description: "obra/superpowers skills provider plus a system-prompt bootstrap for DeepSeek Harness." kind: "package-bundle"

@firefly0621/dsh-superpowers

English | 中文

Summary

Bring the obra/superpowers workflow kernel and its fused extensions into a DeepSeek Harness profile as one skill catalog plus a standing skill-first bootstrap. Use it when product sessions should move from intent through plan, test, debug, review, and verification in a repeatable order. It is opt-in and ships no harness peers.

Table of Contents

• Summary
• Model Experience
• Known Limitations and Deferred Work
• Dev Note

Package identity

Installable DeepSeek Harness plugin bundle that brings obra/superpowers into product sessions: a 35-skill catalog split into core, advanced, product, and authoring bundles, adapted to dsh tooling, plus a SessionStart-equivalent bootstrap. The default profile loads the 19-skill core bundle (diagnosing-superpowers stays out until a deep dsh port).

Publish name: @firefly0621/dsh-superpowers (version tracks the harness family, currently 0.1.0-rc.44). Source of record for this fork package: github.com/oThTJx/dsh-superpowers. This package is opt-in and is not part of the official main dsh-base composition.

What it restores

| Superpowers behavior | dsh realization | |---|---| | Skills library under skills/*/SKILL.md | ctx.skills provider superpowers (packaged skills/ directory, dsh-adapted) | | SessionStart hook injecting using-superpowers | system-prompt/assemble section skill:superpowers with the same <EXTREMELY_IMPORTANT> framing (every non-subagent request); using-superpowers is user-only in the skill catalog | | Platform references/*-tools.md | bundled skills/using-superpowers/references/dsh-tools.md appended inside that section | | Subagent skip (SUBAGENT-STOP) | no full using-superpowers bootstrap when session.header.origin === 'subagent'; injects a short skill:superpowers-worker preamble instead | | Model loads other skills on demand | existing dsh-tool-skill catalog + skill tool |

The skill bodies ship as dsh-adapted prose under skills/ (no vendored source, no git submodule). Foreign-harness tool names and platform references were rewritten for dsh, and the superpowers: skill-name prefix was removed. The published tarball includes skills/, assets/NOTICE.md, and the offline scorer scripts/score-skill-first.mjs. NOTICE records obra/superpowers (2ce8aa15, v7.0.1), mattpocock/skills (5beb35b0), addyosmani/agent-skills (aeac41d4), and bmad-code-org/BMAD-METHOD (61e01f82, with trademark notice) as review pins.

Capabilities that overlap are fused into one owning skill. When an upstream source has material that does not justify a second skill directory, it ships as a reference pack under the owner and is loaded only at the step that needs it. The catalog therefore stays at 35 skills while the fused behavior keeps growing.

Bundles

The package ships four bundles. core is always loaded; the others are opt-in through the bundles config.

| Bundle | Default | Skills | Purpose | |---|---|---|---| | core | yes | 19 | The day-to-day workflow: brainstorming, planning, execution, TDD, debugging, review, verification, finish, domain/design, security, sources, constraints | | advanced | no | 6 | Subagent-driven execution and parallel dispatch, context engineering, prototype, wizard, grilling | | product | no | 8 | Idea, PRFAQ, PRD, spec, UX, retrospective, course correction, and multi-role debate for multi-stakeholder product work | | authoring | no | 2 | writing-skills and dsh-harness-contributor for skill and repository maintainers |

Core is always included, even when bundles names only an optional bundle. A bare install therefore sees 19 skills; a full install sees 35. The model-visible core catalog is 16 skills after using-superpowers, handoff, and wait-what stay user-only.

Matt skills and fused extensions

Superpowers remains the workflow kernel (brainstorming → writing-plans → TDD / debugging / review / verify). Selected mattpocock/skills, addyosmani/agent-skills, and BMAD-METHOD practices ship as one fused catalog: every skill declares at most one bundle label, and that label is the only catalog metadata this provider publishes.

