Agent Blackbox: Recording Commands Without Keeping Raw Output by Default

When an agent claims success, a reviewer needs the command and its result. Agent Blackbox captures a process while it runs, then turns the output into a local diagnostic record. The current privacy default is materially different from the original article: both persisted output and console output are sanitized unless raw capture is explicitly requested.

Install from a reviewed checkout

git clone https://github.com/godhiraj-code/agent-blackbox.git
cd agent-blackbox
python -m pip install .
agent-blackbox --version
agent-blackbox demo

Record a small command first

agent-blackbox record --name "smoke" --out-dir runs -- python -c "print('recording proof')"
agent-blackbox run-doctor --runs-dir runs --out exports/latest-run-postmortem.md

run-doctor selects the newest matching recording. For a specific incident, use diagnose-run with the exact recording path. Avoid assuming that “newest” identifies the task you intended when several processes write to the same directory.

The privacy boundary

Raw source transcripts can still contain private data. Redaction is defense in depth, not a proof of anonymity. Review exports before sharing and keep raw captures out of Git and shared CI artifacts.

Capture is not deterministic replay

A subprocess recorder sees what the target process emits. It cannot reconstruct hidden internal actions, prove that a command was safe, or turn a nondeterministic agent into a deterministic system. Buffering and scheduling also affect stream order; the earlier sub-millisecond fidelity claim was not justified.

Use diagnostic exits deliberately

Analysis distinguishes healthy results, diagnosed incidents, and invalid or unsupported evidence. Invalid evidence must not become a clean run. Use the documented CI modes and nonzero exit codes for gating; use report-only overrides only when that is an intentional integration choice.

Turn evidence into the next check

Compare the final claim with the command output and the actual changed state. If output shows that verification was skipped, add the missing check to the harness. A recorder helps identify the gap; it does not close it by itself.

Current project documentation