deep-plan（深度计划）

Deep Planning Skill

Orchestrates a multi-step planning process: Research → Interview → External LLM Review → TDD Plan

CRITICAL: First Actions

BEFORE using any other tools, do these in order:

1. Print Intro and Validate Environment

Print intro banner immediately:

⚠️  CONTEXT WARNING: This workflow is token-intensive. Consider compacting first.

═══════════════════════════════════════════════════════════════
DEEP-PLAN: AI-Assisted Implementation Planning
═══════════════════════════════════════════════════════════════
Research → Interview → External LLM Review → TDD Plan

DEEP-PLAN starts by running `validate-env.sh`. This script:
  - Checks env for external LLM auth values
  - Validates external LLM access by running tiny prompt(s) programmatically

SECURITY:
  - `validate-env.sh` reads secret auth values in order to validate LLM access
  - It never publishes these values or exposes them to claude

 Note: DEEP-PLAN will write many .md files to the planning directory you pass it

CRITICAL: Locate plugin root BEFORE running any scripts.

The SessionStart hook injects DEEP_PLUGIN_ROOT=<path> into your context. Look for it now — it appears alongside DEEP_SESSION_ID in your context from session startup. Use it as plugin_root for all script paths.

If DEEP_PLUGIN_ROOT is in your context, run validate-env.sh directly:

bash <DEEP_PLUGIN_ROOT value>/scripts/checks/validate-env.sh

Only if DEEP_PLUGIN_ROOT is NOT in your context (hook didn't run), fall back to search:

find "$(pwd)" -name "validate-env.sh" -path "*/scripts/checks/*" -type f 2>/dev/null | head -1

If not found: find ~ -name "validate-env.sh" -path "*/scripts/checks/*" -path "*deep*plan*" -type f 2>/dev/null | head -1 Then run: bash <found_path>

Parse the JSON output:

{
  "valid": true,
  "errors": [],
  "warnings": [],
  "gemini_auth": "api_key",
  "openai_auth": true,
  "plugin_root": "/path/to/plugin"
}

Store plugin_root from the JSON output - it's used throughout the workflow.

2. Handle Environment Errors

If valid == false:

• Show the errors to the user

If errors are critical (uv not installed, plugin root not found):

• Stop the workflow. User must fix these before proceeding.

If errors are ONLY about missing LLM credentials (gemini_auth is null AND openai_auth is false):

AskUserQuestion:
  question: "No external LLMs configured. How should plan review be handled?"
  options:
    - label: "Use Claude Opus for review (Recommended)"
      description: "Launch an Opus subagent to review the plan"
    - label: "Exit to configure LLMs"
      description: "Stop to set up Gemini/OpenAI credentials"
    - label: "Skip external review"
      description: "Proceed without any external plan review"

Store the choice as review_mode:

• "Use Claude Opus" → review_mode = "opus_subagent"
• "Skip external review" → review_mode = "skip"
• Default (LLMs available) → review_mode = "external_llm"

Environment validated:
  Gemini: {gemini_auth or "not configured"}
  OpenAI: {openai_auth ? "configured" : "not configured"}
  Review mode: {review_mode}

3. Validate Spec File Input

Check if user provided @file at invocation AND it's a spec file (ends with .md).

If NO @file was provided OR the path doesn't end with .md, output this and STOP:

═══════════════════════════════════════════════════════════════
DEEP-PLAN: Spec File Required
═══════════════════════════════════════════════════════════════

This skill requires a markdown spec file path (must end with .md).
The planning directory is inferred from the spec file's parent directory.

To start a NEW plan:
  1. Create a markdown spec file describing what you want to build
  2. It can be as detailed or as vague as you like
  3. Place it in a directory where deep-plan can save planning files
  4. Run: /deep-plan @path/to/your-spec.md

To RESUME an existing plan:
  1. Run: /deep-plan @path/to/your-spec.md

Example: /deep-plan @planning/my-feature-spec.md
═══════════════════════════════════════════════════════════════

Do not continue. Wait for user to re-invoke with a.md file path.

4. Setup Planning Session

First, check for session_id in your context. Look for DEEP_SESSION_ID=xxx which was set by the SessionStart hook. This appears in your context from when the session started.

