Agent Stack Loops
The path one piece of work takes from request to merged code in six agent stacks, drawn with one grammar.
Each stack is one lane, read left to right. Point at a step or a labelled connector and its details open in the panel on the right. Click to pin them; click again to release. Esc closes the panel. The address of the page records a pinned selection, so a link to it shares it. The two icons in a step list what runs it.
Superpowers
obra/superpowers @22f0b46Jesse Vincent · Claude Code, about 15 other harnesses
/brainstorming, /using-git-worktrees, /writing-plans, /subagent-driven-development or /executing-plans, /test-driven-development, /requesting-code-review, /finishing-a-development-branch. Each skill names the next one in its own text. The skills differ from the README: the worktree is created after the plan, per-task review uses its own prompt, and whole-branch review is not in the README list.pstack
cursor/plugins @2f87c85poteto (Cursor) · Cursor, with a Claude Code adapter
/poteto-mode routes a request to one of 22 playbooks; there is no single pipeline. This lane is our synthesis of the feature, bug-fix, opening-a-pr, babysit and shipping playbooks./architect, four-item checkpoint, delegate, verify, small commits, /interrogate if contested, open PR ( feature). babysit, land and /reflect are separate requests. The guide follows the same chain: "Read the pages in order the first time." The author refuses plans and specs, not process: "the best spec is code."AI Hero skills
mattpocock/skills @5863806Matt Pocock · any agent (Claude Code, Codex, Copilot, Gemini CLI)
/ask-matt, names the next command and stops./grill-with-docs → /to-spec → /to-tickets → /implement → /code-review → /retro" ( docs/engineering/implement.md). Spec and tickets are skipped for single-session work; /retro is optional after a smooth build; prototype and /pr sit outside the main flow. A static table of steps would "be wrong for most situations" ( docs/engineering/ask-matt.md)./to-tickets can also run from the conversation alone, with no spec./implement per ticket (clear context between each) or /implement-spec for the whole graph.Compound Engineering
EveryInc/compound-engineering-plugin @0f7ae5eEvery (Kieran Klaassen, Trevin Chow) · Claude Code, Cursor, Codex and 11 more
/lfg command runs; there Brainstorm is an optional route and Compound runs only when the work corrected a wrong belief./lfg, which runs Compound before the PR; the /ce-work guide puts it after shipping. Typed by hand, nothing enforces any of these orders.Sisyphus Labs · its own omo binary; OpenCode and Codex plugins
/ulw-plan, then /ulw-execute. The OpenCode edition has the same step order with different reviewers and verifier./ultrawork … The agent figures everything out." The path drawn here is offered as "Want more control? Run /ulw-plan for interview-based planning, then /ulw-execute" ( overview.md). Code restricts when the plan-review agents can be spawned, and re-injects a continue prompt while the plan has unchecked boxes. Nothing in code checks the step order./ulw-execute separately.Roberto Mello · Claude Code, OpenCode, Gemini CLI, Cortex Code
/lavra-design, /lavra-work, /lavra-ship, with /lavra-qa optional for web apps ( README.md). Inside /lavra-design the phases chain: "Brainstorm → Plan → Research → Revise → CEO Review → Eng Review → Lock". Review and the fix loop run inside /lavra-work. /lavra-quick is a shortcut for small tasks./lavra-work after phase 6 locks the plan./lavra-ship after QA.Hyper planning loop
this project · proposed, not built
Planning questions each stack answers
Explicit: a named artifact or section, with a gate. Implicit: done inside another step, with nothing to check.
