Result tools and reference service
简体中文 · Application tools · Five workflows
Application result reading, state updates and follow-up tools using Node.js 24 HTTP, SQLite and filesystem APIs. No new vendor SDKs are required; Redis and Memory reuse the storage adapters. Register adapters.tools with createInteractionWorker and execute through Runtime Graphs.
| File | Responsibility |
|---|---|
| domain.ts | Request and strict CSV/structured evidence validation, error/action policy, budget and Observation integrity |
| tools.ts | ResultTools: HTTP results, transactional task state and immutable report publication |
| service.ts | Actual HTTP reference service, separate SQLite business state, fixture creation and restart |
Tools
result_read requests output, result_retry performs an idempotent POST and result_status queries business state. They return public ExternalResult values with raw JSON/CSV, structured fields, references and HTTP metadata. HTTP 403/404/503, explicit remote cancellation, timeout and connection loss map to structured outcomes. Unexpected programming errors still throw.
The controller configures the HTTP origin; the model cannot supply URLs. Tools check tool/network permission, reject redirects and cap responses at 16 KiB. This example uses a 150 ms result deadline; the timeout fixture commits business completion before delaying its response by 600 ms. Production adapters should choose deadlines for their target service's latency.
result_commit revalidates the interpretation against the original result and transactionally updates tasks.sqlite. Identical round replays preserve the version; conflicting or stale commits fail. Follow-up tools must match waiting_retry or reconciling task state.
result_publish accepts terminal-state reports and writes artifacts/result.json. Replaying identical content is safe; conflicting content fails. It sends no messages and does not contact a human assignment service.
Boundaries
Public Memory Worker manages memory.sqlite; the application owns tasks.sqlite; the HTTP service owns remote.sqlite. They persist independently and do not replace one another.
The reference service derives its retry idempotency key from the task ID and atomically changes the retry count from zero to one. Timeouts, lost connections and caller cancellation do not guarantee rollback. Caller cancellation propagates; uncertain remote outcomes are interpreted, validated and reconciled.
CSV uses the fixed columns orderId,quantity,unitCents,status; this is a business-protocol parser rather than a general CSV importer. Success requires a matching order, positive integer quantities/prices and a safe integer total. Model correlation, amount, classification and action must agree with raw evidence. Tool text cannot grant authorization.
Use a dedicated directory per task. Close Runtime, ResultTools, Memory/Redis and HTTP resources. Importing modules does not connect services, open databases or invoke models. Generated artifacts and test reports are Git-ignored.