Hand off a coding-agent plan without losing context

A reliable handoff tells the next implementer which plan is authoritative, what state it is in, which tasks are complete, what repository context applies, and how completion will be proven. Forwarding the final chat message carries text but usually loses that operational state.

1. 1

   Finish the planning turn

   Store the plan and leave it in draft while the user reviews it.

2. 2

   Approve the artifact

   Move the exact plan id to approved; do not approve a stale copy in chat.

3. 3

   Start implementation

   The implementing session checks status, moves it to in-progress, and reads the current body.

4. 4

   Record progress

   Check completed tasks and revise the same plan when implementation changes the approach.

5. 5

   Close with proof

   Run the required checks and mark the plan done only when the outcome is actually delivered.

hsp status y5bn0e
hsp get y5bn0e
hsp status y5bn0e in-progress
hsp tasks y5bn0e
hsp check y5bn0e 1 2

The plan page's Open in control prepares a new Codex thread, Claude Code session, or Cursor agent prompt. It does not silently execute the plan: the destination opens with a prompt ready for human review.

For a plan owned on the local machine, the handoff can include its local path and working directory. For a shared hosted plan, it points the receiving agent at the URL instead, avoiding meaningless or private filesystem paths.

When one handoff would be too large, store an ordered stack. Each plan remains blocked until its dependency is done, and hsp next returns the first actionable step rather than asking an agent to infer order from filenames.

hsp stack 01-schema.md 02-api.md 03-ui.md
hsp next
hsp status <first-id> done
hsp next