What problem does it solve? Agents that publish outward (posts, messages, API writes) face an unsolvable ambiguity: when a request times out or the worker crashes, nobody knows whether the bytes landed. Naive retry logic then posts the same thing twice, and a duplicate side effect cannot be undone. This Skill provides a complete design for an at-most-once delivery system where unknown outcomes are reconciled by looking at the platform exactly once, never by blind resending. ## Core Features & Use Cases - Six-state delivery state machine: pending, authorized (in-flight with lease), published, failed (the only retryable state), indeterminate (never retryable), and parked (terminal, human-owned), with strict transition rules. - Separate intent and attempt rows: one durable delivery intent keyed by a stable delivery key, with append-only immutable attempt rows created only by a single begin() path. - Seven safety mechanisms: before-byte durable intent writes, final pre-send authorization, stale-worker fencing with sequence-backed generations, read-back verification, duplicate-text exclusion windows, and append-only audit history. - Idempotency identity rules: four separately named identities (source version, claim, delivery key, attempt token) with precise rules on what goes in each hash, including excluding timestamps and normalizing content before hashing. - Use Case: You are building an agent that auto-posts to social platforms. Use this Skill to design the publishing layer so a crash mid-send results in a reconcilable indeterminate state instead of a duplicate post. ## Quick Start Ask the agent to design or review the publishing layer of your agent using the at-most-once state machine so a crashed or timed-out send never produces a duplicate post.
