✨ pi-zro-provider

3+ models through Zro

GLM-5.2, Kimi K3, and DeepSeek V4 / V4.1 Flash — run through the Zro inference endpoint with pi.

Features

• 6+ AI Models — GLM-5.2 (default), Kimi K3, DeepSeek V4 Flash, and DeepSeek V4.1 Flash, straight from Zro's production catalog
• Native reasoning effort — every model ships an endpoint-verified piLevel → level-id map (glm-5.2, glm-5.3-flash, deepseek-v4-flash-0731: full off–max ladder; deepseek-v4.1-flash: low/high/max — catalog has no none, and xhigh/max both send max; kimi-k3: low/high/max — the proxy rejects medium there), so /thinking low etc. maps exactly to what the Zro proxy accepts. Quirks: glm-5.3-flash silently disables thinking at low, so minimal/low map to minimal; glm-5.3 mirrors flash's full ladder — Z.ai officially removed none for the 5.3 family, but the proxy accepts it and streams cleanly (occasionally a short chain-of-thought preamble surfaces in content before the answer), and xhigh/max both send max
• OpenAI-compatible API at https://zro.moonmath.ai/v1
• Official catalog sync from Zro's /api/cli/models endpoint — same one zro models uses
• Quiet live model sync — the catalog refreshes in the background, at most hourly and by one pi process at a time (details)
• Zro login reuse — if you've run zro login, the extension picks up ~/.config/zro/credentials.json automatically (no duplicated keys)
• Footer status widget — session spend/requests and account plan / available spend / usage packs / 30-day activity at a glance

Available Models

| Model | Type | Context | Max Tokens | Input Cost | Output Cost | |-------|------|---------|------------|------------|-------------| | Auto | Text | 1.0M | 131K | — | — | | DeepSeek V4.1 Flash | Text + Image | 1.0M | 384K | $0.14 | $0.28 | | Dolly 1 Security | Text | 1.0M | 64K | — | — | | GLM-5.3 | Text | 1.0M | 131K | $1.40 | $4.40 | | GLM-5.3 Flash | Text + Image | 1.0M | 64K | $0.15 | $0.50 | | Kimi K3 | Text + Image | 1.0M | 131K | $2.50 | $12.00 | Costs are per million tokens. Billing is metered by your Zro plan and usage packs — see zro.moonmath.ai for pricing.

Installation

Option 1: Using pi install (Recommended)

Install from npm:

pi install npm:pi-zro-provider

Or directly from GitHub:

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

Then make sure you're logged in (or set a key) and run pi:

# Recommended: reuse your zro CLI login
zro login

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

pi

Get your API key from zro.moonmath.ai.

Option 2: Manual Clone

1. Clone this repository:

   git clone https://github.com/monotykamary/pi-zro-provider.git
   cd pi-zro-provider

2. Make sure you have a Zro credential (any of the Authentication options).

3. Run pi with the extension:

   pi -e /path/to/pi-zro-provider

Usage

After loading the extension, select a model with:

/model zro glm-5.2

Or start pi directly with a Zro model:

pi --provider zro --model glm-5.2

Reasoning Effort

All Zro models are reasoning models. Control thinking depth with /thinking or the CLI:

pi --provider zro --model glm-5.2 --thinking xhigh

The available levels are model-specific and come straight from Zro's catalog:

| Model | pi levels | Sent as reasoning_effort | |-------|-----------|----------------------------| | glm-5.2 | off, high, max | none, high, max | | glm-5.3 | off, minimal, low, medium, high, xhigh, max | none, minimal, minimal, medium, high, max, max | | kimi-k3 | low, high, max | low, high, max | | deepseek-v4-flash-0731 | off, low, high, max | none, low, high, max | | deepseek-v4.1-flash | low, high, xhigh, max | low, high, max, max |

Levels map one-to-one through thinkingLevelMap, so off sends Zro's none token instead of silently dropping the field.

Footer Status

A Neuralwatt-style status line sits below the editor. It appears after the session's first Zro turn completes (never before — no half-empty line on fresh sessions or other providers), refreshes its balance when the agent run fully settles, and makes no status-related API calls in sessions that never use Zro. The account prefetch at session start runs only in interactive sessions (not pi -p or child agents) and is shared across pi processes: one fetches /api/cli/status and writes ~/.pi/agent/cache/zro-account.json, and sessions starting within the next 60 seconds render from that file instead:

⚡ $0.42 · 7 req                   Pro ◆ $12.34 avail · $2.50 pack · 1.2k req/30d
└─ session spend+requests ─┘     └─ plan · available spend · packs · activity ──┘

The left side tracks what this session has done — request count, the spend Zro attaches to a response when it reports one, and elapsed time (raw token totals stay in pi's own footer, which can't show Zro cost since its catalog carries no pricing). The right side shows your plan name, total available spend, usage-pack balance, and 30-day request/token activity from Zro's /api/cli/status endpoint — the same account view zro status prints — plus a request-rate atom from response headers when present. The right side compresses progressively as the terminal narrows, and turns to a warning color at/below the lowBalanceUsd threshold.

