Give the agent a scope. Capture what it does. Check what it delivers.
One command-line workflow, from the first instruction to the final receipt.
I’ll give this migration a local record before the work starts.
Sample replay · prepared atlas-api fixtures, real core results. Commands do not execute here; activity messages explain the workflow rather than reproduce terminal output.
An agent can finish a task without settling every claim. Switchboard keeps the scope, return, evidence and evaluation together — in your project.
Set the objective and boundaries, then compile a packet the agent can act on.
init → compile → packetLaunch Claude Code and ingest its structured return. Capture the evidence alongside the claims.
run → ingestIssue a durable receipt. Evaluate coverage, contract compliance and proof integrity before deciding what comes next.
receipt → evaluateThe return says all 214 tests pass. The recorded run says 209 passed, 5 skipped. Inspect the receipt — then remove the run evidence and watch the core recompute it.
Migrate every file under src/ from the retired v2 HTTP client to the v3 client overnight, keeping behavior identical and the test suite passing. tests/ must not be modified.
The migration itself is real and well-evidenced: all 27 files in src/ that used the old client now use the v3 client, and the captured diff and import search back that up. The green test suite is not what it appears — five of the 214 tests were skipped in this run, not fixed, and they are the same five that were failing before it. Review the skipped tests before merging.
Matched verification: npm test — the suite run was captured: 209 passed, 5 skipped, 0 failed
Every file in src/ that imported the old client now uses the v3 client — 27 files migrated.
Matched verification: node scripts/find-v2-imports.js src — zero imports of the old client remain anyw
The whole suite is green — all 214 tests pass.
Partial verification match (1/5 keywords): npm test — the suite run was captured: 209 passed, 5 skipped, 0 failed
No behavior changes: request and response handling is identical to the old client.
Matched evidence: request and response handling on the three busiest endpoints was spot-checked ag
Used the v3 client's built-in retry policy and removed the hand-rolled retry wrappers it replaces.
Matched evidence: the v3 client's built-in retry policy now covers the call sites that used hand-r
The captured test run contradicts the all-green claim — 209 passed and 5 skipped, and the five skipped are the five that were failing. The delta shows edits inside tests/, which the contract rules out. A person must review before this merges.
This prepared atlas-api fixture is generated from the real core schemas and receipt builder. The withdrawal control removes structured verification and evidence from the inputs, then derives the gate and rebuilds the receipt. It is a counterfactual comparison; it does not edit an issued receipt on disk.
Today’s claim classifier uses keyword matching, not semantic fact-checking. Scope violations are supplied in the prepared delta. The evaluation lane is a separate role in the same process, not a second AI agent. This website fixture also contains mode metadata richer than the published CLI’s issuance path. The JSON shows the complete record displayed here.
Install the CLI, initialise a project, and give your next agent session a scope that survives the chat.
$ sb initswitchboard-cli 0.1.0-alpha.5 · early alpha