Plan Keeper Task Planner

The Problem

Long AI-agent tasks often lose context across sessions: goals fade, progress lives only in chat, past decisions are hard to locate, and imported skill dependencies are scattered. Plan Keeper externalizes task state into project files: current-plan.md stores the active goal, steps, and status; .atomcode/records/ stores progress logs, changes, snapshots, and dependencies; .atomcode/index.md builds a searchable index with [🔖 tag] anchors instead of forcing full record reads.

How It Works

• Initialization: checks for .atomcode/skills/plan-keeper/skill.md; if absent, creates the project skill file, current-plan.md, and records/, then scans global/local skills for selective import.
• Import management: imports.skills and imports.records declare dependencies, with deduplication, project-level overrides, a 3-level depth cap, and cycle detection; commands such as add, remove, verify, and reload manage load state.
• Execution rules: a plan entry is required before producing files; completed steps update both current-plan.md and records/; changes are appended rather than overwritten; pauses create snapshots and blocking dependencies create [🔖 dependency: ...] entries for later lookup.

Boundaries

It is best understood as a planning and memory protocol for agents, not a standalone project-management backend. Its usefulness depends on accepting the .atomcode/ and current-plan.md file layout; manual edits, deleted indexes, or cross-tool state sharing may require extra validation of tags, permissions, and file consistency.