Shadow

Shadow is you, one step down. You shape intent; Shadow does what you'd otherwise type at an AI agent — opens the board, picks the row, builds the prompt, writes the proof. It survives what loses work: a chat dying, a machine change, ten conversations at once.

Install

git clone https://github.com/firstbitelabsllc/shadow.git && cd shadow
bash install.sh && shadow doctor

Git, Bash, Python 3.10+, one native host (Claude Code, Codex, or Cursor). No Node, no npm — the clone is the install, git pull is the update. Then paste the standing goal from host integration into your hosts' instruction files.

The Claude mount also activates the Brief contract's Stop hook: the mounted directory carries .claude-plugin/plugin.json, so Claude Code loads it as the plugin shadow@skills-dir and reads hooks/hooks.json from it. Run /reload-plugins or restart to pick it up in a live session.

Use

shadow status                  # the board — same list from any directory, any machine
shadow amp                     # one goal block that POINTS at the plan, ≤4k chars
shadow throw --task '~ab12'    # claim a row before handing it to another seat
shadow accept --row '~ab12'    # rerun the proof in a clean checkout, then flip the row

Also init --here, lint, browse, host run, doctor.

Three ideas

One plan, in git. Each project's truth is its own PLAN.md, so Shadow opens the same board anywhere — an empty directory falls back to your portfolio instead of pretending nothing exists. Chat is a projection; the plan is the memory.

The goal is a pointer. No prompt carries ten projects in 4,000 characters, so amp doesn't try: authority ref, one resume row, its proof, and the rule that when block and plan disagree, the plan wins.

No proof, no completed. Every task carries a proof: that can refuse bad work — cmd, read, or gate <owner>. accept reruns it in a clean checkout and is the only path that flips a task.

What it won't do

No database, daemon, scheduler, or transcript memory — a surface that could mutate a row outside PLAN.md is banned. Native hosts keep auth and model choice. host run runs one sealed task and never launches work itself. accept reruns only cmd proofs; read and gate are judgments a person re-observes. A receipt is evidence, not acceptance.

Develop

scripts/shadow-python.sh -m unittest discover -s tests -p 'test_*.py' — the method is one page, AGENT.md, enforced by the grammar.

Quick start · Commands · Amp · Privacy · MIT.