ls-arbiter
Push decisions to Arbiter Zebu for async human review. Use when you need human input on plans, architectural choices, or approval before proceeding.
SKILL.mdls/skills/ls-arbiter/SKILL.md
GitHubArbiter Skill
Use this skill when work is blocked on a human decision: plan review, architecture tradeoffs, release approvals, or a batch of related choices that should be answered asynchronously.
Do not use it for simple questions you can answer from the repository, urgent real-time decisions, or approvals that require direct user confirmation in the current chat.
Active Contract
This skill ships a Python queue helper:
python3 scripts/arbiter_cli.py push '<plan-json>'
python3 scripts/arbiter_cli.py status <plan-id>
python3 scripts/arbiter_cli.py get <plan-id>
python3 scripts/arbiter_cli.py await <plan-id> --timeout 3600
The helper writes and reads markdown plans in ~/.arbiter/queue/:
| Directory | Purpose |
|---|---|
pending/ |
Plans waiting for review |
completed/ |
Plans marked complete with answers |
notify/ |
Optional notification files for agent heartbeat checks |
The helper does not run the Arbiter Zebu bot, contact Telegram, or replace the reviewer. It only manages the filesystem queue contract that Arbiter Zebu consumes.
Prerequisites
- Install framework Python dependencies:
uv sync --locked --no-dev
- Ensure an Arbiter Zebu bot or compatible reviewer watches
~/.arbiter/queue/. - Keep
~/.arbiter/queue/pending,completed, andnotifyreadable by the agent process.
If you intentionally installed the upstream Arbiter CLI, you may use it directly, but this skill documents and tests the bundled Python helper as the local supported surface.
Push A Plan
Create one JSON object with a title and a non-empty decisions array:
python3 scripts/arbiter_cli.py push '{
"title": "API Design Decisions",
"tag": "nft-marketplace",
"context": "Implementation is blocked until these choices are made.",
"priority": "normal",
"notify": "agent:swe2:main",
"decisions": [
{
"id": "auth-strategy",
"title": "Auth Strategy",
"context": "How should admin users authenticate?",
"options": [
{"key": "jwt", "label": "JWT tokens", "note": "Stateless"},
{"key": "session", "label": "Server sessions", "note": "Central revocation"}
],
"default": "session"
},
{
"id": "database",
"title": "Primary Datastore",
"options": [
{"key": "postgresql", "label": "PostgreSQL"},
{"key": "mongodb", "label": "MongoDB"}
],
"allowCustom": true
}
]
}'
Successful output:
{
"planId": "abc123",
"file": "/home/user/.arbiter/queue/pending/agent-nft-marketplace-abc123.md",
"total": 2,
"status": "pending"
}
JSON Fields
| Field | Required | Description |
|---|---|---|
title |
Yes | Plan title |
decisions |
Yes | Non-empty list of decision objects |
tag |
No | Project or topic filter; defaults to general |
context |
No | Background for the reviewer |
priority |
No | low, normal, high, or urgent; defaults to normal |
notify |
No | Session key for heartbeat notification |
agent |
No | Agent ID; defaults from ARBITER_AGENT, AGENT_ID, or USER |
session |
No | Session ID; defaults from ARBITER_SESSION or AGENT_SESSION |
Decision objects require id, title, and a non-empty options array. Each option requires key and label; note is optional. Set allowCustom: true if the reviewer may answer with free text. push rejects plan or decision completion fields and always writes each new decision as status: pending with null answer and answered_at values.
Check Status
Use a plan ID or the newest plan with a tag:
python3 scripts/arbiter_cli.py status abc123
python3 scripts/arbiter_cli.py status --tag nft-marketplace
Output includes status, total, answered, remaining, and per-decision answer state derived from the frontmatter decisions list. A decision is answered only when its answer is a non-empty string matching an option key, unless that decision has allowCustom: true.
Get Answers
After the reviewer records a valid answer in frontmatter for every decision:
python3 scripts/arbiter_cli.py get abc123
python3 scripts/arbiter_cli.py get --tag nft-marketplace
If any answer is missing or invalid, the command exits non-zero and reports the remaining decision count. A file location or plan-level status: completed value does not override answer validation.
Await Completion
For blocking workflows, poll until completion or timeout:
python3 scripts/arbiter_cli.py await abc123 --timeout 3600 --interval 30
Use await only when the current workflow can safely block. For routine agent loops, prefer heartbeat polling.
Heartbeat Integration
Add a heartbeat or session-resume check that:
- Looks for files in
~/.arbiter/queue/notify/for the current session. - Runs
python3 scripts/arbiter_cli.py get <plan-id>orget --tag <tag>. - Resumes blocked work when answers are complete.
- Archives or removes handled notification files according to the local Arbiter policy.
Troubleshooting
| Issue | Check |
|---|---|
python-frontmatter is required |
Run uv sync --locked --no-dev from the LocalSetup source checkout |
| Plan not visible to Arbiter | Confirm the generated file is under ~/.arbiter/queue/pending/ and has YAML frontmatter |
get reports pending |
Check for missing or invalid per-decision answers in frontmatter; moving the plan to completed/ does not determine completion |
| Tag finds the wrong plan | Use the explicit planId; tag lookup returns the newest matching plan |
Reference
See references/overview.md for the filesystem lifecycle and generated file shape. The template in templates/decision.md mirrors the bundled CLI output.