Configuration

Edit ~/.pi/agent/extensions/zro.json or run /zro-status:

| Setting | Values | Default | |---------|--------|---------| | session | widget | statusbar | off | widget | | account | widget | statusbar | off | widget | | hideOnOtherProvider | true | false | true | | lowBalanceUsd | number | null | 10 | | glyphs | auto | unicode | ascii | auto |

Non-interactive toggles:

/zro-status session widget|statusbar|off
/zro-status account widget|statusbar|off
/zro-status hide true|false
/zro-status lowBalance 25|off
/zro-status glyphs auto|unicode|ascii
/zro-status refresh
/zro-status reset

glyphs: "auto" swaps the emoji footer glyphs (bolt, gem, warn, separator) for ASCII equivalents on legacy terminals (mintty/Cygwin), whose cell-width tables can disagree with the extension's width math. On those terminals a full-width widget line can wrap physically, which desynchronizes pi's renderer and leaves stale rows behind. The widget also never paints the terminal's last column, and clamps an explicit unicode choice to ASCII on legacy terminals; the statusbar is not edge-padded and always honors the exact choice.

Authentication

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

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

   { "zro": { "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. Environment variable — Set ZRO_API_KEY
3. zro login reuse — If you've already run zro login, the extension reads ~/.config/zro/credentials.json (XDG_CONFIG_HOME aware) automatically. No extra configuration needed.
4. Runtime override — Use the --api-key CLI flag

Get your API key from zro.moonmath.ai.

Environment Variables

| Variable | Required | Description | |----------|----------|-------------| | ZRO_API_KEY | No | Your Zro API key (env var fallback) | | ZRO_ENDPOINT_ROOT | No | Override the endpoint root (default https://zro.moonmath.ai) | | ZRO_AUTH_URL | No | Override the authentication/API host (used by zro login; the extension resolves catalog/status from it too) | | PI_PROVIDER_SYNC_TTL_MS | No | Model sync freshness window in ms (default 3600000; 0 = revalidate every session start) |

Configuration

Add to your pi configuration for automatic loading:

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

Catalog and Compat Settings

Model metadata matches the transform Zro's own pi adapter applies to its /api/cli/models catalog:

• canonical ids, display names, context windows, and output caps
• thinkingLevelMap built from each model's published reasoning levels (piLevel → provider level id)
• compat.supportsReasoningEffort: true and maxTokensField: "max_tokens" (plain openai-completions, no custom thinkingFormat)
• supportsDeveloperRole: false
• all text-only inputs

patch.json is reserved only for a verified provider regression and is currently empty.

Patch Overrides

patch.json is applied on top of models.json only for verified endpoint corrections. It is currently empty because every live field comes from Zro's canonical model catalog.

Model sync

Startup never waits on the network. pi starts with the cached catalog (~/.pi/agent/cache/zro-models.json, falling back to the bundled models.json) and revalidates against /api/cli/models in the background — only when:

• a Zro API key is configured (auth.json, ZRO_API_KEY, or a zro login; no key, no request),
• the cache is older than 1 hour, missing, or was fetched with a different API key (so a new zro login refreshes it right away), and
• no other pi process is already refreshing it (a zro-models.lock file next to the cache).

So /new, /reload, pi -p and fan-out child agents share one cache instead of each calling the API; a session that skips the fetch still picks up a cache another process refreshed. A failed refresh is retried after 5 minutes at the earliest. PI_PROVIDER_SYNC_TTL_MS changes the window (0 revalidates on every session start). Only a hash fingerprint of the key is stored (zro-models.sync.json), never the key itself.

The footer's startup account prefetch follows the same rules with its own 60-second window (zro-account.json, see Footer Status); balance polls after each agent run are unchanged.

Updating Models

Run the update script to fetch the latest models from the Zro catalog:

node scripts/update-models.js        # uses zro login credentials
# or
ZRO_API_KEY=your-api-key node scripts/update-models.js

This will:

1. Fetch models from https://zro.moonmath.ai/api/cli/models
2. Regenerate models.json as pure metadata from the live catalog
3. Apply overrides from patch.json only when building the README
4. Remove custom models now available upstream from custom-models.json
5. Reconcile delisted models through the 14-day deprecated-models.json grace layer
6. Update models.json and the README model table

License

MIT