.

Expected Behavior:

When auto: false is set in the configuration OR OPENCODE_DISABLE_AUTOCOMPACT=true is present in the environment, the automatic compaction sequence must never be triggered under any circumstances. It should defer to the external context manager, pass the raw context to the provider, or throw a standard overflow error if actually full.

Actual Behavior:

The CLI ignores the boolean off-switches and environment variables, triggering compaction automatically.

Plugins

No response

OpenCode version

1.17.7

Steps to reproduce

No response

Screenshot and/or share link

The context window of GPT 5.5 is though 1.05M however, even if it's overridden (to fix the correct one over the cache) it treats the incorrect 400k as the target calculation, not the opencode.json override one.

Operating System

No response

Terminal

No response