Bug: DeepSeek "Input token exceed the limit" (400 quota_limit_reached) not classified as context overflow - auto recovery never triggers

[fantasywinter207]

Bug: DeepSeek Input token exceed the limit (HTTP 400, code quota_limit_reached) is not classified as context overflow, so automatic compaction recovery never triggers

现象 / Symptom

When a conversation exceeds the model context window, the DeepSeek API rejects the request with:

{"message":"Input token exceed the limit (request id: ...)","type":"api_error","param":"","code":"quota_limit_reached"}

The harness classifies this as INVALID_REQUEST (the 400 fallback), the turn fails, and a manual /compact afterwards fails with "Compaction could not produce a useful summary", because the history is already too long for the summarizer to read. I hit this against the official API through an OpenAI-compatible proxy (pi-ai provider).

根因 / Root cause

isContextWindowExceededError in packages/llm/llm/src/error.ts recognizes wording like context length exceeded, maximum context length, and too large for the model context, but not DeepSeek's terse production message Input token exceed the limit (the quota_limit_reached code name is reused for this failure). So both the llm-deepseek and llm-pi-ai adapters normalize it to INVALID_REQUEST.

The consequence is more than cosmetic: compaction-basic's automatic overflow recovery (agent/request-error listener) keys on CONTEXT_WINDOW_EXCEEDED, so the prune → compact → retry path never runs and the turn just dies with 400.

建议修复 / Suggested fix (already prepared)

Add an INPUT_TOKEN_EXCEED_LIMIT pattern to the classifier:

\binput[\s_-]tokens?\s+exceed(?:s|ed)?\s+the\s+limit\b

Requiring the limit to follow exceed immediately keeps usage-limit phrasings such as input tokens exceed the monthly usage limit out of the branch. The fix is implemented and verified locally (classifier unit tests + llm-deepseek mock-server test with the real 400 body + llm-pi-ai stream test; error.ts stays at 100% coverage). It is available for reference on a fork branch: fantasywinter207/deepseek-harness branch fix/classify-deepseek-input-token-exceed-limit (commit 6efce3a26e), in case the team wants to pick it up.

Happy to adjust anything if the team prefers a different wording/approach.

[denial123789]

I traced this against rc.7 (99f6f02). The important boundary is the raw provider body versus the normalized Harness error:

1. DeepSeek returns HTTP 400 with code: "quota_limit_reached" and message Input token exceed the limit.
2. The direct adapter maps HTTP 400 to INVALID_REQUEST; the shared text classifier does not currently recognize that exact phrase.
3. compaction-basic only runs its automatic prune / compact / retry path when agent/request-error.code === "CONTEXT_WINDOW_EXCEEDED".

That explains why this fails immediately even though it is physically a context overflow. It also explains why trying /compact afterward may fail: the summarizer must still read an already oversized history.

A narrow regression should include both sides:

• positive: Input token exceed the limit becomes CONTEXT_WINDOW_EXCEEDED
• negative: account or monthly usage-limit wording stays a quota/account error

Operationally, I would capture the raw body plus normalized code, lower requested output only when there is actual headroom, try /compact once, then move to a new Session with a concise handoff if summarization also overflows.

I wrote up the evidence chain and bounded recovery flow here: https://sandbaseai.github.io/deepseek-harness-handbook/context-window-overflow.html

Independent community handbook, not an official DeepSeek AI project.

[fantasywinter207]

Thanks for the independent confirmation — this matches what I found exactly, and it's good to have the evidence chain written down.

On the regression matrix: the fix I prepared already covers both sides you listed.

• positive: the exact Input token exceed the limit (request id: …) message (and the adapter-joined quota_limit_reached api_error … detail) → CONTEXT_WINDOW_EXCEEDED, verified end-to-end against the real 400 body in the llm-deepseek adapter mock-server test, and through the llm-pi-ai stream translation;
• negative: input tokens exceed the monthly usage limit stays out of the context-overflow branch, so account/usage-limit wording is not misclassified.

The change is one additional pattern in isContextWindowExceededError (packages/llm/llm/src/error.ts), requiring the limit to follow exceed immediately. It's available on fantasywinter207/deepseek-harness branch fix/classify-deepseek-input-token-exceed-limit (commit 6efce3a26e) if the team wants to pick it up; error.ts remains at 100% statement/branch/function/line coverage with the owning tests.

Agreed on the operational flow, with one nuance worth adding: automatic recovery can only help before the history is already over the window. Once the summarizer itself cannot read the history, a new session with a concise handoff is the only way out — exactly as you wrote.