omp-superpowers

obra/superpowers adapted for omp (Oh My Pi).

Upstream superpowers is a set of workflow skills (brainstorm → plan → implement via subagents → review → finish). It ships a Pi/omp compatibility layer that is wrong on omp: it tells the agent there is no subagent tool and no todo tool, and its Model Selection rules use a model: field the omp task tool does not have. Under those rules every subagent silently inherits the slowest configured model over the widest tool set — in a real 157-subagent session that was where most of the wall-clock went.

This fork keeps the skills and fixes the omp side:

• the injected tool mapping tells the truth about task, todo, read/grep/glob
• five omp agents (agents/sdd-*.md) give every subagent-driven-development seat its own model, thinking level, and tool set
• subagent-driven-development is rewritten around them: three-round fix loop, structured reviewer output, implementers one at a time

Install

omp plugin install npm:@loneexile/omp-superpowers

Releases are versioned <upstream version>-omp.<n>: 6.4.2-omp.1 is upstream v6.4.2 plus this fork's first iteration on it. CHANGELOG.md lists each release. To move to another one, uninstall and install it pinned:

omp plugin uninstall @loneexile/omp-superpowers
omp plugin install npm:@loneexile/omp-superpowers@<version>

Unreleased main installs from GitHub under the same plugin name, pinned to a commit: omp plugin install github:LoneExile/omp-superpowers#<sha>. Installs made before 6.4.2-omp.1 used the plugin name superpowers; run omp plugin uninstall superpowers before installing this package.

npm and github: installs keep each seat's model. For a marketplace or --plugin-dir install of the repo root, the repo carries .omp-plugin/plugin.json: omp 18.4.10's agent discovery treats a root that has .claude-plugin/plugin.json as Claude dialect, and drops every agent's model: line, unless that file is present. This was read from omp's source; a marketplace install has not been exercised.

Verify it loaded — the roster of your task tool should now include the five sdd-* agents:

omp -p --no-session --no-skills --model anthropic/claude-haiku-4-5 \
  'List the agent type names your `task` tool offers, comma-separated, nothing else.'
# scout, reviewer, security-reviewer, sonic, task, sdd-implementer, sdd-reviewer, sdd-rereviewer, sdd-escalation-implementer, sdd-final-reviewer

Agent definitions are read per dispatch, so a reinstall takes effect on the next task call in a running session. The injected tool mapping and the skill catalog are read at process start — restart omp after upgrading to pick those up. The bootstrap is injected at the start of each session, including /new, /resume and /fork, until that session's first agent run ends, and again after compaction. Main session only; subagents never get it.

Using it

The skills drive themselves once loaded. The usual path:

1. read skill://brainstorming — turn an idea into a design
2. read skill://writing-plans — produce docs/superpowers/plans/<date>-<name>.md
3. read skill://subagent-driven-development — execute the plan task by task via subagents. If you choose Native at the plan handoff, executing-plans runs instead: the session does every task itself, and only the final review is a subagent (sdd-final-reviewer)
4. read skill://finishing-a-development-branch — merge / PR / discard

On omp the agent invokes a skill by reading it (read skill://<name>); you can also invoke one explicitly with /skill:<name>.

Vibe mode. After /vibe the session is a director of fast/good workers and has no task, bash or file-writing tool, so it cannot run subagent-driven-development or executing-plans itself. Both skills tell it to hand the whole run to one good worker, which dispatches the sdd-* seats through its own task tool. Each worker brief names the tests to run, because omp 18.4.12 workers skip them otherwise. Vibe refuses /move and /wt, so create the worktree and /move into it first; for a long plan set task.softRequestBudget: 0. Details are in the Vibe mode section of pi-tools.md.

What subagent-driven-development does per task

flowchart TD
    B["task-brief: one task's text from the plan"] --> I
    I["sdd-implementer: TDD, commit, report file"] --> P["review-package: git diff -U10 to a file"]
    P --> R["sdd-reviewer: spec + quality, structured result"]
    R -->|"approved"| D["ledger: Task N complete"]
    R -->|"Critical or Important findings"| F
    F["fix round 1 or 2: resume sdd-implementer"] --> RR["sdd-rereviewer: verdict per finding"]
    RR -->|"all addressed"| D
    RR -->|"still open, round 3"| E["sdd-escalation-implementer: fresh context, strongest tier"]
    E --> RR
    RR -->|"still open after round 3"| A["controller adjudicates and parks with a ruling"]
    A --> D
    D --> N{"more tasks?"}
    N -->|"yes"| B
    N -->|"no"| W["sdd-final-reviewer: whole-branch, once per plan"]
    W --> X["finishing-a-development-branch"]

    classDef seat fill:#1e3a8a,stroke:#3b82f6,color:#dbeafe
    class I,R,RR,E,W seat

Blue nodes are the five omp agents this fork ships. A reviewer that cannot read its package yields the package_gap verdict value (with what it tried) instead of inventing one, and the controller regenerates and re-dispatches.

