Agent-Skills.md

Agent Skills: tmux-verify

The outer "is this TUI feature ACTUALLY done?" proof loop for the ainb terminal app. Use when the user asks to "verify the TUI is actually done", "prove this TUI feature works", "validate per tmux-verify", "vhs verify each journey", "frame-truth check the UI", "is the diff actually shown correctly", or to define the acceptance contract for a TUI goal. Orchestrates 6 gates — (1) comprehensive tmux-ui-tripwire passes, (2) a vhs recording per user journey whose frames are READ with /media-processing to assert the EXACT user-visible outcome (not "screen shows something"), (3) one recording per journey, (4) fix-loop until acceptance + tripwire green, (5) a continuous /explain-to-me explainer (tests + commits + GIFs) hosted on /here-now, (6) final HTML-solid validation with the human as the last gate. References (does not reimplement) tmux-ui-tripwire, /media-processing, /explain-to-me, /here-now.

UncategorizedID: stevengonsalvez/agents-in-a-box/tmux-verify

Repository

stevengonsalvez/agents-in-a-box
stevengonsalvezLicense: MIT
186

Install this agent skill to your local

pnpm dlx add-skill https://github.com/stevengonsalvez/agents-in-a-box/tree/HEAD/.claude/skills/tmux-verify

Skill Files

Browse the full folder contents for tmux-verify.

Download Skill

Loading file tree…

.claude/skills/tmux-verify/SKILL.md

Skill Metadata

Name
tmux-verify
Description
The outer "is this TUI feature ACTUALLY done?" proof loop for the ainb terminal app. Use when the user asks to "verify the TUI is actually done", "prove this TUI feature works", "validate per tmux-verify", "vhs verify each journey", "frame-truth check the UI", "is the diff actually shown correctly", or to define the acceptance contract for a TUI goal. Orchestrates 6 gates — (1) comprehensive tmux-ui-tripwire passes, (2) a vhs recording per user journey whose frames are READ with /media-processing to assert the EXACT user-visible outcome (not "screen shows something"), (3) one recording per journey, (4) fix-loop until acceptance + tripwire green, (5) a continuous /explain-to-me explainer (tests + commits + GIFs) hosted on /here-now, (6) final HTML-solid validation with the human as the last gate. References (does not reimplement) tmux-ui-tripwire, /media-processing, /explain-to-me, /here-now.

tmux-verify

Overview

tmux-verify is the OUTER acceptance harness: it proves a TUI feature is actually done as a user would experience it. It does NOT write the automated test (that is tmux-ui-tripwire) and does NOT reimplement media tooling (that is /media-processing). It orchestrates both plus a live explainer into one repeatable contract.

Use it two ways:

  • As a contract — a goal/plan says "validate per /tmux-verify" and inherits all 6 gates.
  • As a runner — execute the gates for a feature end-to-end until green.

The non-negotiable principle: read the frame, don't blank-check. A passing screen is one whose pixels/text say the correct thing for that journey — not merely that something rendered.

The 6 gates

                         ┌──────────────────────────────────────────────┐
 G1 automated  ───────▶  │ tmux-ui-tripwire: comprehensive, VT100-true,  │
 (ref skill)             │ passes in the real TUI, non-flaky ×3          │
                         └──────────────────────────────────────────────┘
 G2+G3 visual            ┌──────────────────────────────────────────────┐
 (THIS SKILL OWNS) ───▶  │ per journey:  .tape ─▶ vhs ─▶ gif+mp4         │
                         │              ─▶ ffmpeg frames ─▶ READ + assert│
                         │ one recording per journey                     │
                         └──────────────────────────────────────────────┘
 G4 fix-loop   ───────▶  acceptance(cargo test) + tripwire green; unbounded
                         until green or a genuine hard blocker (then log + go on)
 G5 explainer  ───────▶  /explain-to-me ─▶ /here-now  (stable slug, republish
                         each phase: progress · test/clippy/tripwire output ·
                         commit log · embedded per-journey GIFs)
 G6 final      ───────▶  HTML solid (fetch URL: links resolve, GIFs load,
                         no empty sections) ─▶ HUMAN validates → done

Gate-by-gate playbook

G1 — Automated gate (defer to tmux-ui-tripwire)

Write/repair the feature's crates/ainb-core/tests/tripwire_*.rs per tmux-ui-tripwire. Bar to clear:

  • Comprehensive: asserts VT100 truth (rendered cells, colors, bg attrs, exact counter text) — NOT substring-OR on chrome strings that pass while the feature is broken.
  • Green + non-flaky: cargo test -p ainb-core --test tripwire_<feature> passes 3 runs in a row.
  • Respect the documented traps: AMFI SIGKILL (exit 137), first-run-wizard key-swallow, EnvFilter crate-name drift, weak assertions. See that skill's references/.
  • tmux safety: only ever tmux kill-session -t <exact-name>. Never kill-server, pkill, killall, or any wildcard.

