fix(ENG-12518): fix field filter and batch sentence classification

[hiskudin]

Context

Two issues causing high Tier 2 latency:

1. extractStrings() field filter was broken — the traverse function had a fallthrough case that collected every leaf string under non-matching keys, making tier2Fields and the riskyFieldNames fallback from Tier 1 completely ineffective. For Gmail emails, this meant classifying ~58 sentences (MIME headers, DKIM signatures, base64 blobs) instead of ~4 (just the snippet field identified by Tier 1). At ~35ms per ONNX inference on server CPU, this caused ~2s latency per request instead of ~140ms.

2. classifyBySentence() called classify() per sentence in a loop — submitting N separate session.run() calls to the libuv thread pool. Under concurrency, these compete for threads and stack up. Switched to classifyBatch() which pads all sentences into a single batched tensor and runs one session.run() call.

Changes

• src/core/prompt-defense.ts: Fix extractStrings() to skip strings under non-matching keys when field filtering is active. Handle bare string input as edge case.
• src/classifiers/tier2-classifier.ts: Refactor classifyBySentence() to collect all sentences first, then call classifyBatch() once instead of classify() N times.

Checklist

• All 174 tests pass
• Lint clean
• Verified locally with real Gmail payloads: scores identical, latency reduced
• Concurrent test (20 parallel calls): no stacking, all complete in ~3.8s total

Post-merge

• release-please bumps version
• Bump @stackone/defender in unified-cloud-api and connect-handler

🤖 Generated with Claude Code

[cubic-dev-ai]

No issues found across 1 file

_{Auto-approved: Standard fix for sharing state across CJS/ESM modules using globalThis. Minimal risk, addresses significant performance overhead without changing core logic.}

[Copilot]

Pull request overview

Updates the ONNX classifier’s session/promise caching so it is shared across CommonJS and ESM evaluations within the same Node.js process, preventing duplicate model loads when the package is consumed in dual-module-format setups.

Changes:

• Moved ONNX session cache and in-flight load promise tracking from module-level Maps to globalThis-backed singletons.
• Added a dedicated SessionCacheEntry type and expanded inline documentation describing the CJS/ESM sharing rationale.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[cubic-dev-ai]

1 issue found across 2 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/config.ts">

<violation number="1" location="src/config.ts:176">
P3: The comment claims the Tier 2 field list adds "data" and "value", but the array only includes "snippet". Either add the missing fields or update the comment to avoid misleading configuration guidance.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[cubic-dev-ai]

0 issues found across 2 files (changes from recent commits).

_{Requires human review: Modifies core filtering and classification logic in a security-sensitive component (Prompt Defense). Such changes to input traversal and execution batching require human verification.}

[cubic-dev-ai]

1 issue found across 3 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/core/prompt-defense.ts">

<violation number="1" location="src/core/prompt-defense.ts:71">
P2: Arrays at the root now bypass field filtering and collect every string. If tool results are arrays of objects, tier2Fields is ignored and non-matching fields get classified again. Keep arrays in the filtered traversal path.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

0 issues found across 2 files (changes from recent commits).

_{Requires human review: This PR modifies core security/filtering logic in the prompt defense system and introduces batch processing for ONNX classification, which requires human validation to ensure no bypasses.}