lope

Use when configuring, installing, verifying, or troubleshooting optional Headroom MCP compression with Lope or Lope-installed agent hosts. Covers opt-in Headroom MCP registration, large validator/tool-output compression, Claude/Codex/Cursor MCP status checks, and the `headroom-ai[mcp]` dependency when the user asks for it.

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.

Use Lope when cross-model perspective helps: multi-phase sprints, flow graphs, ask/review/vote/compare/pipe, team, lope memory, lope deliberate, gate/check, update, or Headroom setup. Trigger on 3+ phases, consequential multi-file work, second opinions, artifact review, consensus/SARIF, ADR/PRD/RFC/build-vs-buy/migration, A/B or yes/no choices, pipes, roster changes, updates, or Headroom. Skip trivial single edits, pure conversation, urgent firefighting.

Run a declarative GRAPH workflow where AI agents negotiate autonomously — no human gates. A flow is a DOT graph: nodes are agent turns, ensemble reviews, shell verify-steps, or judge/routers; edges carry conditions and loops. Each node dispatches into lope's existing multi-CLI executors (any CLI implements, the ensemble votes). Bounded by per-node and graph-wide visit caps so an unsupervised run can never loop forever. Use when the user wants to shape HOW agents collaborate (fan-out proposers, consensus, fix-loops) as editable, reviewable, version-controlled process — not a fixed linear sprint.

Manage the validator team (add, list, remove, smoke-test) via CLI flags. Trigger on any user request about adding, removing, configuring, enabling, disabling, or testing a validator/teammate/CLI on lope — including 'add openclaw to lope', 'remove ollama from the team', 'list validators', 'is my new provider working', 'hook up my mistral pod', 'what CLIs are configured', OR the user pasting a curl command and saying 'add this'. Works for subprocess binaries (any local CLI), OpenAI-compatible HTTP endpoints (pods, gateways, cloud APIs), and pasted curl commands via `--from-curl`. The LLM translates natural language into the right invocation — the end user never edits JSON.

Draft a sprint doc via multi-round negotiation with AI validators. Lope sends your plan to other AI CLIs (Claude, Gemini, OpenCode, Codex, Mistral Vibe, Aider, Ollama, Goose, llama.cpp, Open Interpreter, Copilot, Amazon Q, or any HTTP API) for independent review. Majority vote. No single-model blindspot. Works for engineering, business (marketing, finance, ops), and research domains.

Print the complete lope reference into the current session — all modes, flags, domains, env vars, slash-command vs natural-language invocation per host, troubleshooting, and hard rules for agents. Use when the user asks 'how does lope work', 'what can lope do', 'what are lope's flags', or anything that needs authoritative lope documentation. Prefer this over guessing from memory.

Run a sprint with zero-human swarm orchestration. First select implementation agents and escalation agents, then Lope executes phases without further human input. Use when the user says implement the whole sprint, stay out of the loop, use Claude/OpenCode or specific CLIs for escalations, or wants autonomous sprint backlog completion.

Run an Agent-Order-style council deliberation on a decision artifact. Six built-in templates: ADR, PRD, RFC, build-vs-buy, migration-plan, incident-review. The 7-stage protocol — independent positions, anonymized critique, revision, synthesis, rubric review, minority report — turns a scenario file into a structured artifact that has survived peer scrutiny. Use when the user asks 'should we adopt X?', 'review this ADR / PRD / RFC', 'build vs buy this capability?', 'plan the migration', 'run the incident review'. Pure adversarial reasoning — no source-file modification, no git mutations.

Ask every configured validator the same question and return N independent answers — one per model. No sprint framing, no phases, no verdict parsing. Use for any multi-perspective query: 'what do these 5 CLIs think about X?', 'compare how 3 models solve this problem', 'get a second opinion across models before I commit'. Works for any domain (engineering, research, writing, business).

Query the persistent Lope finding store (SQLite, populated by `lope review --remember`). Use when the user asks 'is this a recurring issue?', 'which files keep getting flagged?', 'show me everything Lope knows about auth.py', 'what was that finding from last week?', or 'how many findings has Lope stored?'. Five subcommands: stats, search, file, hotspots, forget. All local, stdlib only, opt-in — `LOPE_MEMORY=off` disables the store entirely.

Read stdin as the prompt, fan out to every configured validator, print per-model answers. The composable shell verb — anything that produces text on stdout can feed lope in a pipeline. Per-validator isolation by default: one timeout doesn't kill the run. Pass --require-all for strict exit-non-zero-on-any-error semantics.

Fan out a file review to every configured validator and collect N independent critiques — one per model. Use for cross-model code review, doc review, contract review, resume review, or any review-shaped task where multiple perspectives catch what a single model would miss. Optional --focus flag targets a specific concern (security, perf, tests, clarity, etc.).

Compare two files across every validator. Each model picks which file is better against explicit criteria (--criteria flag). Tally the picks; print the winner. Use for A/B review, before/after diff evaluation, migration decisions, bake-offs. Criteria is mandatory in the prompt — 'better' is never model-invented.

Send a question with a fixed list of options to every validator. Each one replies with exactly one option label. Tally the picks; print the winner. Use for decisions where you want multi-model consensus on a pre-defined set of choices — 'should we ship X or Y?', 'pick 3.12 or 3.13', 'yes / no / needs-more-info'. Per-validator isolation: one timeout doesn't kill the tally.

Generate a scorecard from sprint execution results — per-phase verdicts, confidence scores, duration, overall status. Optionally writes to the lope journal for historical tracking.

Execute a sprint phase-by-phase with validator-in-the-loop retry. For each phase: implement (code, deliverables, research, etc.) → AI validators review → retry on NEEDS_FIX → advance on PASS. Works with any domain (engineering, business, research).