Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

The google/gemma-4-26b-a4b-it model incorrectly displays a 1024k context window in 'openclaw models list' due to a registry lookup failure triggering an incorrect template fallback.

Steps to reproduce

1. Add the following model to the google provider in models.json: { "id": "gemma-4-26b-a4b-it", "contextWindow": 262144, ... }
2. Run openclaw models list.
3. Observe the context window column for google/gemma-4-26b-a4b-it.

Expected behavior

The model should display 256k (or 250k if coarse-rounded), matching the contextWindow value defined in models.json.

Actual behavior

The model displays 1024k. This occurs because ModelRegistry.find() fails to match the ID 'gemma-4-26b-a4b-it', causing the system to fall back to the Google provider's forward-compat logic, which incorrectly maps gemma- prefixed models to the gemini-3-flash-preview template (1M window).

OpenClaw version

2026.4.14

Operating system

Linux (Ubuntu 24.04)

Install method

npm global

Model

google/gemma-4-26b-a4b-it

Provider / routing chain

openclaw -> google

Additional provider/model setup details

The issue is specific to IDs containing non-standard segments (e.g., '-a4b-'). For comparison, google/gemma-4-31b-it matches correctly and displays the expected 250k. The root cause is a lookup failure in the @mariozechner/pi-ai ModelRegistry, which then triggers an incorrect template override in the Google provider.

Logs, screenshots, and evidence

Run results from isolated test script using dist/pi-model-discovery-tCHcDn8S.js:

--- Registry Find Results ---
gemma-4-31b-it            => ✅ FOUND (ContextWindow: 256000)
gemma-4-26b-it            => ✅ FOUND (ContextWindow: 256000)
gemma-4-26b-a4b-it        => ❌ UNDEFINED
----------------------------

Impact and severity

Affected: Users utilizing Gemma 4 models with specific identifiers. Severity: Low (Display bug). Consequence: Misleads users regarding the model's actual context window capacity, potentially leading to unexpected truncation when exceeding 256k.

Additional information

This is a cascaded failure: first, @mariozechner/pi-ai's ModelRegistry.find() fails to match the ID 'gemma-4-26b-a4b-it' (returning undefined), and second, the Google provider's forward-compatibility logic incorrectly maps any 'gemma-' prefixed model to the gemini-3-flash-preview template (1M window).