qprompt-cli 0.1.0

Flight recorder for LLM apps

qprompt-cli

Flight recorder for LLM apps: capture how a question was parsed, what tools were called, what evidence was used, and what risks remain.

What this tracks

• question, model, timestamp, trace_id
• parser output (intent, entities, assumptions)
• request envelope (messages, tools available)
• tool calls (name, redacted input, output summary, status/error)
• response metrics (latency, token estimates)
• evidence and answer audit risk flags

What this does not track

• Hidden internal chain-of-thought for hosted closed models
• Neuron-level/attention internals from provider backends

Install (local)

python -m pip install -e .

Python import:

from qprompt import Tracer

CLI

qprompt run "why did revenue drop in March?"
qprompt list
qprompt show <trace_id_or_path>
qprompt diff <trace_a> <trace_b>

Traces are saved under:

.traces/YYYY-MM-DD/trace_<uuid>.json

Schema

• JSON schema file: src/llmtrace/trace_schema.json
• Runtime trace builder/validator: src/llmtrace/schema.py

Example output (run)

• Parsed intent and entities
• Tools used and summaries
• Evidence list
• Risk flags

Next integration step

Replace the mock callable in Tracer.chat(...) with your real model provider call and preserve returned usage token fields.