Session skills (grilling, handoff, wait-what) are absent from the model <available_skills> catalog and refused by the model skill tool. Users invoke them with a whitespace-bounded /name token in a user message (for example /grilling); dsh-tool-skill injects the skill body as instructions. using-superpowers is also disable-model-invocation: true, but its standing copy arrives through the system-prompt bootstrap rather than that gesture (users may still /using-superpowers to reload the body without the adapter).

Obra skills that absorbed merges (keep on obra sync): brainstorming (inline Architectural grilling; Not Doing; Always/Ask/Never; reject aim-past-100), using-superpowers (≤5-line intent fork; anti-stack; tiny-edit; implementation priority; contextual gates), test-driven-development, systematic-debugging, requesting-code-review (Standards ∥ Spec + checklists + review lenses + doubt pass), executing-plans (slice discipline + impact map), refactoring (clarity + structural modes), source-driven-development (research mode), codebase-design, writing-plans, handoff, verification-before-completion.

Implementation and quality extensions: executing-plans owns the task loop, slice discipline, and impact map; refactoring owns clarity and structural modes; source-driven-development owns API-grounded work and the research mode; requesting-code-review owns post-hoc review and the in-flight doubt pass; subagent-driven-development owns per-task dispatch and parallel-dispatch reference. dsh-harness-contributor and explaining-changes live in the authoring and advanced bundles and do not enter the default core catalog.

bundle is catalog metadata only — not a router. Invocation frontmatter is parsed privately by this provider (plugin-only; upstream harness packages unchanged).

Fusion map

Each skill owns one capability, and every selected source that covers that capability feeds its body. When two sources overlap, the package keeps one skill and merges both contributions — it never ships two peers for one job, and it does not drop a selected source's behaviour because another source moved first. Unselected capabilities are listed as out of scope in the debt table.

| Skill | Sources fused into it | It owns | |---|---|---| | brainstorming | obra spine · BMAD product scheme · agent-skills anti-patterns | intent → approved design, size-based gates | | writing-plans | obra plan format · agent-skills task breakdown · Matt spec-to-tickets | approved spec → ordered plan | | test-driven-development | obra Iron Law · agent-skills trigger breadth | RED-GREEN-REFACTOR loop | | subagent-driven-development | obra controller · dsh settlement wait · parallel dispatch | per-task dispatch, review, and parallel investigation | | executing-plans | obra inline execution · agent-skills slicing rules · dsh script wiring | inline plan execution, slice discipline, impact map | | systematic-debugging | obra four phases · agent-skills error recovery · observability and performance | root cause before any fix | | refactoring | agent-skills clarity body · dsh structural extension · agent-skills deprecation/migration | clarity and structural modes, retirement | | requesting-code-review | obra Standards ∥ Spec · agent-skills quality body · BMAD review lenses · in-flight doubt pass | review request, lenses, and disposition | | verification-before-completion | obra gate · agent-skills browser and CI evidence | proof before claiming done | | finishing-a-development-branch | obra finish · agent-skills shipping, git, CI/CD · Matt PR body | integrate, release, roll back | | codebase-design | Matt deep modules · agent-skills interface contracts · BMAD architecture spine | module, interface, and architecture decisions | | domain-modeling | Matt glossary · agent-skills ADRs and docs | model and decisions | | grilling | Matt design tree · agent-skills interview · BMAD advanced elicitation | user-requested interview | | forge-idea | BMAD distillation · agent-skills idea refine | idea verdict | | security-and-hardening | agent-skills controls · dsh untrusted-output stance | threat model and boundaries | | source-driven-development | agent-skills API facts · research mode · BMAD deep recon | docs-grounded API facts and cited investigation | | constraint-driven-development | agent-skills | project quality floor | | context-engineering | agent-skills hierarchy · dsh budget model · BMAD project context | context budget, packing, project instructions | | ux-spec | BMAD spines · agent-skills frontend acceptance and engineering | UX acceptance and implementation contracts |

The map is an inventory, not a lock: a skill may gain a source, but it never splits into one peer per source, and a merged section is never replaced by whichever upstream moved last.

