openclaw - 💡(How to fix) Fix [Bug]: sessions_spawn model parameter ignored due to write/read field mismatch [1 comments, 2 participants]

Description

When sessions_spawn is called with an explicit model parameter (e.g., "ollama/kimi-k2.6:cloud"), the subagent ignores the parameter and uses the agent's default model instead. The spawn-supplied model never reaches the inference path.

This is caused by a field-name mismatch: the spawn writer populates entry.model and entry.modelProvider, but the agent-command handler reads from entry.modelOverride and entry.providerOverride. The handler's override-detection check returns false, and the spawn falls through to the agent's default model.

Version

OpenClaw 2026.4.26 (be8c246) (also reproduces in 2026.4.25)

Source locations

Writer — dist/subagent-spawn--QfZonyO.js lines 339-345 (function buildDirectChildSessionPatch):

if (typeof patch.model === "string" && patch.model.trim()) {
    const { provider, model } = splitModelRef(patch.model.trim());
    if (model) {
        entry.model = model;
        if (provider) entry.modelProvider = provider;
    }
}

Reader — dist/agent-command-BxeGYLZy.js line 511:

const hasStoredOverride = Boolean(sessionEntry?.modelOverride || sessionEntry?.providerOverride);

Subsequent reads at lines 536-537 and 559-560 only consult modelOverride and providerOverride. Since the writer never populates those fields, hasStoredOverride evaluates to false and the override path is skipped entirely.

Reproduction

1. Configure an agent with a primary model that differs from a model defined in models.providers.ollama.models (e.g., primary = minimax-portal/MiniMax-M2.7-highspeed, also configure kimi-k2.6:cloud under the ollama provider).
2. Call sessions_spawn with model="ollama/kimi-k2.6:cloud" and any prompt.
3. Inspect the subagent session JSONL for the model_change event.

Expected: provider: "ollama", modelId: "kimi-k2.6:cloud" Actual: provider: "minimax-portal", modelId: "MiniMax-M2.7-highspeed" (the agent default)

The session metadata also shows api: "anthropic-messages" instead of api: "ollama", confirming the request was served by the wrong provider.

Local fix that resolved it

Modified buildDirectChildSessionPatch to also write to modelOverride and providerOverride:

--- a/dist/subagent-spawn--QfZonyO.js
+++ b/dist/subagent-spawn--QfZonyO.js
@@ -339,6 +339,8 @@ function buildDirectChildSessionPatch(patch) {
     if (typeof patch.model === "string" && patch.model.trim()) {
         const { provider, model } = splitModelRef(patch.model.trim());
         if (model) {
             entry.model = model;
+            entry.modelOverride = model;
             if (provider) entry.modelProvider = provider;
+            if (provider) entry.providerOverride = provider;
         }
     }
     return entry;
 }

After this change, sessions_spawn with model="ollama/kimi-k2.6:cloud" correctly routes to the Ollama provider. Verified via subagent session metadata showing api=ollama, provider=ollama, model=kimi-k2.6:cloud.

Suggested upstream fix

Align the field names. Either standardize on modelOverride/providerOverride as the canonical fields and have buildDirectChildSessionPatch write to those exclusively, or have the agent-command handler check both model/modelProvider and modelOverride/providerOverride as fallbacks.

Why this matters

This bug causes silent model substitution. The user requests one model, the system returns output from a different model, and there is no error or warning. For workflows that depend on specific models for quality, cost, or compliance reasons (e.g., routing writing tasks to a thinking-mode model while general chat uses a faster model), this produces unpredictable results that are difficult to diagnose without source-level investigation.