AGENT PLAN
What the planner proposed
…
RECORDED MODEL TRACE · PROVIDER KEYS NOT PRESENT
Overview Agent I/O
See what the agent is told — and what it answers
This public page shows a committed, redacted run of the planner and verifier models. The trace records the configured native OpenAI or Anthropic client and model IDs for that run. You get system + user prompts, tool-catalog context, JSON outputs, then the deterministic policy gate (not the model) deciding authorize vs deny.
No GitHub / Slack / PagerDuty / Jira / DataHub mutations in this trace. API keys never appear here.
FLOW
How the agent acts (one turn)
Sense is tools/catalog. Plan and critique are LLMs. Gate is ordinary Python. Tools stay held in this trace so you can inspect I/O safely.
LLM INPUT / OUTPUT
Every model call: prompt in → JSON out
Expand a card to read the system prompt, user instructions, and the model’s structured reply. Context is summarized for readability; full tool catalog remains visible when present.
AFTER THE MODELS
Sealed plan + policy decision
POLICY GATE
What deterministic code decided
…
REGENERATE LOCALLY
Test-run with OpenAI or Anthropic
# OpenAI (default) export OPENAI_API_KEY=… export LEDGERLENS_LLM_PROVIDER=openai export LEDGERLENS_PLANNER_MODEL=gpt-4o export LEDGERLENS_VERIFIER_MODELS=gpt-4o-mini,gpt-4-turbo # Or Anthropic # export ANTHROPIC_API_KEY=… # export LEDGERLENS_LLM_PROVIDER=anthropic # export LEDGERLENS_PLANNER_MODEL=claude-sonnet-4-20250514 # export LEDGERLENS_VERIFIER_MODELS=claude-3-5-haiku-latest,claude-sonnet-4-20250514 uv run python scripts/run_agent_io_trace.py --force # writes src/ledgerlens/static/agent-io-trace.json # then refresh this page
Script: scripts/run_agent_io_trace.py ·
Recording client: RecordingJsonClient in
model_runtime.py
MORE