Pause Mission

Pause an SDD mission into its plan

An SDD mission's durable handoff is its plan brief at .agents/plans/<cr-ref>.plan.md. A pause writes this session's state back into that plan so any next session can pick the mission up by reading it — no special tool required (resume-mission is one convenient reader; the plan stands on its own). Capture enough to continue, nothing to relitigate.

Procedure

1. Honor the focus. If the user named what the pause should emphasize, scope the checkpoint to it; otherwise checkpoint the whole live frontier.

2. Locate the plan — scaffold it if missing. Find .agents/plans/<cr-ref>.plan.md. If none exists, create a minimal one (frontmatter todos + a ## NEXT anchor) so the checkpoint has a durable home; a pause always leaves a plan behind.

3. Update the todo statuses. In the frontmatter todos, set each touched todo to its true status (pending | in_progress | completed). Mark the one you were on in_progress; never mark a todo completed unless it truly is — partial work stays in_progress with a one-line note of what remains. Add a todo for any new work or decision that surfaced this session.

4. Rewrite the ## NEXT — resume here anchor (create it if absent) so a cold reader can start in under a minute. Order it for pickup, action first:

   • The next action — first, concrete, runnable. One line the resumer can act on immediately: the file or unit to touch and the skill or command to invoke for it — e.g. "build <unit>'s spec + suite via start-mission, then run the project's spec check," never a vague "continue authoring."
   • Blocking decisions — open scope calls and unresolved <!-- open: --> markers, each with its options, so the next session resolves rather than rediscovers them.
   • Findings the commits won't show — only what isn't obvious from the diff and changes the next move. Skip the play-by-play.
   • A pointer to the working method / resolved decisions already in the plan body ("do not relearn — see ## Resolved decisions").

   Lead with the action; put history and findings below it. A resumer needs what to do first and what happened second.

5. Reference, don't restate. Point to commits (SHAs), files (repo-relative paths), design rules, and issues by reference — never paste their contents. The plan is an index to the work, not a copy of it. A file the brief depends on must live in the repo: if the source is a machine-local doc (a plan-mode plan under ~/.claude/plans/…), bring it in as a sibling <cr-ref>.design.md and reference that, never the external absolute path.

6. Keep it commit-message-grade — the safe-to-publish floor. The plan is tracked and committed, so it carries the same floor as the combat log (combat-log-governance): never committed — secrets, keys, prompts, PII, absolute paths, OS usernames/hostnames, $HOME/$USER, or any machine-local reference outside the repo. Describe the decision or its class, not literal payloads or machine-local locations.

7. Commit the checkpoint. Before committing, run the plan-brief leak guard over the brief — node "<sdd-skills>/check-plan-safety/scripts/check-plan-safety.mts" --path <brief> --check (the check-plan-safety engine). A hit blocks the commit: scrub the machine-local reference (bring the content in-repo, reference it repo-relative) and re-run. Then commit the plan update (docs:) so the pause is durable in git and the working tree is clean. Leave no uncommitted work behind.

Reconcile-forward checkpoint

When the mission's live state is "this CR's work is already merged, not something this session ran live" (e.g. resuming a plan and finding its commits already landed on main), treat it as a checkpoint whose input is git/ledger evidence instead of a live conductor session — same procedure above, plus one gate before declaring the plan retirement-ready:

• Only for a CR-bearing mission. A mission that opened a real CR and reached status: approved or implemented gets this check. A non-CR mission (the start-mission escape hatch — no suite-relevant behavior, no CR opened, no gate invoked) needs none; mark it complete on its own merits.
• Run the gate floor. Run checkGateFloor (plugins/sdd/skills/spec-gate/scripts/ check-spec-state.mts, wired into pnpm verify:specs) before marking the plan retirement-ready. A clean floor → mark it. A violation (the status is approved/implemented but the ledger has no matching gate approve line) → do not mark the plan retirement-ready on git evidence alone; record the violation explicitly in ## NEXT so a human resolves it (either the ledger line was genuinely never written, or it needs relocating/re-checking).

Write scope. A checkpoint writes only the plan brief — the todos, the ## NEXT anchor, and (with --approve) the top-level status. It never touches spec.md's status or approval: the mission's contract lifecycle is the gates', not the checkpoint's.

Clear for headless dispatch — --approve

A plain pause never touches the brief's top-level status. When invoked with --approve, the checkpoint additionally sets the brief's status: approved — the human review act that clears the mission for the gateway's headless dispatch queue (the go-signal dispatch selects on; the enum + the three-way status distinction live in the SDD provenance-model). Rules:

• Human act only. Setting approved is a review decision — a person has read the brief (todos, ## NEXT, the run-level leash) and cleared it. A headless automaton never self-approves: with no user channel it may checkpoint progress but must refuse --approve and leave status unchanged (the same positional-authority rule that reserves a human-ratified gate verdict).
• Idempotent. Approving an already-approved brief leaves it approved.
• Flag-only. It writes the one frontmatter field; it does not dispatch the mission (that is the gateway) and does not touch the run-level leash — approved says run it, the leash says how far.

What a good pause is not

• Not a transcript. Capture decisions and the next action, not a play-by-play.
• Not a duplicate. If it already lives in a commit, an ADR, the spec, or an issue, link it.
• Not buried. The next action leads the anchor; if a resumer has to hunt for where to start, the pause failed.
• Not self-referential. The plan never tells its reader to "invoke resume-mission" or explains how to resume — a session reading the plan is already past that, and the ## NEXT — resume here heading orients a human on its own. Open ## NEXT with the work.
• Not final-sounding. A pause is a resumable checkpoint — leave the live frontier and the open questions explicit, not smoothed into a summary that hides where to begin.

The plan is self-sufficient: any session that opens it can continue. resume-mission reads this back conveniently, but is not required.