Z.ai coding plan

Route Stella at the GLM coding-plan endpoint — the flat-quota subscription setup, the exact precedence chain, and a full pipeline profile that stops metering the worker.

If you subscribe to a Z.ai GLM coding plan, your quota lives on a dedicated endpoint — https://api.z.ai/api/coding/paas/v4 instead of the pay-per-token https://api.z.ai/api/paas/v4. Same key, same models, different billing: a flat subscription instead of a meter. This page is the complete setup.

Route Stella at the plan

Two equivalent routes — pick one:

The env toggle (quickest — one line in your shell profile):

export ZAI_API_KEY="..."
export ZAI_GLM_CODING_PLAN=1

The settings route (durable, provider-agnostic mechanism). A subscription is personal, so this belongs in user scope:

~/.config/stella/settings.json

{
  "providers": {
    "zai": {
      "base_url": "https://api.z.ai/api/coding/paas/v4"
    }
  }
}

Don't set both, and know the full precedence chain — later entries are silently overridden by earlier ones: --base-url / STELLA_BASE_URL > ZAI_GLM_CODING_PLAN=1 > settings.json base_url > the built-in default. A stray STELLA_BASE_URL in your environment routes you off the plan without a word; a leftover ZAI_GLM_CODING_PLAN=1 overrides the settings file. And if you put the base_url in a project's .stella/settings.json instead of user scope, the trust boundary drops it unless STELLA_TRUST_PROJECT=1 is set.

Verify before trusting it:

stella config    # Base URL must read https://api.z.ai/api/coding/paas/v4

The full profile

With the plan carrying the worker, route the whole pipeline accordingly. Two sensible shapes:

Plan worker, API judge — the balanced shape with the worker's marginal cost near zero. The judge stays cross-family on a metered key you barely touch (a verdict is a few cents, and --test-command runs often skip it entirely):

~/.config/stella/settings.json

{
  "providers": {
    "zai": { "base_url": "https://api.z.ai/api/coding/paas/v4" }
  },
  "agent_engine_config": {
    "default_model": "zai/glm-5.2",
    "pipeline_judge_model": "anthropic/claude-fable-5",
    "pipeline_triage_model": "zai/glm-5.2",
    "effort_auto": "on",
    "reasoning_auto": "on"
  }
}

All-in on the plan — every role on GLM, nothing metered. The judge is same-family with the worker, so give it a strict prompt and prefer arming --test-command so verification is deterministic rather than a same-family opinion:

~/.config/stella/settings.json

{
  "providers": {
    "zai": { "base_url": "https://api.z.ai/api/coding/paas/v4" }
  },
  "agent_engine_config": {
    "default_model": "zai/glm-5.2",
    "pipeline_judge_model": "zai/glm-5.2",
    "pipeline_triage_model": "zai/glm-5.2",
    "agents": {
      "judge": {
        "prompt": "You are a strict, evidence-first code judge. Demand proof.",
        "effort": "high"
      }
    }
  }
}

default_model is what routes the pipeline's execute turns — see model precedence.

Fleets on a flat quota

The plan's economics invert the usual caution: fleet workers all draw on the same subscription, so parallel width costs attention, not dollars:

stella fleet --plan tasks.toml --max-concurrency 4 --watch

Stella's telemetry prices tokens at the catalog's list rates — on a coding plan the $ figures in stella stats are notional (you pay the flat subscription, not per token). Token counts, step counts, and $/resolved comparisons between models remain meaningful; absolute dollars for zai/* do not. --budget caps are enforced against those notional prices too — set them generously or not at all for plan-routed runs.

Z.ai provider reference — models, pricing, thinking mode
Single-key setups — the same idea for other providers
Balanced profile — the metered version of this shape

Route Stella at the plan

The full profile

Fleets on a flat quota

Related

On this page