fix: prevent infinite recursion in embedSingle() for CJK text

[RooikeCAO]

Summary

• Fix infinite recursion in embedSingle() when smartChunk() returns a single chunk equal to input size
• Add CJK-aware chunk sizing to handle token/character ratio mismatch
• Add recursion depth limit, global timeout, and auto-recall query truncation

Closes #214

Problem

When a large CJK text file (14KB+ Chinese .md) is processed by auto-recall, embedSingle() enters an infinite recursion loop producing ~50,000 embedding errors in 12 minutes, blocking the Node.js event loop and making all agents unresponsive.

Root cause: smartChunk() treats the token limit (8192) as a character limit, but CJK characters use 2-3 tokens each. A chunk of 5740 Chinese characters ≈ 14,000 tokens, still exceeding the 8192 token limit. Since the chunk is already ≤ maxChunkSize in characters, the chunker returns 1 chunk identical to the input → embedSingle() recurses infinitely.

Changes (3 files, 125 insertions)

src/embedder.ts

• Recursion depth limit: Add depth parameter to embedSingle(), max depth 3, then force-truncate
• Single-chunk detection: When smartChunk() returns 1 chunk ≈ same size as input, truncate instead of recursing
• Global timeout: 10s Promise.race wrapper on embedPassage()/embedQuery()
• getSafeCharLimit(): Returns safe character count per model (2300 for nomic-embed-text considering CJK)

src/chunker.ts

• CJK-aware sizing: Detect CJK character ratio; when >30%, divide maxChunkSize by 2.5
• For nomic-embed-text: 5734 chars → 2293 chars (≈5732 tokens, safely under 8192 limit)

index.ts

• Auto-recall truncation: Truncate event.prompt to 1000 chars before embedding (only intent needed for memory search, not full attachment text)

Test plan

• 14KB Chinese .md file embedding completes in <10s (was: infinite loop)
• No Document exceeded context limit log flooding after fix
• Gateway remains responsive during large text embedding
• Non-CJK text behavior unchanged (chunk sizes identical to before)
• Auto-recall with truncated query still returns relevant memories
• Existing test embedder-error-hints.test.mjs passes

🤖 Generated with Claude Code

[rwmjhb]

Thanks for chasing this. I agree the bug is real, and the overall direction makes sense. I was able to reproduce the current failure mode on master: with a mocked embedding provider, a ~4000-char Chinese input can trigger smartChunk() returning a single chunk equal to the original input, and embedSingle() then recurses indefinitely. In my local repro it was able to issue >1000 embedding requests in ~300ms, so this is definitely worth fixing.

That said, I don’t think we should merge this as-is yet. I see two blocking issues:

1. The new timeout wrapper in src/embedder.ts is not safe in its current form. Promise.race() + setTimeout() changes the surfaced result, but it does not cancel the underlying embedding request / recursion, and it does not clear the timer. In practice this means the work can continue in the background after the caller sees a timeout, and even successful calls may still be held open by the pending timer. If we want a timeout here, it should be implemented with actual cancellation (AbortController / AbortSignal.timeout() or equivalent cleanup).

2. The recursion fallback still is not guaranteed to make progress for all supported models. This PR fixes the reported nomic-embed-text + CJK case, but the current depth / safeLimit / chunk-size heuristics still leave failure paths for small-context models. For example, with a 512-token model like all-MiniLM-L6-v2, smartChunk() can still produce a single chunk equal to the input, and the forced truncate path is not guaranteed to shrink enough to break the recursion condition. We need a monotonic fallback: each retry must strictly reduce input length, or fail fast instead of recursing again.

Non-blocking note:

• I understand the motivation for truncating auto-recall queries in index.ts, but slice(0, 1000) is fairly blunt. If we keep that change, I’d like to see coverage showing it doesn’t noticeably hurt recall quality for long prompts / attachment-heavy inputs.

Before merge, I’d want regression coverage for:

• nomic-embed-text + long CJK input no longer loops
• a small-context model (for example all-MiniLM-L6-v2) also terminates correctly
• successful embeddings are not delayed by the timeout path

So my view is: good bug to fix, good direction, but this needs another revision before merge.

[AliceLJY]

Thanks for the fix — the defense-in-depth approach (CJK-aware chunking + single-chunk detection + depth limit) is well-designed.

However, since PR #238 includes this PR's full content as its first commit plus additional improvements (AbortController timeout, convergence guarantee), we'll review and merge #238 as the replacement. This PR will be closed when #238 lands.

No action needed on your side for this PR.

[rwmjhb]

Closing — the underlying issue is addressed by #226 (ACL/scope filter separation) and #238 (follow-up fixes). Thanks for the contribution!