The SDD seats

| Seat | agent | shipped default · thinking | falls back to | tools | |---|---|---|---|---| | implementer (every task, fix rounds 1–2) | sdd-implementer | claude-sonnet-5-5 · medium | @task | read, write, edit, bash, grep, glob, lsp, ast_grep | | task reviewer | sdd-reviewer | claude-sonnet-5-5 · high | @task | read, grep, glob | | scoped re-review | sdd-rereviewer | claude-sonnet-5-5 · low | @task | read, grep | | fix round 3 / ruled strongest-tier task | sdd-escalation-implementer | claude-opus-5-5 · max | @slow | same as implementer | | final whole-branch review (subagent-driven or Native) | sdd-final-reviewer | claude-opus-5-5 · max | @slow | read, grep, glob, bash, lsp, ast_grep |

lsp and ast_grep are listed for the seats that benefit, but omp only activates them in subagents when task.enableLsp / astGrep.enabled are on (both default off).

omp takes the first list entry whose provider has credentials. If the pin has none, the seat runs on your modelRoles.task / modelRoles.slow model at that role's own level. An unset @task drops out, and the seat then silently runs on the controller's model; an unset @slow expands to omp's built-in strong-model ranking. The list is also the retry chain on rate-limit and usage-limit errors, gated by retry.modelFallback (default on); there is no switch on context overflow. Turn on task.showResolvedModelBadge, or read session_init.resolvedModel in a subagent transcript, to see what actually ran.

The shipped defaults are Anthropic: sonnet for the seats that run on every task, opus for the two that run once per plan or once per stuck task. The tiers are a thinking ladder as much as a model ladder — re-reviews reason at low, escalation and the final review at max.

The two reviewer seats have no bash, edit, or filesystem write: the review package (git diff -U10 written to a file) is their whole view of the change, and they return a structured result (spec_compliance, task_quality, findings[], cannot_verify[]; finding_verdicts[], new_breakage[], round_verdict) plus package_gap when they could not read the package. Note the honest boundary: an agent's tools: list limits built-in tools only. Extension tools and MCP tools (mounted as xd:// devices and run through a device-only write) still attach to every subagent; memory tools attach only when the agent's tools: list names them. On omp 18.4.10 a read/grep/glob reviewer ran a shell command through an MCP device. Read-only is by instruction, not a sandbox.

Changing models — no file edits

omp resolves each seat's model and thinking level in a fixed order. The override map is the layer meant for you:

flowchart LR
    subgraph model["model"]
        M1["task.agentModelOverrides"] -->|"unset"| M2["agent file model list: pin, then @task or @slow"]
    end
    subgraph level["thinking level"]
        L1["explicit :level on the winning model string"] -->|"none"| L2["agent file thinkingLevel:"]
        L2 -->|"none"| L3["model default"]
    end
    model --> level

    classDef you fill:#1e3a8a,stroke:#3b82f6,color:#dbeafe
    class M1 you

Set the override in ~/.omp/agent/config.yml, or interactively with /agents in any session (it edits the same map). It applies on the next dispatch; delete a line to fall back to the agent file.

Example — every seat on one flat-rate model, mapped onto that model's own thinking ladder:

task:
  agentModelOverrides:
    sdd-implementer: opencode-go/deepseek-v4.1-flash:high
    sdd-reviewer: opencode-go/deepseek-v4.1-flash:high
    sdd-rereviewer: opencode-go/deepseek-v4.1-flash:low
    sdd-escalation-implementer: opencode-go/deepseek-v4.1-flash:max
    sdd-final-reviewer: opencode-go/deepseek-v4.1-flash:max

Use the exact catalog id and a level the model actually offers. omp models ls <provider> shows each model's ladder. A level that is not on it is silently clamped to the nearest valid one — deepseek-v4.1-flash offers only low, high, max, so :medium would run at low and :xhigh at high, collapsing the ladder without any error. Check session_init.resolvedModel in a subagent transcript if in doubt: it shows the level that actually ran. Catalog ids also get renamed (this one was deepseek-flash a day earlier, and the old string now matches an empty placeholder row).

