# Sw Loop > Run explicitly requested swarm workflows for feature delivery. - Skill: `kevinslin-skills/sw-loop` (Agent Skill) - Install (CLI): `npx skillmds@latest add kevinslin-skills/sw-loop` - Raw SKILL.md: https://api.skillmd.com/api/skills/kevinslin-skills/sw-loop/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: kevinslin (https://skillmd.com/u/kevinslin-skills) - Updated: 2026-09-10 - Page: https://skillmd.com/skills/kevinslin-skills/sw-loop --- # Sw Loop ## Overview Use this skill to run multi-agent feature work as a controlled swarm. Keep orchestration disciplined: define the critical path, delegate bounded sidecars, integrate results centrally, and do not let review or verification turn into unowned follow-up. Treat the active milestone stack as sticky session context. Once a thread has an approved bundle, keep that bundle in the thread state and only restate it when the stack or phase changes. ## Entry Gate 1. Restate the goal, required deliverable, and completion condition. 2. Confirm whether the user wants full execution or only planning/review. 3. Treat explicit decisions in the current task, parent handoff, or approved spec as settled. If older architecture disagrees, record the narrow approved delta and update only authorized repository-owned material; do not reopen the decision or change unrelated canonical sources. 4. Stop after the spec-and-review phase when either of these is true: - ambiguity remains that affects implementation shape - the user asked to review the plan before coding 5. Prefer `$dev.loop` alone for simple single-threaded work. Use this skill only when multiple agents materially help. ## Checklist Gate - Create a swarm checklist before starting the loop. - Before writing any durable spec, checklist, flow doc, proof, review artifact, runbook, or long-lived project note, invoke `$mem` to resolve the intended base, root, schemas, and file rules. Do this by artifact intent, not by path shape: `$mem` roots may be anywhere and may not contain `.mem`. - Store the checklist in the active spec folder when using `$mem schema` `specs`. - Store the checklist in a temporary folder for all other workflows. - Include every required gate: spec, one-pass spec review, implementation, test-quality cleanup, review swarm, review fixes, final evidence-backed simplification audit, verification, implementation flow doc, PR push, and any user-requested stopping condition. - Translate explicit user completion requirements into checklist rows before starting. If the user names a live proof suite, negative cases, inline screenshots, CI green, a PR push, or a "do not stop until" condition, those are required rows, not optional follow-up notes. - Check off items one by one as each gate is actually complete. Do not bulk-check items at the end. - Do not consider the loop done until every checklist item is checked off. ## Swarm Layout - Manager track: use `$sw-ctrl spec` to own intake, decomposition, delegation, monitoring, and integration. - Spec track: use `$specy` to create a feature spec with milestones, risks, and explicit verification targets. - Pre-implementation review: run one `$dev.review` pass against the spec and apply straightforward improvements before coding. Do not loop on the spec. - Implementation track: use an available coding subagent when bounded delegation is justified; otherwise implement locally. Keep `$dev.loop` responsible for the approved plan, phase gates, and integration. - Test-quality cleanup: after implementation, remove low-value branch tests or replace them with higher-level outcome-focused coverage before review. - Post-implementation review swarm in parallel: trigger:loop `$dev.review` passes for code, explicit simplification/deslop, documentation, and dead code. - Final simplification gate: after review fixes, audit the current diff for one canonical owner, unnecessary mechanisms, duplicate tests, and avoidable churn. - Verification track: run only the local Verify phase of `$dev.loop` in a separate subagent after review fixes land. Do not push or create a PR during this phase. - PR push track: after verification succeeds and the implementation flow doc is written and validated, run the separate `$dev.loop` Push phase, including `trigger:push-pr`, unless the user explicitly said not to push. ## Workflow ### 1. Orchestrate - Adopt the `$sw-ctrl` manager role immediately. - Identify the critical path and keep the immediate blocker local. - Delegate only sidecar work that is concrete, bounded, and non-overlapping. - Keep a short session header for the current milestone: goal, active skill bundle, current phase, and unresolved blockers. Reuse it across turns instead of re-announcing the full bundle each time. ### 2. Plan and Gate - Use `$specy` to create a feature spec first unless one already exists. If the spec is durable project memory, route the artifact root through `$mem` before creating or updating the spec folder. - Preserve every approved outcome while choosing the smallest end-to-end design: one source of truth, actual lifecycle ownership, existing contracts and SDK types, and distinct discriminators for distinct concepts. Defer speculative resources, drivers, compatibility paths, and coordination mechanisms. - Keep source-backed security, isolation, authorization, and immutable-state invariants intact. Simplification removes unnecessary mechanisms, not approved capabilities or required protections. - Run one `$dev.review` pass against the spec and incorporate actionable recommendations. Do not loop on the spec. - Stop and return the spec to the user when ambiguity remains or the user asked to review the plan before implementation. ### 3. Implement - Use an appropriate available coding subagent when code-writing work can be delegated with bounded, non-overlapping ownership; otherwise implement locally. Give each delegated agent the approved plan, explicit file/module ownership, expected tests, and the relevant `$dev.loop` phase context. - Keep `$dev.loop` as