Skill

Vibejar — agent bug queue

Vibejar is the screenshot tool that queues bugs for coding agents. Testers capture on phone/web, annotate, send to a jar. You list open work, claim one task at a time, fix from the shot + note, open a PR, mark status.

Do not invent the protocol. The stable contract is https://vibejar.com/contract.md. The CLI is the only transport.

Before first use

If the CLI is already installed, verify it:

bun ~/.vibejar/cli.ts whoami

If this skill came from the Claude plugin and the CLI is missing, install from the plugin files the user already approved:

VIBEJAR_SOURCE_DIR="${CLAUDE_PLUGIN_ROOT}" \
  bash "${CLAUDE_PLUGIN_ROOT}/install.sh"

For other hosts, follow https://vibejar.com/connect.md. Do not pipe a remote script straight into a shell: download it, read it, then run the reviewed local file.

If whoami says the machine is not paired, continue to Setup below.

CLI entrypoint

Always:

bun ~/.vibejar/cli.ts <command>

Never assume a global vibejar binary unless the user installed one.

Setup (only if the CLI is missing or whoami is unpaired)

# Download, inspect, then run — never curl | bash.
installer="$(mktemp)"
curl -fsSL https://vibejar.com/install.sh -o "$installer"
cat "$installer"

Read the complete file. If it matches the documented install (Vibejar CLI, its two dependencies, the Vibejar skill, and skill-directory links), run:

bash "$installer"
rm "$installer"
bun ~/.vibejar/cli.ts whoami

Pair with the phone app (shares the phone's jars):

bun ~/.vibejar/cli.ts pair <token> --name "Grok Build"   # or "Claude Code" / "Codex"

Jar link only (vibejar.com/j/...): no pair needed — fork mints identity.

Commands (contract)

Command	Output
whoami	{ "id", "email" }
jars	lines: <slug> <open> open / <total> total
fork <slug-or-url>	new slug in this account (never mutates the sender)
list [jar-slug] [--state …|--all]	read-only JSON array of captures (default: todo). Aliases: ls, todo
claim <id>	claim one capture by id → { id, note, shot, jar }
status <id> <state> [--pr <url>]	states below (todo releases a claim)
watch [jar-slug]	streams new capture: <id> <note>
self-update	refresh CLI + skill from vibejar.com

There is no drain. Always list, then claim the id you are about to fix.

States

todo → claimed → fixing → review → done (or failed).

• list never changes state
• claim moves one capture to claimed
• set fixing while working if the run is long
• review --pr <url> when a PR exists
• done only after the human approves the PR
• failed when a human must look
• Wrong claim: status <id> todo immediately

Default workflow (always this)

1. whoami — install or pair only when needed.
2. jars — see open counts; pick the jar for this repo (or inbox if asked).
3. list [jar] — read notes; decide which captures you will fix in this session.
4. For one capture at a time:
   • claim <id> for the task you are starting now (not a batch of claims up front).
   • Read the image at shot (annotated screenshot is ground truth). Note is context.
   • Reproduce → fix in the correct repo → open one PR.
   • status <id> review --pr <pr-url> with before/after proof when visual.
5. Tell the user the PR URL; they mark done (or you do after explicit approval).
6. Repeat claim for the next chosen id from the same list, or re-list if the queue may have changed.
7. Captures you do not want: leave them todo. Never claim them "for later."

Filtering by project

Inbox jars mix products. From list output, keep only rows that match the current repo / user target (note text, jar slug, screenshot context). Claim those only.

Rules

1. List first. Claim second. Never claim without having listed (or the user pasting a specific id).
2. Claim only what you are working on now — one at a time.
3. Image first. Circle/arrow marks what the tester meant.
4. Separate bugs from praise/suggestions — fix bugs; mention the rest in summary.
5. One PR per fix. Do not batch unrelated captures into one PR.
6. claim is atomic — safe across parallel agent sessions for that id.
7. Never edit the sender's jar; fork first when given a shared link.
8. Do not claim done without human approval of the PR.
9. Mistaken claim → status <id> todo before ending the session.

Project ↔ jar mapping

Jars are projects (e.g. glp3, journeys, inbox-…). Prefer the jar that matches the current repo. If unclear, jars + ask, or list the inbox and filter without claiming unrelated rows.

Nightly / autonomous loops

Vibejar is the queue, never the runner. Your scheduler owns the loop:

# example: list once, then claim+hand-off each matching id
ids=$(bun ~/.vibejar/cli.ts list my-app-jar | …)  # parse JSON, filter
for id in $ids; do
  bun ~/.vibejar/cli.ts claim "$id"
  # hand claim JSON to a coding agent session for this repo
done

Troubleshooting

Symptom	Fix
command not found: bun	Install bun: https://bun.sh
CLI missing	Download and review https://vibejar.com/install.sh, then run the local file
empty jars after pair	Confirm pair token; pass --name of this agent
skill not auto-invoking	Ensure install ran (skill lives under ~/.agents/skills/vibejar/)
stale protocol	bun ~/.vibejar/cli.ts self-update (only when asked or a newer version is reported)
stuck in claimed	status <id> todo to release
drain is removed	Use list then claim <id>