Reading traces

trace.sh reads a cortex trace. Six views, each one jq filter.

./trace.sh <view> <trace.jsonl> [arg]

Views

timeline

One line per thing that happened, with just enough payload to see the shape of a run. Streaming token noise is dropped — it's the same text the assistant_end line carries in full. Great for getting the overall shape of a run at a glance.

./trace.sh timeline runs/report-final/trace.jsonl

prompt

Every message of one model call, exactly as the provider saw it. Defaults to the last main-model request; pass a call number for a specific one, or pruner for the curator's last.

./trace.sh prompt runs/report-final/trace.jsonl          # last main call
./trace.sh prompt runs/report-final/trace.jsonl 3        # call #3
./trace.sh prompt runs/report-final/trace.jsonl pruner   # the pruner's last call

system

The system message alone — the prompt under test, plus whatever axon appends (including the tool catalog). This is the view for prompt engineering: it shows the composed system prompt the model actually received.

./trace.sh system runs/report-final/trace.jsonl

tools

What the agent actually did, in order, with full arguments and a clipped result. This is the view that shows a wrong approach fastest.

./trace.sh tools runs/report-final/trace.jsonl

usage

Token spend per call and in total. Reported by the provider, not estimated.

./trace.sh usage runs/report-final/trace.jsonl

errors

Every error, tool error and wire error, with timestamps.

./trace.sh errors runs/report-final/trace.jsonl