Proposal (rationale)

OpenHands is the largest repo in the corpus and is scheduled last for that reason: it is where the template meets the hardest case, after five cheaper repos have hardened it.

It is in the corpus because it is a product, not a library. The other five hand you primitives and leave the hard parts — where does the code actually run, what happens when the agent breaks something, what does a user see mid-run, how do many runs coexist — to the person integrating them. OpenHands has to answer all of them, in public, for real users. Those answers are the parts a library’s design deliberately omits, and they are precisely the parts rungs’ scaffolded repos hit.

The corpus’s closest existing analogue is rift-forge’s 401 branches and 51 live worktrees — one operator’s hand-rolled answer to running many agent sessions without them colliding (synthesis.md). A product that solves the same problem with containers rather than worktrees is the natural check on whether the worktree pattern rungs ships is a genuine solution or an artefact of not having infrastructure.

The specific questions:

• What is the isolation boundary, and what crosses it? A sandbox is a statement about what the agent is not trusted with; the exception list is the interesting part.
• What happens when the agent breaks the environment, and how is that recovered?
• How does a repository get in and changes get out — clone, mount, branch, patch, PR? This is the same surface rungs init writes into, from the other side.
• What does the user see and control while a run is in progress? The private corpus has no answer beyond “read the terminal”, and long-running work is where that fails.
• How do concurrent runs stay separate, and what does that cost per run? The direct comparison with worktrees.
• What is in the product that is not in any of the five libraries? The residue is the answer to “what does a library leave you to build”, and it is the single most useful thing this extraction can produce.

Expectations from the project’s positioning, not evidence. Confirm the URL, the license and the current architecture at a pinned commit; this project has renamed and restructured before, so old write-ups about it are not a source.

Decision

accepted — 2026-08-15. The user directed the remaining WI-009 children to proceed sequentially; WI-015 is complete and this is the next planned child.

Plan

Requirements

• docs/research/frameworks/openhands.md on the WI-010 template, eight sections answered.
• Snapshot pins a commit SHA, license, read date, per-count commands, and an explicit read boundary — what was read and what was not. A partial read of a large repo that says so is evidence; one that does not is a claim about the whole repo.
• Sections 4 and 5 carry the sandboxing and concurrency answers with file-level evidence.
• A short “what the product has that the libraries do not” list, derived from the other five extractions being complete — this is why the item is last.
• Section 8 cites pattern ids; candidates go to WI-017.

Impacts

• One new document; one row in the frameworks index; site route and links.
• Bears directly on the concurrency and worktree patterns in pattern-catalog.md — read, not edited; WI-017 argues any rung change.

Approach

Bound hard, and say where the boundary is. Trace one run from request to committed change, and read the isolation and concurrency machinery around it. Everything else is out of the read.

Do not run it. The extraction is a source read. Standing up the product would answer different questions at a much higher cost, and a screenshot is not evidence about architecture.

Write the residue list last, after the other five documents exist, and cite them by id.

Acceptance criteria / tests

1. All eight sections answered; Snapshot carries SHA, license, date, commands and the read boundary.
2. The isolation boundary is described with what crosses it, traced to named files.
3. The concurrency answer is stated in terms comparable to the worktree pattern, without asserting which is better — that verdict is section 8’s, with its reasoning.
4. The residue list cites the five sibling documents by path.
5. Opinion marked as opinion; every other claim carries a path or a quote.
6. rungs check passes; the site builds with links resolving.

Out of scope

• Running or deploying OpenHands, and anything that would require it — performance, resource cost, model behaviour.
• The hosted product and anything not in the repository. Recorded as not inspectable.
• The frontend beyond the run-control surface, which is the only part the corpus question needs.
• Changing the worktree pattern’s rung or contents — WI-017, and a module change is its own item beyond that.

Execution

Branch feature/WI-016-extract-openhands, cut from main 2026-08-15. The source read pinned Agent Canvas at dc99e98615de4ace821692773b00a7f50d476e50 and software-agent-sdk at 46ad3d43dc385b2e7975c0935f157153930ebb16.

Review

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

1. Pass. openhands.md answers all eight template sections. Its two-repository Snapshot records both pinned SHAs, MIT licences, read date, seven measured path counts with commands, and an explicit boundary that excludes hosted Cloud, automation internals, provider transports, most frontend surfaces, and running the product.
2. Pass. Section 4 distinguishes the default direct-host LocalWorkspace, the all-in-one Canvas Docker deployment, and the optional per-workspace SDK container. It traces filesystem access, absolute editor paths, mounted host projects, selected environment/network/port crossings, container lifecycle, and a bounded absence check for resource/read-only flags.
3. Pass. Section 5 compares Canvas’ default per-conversation Git worktree with the shared host and optional container boundaries without ranking them. The 16-conversation stress test is bounded to the scheduling, persistence, leakage, wall-time, and RSS properties its fixture can establish; real tool/model capacity and cost remain unclaimed.
4. Pass. Section 7’s residue list links all five sibling extractions and identifies the composed backend control, persistence/streaming, repository/worktree, live run-control, and packaging surfaces found in the product layer.
5. Pass. Judgement is labelled Opinion. Measurements are dated, absence claims name their search boundary, implementation claims use pinned file permalinks, and current counter-evidence replaces the item’s initial container-versus-worktree expectation.
6. Pass. node src/cli.ts check passed 20/20 gates. In site, npm run build generated 61 pages and npm run check reported 0 diagnostics and 530 internal links with 0 broken.