dsh-zcode-start

English · 简体中文

Use your Z.ai / ZCode Start Plan quota directly inside DeepSeek Harness — GLM models as a normal provider, plus plan claiming, quota refresh and sign-in, all from a settings page.

One-line description: Bring ZCode's Z.ai / Start Plan quota into DeepSeek Harness as a native GLM model provider, and manage the quota from DSH.

Features

Area	What it does
Model provider	Registers zcode-start-plan with GLM-5.3-Flash and GLM-5.3 (1M context, 128k output), tool calling, streaming reasoning
Reasoning levels	off / low / high / max, sent as output_config.effort exactly like the official client
Settings page	Account and JWT status, quota progress bars, model capability table (from the API, not hard-coded), expandable request log, device-code sign-in / sign-out
Claimable plans	Manual check plus a 30-minute background probe that raises a corner card when a gift plan is available
Tools	zcode_quota, zcode_claim, zcode_reset, zcode_chat, zcode_login
Reliability	Automatic retry with exponential backoff (10 attempts, 5-minute budget) for rate limits, gateway errors and empty streams
Error reporting	28 gateway business codes plus HTTP fallbacks, each mapped to a readable message and a retry decision
Privacy	Outbound scrubbing keeps harness identity out of the upstream prompt; credentials never leave the machine

Requirements

• A DSH installation with a profile.
• The ZCode desktop client installed and signed in at least once. The plugin reads its local credential store (~/.zcode/v2/credentials.json) and the device id from ~/.zcode/v2/telemetry-state.json.
• An account holding an active Z.ai / Start Plan entitlement.

Install

dsh plugin add github:AkutagawaKoyori/dsh-zcode-start

With the plugin_manager tool:

{ "action": "install_bundle", "target": "github:AkutagawaKoyori/dsh-zcode-start" }

Local checkout:

{ "action": "install_bundle", "target": "/absolute/path/to/dsh-zcode-start" }

The repository declares dsh.bundle.patch in package.json with cordis.patch.yml beside it, which is what makes dsh plugin add work. Then open Settings → ZCode in the Web UI; restart the harness if the provider does not appear in the model list.

Usage

1. Settings → ZCode shows the account, JWT status, quota bars and the request log.
2. Pick ZCode (Z.ai Start Plan) as the provider, then GLM-5.3-Flash or GLM-5.3 and a reasoning level.
3. Use Check claimable to look for gift plans; the background probe also notifies you.

Privacy

• Credentials are read from the local ZCode store on demand and are never written to this repository, to logs, or to commit history.
• The request log lives in ~/.dsh/zcode-plan-logs/ and is redacted before writing: every authorization / x-api-key / *token* header or body field becomes ***.
• Outbound text is scrubbed (DeepSeek Harness → ZCode, harness package and command names neutralised) before it reaches the gateway.
• This plugin adds no telemetry.

Notes

• identity.json holds the official ZCode system-prompt envelope the gateway expects; a missing or malformed envelope makes it answer code 3012. It contains no user data — only role text with placeholders filled at runtime.
• claim needs a live activity; reset needs a paid Coding Plan. Otherwise the gateway's own message is surfaced unchanged.
• Quota buckets appear only after the server provisions a granted plan. Until then the settings page shows the granted total with "用量待统计".

License

MIT — see LICENSE.