Resume Session

Reconstruct where a prior OMP session left off using the resume_session tool for
ALL discovery and reading, with two mandatory stops: the user chooses the
session, and the user confirms before any work resumes.

Non-negotiable rules

MUST Use resume_session for everything. Never identify sessions by reading .jsonl files under ~/.omp/agent/sessions, running cat/tail/grep on transcripts, or running git log yourself.
MUST Load exactly one session -- the one the user picks. Never read a second session's transcript.
MUST The two STOP gates below are hard. Until gate 1 is cleared, only mode: "list" is allowed. Until gate 2 is cleared, do not read files, run git, or start work.
MUST Session cwd in a different worktree than yours → confirm target worktree with the user before reading or editing any files.

Workflow

1. List sessions -- your first and only action so far. Call
   resume_session with mode: "list" (defaults to the git repo root of the
   session cwd; pass path for another repo). It prints a newest-first summary:
   id, last-active, turns, branch, worktree:, title, and a ↳ left off: line.

   • Worktree-aware by default. Enumerates every worktree (git worktree list)
     and matches transcripts by their recorded cwd. Pass worktrees: false to
     scan only the current checkout.
   • Git-activity overview. When the repo has more than one worktree, also
     prints a "Worktree git activity" block (most-recently-committed first) with
     branch, last-commit time + subject, and a ✎ dirty mark. Pass git: false to skip.
   • Branch label = the branch the session worked on (recovered from its
     transcript), not the checkout's branch now. If the worktree has since
     switched, the row shows [worked-on → worktree now on current] -- a drift
     warning that on-disk files no longer match that session. Don't read the
     second branch as the session's work.

2. STOP. Present the list and let the user choose. Show the newest few rows
   including worktree: and ↳ left off: lines, and ask which to resume.
   You may recommend, but wait for their answer -- do not pick for them.

   • Only exception: if the user already gave a session id, skip to step 3.

3. Read that one session. Call resume_session with mode: "read" and
   session: "<id>" (newest 8 turns, filtered, newest-first; an id prefix is
   enough). Anchor on the Latest plan / todo state block. Stop reading when
   you can state what was being done and what remains. Page back with
   offset + turns if still unclear. Never open another session.

4. STOP. Summarize, surface ambiguities, and ask. State: the goal, the last
   action, current todo/plan state, branch/cwd, and what is incomplete. List
   unrecorded decisions, half-done work, or paths that may be stale. Ask for
   confirmation and any new direction. Wait.

5. Resume. Only after the user confirms: run a quick reality check
   (git status, branch, referenced files exist), then continue from the agreed
   next step.

Notes

• Current user instructions override anything in the transcript; it is evidence of the past.
• Do not silently re-run destructive or outward-facing actions (commits, pushes, deploys) -- reconfirm first.
• Each call prints an estimated uncached-token cost; report the total used vs. full transcript size.
• Set include_thinking: true only when the text record shows a logic gap, incomplete sentence, or unexplained branch that tool calls alone cannot resolve.
• A compaction marker in the window means earlier turns were summarized away in the original run; the summary is the only record of them.
• This skill resumes a session transcript; it does not read saved handover files.

See references/transcript-format.md for store locations, record schema, and filtering/paging.