Proposal (rationale)

Phase 1 extracted four repositories, all built by one operator. That is the corpus’s strength and its ceiling: synthesis.md §1 argues the four are one learning curve, which makes the arc legible but means every convergence in the catalogue is a convergence of one person with themselves. “Four repos converge on X” is the strongest claim the repo makes (CLAUDE.md — extraction discipline #4), and it is weaker than it reads when the four share an author.

Six public repositories, built by six independent teams under real adoption pressure, are the cheapest available test of which extracted patterns are portable and which are one operator’s habits.

This is a second axis, not more of the first

The existing corpus is read for workflow — how work was proposed, decided, executed, verified and remembered. research/README.md closes by stating the non-goal explicitly: it does not judge the source repos’ products, architectures, or code quality.

The six repos above are being read for exactly that — architecture. Their contributor workflow is incidental; their agent design is the subject. Reading them into the same directory under the same template would quietly redefine what docs/research/ means, and the next reader would find architecture claims filed as workflow evidence. So the corpus gets its own home and its own template, and the research index gains one line saying there are now two axes. That is WI-010, and it is a prerequisite, not overhead — Phase 1’s own lesson was that four repos on one template could be compared column by column “rather than read as four essays” (roadmap).

Two constraints the first corpus never had

1. The source moves. The four private working trees were surveyed once, on 2026-08-14, and nobody else was pushing to them. These six are under active development, so a date is not enough to make a count reproducible. Every extraction pins a commit SHA, and the SHA is what a re-reader checks out. This is the evidence rule’s “counts go stale” corollary with the staleness made addressable rather than merely visible.
2. The source is licensed. Quoting a private repo raises no question; quoting six public ones does, and rungs ships what it extracts into every scaffolded repo. So each extraction records the license, and the synthesis states what may be restated as a pattern versus what may only be cited. Nothing is copied into modules/ under this epic — see Out of scope.

Value, and the cost of not doing it

The pattern catalogue is the input to module authoring. Right now a pattern’s rung and its “target module” rest on a single-author sample. The likely outcomes are (a) patterns confirmed by independent teams, which get to state a much stronger claim; (b) patterns that turn out to be habits, which get demoted; and (c) practices six teams use that the four repos never invented — the most valuable category, and the one no amount of re-reading the existing corpus can produce.

Nothing in this epic is evidence yet. The “question it is being read for” column above is an expectation drawn from each project’s own positioning, and expectations are exactly what an extraction is supposed to contradict. Where a repo turns out not to work the way this proposal assumes, the divergence is the finding.

Decision

accepted — 2026-08-15. The user directed work on the research epic to start. The epic is planned through its eight children; each child remains individually decidable and executes on its own branch.

Plan

Filled once accepted. The per-child plans are written in the children.

Requirements

• One template item, six extractions, one synthesis — see children.
• Every child produces a document under docs/research/frameworks/, pinned to a commit SHA.
• The synthesis reconciles findings against the existing pattern-catalog.md rather than starting a second catalogue.

Impacts

• New directory docs/research/frameworks/ (created by WI-010).
• research/README.md gains the two-axis distinction.
• pattern-catalog.md gains sources and possibly rung changes — written by WI-017, not by the per-repo children.
• The site publishes docs/**, so every new file is a wiki route and a link-check surface.
• Possibly an ADR — if the synthesis changes a module’s contents or a pattern’s rung, that is a design decision and gets one. Not decidable before the evidence exists.

Approach

Template first, then one extraction, then fan out. WI-011 (SWE-agent, the smallest) is the template’s first real test and is expected to correct it; Phase 4 produced sixteen corrections to the module format, every one found by writing a module rather than reading the format. The remaining five are independent of each other and may run in any order or in parallel.

Ordered smallest to largest so the template hardens on cheap repos before meeting OpenHands.

Acceptance criteria / tests

1. All eight children reach done.
2. docs/research/frameworks/ holds six extractions on one template, each with a commit SHA, plus an index.
3. Every pattern claim in the synthesis is either traced to a named file in a pinned commit or marked as opinion in the first person.
4. rungs check passes, and the site builds with links resolving.

Out of scope

• Any change to modules/. This epic produces research. Shipping a newly-evidenced pattern into a module is a separate item, opened by the synthesis if the evidence warrants it — the same boundary Phase 2 and Phase 4 kept.
• Vendoring, forking, or copying code from any of the six. Extraction reads; it does not import.
• Judging the six as products — which framework to use, benchmark scores, adoption counts. Neither the question nor something this repo can answer.
• Re-surveying the four original repos. They were measured 2026-08-14; if the new corpus makes a claim about them look wrong, that is a finding against the existing corpus, not a re-run.
• The module registry (Phase 7) — unrelated, despite both touching “third-party”.

Execution

Executed sequentially 2026-08-15 through children WI-010–WI-017. Each non-trivial child planned on main, executed and reviewed on its own feature branch, passed its gates/site checks, and was merged before the next began. The synthesis opened WI-029 rather than changing shipped modules.

Review

Epic review completed 2026-08-15 against every acceptance criterion:

1. Pass. A frontmatter scan reports status: done for all eight children, WI-010 through WI-017.
2. Pass. docs/research/frameworks/ contains one applied template, six framework extractions, their index, and the cross-framework synthesis. Every extraction Snapshot records at least one full commit SHA; the two-repository subjects record both pins.
3. Pass. The synthesis matrix and reconciliation link every extracted mechanism to one of the six documents, which resolves onward to pinned source. Synthesis judgement is marked Opinion in the first person. All 17 candidate ids are adjudicated exactly once.
4. Pass. On the merged main tree, node src/cli.ts check passed 20/20 gates. In site, npm run build generated 63 pages and npm run check reported 0 diagnostics and 662 internal links with 0 broken.

The proposal’s optional ADR was not admitted: WI-017 and the canonical catalogue already own the documentation classification, and no pattern rung changed. Module implementation remains the separate proposed WI-029.