Proposal (rationale)

LangGraph is in the corpus for one question the four private repos never answered: what does an agentic workflow do when it is interrupted?

The existing corpus’s answer to durability is documentary — a work item, a findings register, a handoff note, a session archive. Those exist because a session ends and the next one starts cold, and every one of the four repos independently reinvented some version of it (synthesis.md §2). LangGraph solves what looks like the same problem mechanically: a graph of steps, a checkpointed state object, resume from the last checkpoint, and an explicit interrupt for human input.

If those two are really the same problem, then the documentary patterns rungs ships are a hand-rolled checkpointer, and the framework’s vocabulary — what is state, what is a step boundary, what does resume replay — is a sharper way to describe them. If they are not the same problem, the difference is worth writing down precisely, because it explains why a durable-execution framework does not remove the need for a handoff document.

The specific questions:

• What exactly is checkpointed, and what is deliberately not? A state model is a claim about what matters.
• What is a step boundary, and who decides it — the author, or the runtime?
• What does resume replay — does a resumed run re-execute side effects, and how is that avoided? This is the failure mode a handoff note has too, and nobody in the private corpus named it.
• How is a human interrupt expressed, and what is the state of the world while it waits?
• What does the graph buy over a loop? The honest version of this question includes what it costs, which by §7 of the template must be answered.

Expectations, not evidence. Confirm at a pinned commit; the project is large and moves fast, and anything below about “checkpointers” or “interrupts” is a term to verify, not a finding.

Decision

accepted — 2026-08-15. The user directed the remaining WI-009 children to proceed sequentially; WI-011 is complete and this is the next planned child.

Plan

Requirements

• docs/research/frameworks/langgraph.md on the WI-010 template, eight sections answered.
• Snapshot pins a commit SHA, license, read date, and per-count commands.
• Section 3 (state and persistence) is the load-bearing one and answers all four questions above with file-level evidence.
• Section 8 cites pattern ids and gives a verdict; new practices are named as candidates for WI-017.
• The document states explicitly whether checkpointing and the corpus’s documentary handoff patterns are the same problem — with the reasoning, marked as opinion if it is one.

Impacts

• One new document; one row in the frameworks index; site route and links.
• Likely input to harness-landscape.md’s primitives discussion, but not edited here — cross-document reconciliation is WI-017.

Approach

Bound the reading before starting. This is the largest surface area in the corpus after OpenHands, and an unbounded read produces a summary of the documentation. Pick one non-trivial graph from the repo’s own examples, trace it end to end, and let the eight sections be answered from that trace plus targeted reads. State in the Snapshot what was read and what was not — a partial read that says so is evidence; one that does not is a claim about the whole repo.

Acceptance criteria / tests

1. All eight sections answered; Snapshot carries SHA, license, date, commands, and the read boundary.
2. The four state questions each have a file-level answer or an explicit “could not determine, here is where I looked”.
3. Section 7 names what the graph abstraction costs, not only what it buys.
4. The checkpoint-versus-handoff question is answered, with opinion marked as opinion.
5. Every count is reproducible at the pinned SHA by the stated command.
6. rungs check passes; the site builds with links resolving.

Out of scope

• LangChain, the broader ecosystem, and any hosted platform product. The corpus question is the orchestration model; a vendor’s managed service is not inspectable and is not it.
• Recommending or rejecting LangGraph as a dependency for rungs. rungs is a scaffolder and ADR-0002 already fixes its runtime footprint; if the extraction genuinely suggests otherwise that is an ADR, not a line in a research file.
• Cross-repo comparison — WI-017.
• Editing pattern-catalog.md or harness-landscape.md — WI-017.

Execution

Branch feature/WI-012-extract-langgraph, cut from main 2026-08-15.

• Confirmed the canonical source as langchain-ai/langgraph, cloned it outside this repository, and pinned the clean checkout at 644815f9e5bc52ad8f7a5227a456227e9c3e639b.
• Bounded the implementation read to the Python StateGraph compiler, Pregel runtime and loop, checkpoint base and in-memory saver, interrupt/command types, and nested-interrupt/callback tests.
• Added langgraph.md and promoted the frameworks index row from proposal to pinned extraction.
• Verified with node src/cli.ts check (20 pass), then npm run build and npm run check under site/ (57 routes, 505 internal links, 0 broken; 0 Astro diagnostics).

Review

Acceptance criteria reviewed 2026-08-15:

1. Pass — all eight fixed sections are present; the Snapshot records repository, full SHA, licence, date, four reproducible commands, and the bounded read surface.
2. Pass — section 3 answers checkpoint contents, runtime-owned superstep boundary, resume replay and side-effect exposure, and the durable interrupt state with pinned file/test evidence.
3. Pass — section 7 names state modelling, identity, saver operation, durability, superstep, and replay-safe-side-effect costs.
4. Pass — section 3 explicitly judges checkpointing and documentary handoff as overlapping but different continuity problems, marked Opinion.
5. Pass — every count is the result of a stated git ls-files PowerShell command at the pin.
6. Pass — repository gates and the complete site build/check suite pass as recorded above.