Reference packs

The following capabilities ship behind their owner rather than as extra skill directories. The owner links the pack at the step that needs it, so the model pays for the material only when that branch is reached.

| Owner | Reference | Fuses | |---|---|---| | systematic-debugging | references/observability-and-performance.md | production observability · performance regression workflow | | refactoring | references/clarity.md | behavior-preserving clarity pass | | refactoring | references/structural.md | structural change after repeated fixes | | refactoring | references/deprecation-and-migration.md | deprecation decision · strangler/adapter/flag · expand → contract | | executing-plans | references/slice-discipline.md | slice loop · slicing strategies · scope discipline | | executing-plans | references/impact-analysis.md | short impact map before a cross-boundary change | | ux-spec | references/frontend-engineering.md | component and state ownership · design-system adherence · accessibility · responsive UI | | context-engineering | references/project-context.md | AGENTS.md setup/adopt/refresh/audit · writing for agents | | codebase-design | references/architecture-spine.md | architecture invariants · stable AD IDs · reviewer gate; plus TypeScript entry-point enforcement in the body | | writing-plans | references/spec-to-tickets.md | spec-driven workflow · tracer-bullet tickets · wayfinding · triage | | requesting-code-review | references/review-lenses.md | adversarial · edge-case · verification-gap · claims/deletion checks · walkthrough | | requesting-code-review | references/doubt-driven-review.md | in-flight adversarial review of a decision about to lock in | | subagent-driven-development | references/parallel-dispatch.md | independent parallel investigations and prompt structure | | writing-skills | references/skill-evals.md | baseline/variant/quality/trigger evals · toolsmith pass | | source-driven-development | references/research-mode.md | open-ended cited investigation · research budget | | source-driven-development | references/deep-recon.md | decision-grade research · research firewall · claim sources | | test-driven-development | references/retrofitting-tests.md | end-to-end/API tests for existing behavior | | grilling | references/advanced-elicitation.md | pre-mortem · red team · first principles · questionnaire | | prd | references/product-brief.md | discovery · coaching/fast paths · right-sized brief | | retrospective | references/evidence-review.md | evidence inventory · range attribution · acceptance verdict · follow-through | | constraint-driven-development | references/floor-guard.md | diff-scoped floor guard · five bar-lowering moves · exit 0/1/2 | | security-and-hardening | references/hardening-patterns.md | OWASP patterns · input validation · destructive paths · dependency audit · rate limiting · secrets · data classification · LLM output handling | | explaining-changes | references/walkthrough.md | intent → broad strokes → slices → periphery | | finishing-a-development-branch | references/ci-and-delivery.md | pipeline order · previews · staged rollout · dependency automation | | finishing-a-development-branch | references/pr-description.md | Summary · Evidence · Test plan · Merge danger | | brainstorming | references/technique-catalog.md | working modes · divergence lenses · convergence tests |

Reference packs are part of the owner's contract: tests/superpowers.spec.ts checks that each fusion pack listed above ships with its defining pins, that every reference file is linked from its owner's prose, and that every relative Markdown link under skills/ resolves. A pack is not a second router and does not appear in <available_skills>.

Install into a dsh profile

Requires DeepSeek Harness peers (@deepseek-ai/dsh-skill and related packages), typically from the official CLI/dsh-base stack. Source-workspace workspace:* peers mean the matching DSH release; packed manifests substitute that exact version. DSH checks those peer ranges at install and profile startup and refuses an incompatible pair without an exact version exemption — see Version compatibility.

dsh plugin --profile web add @firefly0621/dsh-superpowers

From a harness source tree:

pnpm dsh plugin --profile web add @firefly0621/dsh-superpowers

Then boot dsh --profile web (or pnpm dsh web). If this fork's dsh-base already mounts id: superpowers, do not add the same row again — disable the base row first, or use a profile that does not include it.

Local / pre-publish check

pnpm run build:lib:host
pnpm --filter @firefly0621/dsh-superpowers pack
pnpm dsh plugin --profile superpowers-demo add ./firefly0621-dsh-superpowers-0.1.0-rc.44.tgz

(pnpm pack writes the tarball to the repository root by default. There is no per-package build script; host lib build produces lib/.)