the implementation workflow owner for plan execution, integration, verification planning, cleanup, and delivery gates. - Keep ownership clear when delegating code changes. Assign files or modules and remind subagents they are not alone in the codebase. - Integrate changes centrally instead of letting multiple agents edit the same surface opportunistically. - If the approved bundle has not changed, continue from the existing session context rather than re-describing the same orchestration stack. ### 4. Clean Up Low-Value Tests - Remove low-value tests added on the branch, especially tests that verify their own monkeypatched behavior. When possible, elevate them to higher-level tests that verify outcomes and are not coupled to implementation. - Record a `keep`, `merge`, or `delete` disposition for every added or modified test. Delete or consolidate each case without a distinct production outcome, preserve requested end-to-end and unique security proof, and rerun the surviving meaningful coverage. - Do not check off test-quality cleanup while a useless test remains or the disposition and rerun lack concrete file-level evidence. - Complete this cleanup after implementation and before starting review. ### 5. Review Swarm Run the initial `trigger:loop` `$dev.review` passes in parallel against one identified source snapshot, with disjoint scopes. After fixes, review the delta and affected behavior; carry forward inspected, still-applicable findings and audit coverage for unchanged areas. Widen a follow-up only for a stated impact or evidence gap, not simply because a fresh reviewer takes over. Review scopes: 1. Regular code review: find correctness issues, regressions, missing tests, and risky abstractions. 2. Simplification/deslop review: identify duplicate ownership, competing sources of truth, parallel implementations, unnecessary adapters, low-value tests, and avoidable generated-artifact or lockfile churn; describe the smaller safe design with concrete files and deletion candidates. 3. Documentation review: find README, flow docs, design docs, specs, or other docs that should change because of the implementation. 4. Dead code review: find obsolete code paths, compatibility shims, stale state, unused functions, or docs that can now be deleted. Require each reviewer to return concrete findings with file references and proposed actions. Require the code reviewer to complete the evidence-backed Simplicity Audit and Test Audit with explicit test dispositions; a bare `CLEAN` does not satisfy either gate. If a proposed fix is straightforward and does not require user input, apply it in the subagent or integrate it locally. If the fix changes product direction, policy, or unclear ownership, bubble it up for human review instead of guessing. After review fixes land, account for the complete final diff in the simplification review: retain applicable audit evidence and recheck affected areas. Do not mark the simplification checklist row complete or start verification while a material avoidable complexity finding remains. ### 6. Verify - After review fixes land, spawn a separate subagent to run only the local Verify phase of `$dev.loop`. Do not push, create a PR, or invoke the Push phase here. - Apply `$dev.loop`'s evidence-reuse and scope rules. Give the verifier the existing proof records and source delta; it checks applicability and runs missing or invalidated checks, including those required by review findings. Independent verification is not an automatic repeat of every successful live run. - Run independent checks in parallel only with non-overlapping mutable resources; keep one owner per shared database or cluster operation. - For live approval/channel suites, require the verifier to report each requested scenario row separately as passed, blocked, or not run, with the artifact or screenshot path for that row when visual proof was requested. - Do not summarize a multi-row suite as complete from one positive scenario, unit tests, or synthetic debug helpers when the user requested live end-to-end proof. - Do not treat verification as implicit. It is a dedicated track with its own owner. ### 7. Document Implementation Flow - Before the `$dev.loop` Push phase or `trigger:push-pr`, create or update a `$specy` `flow-doc` for the primary logic path exercised by the implementation. - Resolve the durable flow-doc target through `$mem` before writing. Store it in the `$mem`-resolved active spec folder when using `$mem schema` `specs`, otherwise use the resolved base and flow-doc route. - The flow doc must describe the changed request path, entry points, key state/config transitions, exit behavior, and validation/proof hooks. - Run the `$specy` flow-doc validator and fix failures before continuing. - Check off the implementation-flow-doc checklist row before PR push. Do not push until this gate is complete. ### 8. Push PR - After verification succeeds, review fixes are committed, and the implementation flow doc is validated, run the `$dev.loop` Push phase, including `trigger:push-pr`. - Treat PR push as required for full execution unless the user explicitly requested planning/review only or explicitly said not to push. - Include the PR URL in the final handoff. If push or PR creation fails, report the exact failure and treat the swarm run as incomplete. - Kick of $babysit-pr after the PR is pushed - $slack-notify me with alert once pr is green ## Manager Rules - Do not send multiple agents after the same unresolved question. - Do not wait idly when a sidecar can run in parallel with local work. - Do not trust raw review output without integration and judgment. - Close stale agents after their results are integrated or discarded. - Keep user updates focused on what is local, what is delegated, and what is blocking. ## Completion Only finish when the spec, implementation, review follow-up, verify track, and PR push are all resolved, or when the ambiguity/plan-review gate explicitly requires stopping for user input.