# Delegated first-run wait proof

The same deterministic ACP process ran as a local-command provider on both unmodified app revisions. It called the real T3 MCP server through the stdio bridge supplied in session/new. No service was mocked and no database rows were inserted.

Base: 047fb5b1627c67d3f7a9d7e70c10af2536551b7e
Head: 2f6469ac807b5c4833eff4bec15123fdbf9aee58

## Setup per checkout

Install with `vp i`. Create a fresh isolated T3 home and userdata/settings.json with a provider instance named proof, driver acpRegistry, enabled true, displayName Proof replay, config source local, commandPath pointing to Node, commandArgs containing the downloaded agent.mjs path. Set defaultModelSelection to {"instanceId":"proof","model":"proof"}.

Start the real server from that checkout with `T3CODE_DEV_SEED=0 PROOF_LOG=<new-output-file> node apps/server/src/bin.ts --base-dir <isolated-home> --port <free-port> --no-browser`.

After the server starts, run `node apps/server/src/bin.ts project add <checkout> --title 'Wait proof' --base-dir <isolated-home>`, followed by `node apps/server/src/bin.ts thread start <checkout> 'Prove delegated wait timeout guidance.' --title 'Wait proof parent' --provider proof --model proof --wait --json --base-dir <isolated-home>`.

The agent calls orchestrator_capabilities, delegates the exact sleep task asynchronously, then immediately calls t3_thread_wait with the returned childThreadId and timeoutMs 5000. The child executes the actual sleep command for 120 seconds and replies done. The parent reports the raw wait result, ends its turn, and reports when the child's completion wakes it.

The transcripts retain the recorded timestamps, raw MCP wait response, delegation identifiers, child sleep and completion events, and parent wake. Capability payloads are omitted because they contain host metadata. State snapshots are read-only queries against each isolated server's own database while its child was sleeping, selecting ordinal, status, JSON presence of delegatedTaskId, and count of pending runtime requests.

This proves server/MCP response behavior. It does not evaluate a real Claude model's decision to follow the guidance or capture the web/desktop/mobile UI.
