fix(cockpit-langgraph): make the subgraphs example demonstrate real nesting - #838
Merged
Merged
Conversation
…esting The `cockpit/langgraph/subgraphs` example advertised subagent delegation it could never perform. `app.config.ts` set no `subagentToolNames`, and the graph added a compiled subgraph as a plain node — which emits a `research:<uuid>` namespace, not `tools:<id>`. The SubagentTracker only registers delegation tool calls, so `agent.subagents()` was permanently empty and the sidebar always rendered "No active subagents". The manual e2e asserted exactly that empty state and nothing else. The parent's edges were unconditional, so nothing decided anything, and both graphs were bare `MessagesState` with no boundary. Kept the plain subgraph-as-node pattern rather than converting to the tool-call path: `cockpit/chat/subagents` already demonstrates tool-call delegation end to end, and this is the repo's only example of LangGraph's actual composition primitive. The website docs already state the rule (docs/langgraph/guides/subgraphs.mdx:114 — "Plain subgraph nodes do not appear in this map"); the example now agrees with them instead of contradicting them. Backend: - `orchestrate` makes a structured routing decision; a conditional edge enters the child graph or skips straight to `answer`. - The child's state (`ResearchState`) has no `messages` key, so it exchanges only `research_topic` / `research_brief` with the parent and can neither read nor append to the transcript. - `answer` is the sole transcript-writing node. Frontend: - Sidebar reads the parent's own state via a typed `agent.value()` instead of the map that can never fill, and says plainly why. - `transcriptNodeNames: ['answer']`. Verified live: without it the message list transiently grows to 3 as the child's tokens merge in and its internal brief renders as a chat bubble, before the parent's `values` event collapses it back to 2. Self-correcting end state, so no final-state assertion catches it. Tests: three e2e specs replacing the smoke test — the research branch populates the sidebar, the child's brief never reaches the transcript (paired with a positive sidebar assertion so it can't pass vacuously), and the direct branch skips the child and still answers. The manual live-LLM spec asserts the branch that ran rather than an empty state. Verified: e2e 3/3 green; production build green; both branches driven against a real model in Chrome, with the transcript boundary checked from the live DOM. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Contributor
blove
added a commit
that referenced
this pull request
Aug 27, 2026
…en Not To' blog post (#839) * docs: design spec for LangGraph subgraphs blog post Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: implementation plan for LangGraph subgraphs post Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(website): add 'LangGraph Subgraphs' blog post * docs: handoff for the GSC-driven blog sequence Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(website): correct attribution + streamSubgraphs claims in subgraphs post Applies blocking review findings: - Attribution is structural (namespace segment IS the tool call id), not a heuristic; the description ladder is unreachable in both shipped graphs. - streamSubgraphs is the LangGraph JS SDK's own option name, not a naming trap; re-aimed at the Python subgraphs=True kwarg. - Test asserts outcome 'interrupted', not phase. - Concede non-observability split reasons, then reframe. - Quote evidence in fences; add controlled-experiment framing for AG-UI. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(website): editorial pass on subgraphs post - Drop the false 'no independent checkpointing' cost: both subgraph children compile bare, while the flat AG-UI graph is the one with MemorySaver. - Handle the checkpoint objection honestly — child checkpointing rides the namespace under a parent checkpointer, which is the thesis again. - State that the AG-UI control group is itself a LangGraph StateGraph, so only one variable differs. - Answer the 'just emit custom events' objection. - Quote the rejected design-doc alternative in a fence. - Trim demonstrative tic (That's 13->7, Here's 7->3), restore Let's transitions, cut duplicated state-isolation and heuristic statements. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(website): account for the examples/ag-ui counterexample Spec review found a fifth graph the post's thesis didn't survive: examples/ag-ui compiles a child StateGraph on a transport that ALREADY emits subagent_activity events, so the split isn't buying observability. It splits because the child has its own ResearchState (topic + iterations) and its own agent->tools loop with an iteration cap. - Add 'So when does a split earn itself?' — two demos on one transport, one flat and one split, with child shape as the variable. - Stop claiming no such case exists in the closing. - Name cockpit/ag-ui/subagents so it isn't read as examples/ag-ui. - Last-resort rung accepts pending OR running (tracker:165-169). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(website): sharpen the split discriminator to control flow Re-review findings: - A custom child state schema doesn't discriminate: examples/chat and cockpit/chat/subagents both define child TypedDicts yet are single-node straight lines. Only control flow separates the earned split. - Drop hand-rolled streaming as a cost of flatness — examples/ag-ui pays it too, so it's a cost of the AG-UI transport, not of staying flat. - Reconcile the two conflicting 'one variable' claims. - 'neither graph we ship' was stale now that three tool-dispatch a child. - Closing no longer claims only two justifications exist. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(website): re-check subgraphs post against the landed cockpit example cockpit/langgraph/subgraphs (#838) now ships a conditional plain-node subgraph whose child state deliberately omits `messages`. - Fix the ladder-unreachability reason: it no longer holds for every graph by way of an empty message list — the plain-node child has no messages key at all. - Cite that example as first-hand evidence for 'isolation is something you design', which until now the post asserted without a concrete instance. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs: refresh blog-sequence handoff for shipped post #1 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This was referenced Aug 27, 2026
