Repository navigation
Agents should check orchestrator_capabilities before saying a provider is unavailable
#39831
satyalyadav
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Problem
A T3 Code thread on OpenCode (
opencode-go/deepseek-v4.1-flash) was asked to consult Opus 5.5 on high through Kiro. The agent said Kiro and Opus were "not connected in this environment." That was wrong. Kiro was enabled, andorchestrator_capabilitiesreturnedkirowithclaude-opus-5.5andreasoningEffort: high. The agent had not called the tool. It assumed from its own subagent allowlist and stale model knowledge that Kiro was unavailable. After the user insisted, it checked, and the delegate ran on Kiro.Proposal
Tighten one line in the orchestration instructions. The instructions already require
orchestrator_capabilitiesbefore the firstdelegate_task, but an agent can decline before it reaches that step. This closes that gap.Current: "Provider or model unavailable: say so."
Proposed: "Never say a provider or model is unavailable until
orchestrator_capabilitiesconfirms it. When the user names a provider or model, check before declining."Scope
packages/provider-core/src/server/orchestrationInstructions.ts.Evidence
opencode-go/deepseek-v4.1-flashorchestrator_capabilities, foundkiro / claude-opus-5.5, and completed the delegate.Related
#23258, #31737, #39747. None covers this check.
All reactions