Run setup-planning-session.py with the spec file, plugin root, review mode, and session ID:

uv run {plugin_root}/scripts/checks/setup-planning-session.py \
  --file "<file_path>" \
  --plugin-root "{plugin_root}" \
  --review-mode "{review_mode}" \
  --session-id "{DEEP_SESSION_ID}"

IMPORTANT: If DEEP_SESSION_ID is in your context, you MUST pass it via --session-id. This ensures tasks work correctly after /clear commands. If it's not in your context, omit --session-id (fallback to env var).

Note: review_mode is from Step 2. If LLMs are available, use external_llm (or omit the flag).

Parse the JSON output:

This script:

1. Validates the spec file exists and has content
2. Creates deep_plan_config.json in the planning directory with plugin_root, planning_dir, and initial_file
3. Detects whether this is a new or resume session
4. Writes task files directly to ~/.claude/tasks/<task_list_id>/
5. If sections/index.md exists, also writes section tasks (positions 22+)

If success == false: The script failed validation. Display the error and stop:

═══════════════════════════════════════════════════════════════
DEEP-PLAN: Setup Failed
═══════════════════════════════════════════════════════════════
Error: {error}

Please fix the issue and re-run: /deep-plan @path/to/your-spec.md
═══════════════════════════════════════════════════════════════

Do not continue. Wait for user to fix the issue and re-invoke.

Common errors:

• "Spec file not found" → User provided a path to a file that doesn't exist
• "Spec file is empty" → User provided an empty file with no content
• "Expected a spec file, got a directory" → User provided a directory path instead of a file

Handle conflict (if present):

If conflict is present in output, this means CLAUDE_CODE_TASK_LIST_ID was set and the task list already has tasks. Use AskUserQuestion:

• Question: "CLAUDE_CODE_TASK_LIST_ID is set to '{task_list_id}' which already contains {existing_task_count} tasks. Proceeding will overwrite these tasks with the deep-plan workflow. Continue?"
• Options:

- "Proceed and overwrite existing tasks" - Re-run with --force flag - "Exit" - Stop here, user can unset CLAUDE_CODE_TASK_LIST_ID and retry

If user chooses "Exit": Stop and tell user to unset CLAUDE_CODE_TASK_LIST_ID If user chooses "Proceed": Re-run setup-planning-session.py with --force flag added.

Handle no task list ID (mode == "no_task_list"):

If mode == "no_task_list", this is a fatal error. The workflow cannot proceed without a task list ID. Use AskUserQuestion:

Question: "Cannot proceed: No task list ID available. The SessionStart hook may not have run. How would you like to proceed?"
Options:
  - "Start a fresh session" (Recommended) - Exit Claude and start a new session
  - "Show troubleshooting steps" - Display the error_details.troubleshooting steps

If user chooses "Start a fresh session":

Please exit this Claude session and start a new one. The SessionStart hook
will capture the session ID on startup.

Command: claude --plugin-dir <plugin_path>

If user chooses "Show troubleshooting steps": Display each item from error_details.troubleshooting and STOP.

DO NOT PROCEED past step 4 if this error occurs.

Verify tasks are visible:

After the script completes successfully, run TaskList to verify the workflow tasks are visible. The output tasks_written shows how many task files were written.

Reading session context: After task writing, the task list includes context tasks with values in their subjects:

• plugin_root=... - extract path after =
• planning_dir=... - extract path after =
• initial_file=... - extract path after =
• review_mode=... - extract value after =

Print status:

Planning directory: {planning_dir}
Mode: {mode}

If mode == "resume":

Resuming from step {resume_from_step}
To start fresh, delete the planning directory files.

If resuming, skip to step {resume_from_step} in the workflow below.

Workflow

Note: All scripts use {plugin_root} from step 1's validate-env.sh output.

6. Research Decision

Read {plugin_root}/skills/deep-plan/references/research-protocol.md for details.

1. Read the spec file (find task with subject starting with initial_file= and extract path)
2. Extract potential research topics from the spec content (technologies, patterns, integrations)
3. Ask user about codebase research needs (existing code to analyze?)
4. Ask user about web research needs (present derived topics as multi-select options)
5. Record which research types to perform in step 7

Always include testing - either research existing test setup (codebase) or ask about preferences (new project).

