leopold

Phase 1 of Leopold. A structured debate that captures the mission and your decision-making, then writes the brief (MISSION, CHARTER, GUARDRAILS, PLAN) that the autonomous run executes.

Launch a local live web dashboard to watch the current project's Leopold run — status, cost meters (context/subagents/forks), event feed, decisions, and a Stop button. Local-only (127.0.0.1), zero dependencies.

Compile a Leopold brief into a dynamic workflow and run it. The plan lives in code instead of one growing context window, every item gets an independent adversarial review, and git stays locked — nothing is committed. The workflow is resumable and shows up in /workflows.

Control plane for the global prompt enhancer: status, on/off, preview a prompt against the gate, and the learn loop that mines your enhancement ledger into proposed prompt-profile rules. Never edits PROMPT-PROFILE.md itself — you review and apply.

The self-improving charter. Mines past Leopold decisions, session corrections, and git history for recurring judgment calls, skeptic-verifies each candidate, and proposes charter amendments. Never edits CHARTER.md itself — you review and apply.

Phase 2 of Leopold. Enters autonomous mode and conducts Claude Code through the plan, deciding from the charter instead of asking, with git locked. The Stop hook keeps it going until the plan is done or a stop condition fires.

Triage a backlog (GitHub issues, bug reports, any queue) as a dynamic workflow: quarantined classifiers over untrusted content, dedupe against what's tracked, severity-ranked report, fix plans for quick wins. Pair with /loop to run it continuously.

Phase 0 of Leopold. One-shot setup that makes a project extract the most from Claude Code: generates CLAUDE.md (/init), teaches Claude how to build/run the app (/run-skill-generator), verifies MCP + permissions, then hands off to /leopold-brief. Run this once per project.

Take the seat back. Cleanly ends the autonomous run at the next turn boundary and writes a final summary. Nothing is interrupted mid-turn.

Diagnose the Leopold install: checks skills, hooks and their wiring, gstack, the driver toolchain, and whether an update is available.

Update the Leopold engine (skills + hooks) to the latest release: pulls the source and re-runs the installer.

Show the state of the current Leopold run: active or not, progress through the plan, decisions logged, and the most recent events.