Enable from this monorepo without npm

This package is opt-in in the harness source tree. Mount the shipped overlay patch (or dsh plugin add):

pnpm dsh web --patch packages/skill/superpowers/cordis.patch.yml
pnpm dsh --profile headless --patch packages/skill/superpowers/cordis.patch.yml "task"

Fork trees that already mount id: superpowers in their base need no extra step; set disabled: true on that row to omit it.

Config

| Field | Default | Meaning | |---|---|---| | skillsRoot | packaged skills/ directory | Absolute skills directory; a relative path fails at apply | | bootstrap | true | Contribute parent using-superpowers (+ dsh adapter) as skill:superpowers, and the short worker preamble as skill:superpowers-worker on subagent-origin sessions | | bundles | ['core'] | Bundles to expose: core is always included; add advanced, product, and/or authoring |

Example full install:

- id: superpowers
  name: '@firefly0621/dsh-superpowers'
  config:
    bundles: [core, advanced, product, authoring]

Update Superpowers

There is no vendored submodule. To re-port newer upstream releases: for each source, diff content baseline → target, then walk every open row in sync debt below, then re-apply the dsh adaptation by hand. Do not use assets/NOTICE.md review pins as the sole diff baseline — those pins record last review, not “skills/ equals that tree”.

Adaptation rules: replace foreign-harness tool names and platform references, remove the superpowers: skill-name prefix, keep the behavioral content verbatim (dsh exceptions only in the table below), and distill Matt / agent-skills / BMAD ADDs (never dump BMAD _bmad runtime, memlog, HTML seals, or trademark personas).

Since the v7.0.1 sync, subagent-driven-development and writing-skills keep their long guidance in references/ and read a file from disk at the step that needs it, so a compacted session can reload one section instead of losing the whole body. Keep that layout when re-porting: a rule that lives inline again defeats it. The same sync removed the executing-plans worked example for the same reason, and rebuilt brainstorming onto the v7.0.1 text (which already carries 961a26d): first message → conversation → playback → size the work, a design doc at every size (a short one for a small change), and a gate that never skips without consent. The dsh decision-support format for architectural clarifying stays inline and states its one-question exception; the older three-path router and its separate gate tables are gone so one gate is stated once. Every packaged SKILL.md stays under upstream's 20,000-byte compaction cut, and the package suite enforces it.

Fusion, not port. Several upstreams feed one skill, and each keeps its own role: obra supplies the behavioral spine, agent-skills supplies ADDs and OPTIMIZE merges, Matt supplies named extensions, BMAD supplies product-discovery distillations. Never replace a merged section with one upstream's version because that upstream moved — re-apply the merge onto the new spine and keep every selected source's contribution (for example brainstorming keeps Not Doing / Always·Ask·Never through the v7 rebuild, and test-driven-development keeps the agent-skills trigger breadth inside obra's description). When the contribution is a distinct capability rather than a replacement section, add a reference pack under the owner; do not create a second skill directory for it. Capabilities outside the selected scope are named as out of scope in the debt table rather than silently counted as ported.

dsh adaptation exceptions (allowed non-verbatim edits):

| Exception | Why | |---|---| | Strip superpowers: / foreign skill-name prefixes | Catalog names are bare in dsh | | Rewrite foreign harness tools → dsh (subagent / subagent_fork, todo_write, dsh-tools.md) | Platform adapter | | writing-plans handoff label Native → Inline | Avoid “native harness” ambiguity; still means same-session executing-plans | | sdd-workspace: refuse marker-less existing dirs (no adopt) | Deliberate ownership tighten vs obra; documented in the script header | | Invoke bundled scripts via bash … | Packaged extractors may strip Unix exec bits | | Keep catalog Do not use negatives / description contracts | Fusion routing gates |

