686 Task Startup Ff0db09a

Task Startup Protocol

When starting work on a task (new or existing):

{todos}

1. Check git status and handle any uncommitted changes

Before starting any task work, ensure a clean working state and sync with remote:

Local Changes

• Check git status {git_status_scope}
• Address EVERY file shown in git status, not just expected files
• Common missed files: CLAUDE.md, sessions/state files, test outputs

If uncommitted changes exist, present them:

[STATUS: Uncommitted Changes] Found uncommitted changes in the repository:

Modified files:

- file1.py: [brief description of changes if known]
- file2.md: [brief description]

Untracked files:

- newfile.txt
- temp/

How would you like to handle these changes?

- COMMIT: Commit all changes with message (git add -A && git commit -m "...")
- STASH: Stash changes to apply later (git stash push -m "...")
- SELECTIVE: Review each file individually (git add -p)

Your choice:

{git_handling}

Remote Changes

• Check for unpulled changes from origin/{default_branch}
• Pull any remote changes BEFORE creating new branches (prevents conflicts)
• If both local and remote changes exist:
  1. Commit or stash local changes first
  2. Pull remote changes
  3. Resolve any conflicts if they arise
  4. Unstash if needed

Note: If resuming work on an existing branch: {resume_notes}

2. Create/checkout task branch{submodule_branch_todo}

Check task frontmatter for branch name{submodule_context}, then create/checkout branches.

Branch Creation/Checkout

1. Start from {default_branch} branch (already clean and synced from step 1)
2. Create the new task branch from the updated {default_branch}

{submodule_management_section}

3. Verify context manifest for the task

Check the task file for a Context Manifest section and report status:

If missing:

[STATUS: Context Manifest] ✗ No context manifest found in task file

Running context-gathering agent to analyze the task and create comprehensive context...

If present:

[STATUS: Context Manifest] ✓ Context manifest found and loaded

• If present, proceed to the next step to load it
• If missing, you must use the context-gathering agent to analyze this task and create a context manifest

4. Gather context for the task

Based on the manifest:

• Read the narrative explanation to understand how everything works
• Note technical reference details for implementation
• Check environmental requirements
• Note file locations for where changes will be made
• Read prescribed file segments (if any)

Verify branch state:

• Confirm you're on the correct branch that matches the task's expected branch
• If the branch doesn't match the task branch, STOP and fix it first

{directory_guidance}

5. Initial Discussion & Planning (After Todos Complete)

Note: This step happens AFTER all startup todos are complete and you've automatically returned to discussion mode.

After gathering context:

1. Analyze the task requirements thoroughly
2. Propose implementation plan with structured format:

[PLAN: Implementation Approach] Based on the task requirements, I propose the following implementation:

□ [Specific action 1] → [Expanded explanation of what this involves]

□ [Specific action 2] → [Expanded explanation of what this involves]

□ [Specific action 3] → [Expanded explanation of what this involves]

To approve these todos, you may use any of your implementation mode trigger phrases: {implementation_mode_triggers}

3. Iterate based on user feedback until approved
4. Upon approval, convert proposed todos to TodoWrite exactly as written

IMPORTANT: Until your todos are approved, you are seeking the user's approval of an explicitly proposed and properly explained list of execution todos. Besides answering user questions during discussion, your messages should end with an expanded explanation of each todo, the clean list of todos, and no further messages.

6. Work Mode

For the duration of the task:

• Discuss before implementing
• Constantly seek user input and approval

Once approved, remember:

• Immediately load your proposed todo items exactly as you proposed them using ToDoWrite
• Work logs are maintained by the logging agent (not manually)

After completion of the last task in any todo list:

• Do not try to run any write-based tools (you will be automatically put into discussion mode)
• Repeat todo proposal and approval workflow for any additional write/edit-based work

Example First Message

"I've loaded the context for [task]. Based on the manifest, I understand we're working on [summary]. The last work log entry shows [status].

Here is the work I think we should accomplish:

• [todo item]: [full explanation]
• [todo item]: [full explanation]
• ...

Use any discussion mode trigger phrase if you approve the proposal or tell me how I should adjust the proposal."