---
description: Workflow memory operating loop — read root state before acting, write state back after meaningful change
globs:
alwaysApply: true
---

# Workflow Memory

This repository keeps agent memory in root markdown files. Follow this loop every session.

## Session Start

1. Read `workflow_state.md` — current phase, plan, blockers, log.
2. Read `project_config.md` — goals, stack, patterns, constraints.
3. Reconstruct the next action from those files before proposing or making changes.

## Loop

One phase at a time: `ANALYZE → PREPARE → IMPLEMENT → VALIDATE`.

- ANALYZE: inspect relevant files and current state before deciding.
- PREPARE: short plan; identify the checks `project_config.md` requires.
- IMPLEMENT: smallest coherent edits that satisfy the task.
- VALIDATE: run the checks appropriate for the repo; report what could not run.
- ROLLBACK: only when the approach must be abandoned and state preserved.

## State Writes

Update `workflow_state.md` only after meaningful change: a phase transition, a completed task, a blocker, or a validation result that changes the next action. No per-message summaries. Preserve headings, keep entries concise.

Update `project_config.md` only for durable changes: goals, stack, patterns, constraints, changelog. No live task notes there.

Never store secrets, tokens, or credentials in memory files. Record only where they load from (env vars, secret manager).

## Exit

When work is complete, report result, changed files, and validation status. No post-completion questions, no optional next-step prompts. Wait for the next user message.