fix(langgraph): replace partial accumulator when final canonical content arrives

[blove]

Summary

Targets Finding C from the live smoke pass: visible answer in <chat-streaming-md> rendered ~1.83× expected length when the AI message content array contained both a reasoning block and a text block.

Root cause

accumulateContent in libs/langgraph/src/lib/internals/stream-manager.bridge.ts falls back to existing + incoming text concatenation when neither side is a strict prefix. Streaming chunks accumulate the visible answer; the final canonical message arrives with full text but with small formatting differences (whitespace / Unicode normalization), the prefix check fails, and the fallback produces visible duplication.

Fix

Narrow heuristic that only triggers on a specific final-canonical shape:

```ts
function isFinalCanonicalReasoningContent(content: unknown): boolean {
// true when content is an array containing BOTH a reasoning/thinking block
// AND a text/output_text block — the shape used for final canonical messages.
}

function accumulateContent(existing, incoming) {
// ...existing prefix-detection logic kept...
if (isFinalCanonicalReasoningContent(incoming)) return incomingText; // NEW
return existingText + incomingText; // legitimate delta-append path preserved
}
```

The legitimate sequential-string-delta path (existing='hello', incoming='world', neither a prefix) is preserved — only the final-canonical-array shape is intercepted.

Verification

• Captured a real chunk sequence from the running examples/chat backend (gpt-5 + reasoning.effort=high, the puzzle prompt) into a JSON fixture (1213 events, 841 canonical text chars). Replay through mergeMessages produces visible text length matching server canonical (within 20 chars).
• Reference research (CopilotKit + Hashbrown source code) showed both alternative consumer libraries use delta-only wire protocols and never need this dedupe; the @Ngaf adapter is the only one trying to recover delta-like behavior from a hybrid snapshot+chunk source. Stays with the current model and hardens the fallback.
• 9 new helper unit tests pin both the legitimate delta-append path and the bug-fix replace path.

Test plan

Verified locally

• `nx run langgraph:lint` — green
• `nx run langgraph:test` — 53 tests pass (44 pre-existing + 9 new)

Pending visual verification

• After merge: re-smoke against the canonical demo with model=gpt-5, effort=high. Puzzle prompt produces a single non-duplicated visible answer.

Spec: `docs/superpowers/specs/2026-05-08-langgraph-streaming-content-dedup-design.md`
Plan: `docs/superpowers/plans/2026-05-08-langgraph-streaming-content-dedup.md`

Note

The captured fixture didn't reproduce the original 1.83× duplication (chunks happened to accumulate byte-for-byte to canonical), so the fix is best-understood as defense-in-depth: the synthetic Phase 4 unit test directly exercises the bug path (existing='partial answer' + incoming reasoning+text array → returns canonical, not appended). Once merged, the next live smoke pass against the demo will confirm whether the original visible duplication is gone or whether a different code path needs investigation.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
cacheplane	Ready	Preview, Comment	May 9, 2026 0:13am