G2 + G3 — Visual gate, one recording per journey (this skill owns it)

For EACH journey (define them first — see references/journeys.md):

  1. Author the tape from scripts/journey.tape.tmpl — isolated HOME (skips the wizard), window size, the exact key sequence, Screenshot at each settle point, Output <journey>.gif + .mp4.
  2. Record + extract + assert with scripts/verify-journey.sh — runs vhs, pulls frames via ffmpeg, OCRs them (tesseract, best-effort), and emits the frame paths + a checklist of expected content.
  3. READ the frames (open the PNGs) and assert the EXACT outcome for that journey — the right diff content, the right file list, word-emphasis on the changed token, the active tab actually changed, Esc landed on the main screen. Recipe + anti-patterns: references/frame-truth.md.
  4. Keep the <journey>.gif (for the explainer) and <journey>.mp4.

Recording-seed gotchas (frames stuck on onboarding / the install modal = a bad seed, not a bad feature):

  • The isolated HOME needs BOTH a version-matched onboarding.toml AND a complete notify install.json. A partial {"prompt_dismissed": true} fails to deserialize into InstallRecord (which also requires agents, hook_script, …), is silently dropped, and the "Get notified when a session needs you?" modal re-appears on top of HomeScreen and swallows your keystrokes. A fixed-timing .tape can't poll-and-resend like a tripwire, so the whole journey then records against the wrong screen. Mirror the tripwire's ainb_plugin_notifyd::dismiss_prompt() shape, or write the full record: {"agents":[],"hook_script":"","claude_plugin_dir":null,"codex_hooks_json":null,"plugin_version":null,"prompt_dismissed":true}.
  • vhs Output and Screenshot paths MUST be quoted when they contain / or - (e.g. any /tmp/ainb-... path) — otherwise vhs's parser splits on them and the tape fails to compile. Key@500ms repeat-with-delay syntax is Type-only; for nav keys use Down + Sleep.

A journey is unverified until a human-or-agent has read its frames and confirmed the content. "Non-blank == pass" is a failure of this gate.

G4 — Fix loop

If any G1 assertion, any G2 frame-check, or any G5/G6 asset is wrong: diagnose → fix → re-run. Unbounded until acceptance tests + tripwire are green and every journey frame-checks clean, or a genuine hard blocker (then log the wall and continue everything parallelizable).

G5 — Continuous explainer (defer publishing to /explain-to-me + /here-now)

Maintain ONE explainer at a stable slug, re-published after every phase/commit. Each republish must contain: current phase + progress, the latest cargo test / clippy / tripwire output, the commit log so far, and the embedded per-journey GIFs. Build it with /explain-to-me (status-report / PR-review template) and publish with /here-now; scripts/publish-explainer.sh checks the dir then prints the exact /here-now invocation. Cadence + required sections: references/explainer.md.

G6 — Final HTML-solid validation

Before declaring done, FETCH the live URL and confirm: every link resolves, every GIF loads, no section is empty/placeholder, the test output shown is the latest green run. Then hand the human the live URL — the human is the final acceptance gate. Do not merge on their behalf unless the goal says so.

Quick start

SK=ainb-tui/.claude/skills/tmux-verify
AINB=$(find ainb-tui/target -name ainb -type f -perm +111 | head -1)   # or: cargo build -p ainb-core

# 1. one journey: render tape, record, extract+OCR frames, print checklist
bash $SK/scripts/verify-journey.sh \
  --name open-g --bin "$AINB" \
  --keys 'Type "G"  Sleep 2s  Screenshot "{OUT}/open-g.png"' \
  --out /tmp/ainb-journeys --expect "Code Review" --expect "skill.rs"
# → then READ /tmp/ainb-journeys/frames/open-g*.png and assert content

# 2. publish/refresh the explainer (after building it with /explain-to-me)
bash $SK/scripts/publish-explainer.sh --dir /tmp/ainb-explainer --slug ainb-warp-diff

Done checklist (all true)

  • [ ] G1 tripwire comprehensive + green + non-flaky ×3
  • [ ] G2/G3 every journey has its own gif+mp4 AND its frames were read + asserted (content correct)
  • [ ] G4 cargo test -p ainb-core + clippy -D warnings green
  • [ ] G5 explainer live on /here-now at the stable slug, current
  • [ ] G6 HTML fetched + solid; live URL handed to the human

References

  • references/journeys.md — define a journey; the worked ainb diff-review journey set
  • references/frame-truth.md — vhs → frames → READ + assert; anti-patterns
  • references/explainer.md — the continuous /explain-to-me → /here-now cadence
  • Sibling: tmux-ui-tripwire (G1). Tooling: /media-processing (vhs/ffmpeg/imagemagick), /explain-to-me, /here-now.
tmux-verify Skill | Agent Skills