planning-with-files-from-github（规划 with files from GitHub）

Planning with Files

Manus-style persistent markdown planning. Filesystem = long-term memory. Context window = working RAM.

Core Principle

Context Window = RAM (volatile, limited)
Filesystem     = Disk (persistent, unlimited)
→ Anything important MUST be written to disk.

When to Use

• Multi-step tasks (3+ phases or 5+ tool calls)
• Research + implementation tasks
• Tasks that span multiple sessions
• Any task where goal-drift is a risk

Skip for: single-file edits, quick lookups, simple questions.

Session Start Protocol

ALWAYS do this first — no exceptions:

1. Run the init script to create the three planning files:

   exec: bash {baseDir}/scripts/init-session.sh "<task description>"

1. Fill in the ## Goal section in task_plan.md with the full task description.
2. Break the task into 3–7 phases and add them under ## Phases.
3. Confirm files exist before proceeding.

If exec is not available, create the three files manually using the templates in {baseDir}/templates/.

The Three Files

task_plan.md — Goal tracker (MOST IMPORTANT)

• Created at session start, never deleted
• Re-read before every major decision
• Update when: phase completes, decision made, error encountered

findings.md — Knowledge store

• Research results, discoveries, technical decisions
• 2-Action Rule: after every 2 view/search/fetch operations → update findings.md immediately

progress.md — Session log

• Actions taken, test results, errors encountered
• Append-only; never rewrite history

Work Loop

BEFORE every major action:
  → Re-read task_plan.md (keeps goal in attention)

EVERY 2 view/search/fetch operations:
  → Save key findings to findings.md NOW

AFTER completing a phase:
  → Update phase status in task_plan.md to "complete"
  → Append phase summary to progress.md

WHEN an error occurs:
  → Log it in task_plan.md under ## Errors table
  → Never repeat the same failing action

BEFORE stopping:
  → Run: exec bash {baseDir}/scripts/check-complete.sh
  → Only stop if all phases show "complete"

File Update Rules

Error Recovery Protocol

ATTEMPT 1: Diagnose & Fix
  → Read error carefully → identify root cause → targeted fix

ATTEMPT 2: Alternative Approach
  → Same error? Try different tool/library/method
  → NEVER repeat the exact same failing action

ATTEMPT 3: Broader Rethink
  → Question assumptions → search for solutions → update plan

AFTER 3 FAILURES:
  → Escalate to user: explain what was tried, share the error, ask for guidance

Completion Verification

Before telling the user the task is done:

1. Run exec bash {baseDir}/scripts/check-complete.sh
2. If any phase is NOT "complete" → continue working
3. Only deliver result when all phases are checked off

Session Recovery

If context was cleared (e.g., /clear), resume by:

1. Reading task_plan.md to restore goal and phase status
2. Reading progress.md to see what was done
3. Reading findings.md to restore research context
4. Continuing from the first incomplete phase

Key Principles

• Plan is Required: Never start a complex task without task_plan.md
• Files are Memory: Context is volatile. Filesystem is permanent. Write important things down.
• Never Repeat Failures: Track what was tried. Mutate the approach.
• Error Recovery = Agentic Behavior: How you handle failures defines task quality.
• Re-read Before Deciding: Attention manipulation — keep the goal visible.