Repository navigation
tool.execute.before: output.args mutations not propagated to tool execution #26909
Description
Mutating output.args in a tool.execute.before hook doesn't affect the tool's behavior.
Evidence: The opencode-workload-router plugin sets output.args.model to override subagent model. Toast fires confirming the routing decision, but the LLM log shows the subagent used the agent-default model instead.
service=llm providerID=wandb modelID=deepseek-ai/DeepSeek-V4-Flash mode=subagent
The [tier-4] prompt mapped to deepseek-ai/DeepSeek-V4-Pro, but the subagent ran Flash.
Tested: args.model (object), args.modelID, args.providerID, args.variant — all ignored.
Expected: output.args mutations in tool.execute.before should propagate to tool execution.
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon May 11, 2026
This issue might be a duplicate of existing issues. Please check:
tool.execute.beforeandtool.execute.afterhooks are declared but never triggered in the runtime)Please close this issue if it is indeed a duplicate of one of the above, or clarify how it differs.
gretel commentedon May 11, 2026
closing as duplicate, sorry.