github / github/app

Telemetry and Cache Explorer completely broken specifically for kimi-k2.7-code model

Aperta
#1,928 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub
Bugs
Lingua principale
Nessun dato sulla lingua
Stelle
2.1k
Fork
153
Metriche di merge delle PR
Nessuna PR unita negli ultimi 30g

Descrizione

### Short summary

_No response_

### Affected version or release

Github Copilot CLI v1.0.70

### Installation context

_No response_

### What happened?

Hey team,

Found a breaking telemetry bug in agent-host-copilotcli that only happens when using the kimi-k2.7-code model. Other models seem to track just fine.

Basically, the token tracking is completely dead, which then cascades and breaks the Cache Explorer dashboard.

1. Zero Token Telemetry Bug
As shown in the first screenshot, the agent successfully runs through 17 turns and 24 tool calls with 0 errors. However, all token metrics (Total Input, Total Output, Total Cached, Total Tokens) are completely stuck at 0. The execution works fine, but the telemetry data is missing.

Image

2. Broken Cache Explorer (Cascading Error)
Because the token counts are missing, the Cache Explorer is completely blind. It shows 0.00% cache hit and throws a fake "Likely cache expiration (~5 min)" warning.

But if you look closely at the timestamps, the gap between Turn 15 and Turn 16 is only 13 seconds (Turn 15 ended at 07:09:51 and Turn 16 started at 07:10:04). There is no way the prompt cache expired in 13 seconds. Both input tok and cached tok fields are completely blank.

It looks like the CLI is either failing to parse the token usage block from Kimi's API response, or it's completely forgetting to pass the caching headers for this specific model.

Image

### Steps to reproduce

_No response_

### Expected behavior

_No response_

### Additional context

_No response_

Guida per i contributori

Apri la guida per i contributori

Direzione di ricerca

Inizia riproducendo la telemetria a zero token con il modello kimi-k2.7-code in Github Copilot CLI v1.0.70, quindi traccia l’analisi dell’utilizzo dei token e i dati passati a Cache Explorer. Il lavoro è completato quando le metriche dei token di input, output, memorizzati nella cache e totali vengono valorizzate correttamente e la dashboard non segnala più falsamente la scadenza della cache.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Valutazione

Stack tecnologico
github
Ambito
observability-sre, tooling
Tipo di issue
Bug
Difficoltà
4/5
Tempo stimato
3-5 giorni
Stato di attività
Tranquilla
Chiarezza
Da chiarire
Idoneità per principianti
38/100

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.