Research mode: lightweight planning without git operations

[chanakyav]

Add a lightweight research/planning mode that runs the agent without creating branches, PRs, or commits.

Use Case

While one autopilot task is implementing a feature, you want to plan the next task. Run a research session that analyzes the codebase, produces a plan, and saves it — so when the first task completes, you can immediately start the next one with full context.

Proposed Design

# Research mode — no branch, no PR, just analysis
autopilot research --prompt "Analyze how pagination works in this codebase and plan how to add it to all API endpoints"
# => task abc123 (mode: research)

# Later, use the research as context for implementation
autopilot start --prompt "Add pagination to all API endpoints" --context abc123

Behavior

• Runs copilot -p with a research-focused prompt (no git operations instructed)
• Agent output saved to session files (same as today)
• No branch creation, no PR, no git checkout
• Task reaches COMPLETE when agent finishes
• Session files can be referenced by future tasks via --context <task_id>

Research Prompt Template

The prompt should instruct the agent to:

1. Analyze the relevant code
2. Identify files that need changes
3. Outline an implementation plan
4. Note any risks or edge cases
5. Write findings to a structured output file

--context flag

When --context <task_id> is passed to autopilot start:

1. Load the research task's session file
2. Prepend it to the implementation prompt as context
3. Agent has full plan before starting implementation

Implementation Notes

• New task_mode: "research" value
• Simplified orchestrator: INIT -> RESEARCH -> COMPLETE (no VERIFY_PR, no review loop)
• No branch locking needed (no git operations)
• Research runs in current working directory (read-only from git perspective)
• Could run in parallel with other tasks safely (no git mutations)

Open Questions

• Should research mode have a timeout? (Default agent_timeout_seconds should suffice)
• Should we support --context referencing multiple research sessions?
• Could this evolve into a general "agent task" framework beyond just research?