github / github/copilot-cli

OTel spans for subagent calls omit billing attributes (github.copilot.nano_aiu, github.copilot.cost), so external cost accounting undercounts actual billing

未关闭
#4,224 4 条评论 1 个 reaction 已指派 0 人 在 GitHub 查看
area:agents
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

### Describe the bug

When a session delegates work to subagents (the `task` tool / custom agents), the OTel spans for those subagent model calls omit **all** billing attributes. The subagent calls consume real AI credits (per the cumulative session display — see #4207, which notes session usage "including usage by the main agent, subagents, and background operations"), but that consumption is invisible to OpenTelemetry, so any external cost accounting built on the documented OTel telemetry (`copilot help monitoring`) systematically undercounts actual billing.

Main-agent `chat` spans carry the full billing envelope:

- `github.copilot.nano_aiu`
- `github.copilot.cost`
- `github.copilot.initiator` (`user` / `agent`)
- `github.copilot.interaction_id` / `github.copilot.turn_id`

Subagent model calls emit spans with correct `gen_ai.*` semconv attributes (`gen_ai.operation.name`, `gen_ai.request.model`, `gen_ai.usage.input_tokens` / `output_tokens` / `cache_read.input_tokens` / `cache_creation.input_tokens`) but carry **none** of the `github.copilot.*` billing attributes — no `nano_aiu`, no `cost`, not even `initiator` or an interaction id. This holds at every level:

1. The subagent's child `chat` spans: no billing attributes.
2. The subagent's own nested `invoke_agent` span (which does carry `gen_ai.agent.name` and rolled-up token counts): no billing attributes.
3. The session root `invoke_agent` span's `github.copilot.nano_aiu` rollup equals the sum of the main-agent `chat` spans' `nano_aiu` **exactly** — i.e. the subagent spend is not folded in anywhere.

Observed within a single trace from one session (models are our configured custom-agent models; the same holds for built-in `task` delegation):

| Span | Model | `gen_ai.usage.input_tokens` | `nano_aiu` | `cost` |
|---|---|---|---|---|
| root `invoke_agent` | claude-fable-5 | 1,963,612 | 506,956,900,000 | 13 |
| main-agent `chat` × 13 | claude-fable-5 | … | present on every one; sums to exactly 506,956,900,000 | 1 each |
| nested `invoke_agent` (custom agent A) | claude-haiku-4.5 | 296,978 | **absent** | **absent** |
| nested `invoke_agent` (custom agent B) | gpt-5.6-terra | 155,061 | **absent** | **absent** |
| subagent `chat` spans | claude-haiku-4.5 / gpt-5.6-terra | 124K–321K each | **absent** | **absent** |

In our sessions the unreported subagent calls account for roughly 10–15% of estimated session cost (rate-card priced from the token counts, which *are* reported).

### Affected version

GitHub Copilot CLI 1.0.73

### Steps to reproduce the behavior

1. Enable OTel export (`COPILOT_OTEL_ENABLED` with an OTLP endpoint, or `COPILOT_OTEL_FILE_EXPORTER_PATH`).
2. Run an interactive session and issue a prompt that delegates to a subagent — e.g. a custom agent configured with a different model, or anything that triggers the `task` tool.
3. Inspect the exported spans for the session trace.
4. Main-agent `chat` spans carry `github.copilot.nano_aiu` / `github.copilot.cost`; the subagent `invoke_agent` and `chat` spans carry token usage but no billing attributes, and the root rollup excludes them.

### Expected behavior

Subagent model calls should carry the same billing attributes as main-agent calls (`github.copilot.nano_aiu`, `github.copilot.cost`), so that OTel-based cost accounting matches what is actually billed. Alternatively, if this is intentional (subagent calls genuinely bill zero credits), it would be great to have that documented in `copilot help billing` / the monitoring docs — though the cumulative session credit display suggests they are billed.

### Additional context

- OS: Windows 11, x86_64, Windows Terminal / PowerShell (also reproduced from sessions on the same version exporting via OTLP to an OpenTelemetry Collector).
- Related issues: #4207 (per-subagent credit breakdown in `/usage` — confirms subagent usage is part of session credit consumption), #1582 (`/usage` undercounts quota consumption; suspects background operations), #2068 (background compaction consumes a premium request), #4107 (verified `nano_aiu` matches the interactive footer exactly for main-agent sessions — the mismatch reported here only appears once subagents are involved), #3778 (OTel cost metric feature request).
- Happy to provide sanitized span exports (main-agent vs subagent, same trace) on request.

贡献指南

打开贡献指南

调研方向

先从 `copilot help monitoring` 中记录的 Copilot CLI telemetry 开始,然后使用启用 OTel export 的 `task` 工具或自定义 agent 重现该问题。比较 main agent 的 `chat` spans 与 subagent 的 `invoke_agent` 和 `chat` spans,包括 session root rollup。完成的标准是:subagent 的 billing attributes 和 usage 得到一致的表示;如果这些调用是有意免费的,则记录 billing behavior。

由索引模型根据 Issue 内容生成。

评估

技术栈
github
领域
cli, observability
Issue 类型
缺陷
难度
4/5
预计耗时
3-5 天
活跃度
活跃
描述清晰度
基本清楚
新手友好度
48/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。