Add --prompt flag to resume for additional instructions

[chanakyav]

Add an optional --prompt flag to autopilot resume so users can inject additional instructions into the fix cycle.

Use Case

You have a PR with preexisting review comments and also need merge conflict resolution or other custom work:

autopilot resume --pr 16173 --prompt "Also fix merge conflicts with main"

Or when resuming a PR where the review comments alone don't capture everything needed:

autopilot resume --pr 42345 --prompt "The reviewer wants us to switch from polling to webhooks"

Current Behavior

resume creates a placeholder prompt "(resumed from PR #N)" that's stored in the DB but never used — the FIX agent only sees the Copilot review comments from PARSE_REVIEW.

Proposed Behavior

• Add --prompt / -p flag to resume (optional, not required)
• Store the prompt in the task record
• In _do_fix, if the task has a non-placeholder prompt, prepend it to the fix prompt as an "Additional Instructions" section
• If no --prompt given, behavior is unchanged (backward compatible)

What the Agent Would See

## Additional Instructions

Also fix merge conflicts with main

## Copilot Review Feedback to Address

### Inline Comments (3)
...

Implementation

1. Add --prompt / -p to resume parser in cli.py
2. Pass it to create_task() instead of the placeholder string
3. In _do_fix (orchestrator.py), check if self.task["prompt"] is a real prompt (not the placeholder) and pass it to fix_prompt() as additional context
4. Update fix_prompt() in prompts.py to accept an additional_instructions param

~20 lines of changes + tests.

[chanakyav]

Addressed in PR #71 (merged). Implemented as --context / -c instead of --prompt / -p to avoid ambiguity with start --prompt, but otherwise matches this proposal exactly:

• autopilot resume --pr N --context "..." passes additional instructions
• Stored in the task's prompt field
• Piped through to fix_prompt() as a task_context parameter (renders as ## Task Context section)
• Backward compatible — no context = unchanged behavior

---

🤖 autopilot-loop