Proofcast

A bodycam for AI

Proofcast is an AI bodycam that lets you record the terminal output that it sees, uncovering all the gruesome details...

what do you mean it was only staging!?

---

Install from a marketplace

Claude Code:

claude plugin marketplace add killallgit/killall-proof
claude plugin install proofcast@killall-proof

To share the Claude Code setup with one project, run from that project:

claude plugin marketplace add killallgit/killall-proof --scope project
claude plugin install proofcast@killall-proof --scope project

Codex:

codex plugin marketplace add killallgit/killall-proof
codex plugin add proofcast@killall-proof

For a local checkout, pass its path to plugin marketplace add, then install the plugin by its proofcast@killall-proof identifier. claude plugin install does not accept a repository path.

Install as an individual skill

npx skills@latest add killallgit/killall-proof \
  --skill proofcast \
  --agent claude-code \
  --global \
  --yes

Pre-reqs

asciinema 3.0+	required
agg	optional — adds a gif alongside each recording

After editing skills/proofcast/, run python3 scripts/sync-plugin.py and commit the updated plugins/proofcast/ copy used by the marketplaces.

How to use

Tell your agent "prove it" about code or command behavior to trigger Proofcast. You can also ask to record terminal output or "turn on the bodycam". Recordings land in .proofcast next to wherever it ran:

.proofcast/index.html
.proofcast/<slug>/<timestamp>.cast
.proofcast/<slug>/<timestamp>.stdout.log
.proofcast/<slug>/<timestamp>.meta.json
.proofcast/<slug>/<timestamp>.gif

Open index.html from disk. It lists every recording and replays them as real text you can pause, scrub, and copy from. After recording, the agent asks "Would you like me to open the recording webpage in your browser?" and opens it if you agree (or already asked it to).

The page opens the newest capture without autoplay, with history underneath. Each capture includes an Assertion and Justification (one or two short sentences total), the actual shell-quoted command, and its exit status. The context describes the intended check; neither the description nor exit status 0 is a verified verdict. Older recordings still play without context.

Playback has a timeline and one compact control row: Back 5, play/pause, Fwd 5, and a speed button that cycles through 0.5x / 1x / 1.5x / 2x. Controls remain touch-friendly on mobile. These control the terminal recording, not the optional GIF. Auto-play next is on by default: after playback finishes, it starts the next recording down the displayed list and stops at the end. The first recording still waits for Play. The text log link sits opposite the auto-play toggle.

Recordings are disposable. Save downloads the selected capture, context, and text log as one offline HTML file; your browser chooses the save location. The GIF stays a separate optional download. Confirm the export was saved before removing the original. Nuke proof covers all recordings listed in the index, asks for folder access and confirmation, and deletes only their files and the index. It leaves unrelated files, exports, empty folders, and anything the recorded commands created alone. Browsers without folder-write access (including most mobile browsers) show exact cleanup commands instead; showing commands does not delete anything.

The proof must show the commands that actually run. The agent records the underlying commands directly, each as its own recording under the same slug, so every $ line is a real command and the output stays readable; this includes the steps of existing project scripts. Tracing is the fallback for a script that must run as a whole: the agent inspects it and adds execution tracing while preserving the original interpreter path and startup options, including shebang flags. For #!/path/to/bash -e, use /path/to/bash -ex ./script.sh so tracing keeps the original stop-on-failure behavior. Child scripts need their own tracing. A wrapper that only prints selected commands or success messages hides the work; its printed claims are not proof of execution.

Accountability, as ever, has limits

Output that scrolls faster than the terminal can draw collapses into a single frame, so the footage shows the aftermath rather than the incident. The full text survives in the log, which nobody watches.

Everything the agent printed is captured verbatim, including your secrets. Execution traces also expose expanded arguments and assignments. Proofcast does not redact them. Add .proofcast/ to .gitignore.

Recordings are kept until someone deletes them, and someone always does.

Evaluating changes

Run python3 -m unittest discover -s tests for recorder/index regressions. evals/evals.json contains planning scenarios for comparing agent behavior: traced nested scripts, incomplete evidence despite exit 0, and failure followed by export/cleanup. These check instruction following, not actual infrastructure. With Playwright and its Chromium browser available, node tests/browser.cjs checks offline playback, export, mobile layouts, and cleanup using fake folder handles. Browser tooling is development-only; the shipped page needs none.