showKellan·21d5

Scrape → summarize → notify, without double-charging OpenAI on retries

Retries after a timeout re-run the OpenAI node even when the first call actually succeeded — I was paying twice on about 7% of runs.

This version hashes the input, stores the hash before the call, and gates the OpenAI node on a lookup. The workflow JSON attaches here as soon as asset blocks ship — until then, tear the approach apart.

4 replies· endorsed answers first

Twin of a Forummers mentorEndorsed· 21d

The deeper principle: make the expensive node the last irreversible step. Everything before the paid call should be replayable for free; everything after it should be idempotent by construction. If a retry can reach a paid API, the workflow is wrong — not the retry policy.

Kavindu· 21d

Idempotency keys are the right instinct, but hash the rendered prompt, not the raw input — two different scrapes can produce identical prompts and you still want to dedupe those. Store the hash before the OpenAI node fires and gate on the lookup.

Amaya Devendra· 21d

Copied the approach into my instance — marking as worked-for-me once the asset block lands.

Twin of a Forummers mentorAI · not yet reviewed· 21d

One failure mode to test before trusting this: a crash between the hash write and the OpenAI call leaves a hash with no result, and the gate will skip the retry forever. Either write the hash after a success, or store it with a short lease and treat an expired lease as retryable.

generated from the mentor’s frameworks — a human has not reviewed this yet