Proposal (rationale)

Six extractions are six documents. The thing that made Phase 1 useful was Phase 2 — the pass that put four repos in one table and turned them into convergences, divergences, eight failure modes and a maturity ladder. Without the equivalent pass, this corpus is a reading list, and a reading list does not change what rungs ships.

This item exists to do one specific job that none of the six extractions may do: decide what the independent evidence does to the existing catalogue. Each extraction is forbidden from editing pattern-catalog.md precisely so that this decision is made once, with all six in view, rather than six times by whoever happened to be writing that day.

Three outcomes are expected, and all three are worth writing down:

• Confirmed. A pattern the four private repos converged on that independent teams also arrived at. It gets a much stronger claim, and the sources named.
• Demoted. A pattern that turns out to be one operator’s habit, or one that the frameworks solve a different and better way. Demotion is the finding here; a catalogue that only ever grows is not being checked.
• New. A practice six teams share that the four repos never invented. The most valuable category and the reason for the whole epic.

There is a fourth outcome the synthesis must be honest about: the two corpora may not be commensurable. The four repos are read for workflow and the six for architecture, and a pattern “confirmed” across that boundary may be two different things wearing one name. Where that is the case, saying so is the finding — silently merging them would put an architecture claim into a catalogue that module authoring reads as workflow evidence.

Decision

accepted — 2026-08-15. The user directed the remaining WI-009 elements to proceed sequentially; all six required extractions are now done, so the final child is unblocked.

Plan

Requirements

• docs/research/frameworks/synthesis.md, structured for comparison first: a six-column table on the template’s own sections, then convergences, then divergences, then what nobody solved.
• A reconciliation table with one row per affected pattern id: confirmed / demoted / new / not commensurable, the evidence, and the resulting change.
• The corresponding edits to pattern-catalog.md — sources added, rungs changed, new patterns admitted — each traceable to a row in that table.
• Where a finding contradicts something in synthesis.md or harness-landscape.md, amend in place with a dated note, the way harness-landscape.md amends synthesis §3.1. Do not leave two documents disagreeing.
• Every claim traced to one of the six documents and through it to a pinned SHA; opinion in the first person and marked.
• Any module change the evidence warrants is opened as a new item, not made here.

Impacts

• New frameworks/synthesis.md; edits to pattern-catalog.md; possible dated amendments to synthesis.md and harness-landscape.md; frameworks/README.md and research/README.md reading order updated.
• An ADR is likely. Changing a pattern’s rung or its target module is a design decision that binds module authoring — ADR criteria 1 and 2. Whether one or several is not decidable before the evidence exists; the requirement is that the choice is stated.
• Site: several routes, and a link surface large enough that the checker is the real test.

Approach

Comparison table first, prose second. The table is what forces the six documents into the same shape and exposes where an extraction quietly skipped a section; writing the prose first would let that pass.

Reconcile by id, not by theme. Every row names a pattern id from the existing catalogue, or proposes a new one. A synthesis that discusses themes without touching ids leaves the catalogue — the actual input to module authoring — unchanged, which is the failure mode this item exists to avoid.

Demotions get the same evidence standard as confirmations. It is easier to add a source than to remove a pattern, and an unchecked catalogue drifts toward everything-is-a-pattern.

No ADR. The admission rule is not met: this work item and the canonical catalogue already own the reconciliation, and reversing a documentation-only classification later is not materially more expensive than making it now. The rejected alternative—an ADR that duplicates the reconciliation table—would create a second authority for the same decision.

Acceptance criteria / tests

1. frameworks/synthesis.md exists with the six-column table, every cell filled or explicitly marked not-applicable with a reason.
2. The reconciliation table covers every candidate pattern named in the six extractions; no candidate is left unadjudicated.
3. Every catalogue edit traces to a reconciliation row; every reconciliation row traces to a document and a SHA.
4. At least one demotion or non-commensurability row, or an explicit evidenced statement that the independent corpus contradicted nothing — stated as a finding either way, because “everything was confirmed” is a result that should be argued rather than assumed.
5. No document in docs/research/ contradicts another; contradictions are amended in place and dated.
6. Module changes appear as new work items, not as edits under this item.
7. rungs check passes; the site builds with 0 broken links.

Out of scope

• Any change to modules/ or to the CLI. This item ends at the catalogue; what the catalogue implies for a module is the next item, and conflating them is exactly the scope creep docs/backlog/README.md §6 forbids.
• Re-reading any of the six repos. If a document is inadequate, that is a defect in that extraction’s item, or a new item — not a re-run folded in here.
• Re-surveying the four original repos — WI-009’s boundary.
• A merged, single catalogue spanning both corpora. The reconciliation may conclude that one is needed; producing it would be a separate item with its own ADR.

Execution

Branch feature/WI-017-framework-synthesis, cut from main 2026-08-15. Compared the six completed extractions without re-reading their source repositories, reconciled all candidates once, amended the workflow synthesis where the second corpus narrowed an older statement, and opened WI-029 for the separately scoped module work.

Review

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

1. Pass. frameworks/synthesis.md opens with a six-framework matrix over template sections 2–7. Every cell is filled; the two absent composition primitives are explicitly not-applicable with bounded reasons, while SWE-agent’s human-in-loop cell distinguishes terminal takeover from a pre-effect approval protocol.
2. Pass. A table-row comparison found 17 unique backticked candidate: ids across the six extraction documents and the same 17 in the reconciliation table, with no missing or extra id.
3. Pass. Every catalogue edit has a matching reconciliation row and source link. The synthesis resolves those links through each extraction’s pinned Snapshot; the catalogue defines the six source abbreviations through the synthesis rather than duplicating pins.
4. Pass. Three candidates and two existing-pattern analogies are not commensurable with the workflow catalogue, and shared-workspace-subagents is demoted/rejected. The synthesis preserves the product findings while refusing to represent UI/live-tail/container implementation as installed rungs features.
5. Pass. The original workflow synthesis now carries dated amendments: OpenHands independently confirms worktrees for Git coordination but not sandboxing or the land protocol, and framework trajectories/checkpoints/events narrow the old “nothing recorded across runs” statement without solving repository-workflow outcome history. No harness-landscape claim was contradicted.
6. Pass. No file under modules/ changed. Proposed WI-029 owns disposition, versioning, fixtures, and upgrade behavior for the affected shipped modules.
7. Pass. 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.