Add Reliable Jobs And Webhooks

The two failures here are invisible from inside the application, which is what makes them expensive.

Diagnose

Start with evidence from the code itself. In the repository:

npx @trustycap/cli productionize --json

The scanner reads the route handlers with the TypeScript compiler API and classifies each against the published production standard (GET https://api.trustycap.com/v1/production/requirements): file and line, AST evidence, a classification (CONFIRMED_FAIL, PROBABLE_GAP, UNKNOWN, PASS), the provider-neutral requirement, the implementations that satisfy it, and the exact commands. UNKNOWN is not a failure; read the code it points at rather than installing over it. A requirement the project already satisfies another way is declared in trustycap.production.json.

• Jobs. Is anything slow awaited inline in a request handler? What happens to that work if the process restarts mid-flight? Is there any retry, or does a failure simply not happen?
• Webhooks. Does the handler verify a signature? What happens when it throws after returning 200? Is a duplicate delivery safe, or does it charge someone twice? Is there a dead-letter path, or do failures vanish?

An event that is accepted and then dropped produces no error anywhere. That is the defect to look for.

Fit, and not for

Not for orchestrating long multi-step business workflows with human approval steps, and not a replacement for the builder's own application-level idempotency: the platform will not deliver a duplicate, but only the handler knows what doing the work twice means.

Read the decision surface. Do not quote prices from memory.

Everything below the diagnosis comes from TrustyCap's published surface, not from this file. Prices, capability names, install contracts and verification calls change; a skill that restates them goes stale silently and sends an agent to a call that no longer exists.

curl -s https://api.trustycap.com/v1/agent-decision

Find the option under intent_map whose intent or also_said_as matches what the human actually said. Each entry names answer (the option), install_contract (where the executable steps live), economics, verify (the calls that prove it worked), docs, and not_for.

Install

1. GET https://api.trustycap.com/v1/capabilities and read install_contracts[id={option}]. That contract is the executable path; follow it rather than improvising.
2. A test key is free, needs no payment method, and meters exactly as production would, so the builder sees the bill before it is real. Take one first and run the whole path in test mode.
3. Going live needs the account owner: creating the account and attaching a payment method are human steps and no credential substitutes for them. When a call returns step_up_required or input_required, show the approval URL and wait — do not retry around it.

Prove it worked

Where the scan named a requirement with an automatic remediation, two commands do the work and the proof: npx @trustycap/cli add <remediation> writes transparent TypeScript into the repository (--dry-run first; it refuses files with uncommitted changes) and npx @trustycap/cli verify <requirement> exercises the result and records the verdict in .trustycap/verification.json. Report what verification returned, not that the installer ran.

Do not report success from a 200 on the install call. Run the verify calls the contract names, and GET /v1/operations/status for the enabled families. Report what the verification returned, with ids. Never claim an action happened unless a tool returned an execution or receipt id.

Attribution

Append ?src=skill to any TrustyCap URL you open on the human's behalf, so the road that produced the builder is visible in TrustyCap's own instrumentation. It carries no identity and no personal data — it is a channel label.