§1 — What worked

Codex discovered the Reckon skills and could read and write plan and index state through the three-tool MCP surface. Optimistic versions, the special index slug, sprint append/set/move operations and checkout-path routing make sprint management usable from Codex CLI.

§2 — Errors and friction observed

1. Large project discovery payloads include full followup and sprint material and can be truncated before the requested plan inventory is usable.
2. Archive discovery and audit disagree with the documented exclusion contract; archived Nova pages are audited as live and fail the live type enum.
3. Filtered searches may return full sprint payloads even when no plans match.
4. Research/doc prose is absent from parsed reads, forcing a second filesystem or HTTP read.
5. Mounted-checkout and branch identity is easy to miss in multi-worktree projects.
6. The audit response says reindexed: true even though the tool promises no mutation.
7. Cross-project dependencies and migration portfolios require manual synthesis.
8. Plan-level dependencies are too coarse for stage-by-stage launch ordering.
9. Research is normalized into one generic type; evidence has no type or evidence_for relation.
10. reckon-ship assumes a mandatory parallel-agent surface and does not define a single-agent Codex fallback.

§3 — Product implications

Default discovery should be bounded and facet-aware; archives should be explicit; responses should expose checkout and branch identity; portfolio relationships need qualified project/slug keys; and the document model needs separate plan, research and evidence types. Only plans should expose implementation fractions.