✳️ pi-aster-provider

Kimi K3, GLM-5.2, GPT-OSS — via Aster

Coding-optimized open weights with lossless 1M context, for pi.

Features

• OpenAI-compatible API — Uses Aster's /v1/chat/completions endpoint
• Coding-optimized models — Kimi K3 (1M context), GLM 5.2, GPT-OSS 120B, and more
• Reasoning models — All chat models think by default; control depth with /reasoning
• Tool use — Function calling works on every chat model (synbad-verified)
• Prompt caching — Discounted cached input on GLM and Kimi K3
• Batch variant — zai-org/glm-5.2-batch, the discounted GLM 5.2 batch lane
• Live model sync — Models refresh from the Aster API in the background

Available Models

| Model | ID | Context | Max Output | Vision | Reasoning | Cache | Input $/M | Output $/M | |-------|----|---------|------------|--------|-----------|-------|-----------|------------| | GLM 5.2 | glm-5.2 | 1.0M | 131K | ❌ | ✅ | ✅ | $1.00 | $4.00 | | GLM 5.2 Batch | zai-org/glm-5.2-batch | 1.0M | 131K | ❌ | ✅ | ❌ | $0.75 | $2.50 | | GPT OSS 120B | gpt-oss-120b | 131K | 33K | ❌ | ✅ | ❌ | $0.15 | $0.60 | | GPT OSS 120B Fast | gpt-oss-120b-fast | 131K | 33K | ❌ | ✅ | ❌ | $0.15 | $0.60 | | Kimi K3 | kimi-k3 | 1.0M | 131K | ❌ | ✅ | ✅ | $2.50 | $12.50 |

Costs are per million tokens.

Non-chat endpoints (e.g. aster/wildflower, per-call search pricing) are intentionally excluded.

Installation

Option 1: Using pi install (Recommended)

Install from npm:

pi install npm:pi-aster-provider

Or install directly from GitHub:

pi install https://github.com/monotykamary/pi-aster-provider

Option 2: With npm

Install from npm:

npm install pi-aster-provider

Option 3: Manual Clone

1. Clone this repository:

   git clone [email protected]:monotykamary/pi-aster-provider.git
   cd pi-aster-provider

2. Set your Aster API key:

   # Recommended: add to auth.json
   # See Authentication section below
   
   # Or set as environment variable
   export ASTER_API_KEY=your-api-key-here

3. Run pi with the extension:

   pi -e /path/to/pi-aster-provider

Then authenticate and run pi:

# Recommended: add to auth.json
# See Authentication section below

# Or set as environment variable
export ASTER_API_KEY=your-api-key-here

pi

Get your API key from asterlab.ai.

Authentication

The Aster API key can be configured in multiple ways (resolved in this order):

1. auth.json (recommended) — Add to ~/.pi/agent/auth.json:

   { "aster": { "type": "api_key", "key": "your-api-key" } }

   The key field supports literal values, env var names, and shell commands (prefix with !). See pi's auth file docs for details.
2. Runtime override — Use the --api-key CLI flag
3. Environment variable — Set ASTER_API_KEY

Get your API key from asterlab.ai.

Environment Variables

| Variable | Required | Description | |----------|----------|-------------| | ASTER_API_KEY | No | Your Aster API key (fallback if not in auth.json) |

Configuration

Add to your pi configuration for automatic loading:

{
  "extensions": [
    "/path/to/pi-aster-provider"
  ]
}

Usage

Once loaded, select a model with:

/model aster kimi-k3

Or use /models to browse all available Aster models.

Reasoning Effort

All Aster chat models are reasoning models (thinking is on by default). Control thinking depth:

/reasoning high

Levels are model-dependent (off … max); pi clamps to the nearest level each model supports and the extension passes the mapped value through as the reasoning_effort request field. Support matrix (probed against Aster):

• gpt-oss-120b / -fast: low medium high — canonically fixed at three levels; thinking cannot be disabled (none is accepted but the model still reasons), and minimal/xhigh/max are rejected with HTTP 400.
• glm-5.2 / glm-5.2-batch: off maps to none (disables thinking), low, medium, high, max. minimal and xhigh are rejected (Z.AI canonical semantics map low/medium onto high behavior).
• kimi-k3: low, high, max (Moonshot canonical — K3 defaults to max thinking and thinking stays on).

API Compatibility Notes

• Reasoning streams as the reasoning_content field (DeepSeek-style), which pi parses natively.
• Both max_tokens and max_completion_tokens are honored; the extension sends max_completion_tokens.
• GPT-OSS models strictly validate prompt + max_completion_tokens ≤ context (131K), so their curated max output is held to 32K; GLM 5.2 and Kimi K3 accept up to 128K completion tokens.
• Image input is not supported by Aster's chat models (text-only).
• gpt-oss-120b (and -fast): Aster's DFLASH speculative decoding rejects grammar-constrained requests — any tool schema sent with strict: true fails with HTTP 400. This extension sets supportsStrictMode: false so pi omits the strict field for these models. Also note the model does not reliably emit parallel tool calls in a single response (it serializes them); sequential tool calling works fine.
• Kimi K3 keeps thinking on at all times (Moonshot canonical); low/high/max map through, max is the model's natural default.

Inference Quality

Verified with synbad (--count 1, plain and --stream):

| Model | Non-stream | Stream | |-------|-----------|--------| | GLM 5.2 | ✅ 15/15 | ✅ 15/15 | | GLM 5.2 Batch | ✅ 15/15 | — | | Kimi K3 | ✅ 15/15 | ✅ 15/15 | | GPT OSS 120B | ⚠️ 12/15 | ⚠️ 11/15 |

All reasoning evals (parsing + multi-turn reasoning preservation) pass on every model. The gpt-oss misses are Aster-side serving quirks — grammar-constrained (strict) tool schemas 400 (mitigated here via supportsStrictMode: false) and serialized parallel calls — not reasoning or history bugs.

Updating Models

To refresh the model list from the Aster API:

npm run update-models

This fetches from /v1/models, updates models.json, and regenerates this README. Idempotent — safe to run repeatedly.

API Documentation

• Aster: https://asterlab.ai
• OpenAI-compatible endpoint: https://api.asterlab.ai/v1
• Models endpoint: https://api.asterlab.ai/v1/models

License

MIT