Measured on a real fix-round re-review with identical inputs: a grok-4.6 seat took 9.7 min / 4 turns, sonnet-5 about 50 s / 1 turn, the DeepSeek flash seat 100 s / 4 turns — same verdict, full structured result, all three.

Rules that are easy to get wrong:

• A bare model (anthropic/claude-haiku-4-5) keeps the agent file's own thinkingLevel. Add :level (…:high) to change reasoning too.
• Delete a line to fall back to the agent file's default for that seat. Overrides apply on the next dispatch — no restart.
• An override replaces the whole model list, fallback included. Append ,@task (for example provider/model:high,@task) to keep one.
• The task tool has no model: field. Never dispatch the bundled task or reviewer agent for an SDD seat — they lack the seat's tools, structured output and no-subagent rule, and run whatever modelRoles.task / modelRoles.slow names.
• effort on a dispatch (when task.enableEffort is on) accepts only "lo", "med", "hi". "hi" maps to the model's top supported level, capped by task.maxEffort; SDD uses effort: "hi" for high-risk task reviews, and for escalation only when a host override lowers that seat.
• Discovery-only model ids (the OpenCode Go catalog, for example) accept :level in config and in overrides but not on the --model CLI flag.

Optional settings that matter

| setting | why | |---|---| | task.isolation.enabled (+ isolation.backend, default auto) | off by default. SDD does not use it: implementers run one at a time, because isolated work returns as an applied patch with no per-task commits, and an isolated subagent cannot be messaged for fix rounds. Older omp builds spelled this task.isolation.mode; check omp config get task.isolation.enabled on your build | | task.enableEffort | exposes per-dispatch effort ("lo", "med", "hi"); "hi" maps to the model's top supported level, capped by task.maxEffort. SDD uses effort: "hi" for high-risk task reviews, and for escalation only when a host override lowers that seat | | task.showResolvedModelBadge (default off) | shows the model each subagent actually ran. A seat whose pinned provider has no credentials silently runs on another model | | task.enableLsp, astGrep.enabled (both default off) | omp only activates lsp and ast_grep in subagents when these are on | | autolearn.autoContinue | omp's auto-learn mints managed skills every session; the whole catalog is injected into every subagent's prompt. Keep it pruned or off — a 2,500-skill catalog was ~220k tokens per subagent turn |

What changed from upstream

• .pi/extensions/superpowers.ts — the injected mapping names task, todo, the sdd-* roster, and the no-model:-field rule; it also re-arms on session_switch, skips subagent sessions, and ignores an agent_end that will auto-continue; it tells a dispatching session to name the checks each subagent must run, since omp 18.4.12 tells subagents to leave builds, tests and smoke runs to the main agent unless their assignment explicitly instructs them; its last paragraph points a vibe-mode director (tools include vibe_spawn but not task) at the Vibe mode section of pi-tools.md
• .omp-plugin/plugin.json — new. Marks the repo root as an omp-native plugin so a marketplace or --plugin-dir install keeps each agent's model: line
• skills/using-superpowers/references/pi-tools.md — same, as the reference doc; it also gains a Checks subsection (the 18.4.12 subagent hand-off rule, and why every dispatch names its checks) and a Vibe mode section: what a director keeps and lacks, how skill actions map onto vibe_* and workers, and the worktree and request-budget limits
• agents/sdd-*.md — new. Each seat's model: is a pin plus a role fallback (@task or @slow), and all set readSummarize: false. sdd-final-reviewer reviews the whole branch for both ways of running a plan. It reads requesting-code-review/code-reviewer.md at run time and applies its What to Check, Calibration, severity definitions and Critical Rules, while carrying upstream's vision-document and Declined-to-judge rules in its own body. Descriptions name roles, never models
• skills/subagent-driven-development/ — Model Selection → Agent Selection; fix-loop cap 5 → 3 with a round-3 escalation seat; a proven-trivial-fix route that replaces a re-review with a one-command proof; implementers run one at a time; the scripts directory resolves via realpath skill://subagent-driven-development/scripts; fix rounds resume the implementer with write agent://<id> (id ledgered) and full results are read at agent://<id>; task review scales with risk through effort: "hi"; the controller checks reviewer results; templates show omp's real { context, tasks: [{ agent, task, solutionSpace }] } wire shape and the reviewers' structured fields; one omp paragraph for vibe mode, where a director (no task or bash) hands the whole run to one good worker that dispatches the seats itself
• skills/executing-plans/ — one additive omp paragraph in Final Review: the task tool has no model: field, so the final review goes to sdd-final-reviewer (its model is what the host maps that agent to), told that no task was reviewed; it also resolves the scripts directory with realpath and reads full results at agent://<id>; one more omp paragraph for vibe mode, where a director hands the whole run to one good worker
• skills/using-git-worktrees/ — additive omp section: /wt is the user's one-line answer to the consent question (moves the session, carries WIP); when the agent creates the worktree itself it uses omp worktree add under ~/.omp/wt/ (the only place omp worktree list/clear manage), knows that command leaves uncommitted changes behind, and asks for /move <path> because cd never moves the tools' cwd. Also: clone-first falls back to a plain checkout without node_modules (no copy-on-write, or in-repo targets); setup and tests use the bash tool's cwd parameter until /move; omp worktree clear --all is warned against, since it force-removes every worktree, uncommitted work included; omp's bash rewrites a literal git worktree add into omp worktree add only without shell expansion; vibe mode refuses /move and /wt, so the worktree and /move come before /vibe, or every worker brief carries the absolute path
• skills/writing-plans/ — one added line in Execution Handoff: in vibe mode both execution approaches run in one good worker the director briefs, not in its own session
• README.md — this file: rewritten for omp (install, seats, overrides, maintenance); upstream's multi-harness README is gone
• tests/pi/test-pi-extension.mjs — covers the extension's lifecycle (bootstrap on start, session_switch, compaction; none for subagents or an auto-continuing agent_end), the action table in pi-tools.md, the seat table against the names in agents/*.md, and the .omp-plugin/plugin.json manifest; the package-name assertion is gone (scripts/check-npm-package.sh checks the published name)
• package.json — the npm identity: @loneexile/omp-superpowers, <upstream version>-omp.<n> versions, a files allowlist (the extension, skills/, agents/), publishConfig, omp/oh-my-pi keywords; main is dropped, since it pointed at the OpenCode plugin this package does not ship
• CHANGELOG.md, .github/workflows/ (CI, npm release), scripts/check-npm-package.sh — new

