finds.dev← search

// the find

OthmanAdi/planning-with-files

★ 25,204 · Python · MIT · updated Jul 2026

Persistent file-based planning for AI coding agents and long-running agentic tasks. Crash-proof markdown plans that survive context loss and /clear, plus a deterministic completion gate and multi-agent shared state on disk. Manus-style. Works with Claude Code, Codex CLI, Cursor, Kiro, OpenCode and 60+ agents via the SKILL.md standard.

A skill/plugin for AI coding agents that persists planning state to three markdown files (task_plan.md, findings.md, progress.md) so sessions survive context resets and crashes. The core idea is sound: filesystem as working memory, with lifecycle hooks to re-inject plan context before tool calls. Aimed at anyone using Claude Code, Cursor, Codex, or a dozen other agents on tasks that take more than one session to finish.

The hook integration is thorough — PreToolUse, PostToolUse, Stop, PreCompact, and UserPromptSubmit all have adapters, so the plan actually gets re-read at the right moments rather than just sitting on disk. The parallel plan isolation (slug-mode directories under .planning/) is a real feature that most similar tools skip, letting multiple concurrent agent sessions share a workspace without clobbering each other's state. The SHA-256 attestation gate in v2.37+ is a practical security addition: tampered plan files are refused at injection, which matters when agents are writing to the same files they read back as instructions. Cross-IDE breadth is genuine — not just README claims, the scripts/ and adapter directories are all present and kept at parity (the changelog documents specific version-sync bugs being fixed).

The benchmark number (96.7%) measures file-pattern fidelity — did the agent create the three files — not whether the approach actually reduces goal drift or produces better outcomes on real tasks. That's a weaker claim than the headline implies. The version history is full of sync-gap bugs where IDE variants lagged the canonical by 9–11 versions; the parity problem keeps recurring even after adding a bump script. The Manus acquisition framing is marketing: this pattern predates Manus, the $2B number is doing heavy lifting, and the actual novel contribution here is the hook wiring, not the three-file idea. Windows support has been noticeably flaky — exec-bit failures, session-catchup silently broken, short-path issues — the fixes are there but the platform clearly gets less testing.

View on GitHub →

// want more like this?

We dig through GitHub every week and send a few repos picked for what you actually care about — each with an honest take like this one.

Get finds in your inbox → Search again →