blove
added a commit
that referenced
this pull request
Aug 28, 2026
…odes (#842) The subgraphs post tells readers that child text lands in the parent transcript unless `filterSubagentMessages` opts out. That holds for a child dispatched from a `@tool` body, but not for the graph shape the post spends its best section on. `filterSubagentMessages` is checked inside a branch guarded by `isSubagentNamespace()`, which matches only `tools:`. A plain subgraph node emits `research:<uuid>`, never enters that branch, and so ignores the option entirely — its tokens merge into the transcript no matter how it's set. The lever for that shape is `transcriptNodeNames`. This matters because the post showcases `cockpit/langgraph/subgraphs`, which is exactly the plain-node shape. A reader who follows it, sees the child's brief appear as its own bubble, and reaches for `filterSubagentMessages` will find it does nothing. Also records the failure mode, which is the expensive part: the leak is mid-stream with a clean end state. The parent's final `values` event rewrites the message list from authoritative graph state, so the stray bubble disappears once the run settles. A final-state assertion cannot catch it — measured against a live model in #838, where the message count transiently hit 3 before collapsing to 2. Verified: website builds; blog + sitemap specs 26/26. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The problem
cockpit/langgraph/subgraphsadvertised subagent delegation it could never perform.The sidebar could never populate.
app.config.tsset nosubagentToolNames, and the graph added a compiled subgraph as a plain node — which emits aresearch:<uuid>namespace, nottools:<id>.SubagentTrackeronly registers delegation tool calls (libs/langgraph/src/lib/internals/subagent-tracker.ts:78-112,265-269), soagent.subagents()was permanently empty and the sidebar always rendered "No active subagents". The manual e2e asserted exactly that empty state and never asserted it populated.No delegation happened. The docstring claimed "the parent orchestrator decides when to delegate," but the edges were unconditional (
orchestrate → research → END). Both graphs were bareMessagesState, so the child inherited and appended to the parent's messages.The decision: honest nesting, not tool-call delegation
cockpit/chat/subagentsalready demonstrates the tool-call path end to end (tasktool +subagent_type+ToolNodeloop + sidebar). Converting this example would have made it a near-duplicate and deleted the repo's only demo of LangGraph's actual composition primitive.docs/langgraph/guides/subgraphs.mdx:114: "Plain subgraph nodes do not appear in this map." The example was contradicting published docs; it now agrees with them.subgraphs, notsubagents.So: keep the plain subgraph-as-node pattern, but make it teach something real.
Changes
Backend (
python/src/graph.py)orchestratemakes a structured routing decision; a conditional edge enters the child graph or skips toanswer. The parent actually decides.ResearchStatehas nomessageskey. LangGraph wires the subgraph node through the keys the two schemas share (research_topic,research_brief), so the child gets a topic, returns a brief, and can neither read nor append to the transcript.answeris the sole transcript-writing node.Frontend
agent.value()instead of the map that can never fill — and the component doc says plainly whysubagents()is the wrong signal here, pointing at Chat Subagents for the tool-call path.transcriptNodeNames: ['answer'].Docs —
python/docs/guide.mdrewritten (the falsestream.subagents()claim at :112 is gone; a<Warning>now states the subgraph-vs-subagent distinction up front). Bothprompts/subgraphs.mdcorrected; the Angular one no longer claims a<chat-subagent-card>per subgraph invocation.Drive-by:
angular/src/index.tscodeAssetPathspointed at a nonexistentsrc/app.component.ts.Tests
Three e2e specs replace the single smoke test. No test asserts an empty state as its only assertion:
The manual live-LLM spec now asserts the branch that ran, not an empty sidebar.
Verification
nx e2e cockpit-langgraph-subgraphs-angular— 3/3 green.nx build --configuration=production— green.[user, assistant]with the brief absent from the live DOM. Direct path: sidebar stays "Direct", answer arrives.messagesevents fromorchestrate, 767 undermessages|research:<uuid>from the child, 633 fromanswer.transcriptNodeNames. Removing it does not fail the aimock e2e — fixture replay is atomic, so the leak is invisible there. Against a live model it is clearly load-bearing: message count transiently reaches 3 (the child's internal brief rendering as its own chat bubble) before the parent'svaluesevent collapses it back to 2. With the option, the count never exceeds 2 at any sampled point. The end state self-corrects, which is why no final-state assertion can catch it — recorded in a comment at the call site rather than claimed as test coverage.Note on guide.md:112
Checked for the separate in-flight correction mentioned in the task: no open PRs, and the only related branch (
blove/langgraph-subgraphs-post) touches website content only. Nothing had landed, so this PR corrects the claim as part of the guide rewrite.🤖 Generated with Claude Code