7. Execute Research

Read {plugin_root}/skills/deep-plan/references/research-protocol.md for details.

Based on decisions from step 6, launch research subagents:

• Codebase research: Task(subagent_type=Explore)
• Web research: Task(subagent_type=web-search-researcher)

If both are needed, launch both Task tools in parallel (single message with multiple tool calls).

Important: Subagents return their findings - they do NOT write files directly. After collecting results from all subagents, combine them and write to <planning_dir>/claude-research.md.

Skip this step entirely if user chose no research in step 6.

8. Detailed Interview

Read {plugin_root}/skills/deep-plan/references/interview-protocol.md for details.

Run in main context (AskUserQuestion requires it). The interview should be informed by:

• The initial spec (from initial_file)
• Research findings (from step 7, if any research was done)

9. Save Interview Transcript

Write Q&A to <planning_dir>/claude-interview.md

10. Write Initial Spec

Combine into <planning_dir>/claude-spec.md:

• Initial input (read the file from task with subject initial_file=...)
• Research findings (if step 7 was done)
• Interview answers (from step 8)

This synthesizes the user's raw requirements into a complete specification.

11. Generate Implementation Plan

Read {plugin_root}/skills/deep-plan/references/plan-writing.md before writing anything.

Create detailed plan → <planning_dir>/claude-plan.md

CRITICAL CONSTRAINTS (from plan-writing.md):

• Plans are prose documents, not code
• ZERO full function implementations - that's deep-implement's job

Write for an unfamiliar reader. The plan must be fully self-contained - an engineer or LLM with no prior context should understand *what* we're building, *why*, and *how* just from reading this document. But it does not need to see full code implementations

12. Context Check (Pre-External Review)

Run:

uv run {plugin_root}/scripts/checks/check-context-decision.py \
  --planning-dir "<planning_dir>" \
  --upcoming-operation "External LLM Review"

Read {plugin_root}/skills/deep-plan/references/context-check.md for handling the output.

• If user chooses "Continue", proceed to step 13
• If user chooses "/clear + re-run", they will restart with fresh context (file-based recovery resumes here)

13. External LLM Review

Read {plugin_root}/skills/deep-plan/references/external-review.md for the full protocol.

Check review_mode from task with subject review_mode=... and follow the appropriate path:

• external_llm → Run review.py script
• opus_subagent → Launch opus-plan-reviewer subagent
• skip → Skip to step 16

14. Integrate External Feedback

Analyze the suggestions in <planning_dir>/reviews/.

Remember that you are the authority on what to integrate or not. It's OK if you decide to not integrate anything.

Step 1: Write <planning_dir>/claude-integration-notes.md documenting:

• What suggestions you're integrating and why
• What suggestions you're NOT integrating and why

Step 2: Update <planning_dir>/claude-plan.md with the integrated changes.

15. User Review of Integrated Plan

Use AskUserQuestion:

The plan has been updated with external feedback. You can now review and edit claude-plan.md.

If you want Claude's help editing the plan, open a separate Claude session - this session
is mid-workflow and can't assist with edits until the workflow completes.

When you're done reviewing, select "Done" to continue.

Options: "Done reviewing"

Wait for user confirmation before proceeding.

16. Apply TDD Approach

Read {plugin_root}/skills/deep-plan/references/tdd-approach.md for details.

Verify testing context exists in claude-research.md. If missing, research (existing codebase) or recommend (new project).

Create claude-plan-tdd.md mirroring the plan structure with test stubs for each section.

17. Context Check (Pre-Section Split)

Run:

uv run {plugin_root}/scripts/checks/check-context-decision.py \
  --planning-dir "<planning_dir>" \
  --upcoming-operation "Section splitting"

Read {plugin_root}/skills/deep-plan/references/context-check.md for handling the output.

• If user chooses "Continue", proceed to step 18
• If user chooses "/clear + re-run", they will restart with fresh context (file-based recovery resumes here)

18. Create Section Index

Read {plugin_root}/skills/deep-plan/references/section-index.md for details.

Read claude-plan.md and claude-plan-tdd.md. Identify natural section boundaries and create <planning_dir>/sections/index.md.

