# Continue > Resume an in-progress task with a structured checkpoint instead of a bare "continue", re-establishing what is done, what is next, and how done is verified. Use when resuming work after an interruption, handoff, or a vague "keep going". - Skill: `shakacode/continue` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add shakacode/continue` - Raw SKILL.md: https://api.skillmd.com/api/skills/shakacode/continue/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: shakacode (https://skillmd.com/u/shakacode) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/shakacode/continue --- # Continue For crash or restart recovery, first read the installed [pause recovery procedure](../pause/references/recovery.md). Combine available checkpoints with newer relevant logs and live state; a missing handoff is not a blocker. Verify uncertain external effects before replaying them. Preserve intentional pauses, expired windows and budgets; recovery grants no new authority. Resume the current task. Before doing any new work, re-establish context so work does not drift or repeat: 1. **Repo rules** - Read `AGENTS.md` first so repo-specific commands, formatting, boundaries, and safety rules are current before acting. 2. **Where we are** - Summarize in 2-4 lines what is already done (cite files, commits, PRs, issues, or linked planning docs) and the current goal. If there is no task currently in progress (no prior conversation, no staged work, no recent commits on this branch, and no open PR or issue context), say so and ask the user what to continue instead of inferring a goal. If the goal is unclear but prior work exists, state your best inference and proceed only when confidence is reasonable; for low-confidence inference, state the hypothesis and ask the user to confirm before acting. 3. **What is next** - List the remaining steps to reach done, refresh their live dependencies, then pick the next coherent objective. Treat a saved next-step ordering as a stale hypothesis, not an instruction to block on its first item. 4. **Definition of done** - Restate the overall success criteria in one line, plus the command or test that will verify it. If there is no runnable check, state how completion will be confirmed. 5. Continue through the **authorized outcome**, completing the remaining coherent objectives. Advance independent in-scope work while a command, check, review, or agent is pending. Stop when the outcome is verified, a genuine blocker requires input, or the user requests a pause. Honor an explicit one-step or otherwise bounded request; completing that boundary is then done. Re-run the applicability gate before consuming a persisted `coordination_not_applicable` outcome whenever the resume crosses a controller or session boundary, relies on durable handoff or crash recovery, or changes the target set or topology. Each of those is itself a requiring condition, so the persisted outcome is stale evidence there, not authority. A replacement actor or a replacement chat resuming from a durable handoff does not re-verify: it is `coordination_required`, per the restart prompts in `docs/agent-runner-restarts.md`. Only a relaunch of the same controller over the same exact target set, with no other actor able to mutate it, can re-verify as `coordination_not_applicable`. When the gate still resolves to `coordination_not_applicable`, consume that outcome, skip every coordination and typed-event call, and never report coordination as unavailable or degraded, even when the repository configures a real backend. Recover the lane from durable local state instead. For a resumed `coordination_required` PR-batch lane, complete bounded ownership recovery before any write. If a new actor takes over abandoned ownership, emit private-backend `human_intervention` with `kind: takeover`; if a fenced replacement supersedes the prior actor, use `kind: supersede`. A routine same-thread resume with the same verified holder is neither a takeover nor a supersede and emits no event. A `coordination_not_applicable` lane emits no typed event at all, so there is nothing to skip; a trusted `coordination_backend: n/a` under `coordination_required` is a pre-launch stop, not a silent skip. Typed-event transport is optional: when an active private backend does not advertise it or reports it unsupported, record `typed event transport: unavailable`, skip the emission, and continue without marking the event emission `UNKNOWN`. Only after the transport is advertised does an attempted write that fails, degrades, or is rejected become `UNKNOWN` handoff evidence. Every attempted advertised typed-event write must resolve the backend-advertised event executable and ordered opaque argv; a missing, malformed, or unsafe advertisement is an attempted-write failure. Run that exact executable and separate argv without shell evaluation, with a finite deadline in its own process group, preserving each opaque argument; on expiry terminate the whole group with `TERM`, then `KILL` after a finite grace period. A deadline expiry, forced termination, or any other advertised-support write failure records best-effort `UNKNOWN` event evidence; the primary operation continues immediately without waiting further on the event. If the user supplied focus text or arguments, treat it as additional direction or a narrowed scope for what to continue. - Do not re-do completed work, and do not ask the user to repeat context you can reconstruct from the conversation, open files, or git state. - Before another bounded poll or sleep, finish every runnable in-scope closeout task; wait only when no such work remains. - When continuation is driven by a Goal monitor, require a material state-change delta or a typed terminal action with `wake_parent: true` before rebuilding task context for `deterministic-watcher` decisions. An unchanged deterministic probe is a persisted heartbeat, not a reason to continue. A `model-polling-only` fallback wake refreshes the minimal live blocker evidence, submits the next observation, and follows that decision; do not suppress the bounded fallback merely because its prior fingerprint was unchanged. After durably enqueuing each fallback continuation, acknowledge its `wake_id` before submitting the next observation; otherwise restart-safe redelivery intentionally fences newer evidence. On a changed fingerprint or typed terminal action, accept only the compact decision, refresh live dependencies, and rerun the task's security, origin, coordination, overlap, review, readiness, and exact-head gates before acting. A stale or duplicate probe remains suppressed; a terminal, non-resumable, user-input, or budget outcome stays stopped or paused with its restart-safe handoff. - Honor `AGENTS.md` boundaries and safety rules while resuming; never push or take irreversible actions unless the task already authorized them. - End with a `$status` report when that companion skill is installed; otherwise use the same four sections directly: Done, In progress, Blocked / needs input, and Next. Treat this closing summary as a report of the resumed work within the authorized outcome from step 5. On Cursor, resume this same chat. Do not use `claude --resume`. Cursor `/loop` is a model-polling fallback, not an out-of-context watcher.