Everything else is upstream, unmodified.

Maintenance

Syncing with upstream

git fetch upstream --tags
git rebase upstream/main
git push --force-with-lease origin main

Expect conflicts in skills/subagent-driven-development/, pi-tools.md, README.md, and package.json whenever upstream touches them — this fork rewrites those files rather than appending to them (README.md is never a partial conflict). In package.json keep the fork's identity and set version to <new upstream version>-omp.1: CI's package check fails until the version's base matches the upstream version in .claude-plugin/plugin.json. agents/, the extension's mapping paragraph, and the omp blocks in using-git-worktrees are additive.

A clean rebase is not the whole sync. Upstream's edits to text this fork replaced never reach the replacement, so read upstream's diff for those files and port what still applies. The v6.4.x sync needed two: bash scripts/... invocations in the SDD text, and code-reviewer.md's new review rules. Upstream edits to code-reviewer.md's What to Check, Calibration, severity definitions and Critical Rules now reach sdd-final-reviewer without porting, because it reads the file at run time. Edits to the vision-document or Declined-to-judge sections still need porting into the agent body, which carries them itself.

Releasing

1. Set version in package.json, move CHANGELOG.md's Unreleased entries under that version, commit, push, and wait for CI.
2. Publish a GitHub release for the tag v<version>; the release workflow publishes it to npm through Trusted Publishing, with provenance:

v=$(node -p "require('./package.json').version")
awk -v h="## [$v]" 'index($0, h) == 1 {f = 1; next} f && /^(## )?\[/ {exit} f' CHANGELOG.md > /tmp/notes.md
gh release create "v$v" --target main --title "v$v" --notes-file /tmp/notes.md

New omp versions

Before trusting a new omp release, re-check the harness facts this fork depends on — against the running binary, not an npm source tree that may lag it. ~/.bun/install/global/node_modules/@oh-my-pi/pi-coding-agent/src routinely lags omp --version (a 17.4.2 tree sat beside an 18.1.17 binary while this README was written, and the isolation setting had been renamed in between). The fork was last verified against omp 18.4.10; its vibe-mode text and the subagent check rule were re-checked against 18.4.12. To read the exact running version's source, pack it into a temp dir: cd "$(mktemp -d)" && npm pack @oh-my-pi/pi-coding-agent@$(omp --version | cut -d/ -f2) && tar xzf *.tgz (source under package/src). The live oracles are omp config list --json (every setting and its default), the task tool's own description in a session (agent roster, item fields), a session_init.resolvedModel line in a subagent transcript (what actually resolved), and grep -a on the binary for an exact string when you need to know whether a code path exists.

License

MIT, inherited from upstream. Not affiliated with obra / Prime Radiant.