planasaurus runs your implementation plan end-to-end — implement, review, fix, repeat — with a deterministic state machine steering and subagents doing the judgment. No external runner. No per-call billing.
/planasaurus:run docs/plans/your-plan.md
The control flow — loops, counters, convergence, when to stop — lives in a tiny deterministic binary. Claude only does what it's good at: writing code, reviewing it, fixing it. Tokens are never spent deciding what to do next.
A compiled state machine decides every step. The model obeys one instruction at a time — it can't lose count across 15 review rounds.
No claude -p fan-out, no spawned processes. Everything runs as in-session subagents under your existing plan.
All state lives on disk in .planasaurus/. Crash, compaction, closed laptop — re-run the command and it picks up exactly where it stopped.
Uses the codex plugin if present, otherwise an adversarial Claude on a different model tier. There is always a second opinion.
Add the marketplace, then install the plugin.
# add the Rayforce marketplace /plugin marketplace add RayforceDB/planasaurus # install the plugin /plugin install planasaurus@rayforce
Use Claude Code's brainstorming + plan-writing flow to produce a plan file with [ ] task checkboxes.
/planasaurus:run docs/plans/your-plan.md — optionally --skip-codex or --max-iterations=N.
It implements each task, runs multi-agent review, takes an external second opinion, fixes what's real, and archives the plan when done.
Requires a git repository. The codex external-review step is optional — without the codex plugin it falls back to an adversarial Claude reviewer.