Summary: Token Usage Analysis in OpenHands I analyzed the OpenHands codebase to answer your question about whether

prompt_tokens

includes cached tokens in the trajectory's

accumulated_token_usage

. Answer:

prompt_tokens

does NOT include cached tokens. Key Findings: 1. Token Separation: OpenHands tracks cached and non-cached tokens separately: ◦ `prompt_tokens`: Contains only non-cached input tokens ◦ `cache_read_tokens`: Contains tokens retrieved from cache (cache hits) ◦ `cache_write_tokens`: Contains tokens written to cache (for providers like Anthropic) 2. Exact Code Location: The token metrics are extracted and computed in: ◦ File:

/workspace/project/OpenHands/openhands/llm/llm.py

◦ Lines: 617-666 ◦ Method:

add_token_usage()

in the

Metrics

class 3. Implementation Details: ◦ Line 617:

prompt_tokens = usage.get('prompt_tokens', 0)

extracts non-cached tokens ◦ Lines 632-639:

cache_hit_tokens

extracted from

prompt_tokens_details.cached_tokens

◦ Line 647:

cache_write_tokens

extracted from

model_extra.get('cache_creation_input_tokens', 0)

◦ Lines 659-666: These are passed as separate parameters to the metrics system 4. LiteLLM Compliance: This implementation follows the LiteLLM documentation standard where: ◦

prompt_tokens

represents non-cached prompt tokens only ◦

cached_tokens

in

prompt_tokens_details

represents cache-hit tokens ◦ Cache creation tokens are tracked separately for accurate cost calculation This separation allows OpenHands to accurately track and calculate costs since cached tokens often have different pricing than non-cached tokens across different LLM providers. _{View full conversation}

Final Summary I have thoroughly analyzed the OpenHands codebase to answer your questions about token usage metrics in trajectories and the web GUI. Answer to Your Questions: 1. Does

prompt_tokens

include cached tokens? No,

prompt_tokens

does NOT include cached tokens. They are tracked separately. 2. Token Usage in Web GUI The web GUI receives and displays the same token structure, maintaining the separation between cached and non-cached tokens. Complete Token Usage Flow: Backend Processing (

openhands/llm/llm.py

, lines 617-666): 1.

prompt_tokens

=

usage.get('prompt_tokens', 0)

- non-cached tokens only 2.

cache_read_tokens

=

prompt_tokens_details.cached_tokens

- cache hits 3.

cache_write_tokens

=

model_extra.get('cache_creation_input_tokens', 0)

- cache writes Data Transmission: • Backend serializes metrics via

Metrics.get()

method (

openhands/llm/metrics.py

, line 203) • Sent to frontend through

llm_metrics.accumulated_token_usage

in action messages • Event serialization handled in

openhands/events/serialization/event.py

, line 120 Frontend Display (

frontend/src/components/features/conversation/metrics-modal/usage-section.tsx

): • Input tokens:

usage.prompt_tokens

(non-cached only) • Cache hit:

usage.cache_read_tokens

(displayed separately) • Cache write:

usage.cache_write_tokens

(displayed separately) • Output tokens:

usage.completion_tokens

• Total:

usage.prompt_tokens + usage.completion_tokens

(excludes cached tokens from total) Key Finding: Both in trajectories and the web GUI,

prompt_tokens

represents only non-cached input tokens, while cached tokens are tracked separately as

cache_read_tokens

and

cache_write_tokens

. This separation allows for accurate cost calculation since cached tokens often have different pricing than non-cached tokens. The web GUI maintains this same structure and displays cached tokens as separate line items under the input tokens section, ensuring transparency about cache usage. _{View full conversation}