Disk-based memory for coding agents that forget the plan

What it does

This project is essentially a filesystem convention plus a bundle of shell scripts that force AI coding agents to write their plans to disk, check progress against those files, and attest completion with SHA-256 hashes. It turns a local markdown file into shared memory that survives /clear, reboots, and handoffs between different agents or IDEs.

The interesting bit

The cleverness is in the boring parts: exhaustive release notes show the authors wrestling with POSIX portability, Windows exec-bit quirks, and keeping a dozen IDE-specific SKILL.md variants in version parity. The skill is mostly glue, but it is obsessively tested glue that closes race conditions with atomic temp-rename and optional flock advisory locks.

Key highlights

• Markdown plans act as crash-proof state that outlives chat context.
• SHA-256 attestation in attest-plan.sh blocks tampering and locks a plan across runtimes.
• Slug-mode and resolve-plan-dir.sh isolate parallel tasks so multiple agents don’t stomp on each other.
• Supports Claude Code, Cursor, Codex CLI, Pi, OpenCode, and others via mirrored SKILL.md adapters.
• Community has already forked it for multi-agent orchestration, interview workflows, and finance agents.

Caveats

• The README is almost entirely release notes and badges; the actual mechanics of the planning protocol are implied rather than explained up front.
• Benchmark claims (96.7% pass rate, 3/3 A/B wins) are presented as badge graphics without reproducible detail in the visible source.

Verdict

Worth a look if you run long, multi-step agentic tasks and need a dead-simple way to persist state across sessions. Skip it if you were hoping for a standalone planner algorithm—this is an orchestration layer, not a reasoning engine.