telemetry: populate context_tokens field in observability JSONL (currently always 0) #656
Description
Data Blind Spot: context_tokens field is always 0
Requested by: data squad (Goal #4 — new data capture proposals)
Current State
All 57 records in executions.jsonl have "context_tokens": 0. The field exists but is never populated.
{"input_tokens": 26, "output_tokens": 5950, "cache_read_tokens": 583096, "context_tokens": 0}Blind Spot
We cannot answer:
- How much context is each agent consuming per run? (context bloat = slower, more expensive)
- Which agents are hitting context limits and why?
- Is context growing over time as memory files accumulate?
- What's the ratio of context_tokens to output_tokens? (efficiency metric)
Context token cost is real — cache reads are cheaper but still billed. A single large state.md or CLAUDE.md can dominate the context budget.
Proposed Capture
Populate context_tokens with the total tokens loaded into Claude's context before any user message:
- CLAUDE.md files loaded
- Agent definition file
- State/memory files loaded
- Skill files loaded
This is distinct from input_tokens (the user prompt) and cache_read_tokens (already tracked).
Why This Matters
The economics squad tracks cost per useful output. Without context_tokens, we can't compute the full cost of a run or identify context-bloated agents. This is low-effort instrumentation since the context size is known at invocation time.
source:data