Content baseline and sync debt

| Source | Content baseline (landed in skills/) | NOTICE review pin | Open sync debt | |---|---|---|---| | obra/superpowers | Spine through v7.0.1 at 2ce8aa15, with brainstorming following the v7.0.1 text (which already carries 961a26d): first message → conversation → playback → size the work, a design doc at every size, never skip without consent, one red-flag table, no three-path router; subagent-driven-development / writing-skills guidance split into references/, the executing-plans worked example removed for the compaction cut, inline task loop kept (scripts/task-start / task-done); review-lens and walkthrough distillations landed in existing owners; bundled scripts carry the post-baseline task-ID, unclosed-fence, Windows-path, and relative---project-dir fixes | 2ce8aa15 (v7.0.1) plus the post-baseline fixes f75c01f, 750ba92, 3735a50, bb3a5aa, fa50d83, 8b0eefc, with 369bd87 re-expressed; 961a26d, 2b2ef34, c7730f7, and 6ff820e are already in 2ce8aa15, and 699edc6 deletes a test harness absent from that tree | (none) for the selected obra spine; diagnosing-superpowers and cross-harness reference material stay in Known Limitations | | mattpocock/skills | GLOSSARY-FORMAT + domain-modeling/handoff/grilling/wizard aligned to 49dd158d1; resolving-merge-conflicts removed; re-read at 5beb35b0 (only chief-of-staff, not shipped); to-spec / to-tickets / wayfinder / triage distilled into writing-plans; writing-for-agents into context-engineering | 5beb35b0 | (none) for selected capabilities; unselected Matt skills (for example ask-matt, teach, retro, in-progress, and misc sets) stay out of scope | | addyosmani/agent-skills | Distilled ADDs; context-engineering OPTIMIZE; security-and-hardening + references/hardening-patterns.md; constraint-driven-development + restored references/floor-guard.md; re-read at aeac41d4 (zod chain order fix landed, TDD trigger breadth fused); observability/performance, deprecation/migration, frontend engineering, CI/delivery, and API/interface material distilled into existing owners | aeac41d4 | (none) for shipped ADDs | | BMAD-METHOD | Distilled product ADDs under functional names (no _bmad runtime); re-read at 61e01f82; correct-course and spec-kernel follow the v7 PO replan.md / spec.md contracts; PRD Essential Spine, graded PRFAQ verdict, DESIGN.md/EXPERIENCE.md, retrospective dispositions, architecture spine, project context, deep recon, review lenses, advanced elicitation, e2e test generation, skill evals, product brief, and evidence-based retrospective distilled into existing owners | 61e01f82 | (none) for shipped ADDs — v7 replaced bmad-correct-course/bmad-spec with the PO replan/spec references; personas/runtime and unselected build/review/customize skills stay out of scope |

Fusion gate checklist (before/after merge):

1. Catalog length stays 35; each skill carries at most one bundle label, the default core catalog is 19 skills, and no skill declares a removed tier label; no second meta-router; no bmad-* skill directory names.
2. Re-assert OPTIMIZE merges (brainstorming Scheme B / Not Doing / Always·Ask·Never; TDD; review Standards ∥ Spec; grilling user-only; executing slice/impact; refactoring clarity/structural pins) after any obra spine overwrite.
3. Same-name overlaps from agent-skills (refactoring, test-driven-development) OPTIMIZE in place — do not add peer directories; merged skills keep their capability as references under the owner.
4. Update SEMANTIC_DESCRIPTION_GROUPS / NEGATIVE_BOUNDARY_BANNED when obra changes skill routing semantics (for example executing-plans inline vs subagent-driven-development).
5. Bump NOTICE review pins when you re-read upstream; update the content baseline / debt table when landed content or debt changes; keep ADD Do not use when… and BMAD exit-state pins.
6. Run pnpm exec vitest run packages/skill/superpowers/tests/superpowers.spec.ts.
7. Confirm executing-plans ships scripts/task-start and scripts/task-done and the SKILL body invokes them (tests failing ≠ tests not recorded).
8. Keep every reference pack linked from its owner and every relative Markdown link resolvable; fusion reference packs in the package suite checks both.
9. Keep the bundle map honest: a skill whose capability is everyday coding belongs in core; product-management, authoring, and opt-in discipline skills stay in their bundles.

