run summary8 passed
Run summary
The command that ran, its status and exit code, timing, and the path to the saved report.
commandpytest -q
statuscompleted · exit 0
duration4.2s
reportreports/run-20260519.html
local-firstgit-awareno hosted dashboard
Proof trails for AI-generated code
Ship AI-generated code with proof, not guesswork.
ProofPatch records the commands, Git changes, failures, policy blocks, reports, and rollback context around AI coding-agent work so reviewers can trust what changed before merging.
Local-first. Git-aware. No hosted dashboard required. Built for Claude Code, Codex, Cursor, and local agent workflows.
ProofPatch terminal
$ proofpatch run --project . --label "demo" -- pytest -q8 passed
HTML report: reports/run-20260519.html
Rollback plan: ready
$ proofpatch workflow run release-workflow.yml --dry-runMode: dry-runRisk: lowNext: review plan, then run approved checks
$ proofpatch workflow report latest --htmlHTML report: workflows/reports/latest.html
The problem
A patch is not proof. The trail around it is.
AI coding agents can generate large, plausible patches in seconds. But the reviewer still has to answer the hard questions: what actually ran, what changed on disk, what failed, what was blocked, and how to recover if it is wrong. Without that evidence, “looks good” is the only review you can give.
ProofPatch captures that evidence as the work happens, so trust comes from a record instead of a hunch.
What ran
The exact commands and workflow steps around the change.
What changed
Git state before and after, file by file.
What failed
Exit codes, stderr, and timed-out runs.
What was blocked
Policy decisions that stopped unsafe commands.
How to recover
A rollback plan for changed and created files.
What to do next
Review prompts pointing at the safest follow-up.
Who it is for
Built for the people on the hook for AI-generated code.
Solo developers
The pain: You let an agent write a big patch, then have to re-read everything to trust it.
With ProofPatch: Get a local trail of what ran, what changed, and what failed before you commit.
Agencies
The pain: Client work shipped with AI assistance needs a paper trail you can stand behind.
With ProofPatch: Hand over reports and rollback context that show the work was reviewed, not guessed.
Review teams
The pain: Plausible AI diffs hide what was executed, blocked, or skipped during the change.
With ProofPatch: Review evidence and a rollback plan alongside the diff, before approving the merge.
How it works
Three steps from agent output to reviewable proof.
01
Wrap the work
Run your command or agent workflow through ProofPatch instead of straight from the terminal.
02
Capture the evidence
ProofPatch records commands, Git state, outputs, failures, policy results, and workflow context as it runs.
03
Review before shipping
Read the generated report and rollback plan, then merge with proof instead of guesswork.
ProofPatch does not replace your coding agent. It wraps the workflow with evidence.
Features
Everything a reviewer needs, recorded locally.
ProofPatch gives developers a local evidence layer for AI-assisted coding: commands, Git changes, failures, policy blocks, reports, comparisons, and rollback context.
Command trail
Every command, label, exit code, and timestamp from the session in one ordered record.
Git evidence
Before-and-after Git state so reviewers see exactly which files the patch touched.
Failure capture
Non-zero exits, stderr, and timed-out runs are kept, not lost in scrollback.
Policy blocks
Risky and disallowed commands are stopped and logged with the reason they were blocked.
Rollback context
A local recovery plan for changed, created, and untracked files if the patch is wrong.
Workflow reports
Markdown, JSON, and HTML artifacts for each run, diff, and review — no hosted service.
Run comparison
Diff two saved runs to see how command, output, status, and Git changes shifted.
Local-only storage
Runs, history, and reports stay in your repo. Nothing is shipped to a remote dashboard.
What you get after a run
Reviewable proof, generated locally — no broken screenshots or hidden state.
Sample output from ProofPatch local reports: a run summary, Git diff evidence, and a rollback plan with policy results. Everything is written to your repo as Markdown, JSON, and HTML.
git evidenceReview
Git diff evidence
Before/after Git state with changed, created, and untracked files flagged for review.
changedapp/auth/session.ts
createdtests/test_session.py
untracked1 file
rollbackplan ready
policy + rollbackBlocked
Rollback plan
Blocked commands with their policy reason, redacted output, and steps to recover local state.
blockedcurl install.sh | sh
reasonnetwork command blocked by policy
output[REDACTED] token hidden
recoverrestore 2 files · drop 1
Quickstart
Try it in a repo you can inspect.
Install locally, run a test command, then inspect the captured run and workflow report. proofpatch is the current supported CLI command.
ProofPatch does not replace your coding agent. It wraps the workflow with evidence.
terminal
git clone https://github.com/5yffb7vk7b-source/agentblackbox.git proofpatch
cd proofpatch
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
proofpatch version
proofpatch run --project . --label demo -- pytest -q
proofpatch latest
proofpatch workflow list
proofpatch workflow run workflows/tasks/release-workflow.yml --dry-run
proofpatch workflow report latest --htmlShip AI-generated code with proof, not guesswork.
Record the commands, Git changes, failures, policy blocks, reports, and rollback context that make a patch reviewable — before it merges.