[waleedlatif1]

Summary

• compactValue re-ran JSON.stringify on every subtree at every level and allocated a promise per node, so compaction was O(size × depth) on every block output, loop iteration/exit, parallel aggregate, and route response — all microtasks, never yielding
• Compaction now returns each node's JSON byte size alongside its value, and a parent derives its size from its children (keys + separators + child bytes) instead of re-serializing. A value is stringified only when it actually spills (its JSON is needed for storage) or when a toJSON makes the sum unsound
• A node falls back to a real JSON.stringify exactly where the old measurement could differ from a sum: its own toJSON, a child with toJSON (which receives the key it is serialized under), and a reject-mode __proto__ key (assignment sets the prototype)
• The walk is synchronous until something spills or block logs compact, so a fitting tree allocates no promises. Children still start in the same order (same seen set order, so shared references resolve identically), and a synchronous throw still observes already-pending siblings so none reject unhandled
• No duplicated rule set: the old compactValue/compactEntries/early-reject functions are rewritten in place with the same branches, reuse checks, reject estimates, manifests, and user-file stripping
• compactSubflowResults sums the aggregate { results } size from the per-result sizes instead of stringifying the whole aggregate

Type of Change

• Improvement (performance)

Testing

• Differential harness (kept out of the repo) against the pre-change serializer: 20,000 generated trees per seed × 3 seeds, plus fixed boundary cases, through compactExecutionPayload, compactSubflowResults, compactBlockOutput, and compactBlockLogs. Compares values, input identity, prototypes, own-key order and enumerability, store calls (json + size), and errors — 0 mismatches. Covers toJSON (own, keyed, on functions), getters, Symbol keys, non-enumerables, holes, -0/NaN/Infinity, BigInt, null-proto, __proto__ keys, cycles and shared refs, Dates, Buffers, class instances, user files, refs, finalBlockLogs, deep chains, lone surrogates, multibyte and escaped keys, negative/zero thresholds, preserveRoot, durable manifests, reject mode. Mutation-checked: each of 8 seeded accounting bugs fails it
• Reject-mode observedBytes match exactly for compactExecutionPayload (the only production reject path). For the concurrent APIs (compactSubflowResults, compactBlockOutput, compactBlockLogs) called in reject mode — which no production caller does — when several branches overflow, which one reports first follows microtask timing, so the byte count can differ; label and maxBytes match
• Bench, median before → after (Bun 1.4.2): plain 1MB rows 19.7 → 5.7ms; 3000 rows + Date each 31.3 → 6.3ms; 3000 class instances 4.2 → 1.7ms; own toJSON per row 17.5 → 6.6ms; finalBlockLogs 17.2 → 9.2ms; 20k user files 7.0 → 6.5ms (base64 kept 81 → 34ms); deep 500 levels 5.4 → 0.3ms; toJSON sibling at every level of 300 (forced fallback each level) 89.6 → 89.3ms; shared ref ×3000 6.7 → 3.3ms; wide 50k keys 15.8 → 12.6ms; 1000 small spills 18.4 → 12.9ms; reject mode 1MB 13.1 → 5.0ms; loop exit 40×1.2MB 1068 → 259ms, 5000×10KB 1510 → 269ms, 500×100KB 1035 → 230ms; block output with child spans 17.9 → 4.5ms. Node/Vitest numbers show the same direction (2–45x, worst case 1.09x)
• New regression tests in serializer.test.ts: oversized child under an object whose toJSON serializes small still spills, and rejects in reject mode; a parent is sized by its child's keyed toJSON; a value exactly at the threshold stays inline and one byte over spills with the exact size (escapes, multibyte, lone surrogates, omitted keys, undefined in arrays). Pass on the old serializer; the first fails on a top-down fast-path prototype, the others fail on the corresponding accounting mutations
• payloads/*.test.ts, loop, parallel, block-executor, executor, engine, execution-core, logging-session, event-buffer suites pass; bun run lint; check:comment-hygiene, check:utils, check:explicit-any, check:source-conventions, check:test-patterns, check:file-names. Full type-check / check:audits / test suite left to CI

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing (new tests pass the test-audit authoring gate)
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated
docs	Skipped		Oct 22, 2026 8:48pm UTC

[cubic-dev-ai]

No issues found across 2 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

<sub>Heads up: you’re close to your flex budget. Increase your flex budget so reviews don’t pause.

Turn on auto-fix | Re-trigger cubic</sub>

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge, with no new actionable issues found.

Summary

The PR sizes payloads bottom-up and avoids promises when no asynchronous work is needed.

• Execution payloads are sized from their children, not re-serialized at every level.

_{Reviews (2) · Last reviewed commit: "test(executor): pin the subflow aggregat..." · Reviewed by Greptile}

[waleedlatif1]

@greptile

[waleedlatif1]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

No issues found across 2 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

<sub>Heads up: you’re close to your flex budget. Increase your flex budget so reviews don’t pause.

Turn on auto-fix | Re-trigger cubic</sub>