conversion-reconcilelisted

# conversion-reconcile ## Agent workflow This is an executable comparison. When two usable exports are available, run the bundled reconciler instead of debating attribution in the abstract. 1. Identify which file is the marketing/analytics platform export and which is the CRM or finance source of truth. Inspect filenames and headers first; ask only when the roles are genuinely ambiguous. 2. Confirm both files contain row-level conversions and a shared identifier. Aggregated daily totals cannot support the join this skill makes. Read [references/input-guide.md](references/input-guide.md) when the user needs an export recipe. 3. Resolve `scripts/reconcile.py` relative to this `SKILL.md` and run it with an available Python 3 interpreter. In Claude Code, `${CLAUDE_SKILL_DIR}` is this skill directory; in other hosts use the skill path supplied by the host. Quote every path. Let key auto-detection run first; pass `--key NAME` when the shared field has another name. 4. Interpret exit codes exactly: `0` means no material divergence shape was diagnosed, `1` means one or more evidence-backed diagnoses were produced, and `2` means the inputs could not support a trustworthy comparison. On `2`, relay the corrective message and stop. 5. Report the join quality and ratio before the ranked causes. Lead with the top cause's evidence and next check, but retain credible secondary causes. Never turn a heuristic confidence weight into a calibrated probability.