Tools

Call counts, error rates, latency, and exactly how many tokens each tool pushed into context.

reads as
Read 71

What it means: Which tools actually dominate this session's activity, by call count.
What to do: A tool called far more than the task seems to need can indicate retrying or inefficient exploration.
Measured: count · from the transcript · exact · how

reads as
ERR 3

What it means: Errors are billed and re-injected into context exactly like successes are.
What to do: The same tool failing with the same input repeatedly is the signal to interrupt and hand the model the fix directly instead of letting it keep retrying.
Advisor: A38 Failure cascade · A47 Turn died
Measured: count · from the transcript · exact · how

reads as
p50 1.4s

What it means: The typical duration for this tool — the number to expect on a normal call.
What to do: Read alongside p95 rather than alone; a high median usually means the tool itself is just slow.
Measured: ms · from the transcript · hook events (after cctop install) · ≈ until hook timings replace it · how

reads as
p95 9.8s

What it means: The tail latency — what occasionally makes a turn feel much slower than usual.
What to do: A high p95 on a tool that also runs in the foreground (long Bash calls, builds, test suites) is a candidate to run in the background instead.
Advisor: A10 Long foreground commands
Measured: ms · from the transcript · hook events (after cctop install) · ≈ until hook timings replace it · how

reads as
LAST 4:02

What it means: How long since this tool was last used — staleness, not activity.
What to do: A tool or MCP server that hasn't been called in a while is still paying for its definitions on every turn; pairs with the MCP call count in Agents & MCP.
Advisor: A07 Idle MCP servers and plugins
Measured: duration · from the transcript · exact · how

reads as
TOKENS→CTX 41.2k

What it means: Exactly how much context each tool's results added — the direct, measurable cause of context growth.
What to do: The biggest single contributor here is almost always the right fix target: narrow the command, delegate to a subagent, or stop re-reading it.
Advisor: A03 Runaway tool results · A04 Re-reads · A25 Exploring in the main context
Measured: tokens · from the transcript · OpenTelemetry · ≈ unless OpenTelemetry is on · how

What it means: Characters the model wrote as tool inputs / 4, per tool — they stay in context like results do
Measured: tokens · from the transcript · always an estimate · how

What it means: Bash calls by the phase classifier's class: explore / implement / test / build-lint / commit / gitread / ops / wait (Bash·test rows)
Measured: count · from the transcript · exact · how

What it means: Failed calls by Claude Code's own taxonomy (Command Failed / User Rejected / Edit Failed / File Changed / File Too Large / File Not Found / Other) plus Content Not Found, Timeout, Tool Not Found and Denied (toolDenialKind)
Measured: count · from the transcript · exact · how

reads as
top ctx: Read src/render.rs 6.1k

What it means: The single largest results by context cost, named individually rather than summed by tool.
What to do: Start with the top entry, not the tool average — one oversized result usually explains most of the growth.
Advisor: A03 Runaway tool results
Measured: tokens · from the transcript · ≈ unless OpenTelemetry is on · how

What it means: API calls since the result landed × its tokens × the cache-read price: what re-reading it has cost so far
Measured: USD · from the transcript · the price table · always an estimate · how

What it means: Deferred tools loaded through ToolSearch per MCP server (matches of its result); each load rewrites the cached prefix
Measured: count · from the transcript · exact · how