Skip to content

ACP usage_update reports size: 65536 for deepseek/deepseek-v4-pro (should be ~1M) #30120

Closed

Description

@swear01

Summary

When OpenCode runs deepseek/deepseek-v4-pro over ACP, the usage_update notification reports size: 65536 (64 K). DeepSeek V4 Pro advertises a 1 M context window, so ACP clients render a misleadingly tight "ctx N / 64 K" indicator and warn the user about running out of context far too early.

Same pattern as #16709 (MiMo-V2-Flash-TEE reporting 32 K instead of 262 K) and #20317 (GitHub Copilot Claude models reporting wrong context limits) — likely a stale or wrong entry in the model registry, but for a different provider/model pair.

Repro

OpenCode 1.15.11 launched with --model deepseek/deepseek-v4-pro:

```
[15:23:38.714] usage_update raw={"sessionUpdate":"usage_update","used":14388,"size":65536,"cost":{"amount":0.005717878,"currency":"USD"}}
```

size: 65536 is exactly 64 K — looks like a default/stale value rather than the model's real window.

Expected

size should reflect DeepSeek V4 Pro's actual context capacity (~1 M tokens). Whatever the canonical source is (models.dev entry, hard-coded provider config, etc.) needs the correct value.

Why this matters for clients

ACP clients (HAPI, Zed, Fabriqa, …) use size directly as the context-window denominator in their status bars. Reporting 64 K when the model can handle 1 M means:

  • the "% used" indicator pegs near 100% on prompts that are nowhere near the real ceiling,
  • compact/handoff warnings fire prematurely,
  • users may waste tokens trying to shrink prompts that were already fine.

Related

Environment

Activity

  1. github-actions commented on May 31, 2026

    @github-actions
    Contributor

    This issue might be a duplicate of existing issues. Please check:

    The ACP size field discrepancy described here (65536 vs ~1M) appears to be a downstream symptom of the same stale model registry entry tracked in those issues.

  2. nexxeln commented on Jun 1, 2026

    @nexxeln
    Contributor

    looks like this was caused by a stale local models.dev cache rather than acp itself. current models.dev reports deepseek-v4-pro at 1m context, and acp just forwards model.limit.context as usage_update.size.

    running opencode models --refresh should fix it

  3. f0d471 commented on Jun 11, 2026

    @f0d471

    test

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions