iflow-cycle

name	iflow-cycle
description	Process many issues hands-off in a row: resolve a queue, then run each through the yolo chain under one up-front confirm. Stops only when input is strictly necessary.
disable-model-invocation	true
issue-flow-version	0.4.2a4

issue-flow — issue cycle (/iflow-cycle)

Follow this skill to process a queue of issues hands-off, one after another, with a single up-front confirmation — the batch equivalent of /iflow-yolo. Each issue runs the full yolo chain (init → plan → start → close yolo, PR auto-merged, switch back to default); the cycle interrupts you only when input is strictly necessary.

Use only when every queued issue is genuinely yolo-fit (small, low-risk, well-specified, test-guarded). A queue of risky changes belongs in the individual commands.

Input — queue spec

• explicit numbers — e.g. 12 15 18.
• label:<L> — every open issue carrying label <L>.
• epic <N> [stage <k>] — the current stage of epic <N> (or stage <k>).
• resume — pick up an interrupted cycle from its state file (see Resuming below).
• onfail:stop (default) / onfail:skip — failure policy (see step 7).
• max:<n> — raise the safety cap (default 10) for this run.
• stay — forward stay to each close so the working copy stays on each issue branch (rarely wanted in a cycle).

Invoke: type iflow cycle in chat, or /iflow-cycle from the slash menu (iflow-cycle also works).

MODEL & EXECUTION DIRECTIVE

Profile: reasoning — Prioritize deep thinking and careful trade-offs over speed or token economy.

In Cursor: switch to a thinking-capable model before invoking this step (not Auto-only).

Keep scope tight to what this step requires.

Resolve project root (multi-root workspaces)

Before any git, gh, or .issueflows/ path operation in this workflow:

Resolution order (stop when unambiguous):

1. Explicit hints in slash input — root:<path>, repo:<folder-basename> (directory name, e.g. cellpy-core), or repo:owner/name.
2. CLI fast path — issue-flow agent resolve [-C <start>] [--from-file <active-file>] [--json]. Use the returned project_root and repo; pass -C <project_root> to other issue-flow agent … subcommands. When the answer came from the workspace registry, the payload sets resolved_via_workspace_default: true.
3. Branch context — exactly one workspace repo whose branch matches ^\d+- → that root.
4. Single scaffold — exactly one .issueflows/ tree visible in the workspace → that root.
5. Workspace default — an issueflow-workspace.toml at the workspace root (created with issue-flow workspace init) may name a default member repo; use it when no scaffold matched above. Tell the user the default was used.
6. Ambiguous → stop and ask; never guess between sibling repos.

After resolution, treat the result as <project_root> and <owner/repo>:

• Git: git -C <project_root> … (or issue-flow agent … -C <project_root> for supported ops).
• GitHub: always gh … --repo <owner/repo> — never rely on gh's implicit cwd default.
• Paths: all .issueflows/… paths are under <project_root>.

When .issueflows/04-designs-and-guides/multi-repo-workspaces.md exists, read it for layout and cross-repo guidance.

Instructions

1. Resolve the queue. Run issue-flow agent queue <spec> --json (numbers, --label, or --epic). Use its queue (ordered), blocked, and skipped_closed output as the source of truth — do not re-derive the order by hand. If it reports a dependency cycle, stop and show it; nothing runs. If the CLI is unavailable, fall back to reading the issues and ordering by Depends on #N lines yourself, but prefer the CLI.

2. Cap check. If the ordered queue is longer than 10 and the input did not pass max:<n> raising the limit, stop and ask the user to confirm a larger run explicitly. Long unattended runs compound risk.

3. One consolidated confirm (the only planned interruption). Present, in normal prose:

   • the ordered queue (numbers + titles), and which issues are skipped (closed) or blocked (open dependency outside the queue) with the reason;
   • that each issue runs the full yolo chain and its PR is auto-merged;
   • the failure policy (onfail:stop, the default, or onfail:skip — see step 7);
   • the default-branch preflight that must hold before starting (clean tree, tests passing). Require an explicit yes; anything else aborts before any work.

4. Write the cycle state file. After the confirm, write .issueflows/01-current-issues/cycle_status.md — the durable record that makes the run resumable and visible to /iflow-status. Include: the queue spec, the onfail policy, an ISO timestamp, and the ordered queue as a checklist with one line per issue (- [ ] #<N> — <title> — pending). Update this file as the loop progresses (see step 5); it is a normal tracking file (not an issue<N>_* group), so the folder sweep never touches it.

5. Per-issue loop. For each issue in order, from a clean default branch:

   • mark it in-progress in cycle_status.md, then create/switch to its <N>-<slug> branch and follow .cursor/skills/iflow-yolo/SKILL.md verbatim — including its own preflight (refuse on default branch, refuse with dirty unrelated changes, tests pass up front) and its consolidated-confirm step, which the up-front batch confirm in step 3 satisfies (do not re-ask per issue).
   • after the yolo close merges and switches back to the default branch, record the outcome in cycle_status.md (- [x] #<N> — <title> — merged <PR-url>) and continue to the next issue.
   • Every yolo safeguard stays in force. A safeguard that trips is a stop condition (step 6), never a guard to skip.