Preserve on obra sync: Matt-added skill directories (domain-modeling, codebase-design, prototype, wizard, grilling, handoff, wait-what and companions); agent-skills / BMAD ADD directories (constraint-driven-development, source-driven-development, context-engineering, security-and-hardening, forge-idea, party-mode, spec-kernel, correct-course, prd, prfaq, ux-spec, retrospective); merged owners (executing-plans slice/impact references, refactoring clarity/structural/deprecation references, requesting-code-review review-lens/doubt references, subagent-driven-development parallel-dispatch reference, source-driven-development research references); bundle assignments in frontmatter; merged sections in brainstorming, using-superpowers (intent fork, anti-stack, tiny-edit, implementation priority, contextual gates), test-driven-development, systematic-debugging, requesting-code-review, codebase-design, writing-plans, handoff, verification-before-completion; private invocation and bundle parsing in src/index.ts; review-pin blocks in assets/NOTICE.md.

CI gate (not README-only): tests/superpowers.spec.ts skill composition contracts asserts bundle is the only catalog metadata label and locks the nineteen-skill core bundle after ADDs. Bundle tests assert the default core catalog and opt-in bundles. fusion reference packs asserts each fusion pack listed above exists with its defining pins. Separate gates assert every reference file is linked from its owner, every packaged Markdown file has paired code fences, and every relative Markdown link under skills/ resolves. catalog description contracts asserts semantic description groups (fail), positive-segment negative boundaries that block route contradictions (fail), and positive trigger overlap among extension skills (warn only). tests/scripts.spec.ts executes the packaged SDD scripts in throwaway repositories. Gated skills include the discipline extension set plus the process spine (brainstorming, systematic-debugging, verification-before-completion, writing-plans, executing-plans, requesting-code-review, subagent-driven-development, test-driven-development). ADDs are not members of those semantic/overlap tables; pairwise Jaccard description collisions warn on the packaged catalog (promote-to-fail is deferred). Dropping a bundled directory, reintroducing a tier label, misassigning a bundle, or leaving a reference pointer broken fails the package suite. The repository's per-file 100% coverage gate also applies to src/index.ts.

Version compatibility

This bundle declares peerDependencies on @deepseek-ai/dsh-* packages. DSH checks those ranges against the running runtime at install and profile startup; an incompatible pair is refused unless the profile grants an exact package@version × dsh-version exemption. Source-workspace workspace:* peers mean the matching DSH release; packed manifests substitute that exact version. See the plugin manager's compatibility reference for version-exemptions, allow-version, and the crash/data-loss risk.

Model Experience

System-prompt bootstrap

What the model sees

On every non-subagent request, system-prompt/assemble prepends section skill:superpowers whose text is buildBootstrapPreamble(...): the framing below, the body of the packaged using-superpowers SKILL.md (frontmatter omitted), then skills/using-superpowers/references/dsh-tools.md. Agentless assemblies omit the section. session.header.origin === 'subagent' gets skill:superpowers-worker from buildWorkerPreamble() instead of the full bootstrap. Compaction cannot drop the guidance because it is reassembled with the system prompt each turn, not stored as a one-shot user message. When the base composition and an agent preset both mount this plugin, the preamble is not repeated: the mount nearest the agent's scope contributes the only section.

Verbatim parent framing

<EXTREMELY_IMPORTANT>
You have superpowers.

## Standing order (before any tool call)

1. Call the dsh `skill` tool with the exact catalog name **before** clarifying, exploring, reading files, or acting — unless every tiny-edit condition in `using-superpowers` holds.
2. Anti-stack: load **one** process skill first; add supporting skills only when that skill delegates them or the user requires them. "Not an automatic workflow" means do not stack — not "skip `skill`".
3. Independent domains in one turn: issue multiple `subagent` / `subagent_fork` calls in the **same** assistant message.
4. After dispatch: finish only local work that does **not** need child results, then **end this turn**. Parent↔child talk uses adjacent `send_message` (Steer; tool returns delivery ack only — replies and `subagent-settled` arrive via inbox). Do **not** loop `list_agents` or status-ping `send_message` ("how is it going?") to wait.

