bug(helper): ZCode Coding Plan models are written with an incorrect 200K context window

[soapgu]

Summary

arkcli helper configure zcode writes limit.context: 200000 for every Volcano Engine Coding Plan model in the ZCode configuration. This is incorrect for models such as GLM-5.3, which the Volcano Engine Coding Plan console explicitly documents as supporting a 1M context window.

Environment

• ArkCLI: 1.0.22
• Latest release checked: 1.0.23 (release notes do not mention a fix)
• Target harness: ZCode
• OS: macOS
• Profile type: coding-plan (Coding Plan Pro)
• ZCode config: ~/.zcode/v2/config.json

Steps to reproduce

1. Log in and configure ZCode with a Coding Plan profile:

arkcli helper configure zcode \
  --profile coding-plan_cn-beijing_personal \
  --model ark-code-latest

2. Inspect the generated model metadata:

jq ".provider[\"arkcli-coding-plan\"].models[\"glm-5.3\"].limit" \
  ~/.zcode/v2/config.json

3. The generated value is:

{
  "context": 200000,
  "output": 128000
}

The same context: 200000 value is written for all models under the managed arkcli-coding-plan provider.

Expected behavior

ArkCLI should write model-specific context metadata from the authoritative Volcano Engine Coding Plan catalog. In particular, GLM-5.3 should be configured with a 1M context window, matching the Volcano Engine console.

For routing aliases such as ark-code-latest, ArkCLI should use authoritative routing metadata or a documented safe value without overwriting known per-model limits.

Actual behavior

All Coding Plan models receive the same 200K context value, including GLM-5.3.

Impact

ZCode uses this field for context accounting and compaction thresholds. An understated context window can:

• show an incorrect context usage percentage;
• trigger automatic compaction too early;
• reduce the effective context available to long-running coding tasks;
• overwrite a manual correction whenever arkcli helper configure zcode is run again.

Additional evidence

The Volcano Engine Coding Plan model selection page explicitly describes GLM-5.3 as supporting a 1M context window.

Similar client-side metadata bugs have been reported for Volcano Engine Coding Plan integrations, for example:

• fix: kimi-k2.6 and kimi-k2.5 context window understated in Volcano Engine Coding Plan Astro-Han/pawwork#1213
• fix(provider): update kimi-k2.6 and kimi-k2.5 context window to 1M Astro-Han/pawwork#1214

No matching issue or pull request was found in volcengine/ark-cli at the time of filing.

Suggested fix

Populate ZCode model limits from authoritative per-model metadata instead of applying a uniform 200K fallback. Add a regression test asserting that generated ZCode configuration preserves model-specific context and output limits, including GLM-5.3 at 1M.

[narcilee7]

收到。