Source: https://reclaimllm.com/news/session-statistics-via-background-task
← Back to News update Mar 18, 2026Reliable Session Statistics via Background Task
We’ve introduced a server-side background task to compute consistent token counts, tool usage, and code impact metrics for every session across all capture sources.Overview
RCLM now automatically calculates detailed statistics for every captured LLM session, regardless of whether you’re using the browser extension, CLI, or automated hooks. Previously, session metadata like token counts and tool usage were only available if the source client explicitly provided them. This often led to incomplete dashboards and unreliable usage analytics. Our new background task ensures that every session is analyzed for its full impact.How it Works
The core of this update is a new background service that identifies sessions with missing statistics and processes them systematically:- Server-Side Analysis: Instead of relying on client-side logic, the RCLM server now reads the full session data and computes statistics directly. This ensures consistency across all platforms.
- Accurate Token Counting: We’ve integrated the
tiktokenlibrary using thecl100k_baseencoding. This provides highly accurate token counts for GPT-4 class models and serves as a reliable industry-standard approximation for other providers like Claude and Gemini. - Expanded Metrics: Beyond simple token counts, we now track several new data points for every session:
- Message and Turn Counts: Understand the depth of each interaction.
- Tool Distribution: A detailed breakdown of which tools (like
grep_search,read_file, or custom tools) were used most frequently. - Code Impact: Automated tracking of
lines_added,lines_removed, and the programming languages involved in each session.
- Automated Cleanup: To keep your workspace clean, any sessions that are found to have zero input and output tokens after analysis are automatically deleted.
Why We Made This Change
As RCLM has grown to support more ways of capturing interactions, maintaining consistent data became a challenge. Client-side computation is prone to inconsistency and adds overhead to the ingestion process. By moving this logic to a background task, we keep the initial session capture fast while ensuring that enterprise-grade analytics are eventually consistent and accurate. This approach also allows us to backfill statistics for historical sessions that were captured before these metrics were introduced.Current Limitations and Feedback
While this significantly improves data reliability, there are a few considerations:- Processing Delay: Statistics are calculated in the background, so there may be a short delay (typically a few minutes) before they appear in your dashboard after a session is first captured.
- Tokenization Variance: While
cl100k_baseis an excellent general-purpose tokenizer, there may be a 5-15% variance compared to the exact counts reported by non-OpenAI providers like Anthropic. - Cost Derivation: At this stage, we are focusing on providing accurate raw primitives (tokens, lines, tools). We haven’t yet introduced direct cost estimation in USD, as provider pricing changes frequently.