CRITICAL: index.md MUST start with a SECTION_MANIFEST block. See the reference for format requirements and examples.

Write index.md before proceeding to section file creation.

19. Generate and Write Section Tasks

Run generate-section-tasks.py to write section tasks directly to disk:

uv run {plugin_root}/scripts/checks/generate-section-tasks.py \
  --planning-dir "<planning_dir>" \
  --session-id "{DEEP_SESSION_ID}"

IMPORTANT: If DEEP_SESSION_ID is in your context, you MUST pass it via --session-id. This ensures tasks work correctly after /clear commands. If it's not in your context, omit --session-id (fallback to env var).

What this script does:

1. Reads sections/index.md to get the section list
2. INSERTs batch and section tasks starting at position 19
3. SHIFTS Final Verification and Output Summary to positions after section tasks
4. Updates all dependencies to reflect new positions

Handle based on result:

• If success == false: Read error and fix the issue (common: missing/invalid SECTION_MANIFEST in index.md, no DEEP_SESSION_ID). Re-run until successful.
• If state == "complete": All sections already written, skip to Final Verification.
• Otherwise: Tasks were written successfully.

Verify section tasks are visible:

After the script completes successfully, run TaskList to see the updated task structure. The output tasks_written shows how many task files were written (section tasks + Final Verification + Output Summary).

Task positions after insertion:

• Position 19+: Batch and section tasks
• Final Verification: Position 19 + section_task_count
• Output Summary: Position 19 + section_task_count + 1

Task list includes batch coordination tasks (subjects like "Run batch 1 section subagents") and section tasks (subjects like "Write section-01-setup.md"). Sections are blocked by their batch task, enabling parallel execution within each batch.

20. Write Section Files (Parallel Subagents)

Read {plugin_root}/skills/deep-plan/references/section-splitting.md for the batch execution loop.

For each batch:

1. Mark batch task in_progress (find by subject "Run batch N section subagents")
2. Run generate-batch-tasks.py --batch-num N → get JSON with prompt_files array
3. Launch Task calls for ALL prompt files in a single message (parallel execution)
4. Each Task: subagent_type="section-writer", prompt="Read {prompt_file} and execute the instructions."
5. Wait for all subagents to complete
6. Verify section files were written (SubagentStop hook writes them automatically)
7. Mark each section task completed (find by subject "Write {filename}")
8. Mark batch task completed
9. If more batches remain, repeat from step 1 with next batch number

Validation After Each Batch:

Hooks execute in isolation - Claude doesn't see success/failure. After subagents complete:

ls {planning_dir}/sections/section-*.md | wc -l

Compare count to expected sections. If any files are missing:

1. Re-run the missing section's subagent
2. If still failing, fall back to manual: parse subagent response JSON and Write the file

21. Final Status & Cleanup

Verify all section files were created successfully by running check-sections.py one final time. Confirm state is "complete".

22. Output Summary

Print generated files and next steps.

Resuming After Compaction

CRITICAL: When resuming this workflow after context compaction, the detailed instructions from this file are lost. The task list is preserved but may not have enough detail. Follow these rules:

1. ALWAYS read the reference file for your current step before proceeding

- Task descriptions include hints like "(read section-index.md)" - follow them - Reference files are in {plugin_root}/skills/deep-plan/references/ - Get plugin_root from task with subject plugin_root=...

1. NEVER skip steps - follow the task list exactly in order

- If a task says "Run generate-section-tasks.py", run the script - If a task says "use section-writer subagents", use subagents (don't write files directly) - You can always re-read in the /deep-plan skill if unsure

1. If message says "MISSING PREREQUISITE" - a required file is missing but later files exist

- This means a step was skipped but later steps ran anyway - Resume from the indicated step and OVERWRITE any subsequent files - Example: if claude-plan-tdd.md is missing but sections/index.md exists, create the TDD plan, then recreate the index (the old index was made without TDD context)

1. Key reference files by step:

- Step 6-7: research-protocol.md - Step 8: interview-protocol.md - Step 11: plan-writing.md - Step 13: external-review.md - Step 16: tdd-approach.md - Step 18: section-index.md (CRITICAL - has required format) - Step 20: section-splitting.md (subagent workflow)