Dramatic increase in token consumption

[cvkmohan]

I have been using opencode for the last 10 days or so. I can say - intensively. However, till yesterday, not once did I reach a stage where we get out of tokens/out of context window. Even with reasonably big operations - somehow opencode managed to run - most of the times - not even reaching 50% of token size. However, from yesterday - after upgrade to 0.2.3x or so - suddenly, the consumption increased very significantly. Every restart getting exhausted in minutes. In fact, with my claude MAX $200 subscription - never reached its limits. In less than an hour - dramatically I am told that my limit gets reset after 4 hours. Has something changed? Or problem is at my end? Raising an issue because - nothing much changed in the way I use - but - I see a dramatic increase in token consumption and context exhaustion.

Just wanted to alert. Close it, if it is not related.

[thdxr]

can you reproduce a situation where the context window gets big faster than you expect a /share a link?

[cvkmohan]

Thanks for the quick response. I will share if it happens. So, no change on that front in code? One difference I can think of is the FigmaMCP that I used. Maybe its output is doing some mischief. Earlier with my Elixir Code - it never crossed 20% - including contexts, tests etc.

[cvkmohan]

https://opencode.ai/s/e7oOMJ1y

[cvkmohan]

I can confirm that token usage/context window usage has dramatically increased. I tried to run the same task in both claude cli and opencode cli - and opencode cli exhausted context window by the time claude cli reached 50%. Earlier it used to be reverse. Something changed. Neither to hurry you nor cribbing. Just wanted to share the information. Please do not mind.

[thdxr]

i looked through your session (if you open browser console in share page it prints out an object with everything)

i don't immediately notice anything suspicious - you can see in each step-finish part it report how many tokens it wrote into the cache

it progressively gets bigger - not sure

[thdxr]

ahhh i think the issue is with every edit we include project level diagnostics and you have a lot of them. let me play with this

[thdxr]

in 0.3.16 i filter out only Error diagnostics. we might remove them all together for project level ones but this is a first step

[ShivaR333]

Facing the same issue. Burnt through around 800K tokens for just a few edits. Is there any way to remove the diagnostics and start from a clean slate?

[adamdotdevin]

@thdxr maybe we make this a flag to disable project diagnostics altogether, but default to errors only enabled

[cvkmohan]

But, this is not a newly introduced feature. I was dealing with a bigger project (same in elixir, phoenix and liveview) a week before this issue - and - opencode was handling the context beautifully. I mean, it reached a maximum of 50% even with a long conversation. Something changed for sure.

[claidler]

@thdxr @adamdotdevin I'm noticing this also with GitHub Copilot as a provider in particular, across various projects seeing the token limit being hit after a few messages. I think it's probably related to diagnostic messages as I'm debugging some large error logs in both cases

[jakubdonovan]

in 0.3.16 i filter out only Error diagnostics. we might remove them all together for project level ones but this is a first step

I wanted to add that I’m also seeing extremely high input token usage on a relatively small SvelteKit project, even after adding an AGENTS.md file that excludes node_modules, .svelte-kit/, and build/ folders.

Since last night, I’ve already burned through $20 in usage, which is far beyond what I expected for this project.

Here’s the exact error I’m now encountering on each prompt:

[!ERROR]
{
  "data": {
    "message": "AI_APICallError: input length and max_tokens exceed context limit: 172775 + 32000 > 200000, decrease input length or max_tokens and try again"
  },
  "name": "UnknownError"
}

I also pulled my Claude usage logs and saw input token counts regularly in the 160k–172k range, all within minutes:

2025-07-20 14:29:30 — 172,448 tokens
2025-07-20 14:29:25 — 172,242 tokens
2025-07-20 14:29:15 — 171,931 tokens
...

I’ve tried compacting the session and deleting recent messages, but I’m now completely blocked due to the context limit being exceeded on every request. As a result, I am unable to do anything.

I’d be happy to share my session log privately if that helps with debugging

[claidler]

Yeah it's worth me stating that one of mine was particularly small also which is why I imagined it may be diagnostics related. I'm wondering if there are restrictions around reading very large files or very large logs. Mine had 3-4 normal files, but some large files after a build that may have been picked up, or some long logs.

[claidler]

Reviewing with AI, it said that there are not per file context limitations - I wonder if there should be. If someone has a large pdf or text file that they need to commit in their project that keeps getting picked up, this would be a point of frustration. I understand not everyone would want this, so it might need to be configurable, but I think we could decide on a sensible default considering any good code project shouldn't have code files longer than a certain amount of lines. Is there anything in place here, and if not, could we put something in place?

[ivanglushko]

This is the issue for me as well, just a simple prompt skyrockets the token usage. Would be nice to see what files / additional info was sent.

