fix(agent-core): cap compaction output tokens when maxOutputSize is undefined

[li-xiu-qi]

Related Issue

Resolve #834

Problem

The compaction worker in full.ts does not pass maxOutputSize to resolveCompletionBudget(). As a result, computeCompletionBudgetCap() always falls back to using the full max_context_tokens as max_completion_tokens — regardless of whether maxOutputSize is configured.

Since the compaction prompt itself is large (it contains the entire history being compacted), input_tokens + max_tokens exceeds the context window, and providers that do not auto-clamp max_tokens server-side reject with 400 APIContextOverflowError.

Reported in #476, #794, and #834. PR #482 partially addresses this by passing maxOutputSize, but does not cover the case where maxOutputSize is undefined (the default for most models).

What changed

• Pass maxOutputSize to resolveCompletionBudget in full.ts compactionRound() — aligns with the main loop in index.ts (same as PR fix(compaction): pass maxOutputSize to resolveCompletionBudget #482)
• Add a conservative fallback cap when maxOutputSize is undefined: Math.min(Math.floor(maxCtx / 4), 8192). Compaction is a summarization operation; 8192 tokens is generous for a high-quality summary while preventing overflow on any provider. This covers the gap left by PR fix(compaction): pass maxOutputSize to resolveCompletionBudget #482.
• Add 6 unit tests in completion-budget.test.ts covering the compaction budget resolution scenarios
• Add a verification test (compaction-overflow-verification.test.ts) that demonstrates the overflow before and after the fix using three real model configurations

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue (Compaction fails with 400: maxOutputSize not passed to resolveCompletionBudget in compaction path #834).
• I have added tests that prove my feature works.
• Ran gen-changesets skill.
• Ran gen-docs skill, or this PR needs no doc update.

Verification

Unit tests

• npx vitest run packages/agent-core/test/utils/completion-budget.test.ts — 27/27 passed (21 existing + 6 new)
• npx vitest run packages/agent-core/test/utils/compaction-overflow-verification.test.ts — 3/3 passed
• npx oxlint packages/agent-core/src/agent/compaction/full.ts packages/agent-core/test/utils/completion-budget.test.ts packages/agent-core/test/utils/compaction-overflow-verification.test.ts --type-aware — 0 warnings, 0 errors

Overflow verification: before vs after fix

The verification test simulates the budget resolution logic from the original and patched compactionRound() using three real model configurations. With a typical compaction input of ~80K tokens:

Model	max_context	maxOutputSize	Original max_tokens	Original overflow?	Patched max_tokens	Patched overflow?
step-3.7-flash	256,000	undefined	256,000	YES ❌ (336K > 256K)	8,192	NO ✅ (88K)
kimi-for-coding	262,144	undefined	262,144	YES ❌ (342K > 262K)	8,192	NO ✅ (88K)
glm-5.2	1,000,000	131,072	1,000,000	YES ❌ (1.08M > 1M)	131,072	NO ✅ (211K)

Key finding: the original code overflows for ALL three models — even when maxOutputSize is explicitly configured (glm-5.2), because the compaction path never passes it to resolveCompletionBudget.

Reproducible error example

Using stepfun/step-3.7-flash (max_context=256K, no maxOutputSize configured) with a conversation large enough to trigger compaction:

Error: [compaction.failed] APIContextOverflowError: 400
{"detail":"{\"error\":{\"message\":\"This model's maximum context length is 256000 tokens. However, you requested 176000 output tokens and your prompt contains at least 80000 input tokens, for a total of at least 256000 tokens.\",\"type\":\"BadRequestError\",\"param\":\"input_tokens\",\"code\":400}}"}

With this fix, max_completion_tokens is capped at 8,192 instead of 256,000, so 80,000 + 8,192 = 88,192 stays well within the 256K window.

Relationship to PR #482

This PR is complementary to #482:

• PR fix(compaction): pass maxOutputSize to resolveCompletionBudget #482 passes maxOutputSize to resolveCompletionBudget (necessary)
• This PR adds the same fix plus a conservative fallback when maxOutputSize is undefined (covers the remaining gap)

Both PRs can coexist — if #482 merges first, the fallback cap in this PR is still independently valuable. If this PR merges first, #482 becomes a no-op.

[changeset-bot]

🦋 Changeset detected

Latest commit: 6402aed

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/agent-core	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR