fix(agent-core): classify context-worded 401/403 as context.overflow instead of auth error

[bowser-bot]

WIP — not ready for upstream. This draft lives only in this fork for development and review. Do not submit to MoonshotAI/kimi-code in the current state.

Related Issue

Related to MoonshotAI#2613

Problem

The Kimi managed subscription intentionally returns 401/403 for capability rejections worded as <model> supports only <N> context. The client treated every 401 as a credential failure, so an over-long context (e.g. an auto-compaction request that itself exceeds a 256K model's window) forced a pointless token refresh and surfaced as -32000 Authentication required over ACP — indistinguishable from a real auth problem.

What changed

The 401/403 detection is added only in the Kimi vendor error module (kosong/src/providers/kimi-errors.ts, beside the existing classifyKimiQuotaError), with a quantity-anchored wording gate: a size must follow "supports only" (256K context, kimi-k3 up to 256K context). Auth/scope rejections that merely mention "context" in prose keep the credential path, and the shared provider-agnostic classification is untouched — nothing changes for non-Kimi providers.

• v1 OAuth wrappers (agent-core/session/provider-manager.ts, node-sdk/kimi-code-model-provider.ts): skip the forced token refresh and throw CONTEXT_OVERFLOW with the provider's message instead of a misleading re-login prompt. errors/serialize.ts maps the same wording to the context.overflow payload code.
• v2 parity (agent-core-v2): the model requester skips the forced refresh and re-mints the error as APIContextOverflowError; classifyApiError reports context_overflow, not auth.
• Compaction (agent-core/compaction/full.ts): propagates context.overflow unwrapped (matching the existing auth-code passthrough) and treats its KimiError form as recoverable, keeping the shrink-and-retry path.
• ACP wire (acp-adapter/session.ts): a turn.ended failure coded context.overflow now rejects session/prompt with -32603 and structured data { code: "context.overflow" } instead of silently resolving end_turn. Only whitelisted public codes cross as data.code; raw provider messages and stacks never cross the wire. Other non-auth codes keep the log-and-resolve path.
• Telemetry (agent-core/turn/index.ts): checks the Kimi-scoped wording before the auth statuses; ordering otherwise unchanged.

Out of scope (awaiting maintainer feedback on the issue): session/set_model pre-flight checks and chunking oversized compaction input. Wordings without a quantity are not matched by design; the Kimi-scoped pattern is the single extension point.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue, or explained the problem above.
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset. — changeset added (minor, both publishable packages; wire-visible behavior semantics change)
• Ran gen-docs skill, or this PR needs no doc update. — behavior fix only, no config/API surface change

Verification

• Targeted suites (8 files) — 312 passed, 1 skipped
• Full regression: kosong + acp-adapter + node-sdk — 2046 passed; agent-core — 4132 passed; agent-core-v2 — 4676 passed, 11 failed (plugin zip/GitHub install fixtures that fail identically on the base commit in this sandbox — unrelated)
• pnpm typecheck clean; oxlint 0 errors on changed files
• End-to-end: OAuth provider + synthetic 401 k3-256k supports only 256K context. → no forced refresh, turn fails context.overflow, ACP client receives -32603 + data.code; genuine-auth negatives (Bearer/scope phrasings) still refresh and stay auth errors.