# 1. Diagnosis The request points to a repair multi-step form state bugs problem whose visible symptom spans more than one step, so the safest starting point is the owning contract rather than a broad workaround. A multi-step form must preserve one coherent user-intent state across navigation, validation, autosave, resume, and final submission, without resurrecting stale fields or duplicating side effects. The immediate goal is to determine where that invariant first stops being true and to keep downstream symptoms from being mistaken for separate root causes. Treat the incident description as the symptom, not as proof of the root cause. In repository/workspace mode, classify every relevant statement as VERIFIED, LIKELY, or UNKNOWN. VERIFIED means directly supported by supplied files, logs, fixtures, configuration, or code that was actually inspected. LIKELY is a hypothesis consistent with that evidence. UNKNOWN means the decision cannot be made without another file, owner rule, or test. Do not claim a production fact merely because it is common for this class of system. The first working hypothesis should remain narrow: inspect step ownership and state shape, then conditional field lifecycle, and then the transition into validation timing and error state. If the evidence shows the fault belongs elsewhere, move the diagnosis instead of forcing the original theory. # 2. Evidence Map Build an evidence map around the concrete scenario: A React enrollment wizard has six steps, conditional questions, autosave, and resume-from-link. Going backward can resurrect hidden fields, and double-clicking Finish occasionally submits twice. Repair the state contract. For each transition, record the input identity, relevant state before the transition, decision or transformation performed, state after it, and the evidence that proves those values. Include autosave and resume semantics, navigation and browser history, and final submit idempotency where they affect the outcome. Preserve timestamps, IDs, versions, or correlation keys when they are part of the supplied evidence, but do not invent missing values. A useful evidence table has columns for boundary, expected invariant, observed behavior, evidence source, confidence, and unresolved question. That structure prevents a later reviewer from confusing a plausible explanation with an observed fact. If two components both appear wrong, identify which one first violates the invariant and which one merely propagates stale or inconsistent state. Check for hidden coupling as well: retries that rerun already-completed work, caches that outlive their source data, validation that happens after side effects, cleanup that assumes success, or UI state that represents a batch as atomic when the backend is item-scoped. Those patterns are especially important because they can make a local fix appear successful while the underlying inconsistency remains. # 3. Contract / Boundary The canonical boundary for this task is: > A multi-step form must preserve one coherent user-intent state across navigation, validation, autosave, resume, and final submission, without resurrecting stale fields or duplicating side effects. Translate that statement into executable or reviewable rules. Define who creates the authoritative state, who is allowed to mutate it, what constitutes successful completion, what constitutes partial completion, and which failures are retryable. Define what downstream consumers may assume and what they must verify themselves. For step ownership and state shape, document the normal path and one boundary case. For conditional field lifecycle, state retry or replacement semantics explicitly. For validation timing and error state, identify the commit or publication point. For autosave and resume semantics and navigation and browser history, state how stale, duplicate, or ambiguous work is detected. For final submit idempotency, define the evidence required before the operation is considered closed. Trace the first component that owns the decision rather than compensating in a caller, UI, retry wrapper, or downstream cleanup path. If an upstream owner already provides a correct signal, preserve it rather than recreating a second source of truth. If no owner exists, add the smallest explicit state or contract necessary to make the transition observable and testable. # 4. Bounded Repair / Review Plan # 5. Deterministic Validation # 6. Unknowns / Handoff No file edit, command execution, production inspection, external-service verification, or passing test is claimed unless it actually occurred in this run. The response is complete when it makes the failure boundary, evidence quality, repair scope, and verification path explicit.