TanStack AI version#
main and t3/prepare-v1-release (1.0.0). @tanstack/ai-sandbox.
Framework/Library version#
Any. Server side, with a harness adapter (@tanstack/ai-claude-code, -codex, -opencode, -grok-build, -acp).
Describe the bug and the steps to reproduce it#
A tool marked needsApproval: true runs with no approval when a harness adapter calls it through the sandbox tool bridge.
createToolBridgeCore().callTool (packages/ai-sandbox/src/tool-bridge.ts, about line 162) looks up the tool and calls tool.execute() directly. It never reads needsApproval. The harness adapters bridge every chat() tool (packages/ai-claude-code/src/adapters/text.ts, toolsForChoice, about line 155).
Steps:
- Define a server tool with
needsApproval: true and an execute() that logs.
- Pass it to
chat({ adapter: claudeCodeText(...), tools: [tool], middleware: [withSandbox(sandbox)] }).
- Ask the agent to call the tool.
- The log line prints. No
tool-approval interrupt is raised.
Expected: the run pauses with a tool-approval interrupt, as it does for a normal chat() adapter. Or the bridge refuses tools with needsApproval with a clear error.
Found in the full review of #1840. docs/structured-outputs/harnesses.md now describes the current behavior.
Your Minimal, Reproducible Example - (Sandbox Highly Recommended)#
See the steps above. The bypass is visible in callTool in packages/ai-sandbox/src/tool-bridge.ts.
Do you intend to try to help solve this bug with your own PR?#
Yes, I am also opening a PR that solves the problem this issue identifies.