Skip to content

Engineering history and agent debriefs

Document type: Historical index, non-normative.

Archetype keeps selected engineering debriefs, readiness inventories, and architecture reports when the path to a capability is itself useful evidence. These records explain what was known, what was proven, and what remained open at a particular revision. They are neither release notes nor an alternative specification.

When a historical record conflicts with current behavior, use this authority order:

  1. the focused normative specification;
  2. executable contracts and evals;
  3. the umbrella Specification;
  4. current guides and examples;
  5. the dated historical record.

Historical records are intentionally not rewritten to make old decisions look inevitable. Corrections should be appended with a date and a link to the newer authority.

Agent debriefs

Date Record Implementation anchor Why it is retained
2026-07-18 From sandbox prototype to resumable coding-agent missions Draft PR #487, commit db9d52a7 Establishes the mission/sandbox/artifact boundaries, retained provider capability, durability evidence, benchmark result, and honest limitations after the application-family refactor.

Readiness inventories

Date Inventory Scope
2026-07-18 Coding-agent mission production-readiness inventory One deduplicated, priority-ordered list combining the original production checklist, architectural seams found during formalization, and observability/evaluation follow-ups.

Architecture and program reports

These older reports remain useful for understanding why later contracts were introduced. Their recommendations may have been superseded.

What belongs here

A record belongs in this section when it binds a consequential implementation or operational result to a date, revision, and evidence set. Routine PR summaries, speculative designs without an implementation anchor, and current normative contracts belong elsewhere.

Every new agent debrief should state:

  • its date, status, scope, revision, and related issues/PRs;
  • what actually ran and what was only proposed;
  • systems of record and failure semantics;
  • quantitative evidence with the workload and caveats;
  • known gaps and the durable issues that own them; and
  • the current specifications that supersede it when behavior evolves.