Verify Vivi

Verify Vivi

Drive the installed vivi executable as a user would. Do not call backend functions directly as proof of CLI or TUI behavior.

Launch

Build the executable once from the repository root:

zig build

Vivi is a short-lived CLI/TUI, not a server. Start every interactive drive in its own VHS or script PTY:

.github/skills/verify-vivi/bin/verify-vivi chat-streaming <run-id>

The TUI is ready when the captured frames show the vivi welcome, Ready context, and the bottom composer with its colored left accent. The helper submits a deterministic prompt, waits for the streamed answer, and sends Ctrl-C to shut down. VHS is only the frame-capture mechanism; its GIF is deleted after extraction by default. VHS and script own their child processes, so no persistent instance remains after the helper returns.

If a manual drive is required, run ./zig-out/bin/vivi chat in a fresh terminal. Press Ctrl-C once for cooperative shutdown. Press Ctrl-C a second time only when the UI is still showing Stopping....

Doctor

Run this read-only health check before any drive when the environment or authentication looks questionable:

.github/skills/verify-vivi/bin/verify-vivi doctor

It requires Zig, VHS, script, Expect, FFmpeg/FFprobe, and GitHub Copilot CLI; builds Vivi; checks the installed Vivi and Copilot versions; and sends a no-tools Copilot prompt that must return VIVI_DOCTOR_OK. A failure means the instance is not worth driving. Copilot CLI does not expose a standalone authenticated-status command, so the minimal prompt is the authentication check.

Drive

Use the executable helper:

.github/skills/verify-vivi/bin/verify-vivi cli-discovery <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-streaming <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-streaming-input <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-async-bash <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-markdown <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-markdown-read <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-ask-user <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-model-menu <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-customization <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-session-resume <run-id>
.github/skills/verify-vivi/bin/verify-vivi chat-shutdown <run-id>
.github/skills/verify-vivi/bin/verify-vivi frame-check <run-id> chat-markdown

<run-id> must contain only letters, digits, dots, underscores, or hyphens. Each command writes to .verify/vivi/<run-id>/, so concurrent runs do not share PTYs or evidence. Multiple Vivi processes may run side by side because each owns its own Copilot child and session, but they share the user's Copilot credential store. Never automate two instances against the same terminal.

The stable user handles are:

• vivi --help and vivi --version for command discovery.
• The context state Ready for an idle chat.
• The context state Responding... while a turn streams.
• Transcript labels You and Vivi.
• The message Stopping... press Ctrl-C again to force exit. after the first Ctrl-C.

Read the feature map before choosing a drive:

cat .github/skills/verify-vivi/features/README.md

Evidence

Evidence lives under .verify/vivi/<run-id>/ and is intentionally ignored by Git. Keep it after cleanup.

For TUI proof, capture:

• *.contact-sheet.png: two-second samples across the full captured session.
• *.frames/frame-*.png: one frame per second for detailed inspection.
• *.frames.txt: capture dimensions, duration, and extracted frame count.
• *.visual-review.md: explicit appearance and state-transition checklist.
• *.terminal.log: raw PTY transcript for machine-readable assertions.
• *.normalized.txt: the same PTY output with terminal control sequences removed for stable text assertions.
• *.assertions.txt: the exact expected marker and command exit status.
• *.tape: the generated VHS input used for the run.

The GIF is not proof. It is an optional presentation artifact for demos or PR descriptions. Set VIVI_KEEP_GIF=1 when running a TUI recipe to retain it:

VIVI_KEEP_GIF=1 \
  .github/skills/verify-vivi/bin/verify-vivi chat-streaming <run-id>

For noninteractive CLI proof, capture stdout, stderr, and exit status in separate files.

A valid proof exercises ./zig-out/bin/vivi, not a backend test seam. Capture the action and the resulting state, not only the final frame. For streaming, the expected answer marker must not occur in the submitted prompt; the helper asks Copilot to transform vivi_stream_ok and asserts the distinct VIVI_STREAM_OK response.

After every TUI drive, open the contact sheet and enough individual frames to check each line in *.visual-review.md. Record the frame numbers and concrete observations, replace every [ ] with [x] for a pass or [!] for a failure, and set status: pass or status: fail. Then run:

.github/skills/verify-vivi/bin/verify-vivi frame-check <run-id> <capture>

The authoritative visual proof is the extracted frame set plus a completed review accepted by frame-check. Normalized terminal text separately proves the expected content. Appearance review must cover readability, borders, overlap, stale content, redraw corruption, and the expected state sequence.

The Copilot service is the production boundary for this feature. Do not replace it with a mock when claiming end-to-end chat verification. Credential-free unit tests remain useful but prove a different layer.

Cleanup

The helper sends the application's documented shutdown keys and waits for the PTY owner it started. Its exit trap kills only that recorded PID if a failed run leaves it alive. It never kills by process name.

Run cleanup explicitly after an interrupted attempt:

.github/skills/verify-vivi/bin/verify-vivi cleanup <run-id>

Cleanup removes only transient PID files and empty generated tapes from an unfinished run. It preserves extracted frames, contact sheets, visual reviews, terminal transcripts, normalized text, assertions, stdout, stderr, and any GIF explicitly retained with VIVI_KEEP_GIF=1. After cleanup, confirm the evidence directory still exists:

test -d ".verify/vivi/<run-id>"

Helpers

.github/skills/verify-vivi/bin/verify-vivi is the only helper. It is executable and accepts:

doctor
cli-discovery <run-id>
chat-streaming <run-id>
chat-streaming-input <run-id>
chat-markdown <run-id>
chat-markdown-read <run-id>
chat-ask-user <run-id>
chat-model-menu <run-id>
chat-customization <run-id>
chat-session-resume <run-id>
chat-shutdown <run-id>
extract-frames <run-id> <chat-streaming|chat-streaming-input|chat-async-bash|chat-markdown|chat-markdown-read|chat-ask-user|chat-model-menu|chat-customization|chat-session-resume|chat-shutdown>
frame-check <run-id> <chat-streaming|chat-streaming-input|chat-async-bash|chat-markdown|chat-markdown-read|chat-ask-user|chat-model-menu|chat-customization|chat-session-resume|chat-shutdown>
cleanup <run-id>

Run it from anywhere inside the repository. It resolves the repository root with Git, builds the existing Zig target, and never installs dependencies.