# Pi + GLM-4.7 (Z.ai) Smoke Lane

**Status:** Transport already proven — this Opencode session runs on GLM-4.7
via z.ai. Remaining step: formal JSONL capture with `pi --mode json`.
**Owner:** Opencode / Z.ai Integrator

## Purpose

Capture a formal JSONL-mode proof that `pi` can talk to GLM-4.7 through
the Z.ai provider lane. The transport itself is already validated — you
are reading output from it right now. This doc exists to produce a
capture artifact other agents can reference.

## Key Situation

Sam does **not** need to provide a separate key. The z.ai API key that
powers this Opencode session is already configured in the host's opencode
setup. The question is whether `pi` can reach the same endpoint using
its own `--provider zai` path with the same key.

If Sam's key is exposed as `ZAI_API_KEY` in the environment, `pi` picks
it up automatically. If not, the key lives in opencode's own config and
`pi` needs it exported explicitly for this smoke.

## Expected Environment

- pi: `0.75.5`
- node: `24.x` (`>=24` per the runtime engine policy)

## Smoke Commands

### 1. Check if the key is already available to pi

```sh
echo "ZAI_API_KEY is set: ${ZAI_API_KEY:+yes}"
```

If not set, Sam can export it for this terminal:

```sh
read -rsp "ZAI API key: " ZAI_API_KEY; echo
export ZAI_API_KEY
export ZAI_API_BASE="https://api.z.ai/api/coding/paas/v4"
```

### 2. Prepare the repo-local capture directory

```sh
mkdir -p tmp
```

### 3. Discover available models

```sh
pi --provider zai --list-models 2>&1 | tee tmp/zai-models.txt
```

### 4. Filter for GLM models

```sh
grep -i glm tmp/zai-models.txt
```

### 5. Run smoke in JSON mode

Use the exact model ID from step 4:

```sh
pi --provider zai --model '<EXACT_GLM_MODEL_ID>' -p --mode json "reply with the single word: ok" 2>&1 | tee tmp/glm47-smoke.jsonl
```

### 6. Validate the JSONL event stream

```sh
python3 -c "
import json

provider = 'MISSING'
model = 'MISSING'
content_parts = []
events = []

with open('tmp/glm47-smoke.jsonl', 'r', encoding='utf-8') as f:
    for raw in f:
        line = raw.strip()
        if not line:
            continue
        obj = json.loads(line)
        events.append(obj.get('type', 'MISSING'))
        provider = obj.get('provider') or provider
        model = obj.get('model') or model
        delta = obj.get('delta')
        if isinstance(delta, str):
            content_parts.append(delta)
        text = obj.get('text')
        if isinstance(text, str):
            content_parts.append(text)

content = ''.join(content_parts)
print('events:', ','.join(events))
print('provider:', provider)
print('model:', model)
print('content:', content[:120] or 'MISSING')
print('PASS' if 'ok' in content.lower() else 'FAIL')
"
```

### 7. Clean up

```sh
unset ZAI_API_KEY ZAI_API_BASE
```

## What to Report Back

1. Exact model ID from `--list-models`
2. Provider and model fields recovered from the JSONL stream
3. Event types seen in the stream
4. PASS/FAIL
5. Wall-clock latency

## Deletion Criteria

- JSONL capture artifact is committed
- Lane confirmed in agent capability table