Git snapshot of notplants.bsky.social/p-lichen-orchestrator; copied for local development and review.
README.md

p-lichen-orchestrator #

An agent-orchestrator project (vendored as the engine/ submodule) that drives a Python reimplementation of lichen.page.

See plans/brief.md for the mission and what lichen.page is.

The agents #

agent role started by
builder builds the reimplementation, phase by phase bare up
correctness-adversary disbelieves the Builder on correctness (works / matches lichen.page) bare up
readability-adversary ensures the code is very readable and easy to understand bare up
feature-completeness occasional audit: is any lichen.page feature missing? on demand

A claimed gate is "done" only when both adversaries PASS it (REVIEW-corr-* and REVIEW-read-*), with no standing ## VETO. The agents coordinate only through git (the work repo's claim(/review( commits + machine-docs/).

Run #

git submodule update --init                       # fetch the engine (first clone)
python3 engine/agents.py status                   # what would run
python3 engine/agents.py up                        # builder + both adversaries + watchdog
python3 engine/agents.py up feature-completeness    # run the coverage audit on demand
python3 engine/agents.py logs builder              # tail a session
python3 engine/agents.py down                       # stop everything

Needs claude on PATH. The Builder works in ./work; the correctness and readability adversaries cold-verify from their own clones ./work-corr and ./work-read; the auditor uses ./work-features. These are clones of a shared work repo — set its remote up before a live run (see the engine's examples/builder-adversary for the two-clone model; here it's three+).

Phases #

  1. bootstrap — running Starlette app + login/session auth.
  2. collab — real-time collaborative editing (pycrdt + WebSocket + sqlite + Yjs client).
  3. features — every remaining lichen.page feature, driven by the feature-completeness audit.

Notes #

  • Two-adversary handoff: claim_pings is a list, so the watchdog pings both adversaries on a claim( (the engine supports a list); the adversaries also poll each wake as a fallback. A review( from either adversary pings the Builder.