state-driven-pipeline-recoverylisted

# state-driven-pipeline-recovery When the pipeline is spinning, the failure is almost never "the agent needs to try harder." It is that **the state name is lying about reality**. ## Step 1 — Smell test for false progress Any of these means stop and investigate rather than retry: | Smell | What it actually means | |---|---| | Worker reports success, but no commit, no state change, no evidence comment | Nothing happened. The report is narrative. | | Ticket is in Coding, but the production entry point bypasses the tested seam | The test proves nothing about the shipped path | | The locked Verification-command references a harness that doesn't exist | The gate was never runnable; nobody ran it | | A foundational protocol/prerequisite is missing, but the ticket pretends the feature lands on air | Structural blocker disguised as an implementation task | | Repeated retries produce **identical** failure output | No new information is being generated — burn without learning | | State says Done, but the gate is red on a fresh checkout | The state was moved without proof | **Retrying an identical failure is not persistence. It is a loop with no exit condition.** ## Step 2 — Re-read reality, in this order Do not touch code yet. 1. The ticket: current state, full body, **all comments** (bounces especially). 2. The plan doc it points at. 3. The **production entry point** — does real traffic actually reach the code under test? 4. The gate: run it. Right now. Record the real output.