| name | lope |
| description | Multi-CLI validator ensemble for AI work. Use for multi-phase sprints, single-shot cross-model checks, autonomous flow graphs, evidence gates, team management, persistent finding memory, council deliberation, and self-update. Any AI CLI implements, any AI CLI validates. 15 built-in CLI adapters plus infinite custom providers via JSON config. |
Lope — multi-CLI validator ensemble sprint runner
Lope is an autonomous sprint runner. One AI CLI implements. Others independently review. Majority vote decides if a phase ships. No single-model blindspot.
Works for any domain: engineering, business (marketing, finance, ops, consulting), research.
The sprint modes
| Mode | Slash command | What it does |
|---|
| Negotiate | /lope-negotiate | Draft a structured sprint doc via multi-round validator review. Validators push back on scope creep, missing criteria, unverified assumptions, until majority consensus. |
| Execute | /lope-execute | Run sprint phases one at a time. After each phase, validators independently vote PASS / NEEDS_FIX / FAIL. NEEDS_FIX retries with specific fix instructions (3 attempts). |
| Implement | /lope-implement | Select implementation/escalation agents once, then run a sprint without further human input. Single-writer safety model. |
| Audit | /lope-audit | Generate a scorecard from the executed sprint — per-phase verdicts, confidence scores, durations, overall status — and append to the journal. |
| Flow | /lope-flow | Run a declarative DOT graph workflow: agent / ensemble-review / shell-gate / judge-router nodes, conditioned edges, fan-out + fix-loops. Autonomous (no human gates), bounded by per-node and graph-wide visit caps. For shaping how agents collaborate, not a fixed linear sprint. |
How to pick which mode to run
- User has a fresh idea and no sprint doc yet →
/lope-negotiate
- User has a sprint doc and wants normal phase running →
/lope-execute
- User has a sprint doc and wants zero-human execution / out-of-the-loop mode →
/lope-implement
- User has already run a sprint and wants the scorecard →
/lope-audit
- User wants agents to negotiate autonomously as a graph (fan-out proposers, consensus, fix-loops) or to reshape how they collaborate →
/lope-flow (lope flow init consensus to start)
- User just wants to know what validators lope found on their machine →
lope status (plain shell)
- User wants to pick which validators to use →
lope configure (plain shell)
Unless the user tells you otherwise, the default flow is: negotiate → implement/execute → audit.
The user probably won't type slash commands
Most users won't remember /lope-negotiate syntax. They'll just describe what they want in natural language — "plan the auth refactor", "negotiate the Q4 campaign carefully", "scope the data migration". Your job as the agent is to recognize those shapes and invoke lope on their behalf.
Trigger words: plan, negotiate, scope, draft, roll out, work through, or phrases like "needs to be right", "don't break things", "let's be careful here".
When you recognize one of those + multi-phase work, construct the goal string from the user's prose and run lope negotiate "<goal>" --domain <engineering|business|research> directly. Do not wait for the user to type a slash command.
See docs/samples.md in the lope repo for 8 end-to-end conversation walkthroughs that show this pattern.
Domains
Pass --domain <name> on negotiate to switch validator role, artifact labels, and review task.
| Domain | Use for | What it checks |
|---|
engineering (default) | code, software, infra, devops | files, tests, acceptance criteria |
business | marketing campaigns, budgets, ops, consulting, finance | deliverables, success metrics, stakeholder alignment |
research | academic work, systematic reviews, studies | artifacts, validation criteria, methodology rigor |
Examples
/lope-negotiate "Add JWT auth with refresh token rotation"
/lope-negotiate "Q2 product launch campaign for SaaS tier" --domain business
/lope-negotiate "Systematic review of LLM alignment techniques" --domain research
Supported validators
15 built-in CLI adapters, auto-detected on $PATH:
Claude Code · OpenCode · Gemini CLI · Codex · Mistral Vibe · Aider · Ollama · Goose · Open Interpreter · llama.cpp · GitHub Copilot CLI · Amazon Q · pi (Traylinx) · Qwen Code · Agy
Plus infinite custom providers via ~/.lope/config.json (subprocess or HTTP). You need at least two different ones for the ensemble to have signal.
Token efficiency — intelligent caveman mode
On by default. Validator prompts get compressed by dropping articles, filler, and hedging, while keeping code, paths, line numbers, and error messages exact. 50-65% token savings per validator call. Disable with LOPE_CAVEMAN=off for tasks that need full prose (external writing, papers).
Dynamic sprint mode
Treat the sprint as dynamic. If during work you discover a better approach or have an "aha" moment, do not silently expand scope and do not ignore it. Raise it with your lope teammates, negotiate whether it belongs, and fold it in if agreed.
When NOT to use lope
- One-off refactors that are a single edit → just do the edit
- Exploratory questions ("what could we do about X?") → have a conversation first, then lope the agreed plan
- Anything where the user has already named the exact files and lines — they want the change, not a sprint
- Urgent fire-fighting (production is broken, fix now) — patch the bug, don't negotiate a sprint
Do not wrap lope
Lope is already a CLI. If the user says "use lope to do X", you invoke lope <mode> <args> in a shell. You do not write a Python wrapper script, a bash harness, or any other scaffolding that imports or calls lope. The whole point of the multi-CLI ensemble is that lope is the harness — you don't need another one. One command, no wrapper.
Install / update / self-check
lope status
lope configure
lope version
lope update
lope update --dry-run
If lope isn't on PATH, prefix commands with PYTHONPATH=~/.lope python3 -m lope or tell the user to add alias lope='PYTHONPATH=~/.lope python3 -m lope' to their shell rc.
Standard installs are git checkouts in ~/.lope. lope update fetches tags, pulls the tracked branch with --ff-only, and reruns ./install so Claude/Codex/Gemini/OpenCode/Cursor/Vibe/Qwen/pi skills stay current. PyPI publishing is not live yet, so use the git checkout path for servers.
