Context window of a run

The other tracks score what an agent produced. context trace reads what the agent held while producing it: how big the prompt was at every model call, what start-up loaded before the agent did anything, how fast the history grew, whether the harness compacted it, and which tool results took the room. It reads a trace file and calls no model.

claude -p "<task>" --output-format stream-json --verbose > run.jsonl
python cli.py context trace run.jsonl --window 200000 --calls

| trace | calls | first | peak | fill | growth | compact | tools | failed | repeat | result chars |
|---|---|---|---|---|---|---|---|---|---|---|
| run.jsonl | 2 | 21,390 | 21,673 | 11% | 283 | 0 | 1 | 0 | 0 | 165 |
run.jsonl: largest tool results (chars): Bash 165
run.jsonl: prompt tokens per call: 21390 21673

What the columns mean

• calls: model calls on the main thread. Claude Code writes one assistant message per content block, and the blocks of one call share a message id, so a call is counted once. Calls a subagent made run in their own window and are counted apart (subagent_calls in --json).
• first: the prompt of the first call, input + cache_creation + cache_read tokens. It is everything start-up loaded (system prompt, tool schemas, memory and guidance files, the skill listing) plus the task.
• peak, growth: the largest prompt, and how far it rose above first: what the history added. fill is peak over --window, shown only when you pass the window, because the trace does not say it.
• compact: compact_boundary events. After one the prompt can drop; the --calls series shows by how much, and --json --calls gives each event's trigger, pre_tokens and (on newer CLIs) post_tokens as compaction_events.
• tools, failed, repeat: tool calls, calls whose result came back as an error, and calls identical in name and input to an earlier call on the same thread that did not fail, with no Bash, Edit, MultiEdit, NotebookEdit or Write call in between. A retry after an error and a Read after an Edit are not repeats; what is left is an agent fetching again what it already had.
• result chars: the size of every tool result, in characters, since the trace gives no token count for them. The line under the table names the tools whose results weighed most.

--json prints a list with one object per trace, also for a single trace. A trace that cannot be read is reported on stderr and left out; the others still print, and the command exits 1.

Comparing context layers

Several traces print as one table, so the layers of a context compare side by side. Run the same task in each condition and pass the traces together:

Condition	What differs	Read
Free	an empty workspace	first is the harness's own baseline
Loaded	+ AGENTS.md / CLAUDE.md, skills, MCP servers	first minus the free run's is what the guidance costs on every call
Execution, without skill	the task, guidance loaded	peak, growth, calls
Execution, with skill	the same task, skill installed	the difference is what the skill cost or saved over the run

Two things make a baseline wrong without looking wrong. Claude Code loads every CLAUDE.md from the working directory up to /, so run the free condition in a directory with no guidance file above it (the system temp dir usually qualifies). And a user-level ~/.claude/CLAUDE.md loads into every run; point CLAUDE_CONFIG_DIR at an empty directory to leave it out. The surface field of --json (the tool, skill and agent counts, and the names of the MCP servers and loaded plugins) shows what start-up offered, so two runs that differ there are not the same condition.