Resuming or handing off — after a crash, session limit, compaction, or model handoff; picking up a HANDOFF or state file; closing a session with work in flight; the SessionStart continuity line. Also on "resume safely" or "write the handoff". Skip for a fresh session with nothing inherited.
Scanned 9/19/2026
Install to Claude Code
npx -y skills add debabsah/superstack --skill superstack-continuity --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Superstack Continuity?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/debabsah-superstack-continuity)More formats (shields.io, HTML) on the badges page.
---
name: superstack-continuity
description: Resuming or handing off — after a crash, session limit, compaction, or model handoff; picking up a HANDOFF or state file; closing a session with work in flight; the SessionStart continuity line. Also on "resume safely" or "write the handoff". Skip for a fresh session with nothing inherited.
---
# superstack-continuity
Sessions die — limits, crashes, compaction, handoffs. Continuity is engineering, not memory: the conversation is scratch; files are the project.
## Resuming: verify before continuing
Never continue from what you remember or what a handoff asserts. First, ground truth:
1. `git status` + `git log` — what actually landed, versus what the last session's record claims landed.
2. The suite (or the project's oracle rows) — the state you inherit is the state that passes now, not the state that was reported.
3. **Inherited state is a claim, not evidence.** A HANDOFF, a task file's `next:` pointer, another session's report, your own memory of "where we were" — audit each against disk before acting on it. Handoffs have asserted hooks that never existed; distrust is the posture that caught it. One inherited class is expired outright rather than audited: an unattended-work go-ahead died with the session it was given to — continuing unattended needs the owner's word in this one (D-81).
4. If a subagent or background task died mid-work, classify what it completed (its commits, its files) before re-dispatching.
## Closing: leave the work resumable
When a session ends with work in flight, the last act is the record, addressed to its actual reader:
- **For your future self / the next model session:** the task file's `next:` pointer updated to the true next action; decisions since last update appended; anything half-done named as half-done.
- **For a different tool or model:** a self-contained pack — constraints, commands, "done looks like" per step — assuming none of your context survives.
- **For the human:** a STATUS-labeled report that leads with what is NOT done.
Pin pause points in commits, not prose ("pausing here" in a commit message beats a paragraph the next session never reads).
## When a plan is active
If `.superstack/plans/` holds an ACTIVE plan, superstack-execute specializes this skill: session close means camping the plan's frontier block (that stamp IS the handoff), and resuming means the torn check — a frontier not stamped `camped` is treated as a crash, and the verify-before-resume ritual runs harder, not softer.
## Mid-work anchors
Long turns write their own lifeline as they go: commit per verified step, decision records appended at re-decide, lessons appended to the overlay when confirmed — so an unexpected death loses minutes, not the session. Multi-tab or cross-machine coordination happens only through durable files, never through assumptions about what the other session did.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!