Proposal (rationale)

WI-009 deliberately fixes six subjects around runtime architecture: the core loop, checkpointed graphs, small primitives, typed/testable agents, enterprise workflows, and a sandboxed product. Its synthesis can therefore answer the question it was designed to answer, but it should not silently be treated as evidence about four materially different boundaries that its subjects do not center:

• long-lived, agent-managed memory and identity;
• repeatable evaluation and metric-driven improvement;
• local, git-native product operation and protocol-extensible harnesses; and
• interoperability between an agent and tools or between independently operated agents.

This epic proposes a follow-on corpus that tests those boundaries without widening WI-009 during execution. It contains four runtime/product subjects, two evaluation/optimization subjects, and two protocol subjects. The split is intentional: forcing all eight through the framework template would make unlike evidence look comparable merely because the headings matched.

The repository names and questions are selection hypotheses, not research findings. Each child must pin source, establish licence, inspect implementation or normative artifacts, and record when the proposed question does not match the source.

Decision

accepted — 2026-08-15. The user directed the follow-on research to start after the recommendations were captured. WI-017 and its parent corpus are done, so the method prerequisite is now unblocked; the epic remains sequenced method → extractions → synthesis.

Plan

Requirements

• One method item, eight subject extractions, and one synthesis; membership is fixed by children.
• Every public source is recorded at a full commit SHA with its licence and read boundary.
• The method keeps the three tracks distinguishable while retaining a small shared comparison spine: source boundary, state, external interaction, human authority, evidence artifacts, cost, and catalogue consequence.
• Per-subject children propose catalogue candidates but do not edit the catalogue; WI-028 reconciles once across all eight.
• Execution begins only after WI-017 records what the original framework corpus already answered, so this epic addresses measured gaps rather than the expectations in this proposal.

Impacts

• New research method and extraction documents under a location selected by WI-019.
• A follow-on synthesis and possible evidence-backed edits to pattern-catalog.md.
• Possible dated amendments to existing synthesis documents when independent evidence contradicts them.
• No module or CLI changes under this epic; any such implication becomes a separate work item.

Approach

Close, then extend. Complete WI-017 first. WI-019 then converts the gaps found there into track templates and read boundaries. The eight extractions may proceed independently after the method is accepted. WI-028 compares within each track before making any cross-track claim.

Prefer a bounded counterexample to another feature inventory. Each child names one mechanism to trace end to end and one strongest counter-evidence question. A subject that duplicates an existing answer is allowed to conclude that it adds no new pattern.

Acceptance criteria / tests

1. WI-019 defines the shared spine and the three track-specific methods before an extraction starts.
2. WI-020 through WI-027 each produce a pinned, licensed, evidence-labelled extraction.
3. WI-028 adjudicates every candidate finding and distinguishes confirmation, contradiction, new pattern, and non-commensurability.
4. Any catalogue edit traces through WI-028 to a pinned source; module changes are separate items.
5. All children reach done; rungs check and the site link check pass at the epic boundary.

Out of scope

• Expanding or reopening WI-009. It remains the fixed six-framework corpus.
• Product rankings, benchmark leaderboards, adoption counts, or model-quality comparisons.
• Hosted behaviour whose implementation is unavailable in the pinned public source.
• Copying source code or licensed prose into modules.
• Changes to modules/, the CLI, or published profiles; those require later work items.

Execution

Not started.

Review

Not started.