| Stack | Requirements | Current state | Design | Task breakdown | Verify against requirements | Learning kept |
|---|---|---|---|---|---|---|
| Superpowers | goals and non-goals in the design doc; no criteria list | "Look Before You Ask" inside /brainstorming | design doc behind a hard human gate | /writing-plans | reviews and tests; no requirement-by-requirement check | ledger deleted at the end |
| pstack | goal restated; a done check is suggested, not required | mechanism confirmed at runtime before design | architect, arena, prototype | four-item checkpoint; plan file only for multi-PR work | against the stated done check, by a non-author; no requirement list to trace | reflect and correct; run by hand |
| AI Hero skills | spec issue with user stories; criteria per ticket | conditional, in /to-spec and /to-tickets: "If you have not already explored the codebase, do so" | test seams and an Implementation Decisions section | /to-tickets, approved by the human | Spec axis of one review pass; no acceptance run | /retro; advisory |
| Compound Engineering | Product Contract with requirement IDs | research in Brainstorm (grounding scout) and Plan (repo-research-analyst) | Key Technical Decisions in the plan | Implementation Units with IDs, files, dependencies; no code checks that every requirement maps to a unit | review flags an unaddressed requirement when the plan is given explicitly; /ce-work, the implementer, re-checks the Definition of Done itself | Compound; conditional under /lfg, otherwise typed by hand |
| omo | ideal-state rows IS-n in the draft; no separate document | Ground: cited findings and GAP-n rows | an approach field; no design artifact | todos in waves with a dependency matrix | final row F4 checks each IS row | automatic memory; no step, nothing to check |
| Lavra | Requirements and Non-Requirements on the epic | research agents run inside Plan; deeper research off by default | Locked Decisions; no design document | child beads, DAG validated | per bead; the epic requirements are not rechecked | memory file; a hook asks, nothing checks |
| Hyper planning loop | Define | Baseline | Design | Decompose | Verify, at every level | not in the proposal |
Mechanisms worth taking, and gaps
| Stack | Worth taking | Gaps |
|---|---|---|
| Superpowers |
|
|
| pstack |
|
|
| AI Hero skills |
|
|
| Compound Engineering |
|
|
| omo |
|
|
| Lavra |
|
|
| Hyper planning loop |
|
My reading
- Every author states an order; none enforces it in code. All six are recommendations in a README, a guide or the skill text. Three hedge: omo defaults to the bare keyword and offers the planned path as opt-in, AI Hero makes its flow conditional and rejects tools that own the process, Compound Engineering publishes four orders that disagree. pstack has no single order, only one per playbook.
- What code does enforce is narrow, and it is never the order or the quality of a step. omo: a hook re-injects a continue prompt while the plan has unchecked boxes (native stops after 8 in a row, OpenCode after 3 rounds without a tool call, Codex has no counter), the planner agent can write only markdown under .omo, plan reviewers spawn only inside /ulw-plan, and on Codex a worker needs an evidence receipt to stop. Superpowers: one script refuses to record an inline task whose tests fail. Lavra: two hooks block a stop. Compound Engineering: a controller and a handshake script guard the optional cross-model path. pstack: watch-pr reports a PR state by exit code and blocks nothing. AI Hero: nothing. Every verdict, round limit and gate drawn on this page is otherwise a sentence the model is asked to follow.
- Current state is a gated step in two of six: omo (cited findings and gap rows) and pstack (mechanism confirmed at runtime before design). The other four research inside brainstorming or planning and write no current-state section to check. This is the gap the proposed Baseline step fills.
- A final check against the requirements exists in two of six: omo (final row F4 against the ideal-state rows) and Compound Engineering (review traces requirement IDs, blocking only when the plan was passed explicitly; the Definition of Done is re-checked by the implementer, not a reviewer). Both depend on requirements that carry IDs. Lavra and AI Hero stop at per-task checks; the Superpowers final reviewer asks only whether all planned functionality is present; pstack checks a done sentence.
- omo is the closest to the proposed Define and Baseline: ideal-state rows, gap rows, every todo closes a gap, every ideal-state row is proven at the end. It has no design artifact.
- No stack runs its loop at more than one level. Lavra (epic, beads) and Compound Engineering (plan, units) have two fixed levels. The recursion in the hyper proposal has no precedent here: not tested, not refuted.
- Five of six have a learning step and none makes it mandatory: the Lavra hook asks and never checks, Compound runs only when a belief was corrected, pstack reflect and the AI Hero retro are separate requests, omo memory is written on compaction or by step count. The hyper proposal has no such step. The strongest forms prefer a check to a note: the Compound Engineering ladder and pstack correct.
- Most specified fix loops are bounded: Superpowers 5 rounds then the controller rules, Lavra 3 then it reports and proceeds, omo native 3 attempts or a repeated failure. The omo OpenCode edition has no retry cap. No program counts any of these rounds; the only counters in code are the omo continuation limits and the Codex evidence gate. The proposed Replan has no bound.
- pstack and omo accept a verdict only from an agent that did not write the code (in omo that is a verifier subagent in the native edition and the orchestrator in the OpenCode edition). The proposal gives Verify to the author of the Define, which is the same rule.