anthropic: cache-aware top-level output_config.effort carrier

[topcheer]

Concept

Anthropic 2026 GA effort parameter lives at top-level output_config.effort (not inside thinking blocks). Official guidance: top-level effort changes do preserve the prompt cache prefix and combining with extended thinking is the recommended pattern (effort steers overall token envelope, budget_tokens caps explicit thinking). SDK v1.68 exposes MessageNewParams.OutputConfig (GA); per-message markers are not expressible yet.

Gap

ggcode maps effort only to thinking.budget_tokens (buildParams L988-991). Mid-conversation effort switches rewrite request-body params — and the repo's own cache_efficiency_monitor lists "Reasoning effort changes" as a cache-storm cause. The cache-safe carrier was missing.

Implementation (internal/provider/anthropic.go, +346)

1. Hysteresis state machine beginEffortTracking: top-level carrier attaches only after effort level is stable across two consecutive requests
   • adaptive effort apply→restore (constant level per request) correctly establishes the carrier; oscillating levels never attach
   • user /effort switch: suppressed once, then rebuilt (one deliberate cache rewrite, not a churn storm)
2. buildParams: emits OutputConfig.Effort only when effortCarrier && conversationEffort == lastCallEffort; comments record the SDK-marker upgrade path
3. Rejection circuit breaker isEffortError: mirrors isThinkingError (4xx/404/422 status gate + anchored phrases); chat and streaming paths retry once on endpoint rejection with CAS carrier teardown, degrading to budget_tokens semantics for the session
4. SetReasoningEffort extended with xhigh/max (carrier-only); CloneWithModel inherits the breaker state, resets the stability window

Tests (zz_effort_carrier_test.go, 213 lines)

Hysteresis matrix, OutputConfig emission gating, isEffortError true/false cases, clone semantics, xhigh/max normalization.

Verification

• go build -tags goolm (worktree + main side) ✅
• -race targeted tests ✅
• provider full package (1.9s) + agent full package (24.6s) regression ✅
• pre-commit hooks (gofmt/vet/build) ✅

Co-Authored-By: ggcode noreply@ggcode.dev

[topcheer]

复审通过 ✅（两重点核销；首轮 CI 失败=mcp SSE 时序 flake，rerun 绿——本地 3 次复跑过+包域零交集）

① KV-cache 前缀影响——hysteresis 设计正确：body 引用的官方语义"top-level effort 不 preserve prefix"（注释同样写明）→ 解法是双连续稳定才挂载体：adaptive-effort 的每轮 set+restore 振荡永远不稳定→永不触载体（budget_tokens 逐轮调制 cache-neutral）；用户显式切换第二次调用重建载体=一次有意的 cache rewrite 而非风暴。effort off 重置窗口（振荡防泄漏）。与 #2445 前缀稳定性约定一致：稳定优先，变更刻意。

② 旧网关拒载回退：isEffortError 锚定短语+4xx 参数类状态码限定（镜像 isThinkingError 形态）；CAS latch 单次降级重试（Chat+ChatStream 双路径，Stream 仅未发射时）；latch 持久+clone 继承（endpoint 级能力记忆）但会话窗口重置（clone 各自 cache prefix 重新学习）——边界划分准确。

SDK 边界诚实：per-message marker 不可表达（v1.68）→重建走 top-level，注释留 flip 点。

5 测试。CI rerun 绿。可合并。