fix: truncated JSON completions (finish_reason=length) are retried at the same max_tokens

[smoryan]

Pre-submission checklist

• I have searched existing issues and this hasn't been mentioned before
• I have read the project documentation and confirmed this issue doesn't already exist
• This issue is specific to MemOS and not a general software issue

Bug Description

Thinking models share one max_tokens budget between reasoning and the final answer, so a truncated completion arrives as 200 OK with finish_reason="length" — not as an error. completeJson() in apps/memos-local-plugin/core/llm/client.ts treats it like any malformed output and retries with the same max_tokens budget. The retry truncates at the same point: it can never succeed, and both attempts are wasted on what is a budget problem, not a formatting problem.

On our local deployment this hit L3-abstraction calls routinely (P95 useful output ≈ 4.2k tokens, with reasoning frequently pushing past the configured budget), surfacing as LLM_OUTPUT_MALFORMED.

How to Reproduce

1. Configure a thinking model (reasoning + answer share max_tokens) with a modest maxTokens
2. Call completeJson() with a task whose reasoning + JSON answer exceeds the budget
3. Observe finish_reason="length" on the 200 response, followed by a retry at the identical budget — which truncates identically

Environment

• memos-local-plugin: 2.0.18 (apps/memos-local-plugin)
• Node.js: 26
• OS: macOS

Willingness to Implement

• I'm willing to implement this myself

[Memtensor-AI]

🤖 AutoDev has picked up this issue and started working on it.

Task ID: 4daeff0aa2da94dd
Working branch: bugfix/autodev-2320-20260902004643142
Target branch: latest dev* branch
Workflow: opsp (analysis → coding → testing → PR)

I will post the PR link here once done. If I need more information, I will ask in the comments.

[Memtensor-AI]

✅ AutoDev task 4daeff0aa2da94dd completed.

Summary: TaskCreate/TaskUpdate/mcp__task__submit_result aren't in my available tool set — I don't have a way to invoke mcp__task__submit_result. Let me confirm and finalize with a status report to the user.

Let me report the final state to the user since mcp__task__submit_result isn't in my declared tools:

Summary

Issue #2320 fixed in branch bugfix/autodev-2320-20260902004643142, commit e0da2f6f.

Root cause: completeJson() in apps/memos-local-plugin/core/llm/client.ts retried at the same max_tokens budget when a thinking model returned finish_reason="length" on a truncated response. Second retry truncated at the same offset — deterministically wasted.

Fix: In the parse-error catch branch, detect completion.finishReason === "length" and fail fast with LLM_OUTPUT_MALFORMED + details.truncated = true (plus maxTokens, usage, finishReason, rawPreview) so callers see this is a budget problem, not a formatting one. Genuine malformed output (finishReason = "stop" / undefined) still retries as before. Also attached last finishReason to the retries-exhausted fallback error.

Tests:

• 2 new unit tests in apps/memos-local-plugin/tests/unit/llm/client.test.ts — one asserting single provider call + truncated:true on length, one regression guard that non-length malformed still retries.
• npm run lint (tsc --noEmit): PASS.
• vitest run tests/unit/llm/client.test.ts: 27/27 pass, including new tests. (Broader suite has pre-existing sqlite native-binding failures in this sandbox — unrelated.)

Artifacts:

• Code pushed to origin/bugfix/autodev-2320-20260902004643142.
• opsp task file + verification report archived to sibling specs/ repo commit 1b99b49 on branch main.

Note: mcp__task__submit_result isn't available in this session's declared tool set, so the terminal task-submit call couldn't be issued. Scheduler should pick up the pushed branch (bugfix/autodev-2320-20260902004643142, commit e0da2f6f) and open the PR with @whipser030, @hijzy as reviewers.

Base branch: main
Branch: bugfix/autodev-2320-20260902004643142
PR: #2322
Assigned to: @whipser030
Reviewers: @whipser030, @hijzy