6. Strictly-necessary-input rule. Between issues the cycle runs unattended. Stop and ask only when: a. tests or lint fail in a way you cannot fix within the current issue's scope; b. a merge is refused or a git pull --ff-only will not fast-forward (divergence); c. the issue spec is ambiguous, contradictory, or turns out not small (yolo's scope check aborts); d. an action would fall outside the confirmed queue (touching an unlisted issue, an unrelated dirty file, a destructive op). Anything else — routine implementation choices, passing tests, clean merges — proceeds without asking.

7. Failure policy (from the onfail: token; default stop). When a stop condition (step 6) trips on an issue:

   • onfail:stop (default) — halt the cycle: finish no further issues, leave the repo on the default branch, clean (the in-flight issue's branch stays as-is for the user to inspect), record the stop reason and the not-reached issues in cycle_status.md, and report. Do not attempt the rest of the queue.
   • onfail:skip — park and continue: record the failure against that issue in cycle_status.md (- [~] #<N> — <title> — failed: <reason>), park its work per .cursor/skills/iflow-pause/SKILL.md conventions (status note + move to 02-partly-solved-issues/), return to a clean default branch, and proceed to the next queued issue. A skip never bypasses a yolo safeguard — it records the trip and moves on.

8. Finish. When the queue is exhausted (or halted), finalize cycle_status.md (mark it - [x] Done) and move it to .issueflows/03-solved-issues/cycle_status_<YYYY-MM-DD>.md so it is archived, not re-detected as in-flight.

9. Batch report. Summarize the whole run: per issue — number, title, PR URL, merge result (merged / queued via --auto / failed-and-skipped / not reached), and duration if tracked; then the queue items skipped (closed), blocked (with blockers), and — on a halt — the stop reason and which issues were not reached. End by reminding the user to run /iflow-cleanup once to prune the merged local branches.

Resuming

/iflow-cycle resume picks up an interrupted cycle:

1. Read .issueflows/01-current-issues/cycle_status.md. If it is missing, tell the user there is no in-flight cycle and stop.
2. Take the remaining issues (those still pending / in-progress) and re-verify them with issue-flow agent queue <original-spec> --json — an issue that has since closed, or become blocked, is dropped/deferred with a note (state can move while a cycle is paused).
3. Do not re-ask the original consolidated confirm for the unchanged remaining items — the batch was already authorized. Ask again only if the re-verified queue differs materially from what was confirmed (new blockers, added issues), and only about the delta.
4. Continue the per-issue loop (step 5) with the same onfail policy recorded in the file.

Parallel dispatch (experimental, opt-in)

By default the cycle is sequential — one issue fully lands before the next starts. When the input passes parallel:<n> and the harness supports background execution, provably independent issues may run concurrently (up to n at a time). This is experimental; the sequential path above is always the default and is never weakened to enable it.

• Only independent issues qualify. Use issue-flow agent queue's independent list — issues with no dependency relation (either direction) to any other queue member. Everything else runs sequentially.
• Harness gate. If you cannot confirm the harness supports background execution (worktrees + parallel agents/subagents), refuse parallel:<n> and run sequentially — never pretend to parallelize.
• Worktree per issue. git worktree add ../<repo>-<N> <N>-<slug> so each issue has an isolated tree; run the yolo work there.
• Serialize merges. Never merge PRs concurrently — the coordinating session merges them one at a time on the default branch, pulling between merges and rebasing/retrying on a non-fast-forward or CI refusal.
• Shared files via the coordinator only. Parallel workers must not each edit HISTORY.md; each leaves its changelog bullet in its issue status file / PR body, and the coordinator appends them in merge order during the serial merge step.

When in doubt, prefer the sequential run — parallel dispatch trades safety for speed and every one of the rules above must hold.

Constraints

• Off-path: /iflow never auto-dispatches to /iflow-cycle; it is an explicit, deliberate batch action.
• Sequential is the default and floor. Parallelism (parallel:<n>) is opt-in and experimental; refusing it must always leave a working sequential run.
• Never weaken a yolo safeguard to keep the cycle moving — safeguards are stop conditions, not obstacles.
• Never run /iflow-cleanup from this skill; batch branch deletion still needs the user to see the merged PRs first.
• One consolidated confirm covers the batch; never silently expand the queue beyond what was confirmed.
• cycle_status.md is the single source of truth for an in-flight cycle: keep it current so resume and /iflow-status stay accurate. It is not an issue<N>_* group, so the folder sweep leaves it alone; archive it (step 8) when the run ends.