| name | ar-taskflow-resume |
| description | Resume an ongoing task from where you left off. Use when user says "ar-taskflow-resume" or "resume task". Asks for task folder path and continues from the last phase. |
| disable-model-invocation | true |
Task Flow Resume
Resume an ongoing task by providing the task folder path.
🧭 Learning routing
Follow the rule at <standards_location>/learning-routing.md (resolve standards_location from .claude/config_hints.json): route any learning to a project rule (docs/ai-rules/), a framework improvement (ar-record-improvement), or conversational-only, never personal auto-memory.
Prerequisites
.claude/skill.config must exist (run ar-init-skills if missing)
Workflow
Trigger: User says "ar-taskflow-resume" or "resume task"
Steps:
-
Read skill.config for paths
-
Ask for specific task folder path:
Give me the path to your task folder (e.g., {tasks_folder}/{namespace}-195-my-task)
-
Read task context:
Once user provides the path, read all available files:
raw_prompt.md (required - if missing, invalid task folder)
prompt-understanding.md (if exists)
executive_summary.md (if exists, 2-3 line digest, surface in your resume summary so the user remembers context fast)
execution_plan.md (if exists)
acceptance_criteria.json (if exists, primary source of "what's left")
execution-summary.md (if exists - prioritize this for state)
4a. Determine current phase:
| Files Present | Phase | Next Action |
|---|
Only raw_prompt.md | 0 | Ask clarifying questions, create prompt-understanding.md |
+ prompt-understanding.md | 1 | Create execution_plan.md |
+ execution_plan.md (no code changes) | 2 | Check branch, run smoke test (step 4b), start coding. If acceptance_criteria.json missing, generate it from the prose AC before coding (pre-v6.1 task). |
+ execution_plan.md (code changes exist) | 3 | Run smoke test (step 4b), then continue coding |
+ ticket.md + pr-description.md | 4 | Run smoke test (step 4b), then ask about commit or archive |
4b. MANDATORY pre-product smoke test (Phase 2+ only):
Before resuming any new code work, prove the previous session left the repo in a working state. Skipping this step is how regressions get buried under new commits.
a. Boot the project using the project's documented dev/start command. Look in this order: AGENTS.md → project README.md → the build/script manifest. If the boot command isn't documented anywhere, ask the user once and note it in execution-summary.md for future resumes.
b. Run the touched-path tests, execute the test files listed in execution_plan.md → "Files to change", plus any test referenced in acceptance_criteria.json → verification. Use the project's test command (test_command/verify.full_command from config_hints.json, else the command documented in the repo), force-rerun if the test runner caches results.
c. Re-check the JSON gate (only if acceptance_criteria.json exists): for every criterion currently marked passes: true, re-run its verification step. If a previously-passing criterion now fails, the prior session left a regression, that becomes the first thing to fix, not new feature work.
If boot fails or any previously-green criterion fails:
🚧 Resume blocked, baseline is not green.
Failures:
- {what failed}
Per the resume protocol, regressions must be recovered before new work.
I'll work on these first. OK?
If everything is green: state it explicitly ("Baseline green: N/N criteria still pass, smoke tests pass"), then proceed to the next failing criterion.
Manual Override: If user says "skip smoke test", note it in execution-summary.md → "Last Action: resumed without smoke test (user override)" so the next resume knows the baseline is unverified.
-
Check git branch status (worktree-aware):
git branch --show-current
- If on feature branch matching task → Good, continue
- If on main → May need to create/checkout branch
Cross-check against execution-summary.md. If it records the worktree fields written by ar-taskflow's Pre-Product Worktree Reconciliation, Worktree, Local Branch, Remote Branch, Reconciliation Result, reconcile them against the current HEAD before doing anything:
- If the task was in a worktree (
Worktree: true) on a branch that differs from the current HEAD, do NOT blindly create a new branch. The task's code lives in another worktree. Warn the user and point them to it:
⚠️ This task was last running in a worktree on branch '{Local Branch}',
but the current checkout is on '{current branch}'.
Its code likely lives in a different worktree. Locate it with:
git worktree list
and relocate there (or use a_g_worktree_* helpers) before resuming -
creating a fresh branch here would fork the work.
- If the recorded Local Branch matches the current HEAD → good, you are in the right worktree, continue.
- If
execution-summary.md has no worktree fields (task started before this was added) → fall back to the plain branch check above.
-
Present resumption summary:
Resuming task: {task_name}
Current state:
- Phase: {phase}
- Branch: {branch_status}
- Last action: {from execution-summary.md if available}
Next step: {what will happen next}
Ready to continue? (yes/no)
-
📤 Push docs on resume: After user confirms, push any uncommitted changes in coding_tasks_root. Follow the full Push-Docs Procedure from the ar-taskflow skill, including pull-rebase and intelligent conflict resolution for TasksSummary/*.md and other shared files.
coding_tasks_root=$(dirname "$(jq -r '.paths.tasks_root' .claude/skill.config)")
-
Continue with task-flow phases:
Once user confirms, continue with the appropriate phase from ar-taskflow:
- Phase 1 → Create prompt-understanding.md
- Phase 2 → Create execution_plan.md
- Phase 3 → Code (follow task-flow Phase 3 rules)
- Phase 4 → Finish (create ticket.md, pr-description.md, commit)
Quick Commands
| Say | Action |
|---|
| "ar-taskflow-resume" | Ask for task path and resume |
| "resume task" | Same as above |
Relationship to ar-taskflow
- ar-taskflow = Start a NEW task (ticket-first or ticket-late)
- ar-taskflow-resume = Continue an EXISTING task
Both skills share the same phase definitions (0-4) and follow the same rules for coding, commits, and archiving.
References
See the ar-taskflow skill for:
- Phase definitions and detailed steps
- Safety rules for commits
- Coding conventions
- Documentation requirements