fix(agent-core-v2): omit completion token cap unless explicitly configured

[sailist]

Requirement or Bug

Supersedes #3978 (same change set, rebased onto latest main, now pushed from an origin branch). Internal incident (2026-09-20): third-party models served on strict OpenAI-compatible stacks got a 400 on every request and fell into an overflow→compaction death spiral.

Bug Reproduction Steps

1. Point an OpenAI-compatible provider at a bare vLLM deployment (validates prompt_tokens + max_tokens > max_model_len before generation, no gateway clamp), serving a model with no max_output_size configured.
2. Send any prompt.
3. The first request 400s (client declares max_tokens = the whole context window, so prompt + cap exceeds max_model_len); later turns send window − last measured usage, still overshooting by the unmeasured delta, so every turn 400s. Each 400 is misclassified as a genuine context overflow, so compaction fires and retries in a loop — users see compaction trigger on every message, earlier and earlier.

Root Cause

When no explicit output limit is configured, the completion-budget chain fell back to capping at the entire context window (hardCap ?? window ?? reservedContextSize ?? 32000) and always encoded that guess as max_tokens/max_completion_tokens. No client-side default can be right for every serving stack; the correct default is to omit the field and let the server allocate the actual remaining window. This is a fundamental fix on the client side (a server-side clamp is out of scope).

Code Changes

• Budget chain: a completion cap is now produced only from explicit configuration (env KIMI_MODEL_MAX_COMPLETION_TOKENS or model max_output_size); otherwise the cap is undefined and the openai / openai-responses / google-genai egresses omit max_tokens/max_completion_tokens, letting the server decide output headroom. The whole-context-window and 32000 fallback chain (reservedContextSize, DEFAULT_UNKNOWN_CONTEXT_FALLBACK, computeCompletionBudgetCap, CompletionBudgetConfig.fallback) is removed.
• kimi provider exception: when no cap is configured, kimi's openai / openai-responses egress computes max(1, model.maxContextSize − usedContextTokens) at the protocol layer (kimiUnsetCompletionTokens), so kimi requests still fill the remaining context window. Nothing is sent when the window or the usage figure is unknown, and an explicit opt-out (<= 0) is honored.
• Anthropic keeps always sending max_tokens from the built-in per-model ceiling table; the fallback for unrecognized models is lowered from 128000 to 64000 (unknown models may be old ones whose server ceiling is below 128000).
• The remaining-window subtraction now counts size (measured anchor + estimated tail) instead of measured, covering messages added between turns and aligning with the human domain.
• Tests adapted in place: budget unit tests reshaped to the omit-by-default contract, wire snapshots drop the default maxTokens field, the compaction unknown-window case asserts omission, trait tests cover the kimi egress value. No new test files; test count not increased. Patch changeset included.

Behavior Changes and Affected Users

Behavior	Before	After	Who relies on the old behavior	Escape hatch
openai / openai-responses / google-genai requests without an explicit output limit	send max_tokens/max_completion_tokens = whole context window (or 32000 fallback)	field omitted; server decides	permissive hosted APIs see no change; strict vLLM stacks previously 400'd	set max_output_size or KIMI_MODEL_MAX_COMPLETION_TOKENS
kimi provider requests without an explicit output limit	send whole-context-window cap	send max(1, maxContextSize − usedContextTokens); omit when window/usage unknown	kimi served models	explicit cap via config; <= 0 opts out
Anthropic requests for models missing from the built-in table	max_tokens: 128000	max_tokens: 64000	unknown/older Anthropic models with lower server ceilings	explicit cap via config
Requests with an explicit cap (env / max_output_size / compaction override)	cap reduced by last measured usage	cap reduced by measured + estimated tail	none (same budget, more accurate accounting)	—
wire llm.request.maxTokens telemetry	always present	absent when no cap is configured	consumers expecting the field on every request	configure an explicit cap

Coverage by module: llm-adapter/model/completion-budget.ts (completionBudget.test.ts), agent/llmRequester/llmRequesterService.ts (loop/tool/plan/config wire snapshots), human/llm/protocol/format.ts + openai/openai-responses/kimi traits (trait.test.ts), human/llm/requester/bases/anthropic/profile.ts (trait.test.ts), compaction override (fullCompaction.test.ts).

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue — N/A: internal incident (2026-09-20), reproduction and root cause above; supersedes fix(agent-core-v2): omit completion token cap unless explicitly configured #3978.
• I have added tests that prove my feature works — existing budget / trait / compaction / wire-snapshot suites reshaped to the new contract.
• The behavior-change table above is complete, and every removed behavior or flipped default is named in the changeset and either has an escape hatch or was explicitly approved by a maintainer in this PR — escape hatch for each row is the explicit cap config.
• Ran gen-changesets skill — .changeset/omit-default-completion-token-cap.md included.
• Ran gen-docs skill — provider docs mention of the old default will be updated in a follow-up (carried over from fix(agent-core-v2): omit completion token cap unless explicitly configured #3978).

[changeset-bot]

🦋 Changeset detected

Latest commit: e059783

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/kimi-code	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

[pkg-pr-new]

pnpm dlx https://pkg.pr.new/@moonshot-ai/kimi-code@e059783

npx https://pkg.pr.new/@moonshot-ai/kimi-code@e059783

commit: e059783

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8e1f67e97b

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Include static prompt tokens in the Kimi remaining-window cap

For a Kimi OpenAI/Responses request with no configured output limit, this computes the cap from usedContextTokens, but the new caller obtains that value from the context-message counter. On a session's first request that counter estimates only the messages; it does not include the system prompt or tool schemas, which are sent separately in the same request. Consequently, a strict Kimi-compatible server receives max_completion_tokens equal to nearly the full window plus the static prompt/tool cost and rejects the initial request for exceeding its context limit. Derive this from the complete outbound request size (or omit the cap until such a measurement exists).

Useful? React with 👍 / 👎.

[sailist]

已知行为，符合预期