Where should safe-write guarantees stop?

[arturict]

I kept running into the same failure mode while building Work SDK: an issue tracker commits a comment, the connection times out, and the job retries. You either duplicate the comment or give up and risk losing the intended write.

The boundary I ended up with is:

• same idempotency key + same normalized intent -> replay the stored receipt
• same key + different intent -> conflict before a provider call
• changed item revision -> throw away the stale plan and prepare again

I wrote down the design and how it differs across GitHub, GitLab, Linear, Jira, and Azure DevOps:
https://work-sdk.vercel.app/guides/agent-safe-work-tracker-writes?src=github-discussion

The part I am still thinking about: when a provider has no native idempotency primitive, how much reconciliation should live in the adapter versus the caller's job system? I would be interested in concrete failure cases from real automations.