Proposal (rationale)

SWE-agent is the smallest repo in the corpus and the one whose subject is closest to rungs’ own: an agent that is handed a GitHub issue in a real repository and has to change code until the issue is resolved. That is the loop every repo rungs scaffolds is trying to make survivable.

Its value here is subtractive. Every other repo in the corpus bundles orchestration, persistence, sandboxing and a product surface together, so it is hard to tell which parts are essential. Reading the minimal one first shows what the irreducible loop is — observe → reason → execute → observe — and turns every additional mechanism in the other five into a question with a cost attached: what failure was this added for?

The specific questions:

• What is one turn? Named file, named function. The four private repos have no answer to this at all — they instrument the human-facing workflow around an agent and treat the loop as opaque.
• What does the agent see? SWE-agent’s public positioning is that the interface the model is given — how a file is shown, how an edit is confirmed, how errors are returned — matters more than the model. If that holds up in the source, it is a direct argument about what a scaffolded repo’s documentation and gate output should look like, since those are also agent-facing interfaces.
• What is the stop condition, and what happens when it never fires? Every gate rungs ships has the same shape.
• Where does it record what it did? The single most-converged pattern in the existing corpus is that work needs a durable place to land.

Everything above is an expectation drawn from the project’s own positioning, not evidence. princeton-nlp/SWE-agent is understood to have moved to a SWE-agent organization; confirm the canonical URL, the license, and the current state before measuring anything.

Decision

accepted — 2026-08-15. The user directed the remaining WI-009 children to proceed sequentially; WI-010 is complete and this is the next planned child.

Plan

Requirements

• docs/research/frameworks/swe-agent.md, on the template from WI-010, all eight sections answered.
• Snapshot pins a commit SHA, records the license, and names the command behind every count.
• Sections 2 and 4 (core loop, tools) trace to specific files and functions, not to the README.
• Section 8 cites pattern ids from pattern-catalog.md; a practice with no matching id is named as a candidate pattern for WI-017 to adjudicate, not added to the catalogue here.
• Template corrections found while writing are made in this item, in WI-010’s files, and listed in ## Execution.

Impacts

• One new document; one row in the frameworks index.
• Possibly edits to frameworks/TEMPLATE.md — expected, and the reason this repo is scheduled first.
• Site route and link surface.

Approach

Read the source, not the paper or the README. The repo’s own account of its loop is a claim about the loop (CLAUDE.md — extraction discipline #1); where the docs and the code disagree, the code is what ran.

Start at the entry point and follow one turn end to end before writing anything, so section 2 is a trace rather than a summary.

Acceptance criteria / tests

1. The document exists with all eight sections answered — none blank, none deferred to another repo.
2. Snapshot carries a commit SHA, a license, a read date, and per-count commands.
3. Section 2 names the file and function where a turn is executed, and a reader at that SHA finds it.
4. Every opinion is in the first person and marked; every other claim has a path or a quote.
5. Section 8 gives a take / take-as-warning / leave verdict with a reason.
6. rungs check passes; the site builds with links resolving.

Out of scope

• Running SWE-agent, or reproducing any benchmark number. The corpus reads architecture; a benchmark score is a claim about model performance and is not something this repo can verify or needs. Any number quoted from the repo is attributed as their measurement, not ours.
• Comparison with the other five. Cross-repo claims belong to WI-017; an extraction that compares is an essay.
• Editing the pattern catalogue. Candidates are named, not added — WI-017.
• Any change to modules/ — WI-009’s boundary.

Execution

Branch feature/WI-011-extract-swe-agent, cut from main 2026-08-15.

• Confirmed the canonical repository as SWE-agent/SWE-agent and pinned 3ea751c087f32b16e039a2233dd6eefecef325d5 before inspecting or measuring it.
• Traced the default turn through DefaultAgent.run → step → forward_with_handling → forward → ToolHandler.parse_actions → handle_action, then followed history and trajectory writes back out of the turn.
• Wrote docs/research/frameworks/swe-agent.md with all eight sections, pinned permalinks, dated PowerShell measurements, two bounded absence checks, and five catalogue verdicts/candidates.
• Updated the frameworks index to point at the pinned extraction.
• No template corrections were required. The distinction between replay and resume fit section 3’s existing prompt, and the ShellAgent takeover path fit section 6 without changing its scope.

Review

Checked 2026-08-15.

1. Pass. docs/research/frameworks/swe-agent.md answers all eight template sections; none is blank or deferred.
2. Pass. Snapshot records the full SHA, MIT licence, read date, and the exact PowerShell command and scope behind each of the three counts.
3. Pass. Section 2 names and links DefaultAgent.run, step, forward_with_handling, forward, and handle_action at the pinned SHA.
4. Pass. Judgement is written in the first person and prefixed Opinion.; implementation claims use pinned permalinks and absence claims give the dated rg command and directory scope.
5. Pass. Section 8 contains five take / take-as-warning verdicts with reasons, including two candidate patterns reserved for WI-017.
6. Pass. node src/cli.ts check → 20 pass, 0 fail, 101 links examined. npm run build && npm run check under site/ → 56 routes, 501 internal links, 0 broken, and 0 Astro diagnostics.