Skip to content

Roadmap

flowchart LR
    V01["v0.1 · in-session engine<br/><small>shipped</small>"] --> DRV["SDK driver<br/><small>shipped · alpha</small>"]
    DRV --> FAN["parallel waves<br/><small>shipped</small>"]
    FAN --> DASH["live dashboard<br/><small>shipped</small>"]
    DASH --> WF["dynamic-workflow engine<br/><small>shipped</small>"]
    WF --> ENH["prompt enhancer<br/><small>shipped · v0.11</small>"]
    ENH --> E2E["headless runtime E2E"]
    E2E --> SBX["sandboxed workers"]
    classDef done fill:#2a9d8f,stroke:#1b6b5f,color:#fff;
    class V01,DRV,FAN,DASH,WF,ENH done;

Shipped

  • In-session engine: /leopold-brief, /leopold-run, /leopold-status, /leopold-stop
  • Stop hook (continuity) and PreToolUse hook (git lock), behavior-tested + red-teamed
  • Brief artifact templates and install.sh with idempotent settings merge
  • SDK driver (alpha): persistent conductor + fresh workers, conductor/worker status protocol, charter-grounded decisions, git-locked canUseTool, notifications — uses your Claude Code auth, no API key
  • leopold doctor, toolchain manager (leopold menu), leopold up (Phase 0)
  • Parallel waves — run --parallel N: dependency-aware scheduler, one worktree per item, staged-patch replay onto the main tree
  • Live dashboard — leopold watch: SSE event stream, cost meters, the decision log, and the native dynamic-workflow phase tree
  • Dynamic-workflow engine — /leopold-workflow (brief compiled into a workflow), leopold workflow (the compiler as tested code; --run experimental headless runtime), /leopold-learn (the self-improving charter + learn_on_finish), /leopold-triage (quarantined backlog triage), plan-by-tournament in the brief
  • Quality panels in the driver — diverse-lens review panel, root-cause hypothesis panel on retries, opt-in smart routing
  • Hardcore CI — CLI smoke of the built binary, shellcheck gate, macOS + Ubuntu matrix, npm provenance, Dependabot, CodeQL
  • Persona testing (0.20–0.21) — synthetic customers: evidence-grounded contract skills, /leopold-persona, and the driver-conducted harness (leopold persona run — journal-before-act, in-code bounds, the persona-guard hook proven live on both harnesses)
  • Prompt enhancer (v0.11) — global UserPromptSubmit hook: weak prompts get a charter-aware structured interpretation from Haiku on the user's own account (off by default, fail-open, anchor veto), /leopold-enhance with a ledger-mining learn loop that proposes prompt-profile rules

Next

  • Scheduled persona cadence (Claude Code scheduled tasks + Codex automations), gated on live verification of both schedulers
  • End-to-end exercise of the experimental workflow --run query shim (a manual workflow_dispatch CI job with a real key — it spends tokens)
  • Worker watchdog for turns that end without a status block
  • Toolchain manager: acceptance-test the ovmem installer on a clean Linux box + macOS
  • ovmem extension: fully-local Ollama/GGUF provider profile (no API key)
  • gstack playbook router as first-class config
  • Sandboxed workers (E2B / Daytona) for isolated parallel execution

Have an idea? Open an issue or a discussion on GitHub.