[raxtor]

Yes, there is something wrong... plain message costs 45k+ tokens... e.g. Kimi K2 is not able work... and is possible somewhere set model provider for OpenRouter?

[rekram1-node]

@raxtor seems like the kimi k2 free configuration is off we will investigate

[timcreatedit]

I also fully can't use opencode with claude anymore, it often maxes out after my first message.

[mrgoonie]

@thdxr I think in v0.9.5 the token count is incorrect:

Screenshot: (26.4K for a simple prompt)

But when I open the shared session, it's showing 252 tokens: https://opencode.ai/s/Whzi8BH0

[thdxr]

do you have any MCP servers enabled? baseline should be around 11k

[mrgoonie]

yes, I have a couple ones, is there a way to see how many tokens that the MCP servers are taking?

[westito]

MCP-s cost a lot of tokens! I enabled just a few. Figma, jetbrains, and it eats up 52k (20%) tokens!!

Figma use A LOT. I always keep it disabled and enable if I need. This is Claude Code, not OpenCode, but you can see context content there.

MCP tools · /mcp
     └ mcp__ide__getDiagnostics (ide): 560 tokens
     └ mcp__context7__resolve-library-id (context7): 874 tokens
     └ mcp__context7__get-library-docs (context7): 835 tokens
     └ mcp__dart__connect_dart_tooling_daemon (dart): 634 tokens
     └ mcp__dart__get_runtime_errors (dart): 631 tokens
     └ mcp__dart__hot_reload (dart): 633 tokens
     └ mcp__dart__get_widget_tree (dart): 569 tokens
     └ mcp__dart__get_selected_widget (dart): 569 tokens
     └ mcp__dart__set_widget_selection_mode (dart): 601 tokens
     └ mcp__dart__get_active_location (dart): 577 tokens
     └ mcp__dart__pub_dev_search (dart): 1.0k tokens
     └ mcp__dart__dart_fix (dart): 675 tokens
     └ mcp__dart__dart_format (dart): 719 tokens
     └ mcp__dart__run_tests (dart): 2.1k tokens
     └ mcp__dart__create_project (dart): 844 tokens
     └ mcp__dart__pub (dart): 773 tokens
     └ mcp__dart__analyze_files (dart): 546 tokens
     └ mcp__dart__resolve_workspace_symbol (dart): 671 tokens
     └ mcp__dart__signature_help (dart): 639 tokens
     └ mcp__dart__hover (dart): 648 tokens
     └ mcp__figma__get_screenshot (figma): 1.1k tokens
     └ mcp__figma__create_design_system_rules (figma): 851 tokens
     └ mcp__figma__get_code (figma): 1.1k tokens
     └ mcp__figma__get_metadata (figma): 1.1k tokens
     └ mcp__jetbrains__execute_run_configuration (jetbrains): 814 tokens
     └ mcp__jetbrains__get_run_configurations (jetbrains): 705 tokens
     └ mcp__jetbrains__get_file_problems (jetbrains): 785 tokens
     └ mcp__jetbrains__get_project_dependencies (jetbrains): 647 tokens
     └ mcp__jetbrains__get_project_modules (jetbrains): 652 tokens
     └ mcp__jetbrains__create_new_file (jetbrains): 765 tokens
     └ mcp__jetbrains__find_files_by_glob (jetbrains): 859 tokens
     └ mcp__jetbrains__find_files_by_name_keyword (jetbrains): 813 tokens
     └ mcp__jetbrains__get_all_open_file_paths (jetbrains): 659 tokens
     └ mcp__jetbrains__list_directory_tree (jetbrains): 755 tokens
     └ mcp__jetbrains__open_file_in_editor (jetbrains): 696 tokens
     └ mcp__jetbrains__reformat_file (jetbrains): 680 tokens
     └ mcp__jetbrains__get_file_text_by_path (jetbrains): 838 tokens
     └ mcp__jetbrains__replace_text_in_file (jetbrains): 1.0k tokens
     └ mcp__jetbrains__search_in_files_by_regex (jetbrains): 873 tokens
     └ mcp__jetbrains__search_in_files_by_text (jetbrains): 871 tokens
     └ mcp__jetbrains__get_symbol_info (jetbrains): 830 tokens
     └ mcp__jetbrains__rename_refactoring (jetbrains): 959 tokens
     └ mcp__jetbrains__execute_terminal_command (jetbrains): 1.0k tokens
     └ mcp__jetbrains__get_repositories (jetbrains): 651 tokens

[rekram1-node]

having context breakdown is good we should have that too

[Nindaleth]

@rekram1-node Is there an issue tracking the context breakdown that I could subscribe to? I can't seem to find it.

[rekram1-node]

Idk if we do tbh can't find one