Repository navigation
Massive token spike causing quota exhaustion in single session #26184
Description
Description
Was using opencode go through opencode tui. Deepseek V4 Pro was selected as the model and the session was about 500k~ tokens. The chat started to slow down and every response took a minute to respond and finally I got quota exceeded.
From my observations session an3G3iVR consumed ~5 hours of quota in ~30 minutes due to input tokens ballooning from ~20k to 500k+ per request.
From usage logs:
- Normal requests (2:58-3:26 PM): 18k-21k input tokens, $0.001-$0.004 cost
- After 3:30 PM: 570k-614k input tokens, $0.98-$1.06 cost per request
15 requests after 3:30 PM cost ~$15 instead of ~$0.05 normal.
Looks like conversation context got stuck in a loop, re-sending entire 500k token history repeatedly.
If there's anything I can do to clarify, please let me know.
Plugins
none
OpenCode version
1.14.40
Steps to reproduce
- Open OpenCode TUI
- Select deepseek-v4-pro model through OpenCode Go
- Work in a session until it grows to ~500k tokens
- Observe: responses slow down to ~1 minute each
- Input tokens balloon to 500k+ per request (should stay ~500k, not multiply)
- Quota gets exhausted in minutes instead of hours
Screenshot and/or share link
Operating System
Arch Linux 6.18.22-1-lts
Terminal
kitty
Reacted by Marko Jovanovic
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon May 7, 2026
This issue might be a duplicate of an existing issue. Please check:
IgorGanapolsky commentedon Jun 6, 2026
The 300× spike (≈$15 for what should've been $0.05) from context re-sending 500k tokens in a loop is exactly the runaway class I work on. Free MIT kit that stops this from torching a Mac/quota: https://github.com/IgorGanapolsky/mac-yolo-safeguards. Besides that session — what's the worst single loop incident you've hit?
Necmttn commentedon Jun 22, 2026
This class of bug needs a per-turn context-growth ledger.
For each request, record request id, session id, model, prompt token count, prior prompt count, delta, cache read/write, compaction state, context segment counts, and whether the same large history block was resent. Then OpenCode can warn before the second or third 500k-token resend instead of only discovering the loop from quota exhaustion after the fact.
Generated with ax - https://github.com/Necmttn/ax
github-actions commentedon Aug 22, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.