github / github/copilot-cli

Bug: Copilot CLI 1.0.82 BYOK silently disables prompt caching (~5x cost)

未关闭
#4,720 0 条评论 1 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

area:models area:networking
主要语言
Shell
星标
11.2k
派生
1.9k
平均合并
14 小时 16 分钟
30 天内合并 PR
6

描述

Summary

Copilot CLI 1.0.82 in BYOK mode sends chat requests that carry no prompt-cache declaration, so prompt caching is effectively disabled for the whole session. Every turn re-sends the entire growing context at full price — provider usage shows cached_tokens=0 and cache_write/cache_creation_tokens=0 on every request across the session, even though the prefix is stable.

Net effect: ~5x cost for normal multi-turn BYOK sessions.

Setup

  • Copilot CLI 1.0.82, BYOK mode (custom OpenAI-compatible /chat/completions endpoint serving an Anthropic model, claude-sonnet-5)
  • Compare baseline: same machine, same endpoint/model, 1.0.80 BYOK → prompt caching works (~97% hit)
  • Also tested 1.0.82 in GitHub-subscription mode on the same machine → caching works, so the build is cache-capable

Expected

Once a conversation prefix stabilizes, the provider should report cached_tokens > 0 and small cache_creation_tokens for the new tail (that's what 1.0.80 BYOK and subscription mode report).

Actual

Every request reports zero cache metrics:

"prompt_tokens_details": {
  "text_tokens": 146022,
  "cached_tokens": 0,
  "cache_write_tokens": 0,
  "cache_creation_tokens": 0,
  "cache_creation_token_details": { "ephemeral_5m_input_tokens": 0 }
},
"cache_read_input_tokens": 0

Over a 150-request session, text_tokens climbed monotonically 118,966 → 285,085 while cached_tokens and cache_write/creation_tokens stayed 0 every time. The zero cache-write/creation is the strong signal: even a cold miss should write the prefix so the next turn can hit; a persistent 0 write means the request never asks for a cache at all.

Evidence

Metric 1.0.82 BYOK (broken) 1.0.80 BYOK (baseline)
Requests 150 320
Input tokens 30,075,042 100,762,401
Cached tokens 0 98,424,636
Cache hit rate 0% 97.7%
Cost $61.21 $28.92

With normal caching, 30.1M input ≈ $11.4; at 0% it costs ≈ $60.2 → ~$49 (~5x) wasted in a single session.

Root-cause hypothesis

The 1.0.82 BYOK request builder omits the prompt-cache declaration (cache_control / ephemeral-cache policy) that subscription mode injects. Because BYOK calls the user's endpoint directly, the missing declaration is only visible to BYOK users.

Severity

High — silent runaway cost on BYOK setups; no CLI error, only a much larger bill. Requires monitoring provider cache hit-rate to detect.

Suggested fix

Make the BYOK path send the same prompt-cache declaration used by subscription mode.

Repro note

I cannot reproduce right now (downgraded to 1.0.80 where caching works). Reproducing requires running 1.0.82 in BYOK against an Anthropic-model endpoint and reading provider cache accounting. Happy to help test a newer build backported if there's a fix candidate.

贡献指南

打开贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

调研方向

定位 BYOK request builder,并将其 request 构造与 subscription path 以及 issue 中描述的 1.0.80 行为进行比较。检查是否存在 prompt-cache declaration,然后针对 Anthropic-model endpoint 执行一个多轮 BYOK 会话,并验证 cache-write 和 cache-read 指标不再为零。

由索引模型根据 Issue 内容生成。

评估

技术栈
shell
领域
api, cli
Issue 类型
缺陷
难度
3/5
预计耗时
1-2 天
活跃度
活跃
描述清晰度
基本清楚
新手友好度
55/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。