**Below is the full content of your `using-superpowers` skill — your introduction to using skills.
For all other skills, call the dsh `skill` tool with the exact catalog name.**

[full using-superpowers SKILL.md body — frontmatter omitted]

## DeepSeek Harness platform adaptation

Your harness is DeepSeek Harness. Follow this adaptation (same role as Superpowers `references/*-tools.md`):

[full dsh-tools.md]
</EXTREMELY_IMPORTANT>

Verbatim worker preamble

<EXTREMELY_IMPORTANT>
You are a Superpowers **worker** subagent.

1. Execute the dispatch brief. Do not invent a second process workflow or re-route the parent task.
2. `using-superpowers` / skill-first orchestration belongs to the **parent**. SUBAGENT-STOP means ignore that standing skill — **not** "refuse every catalog skill".
3. Call `skill` only when the brief names a skill, or a catalog entry is clearly required to finish the brief.
4. Do the assigned work yourself unless the brief says otherwise. Prefer finishing over polling parent state.
5. You may `send_message` your direct parent (and the parent may message you): delivery ack only; put self-contained findings in the message body. Settlement is separate and unconditional.
</EXTREMELY_IMPORTANT>

Token effect

One large system-prompt section on every non-subagent request; skill catalog entries add their usual summary cost when dsh-tool-skill is mounted.

KV Cache effect

Stable prefix for a given skills root and bootstrap toggle: the section text is loaded once at plugin apply and reused. While bootstrap stays enabled and the packaged preamble is unchanged, the rendered text is byte-stable so the warm prefix cache is reused. Changing skillsRoot or disabling bootstrap changes the assembled prompt.

Skill catalog and skill tool

What the model sees

Indirectly through @deepseek-ai/dsh-tool-skill: every model-invocable Superpowers skill name/description in <available_skills>, and full bodies when loaded via skill.

Token effect

Scales with skill count and description caps owned by dsh-tool-skill. The packaged catalog is 35 skills; the default core catalog is 19, of which 16 are model-invocable. Optional bundles add 16 more skills, 15 model-invocable (grilling is user-only). Reference packs add no catalog entries.

KV Cache effect

Same as the skill consumer: initial catalog after the reusable prefix; replacements append.

Known Limitations and Deferred Work

• No runtime invariant companion is published; skills are immutable packaged files and bootstrap is a system-prompt/assemble listener with no independently observable durable event stream. Offline turn scoring lives in scripts/score-skill-first.mjs (first tool/call of each turn vs skill, plus same-turn list_agents counts after a subagent / subagent_fork dispatch); it does not enforce behavior at runtime.
• No upstream hook binary — bootstrap is a Cordis system-prompt/assemble listener, not hooks/session-start.
• A complete persona replaces every section — an agent whose composition registers a complete persona (agent presets) suppresses all prompt sections, including skill:superpowers and skill:superpowers-worker, for that agent.
• Packaged preamble must avoid {{...}} prompt-variable syntax — loadBootstrapPreamble rejects a complete {{name}} group at plugin apply (mirrors skill-always-apply); a custom skillsRoot that introduces one fails loud before the first turn.
• Cross-harness reference content is removed — foreign platform tool refs, Anthropic skill-authoring guides, and harness-specific testing examples are omitted; dsh-tools.md is the sole platform adapter and must stay deep enough for SDD wait/resume/model routing.
• Upstream sync is manual — no vendored submodule; re-porting means diffing against the recorded commits and re-applying the dsh adaptation. A local superpowers/ checkout is for comparison only; keep it outside this repository.
• diagnosing-superpowers out of catalog — obra v6.4.1+ ships a session-diagnosis skill that depends on foreign harness transcript paths and a large prompt/template tree; a shallow port would mislead. A deep dsh adaptation is a separate feature, not open sync debt. Packaged catalog length is 35; default core is 19.
• Optional bundles are opt-in — advanced, product, and authoring skills are packaged but not loaded by default. A profile that needs product discovery or skill authoring must add the bundle name to config; the plugin does not infer it from the task.
• BMAD product skills stay distilled — prd, prfaq, party-mode, retrospective, correct-course, and spec-kernel carry a lean distillation of the current v7 workflow, not BMAD's persona/runtime layer. correct-course and spec-kernel follow the v7 PO replan/spec contracts; the PRD keeps the Essential Spine's load-bearing fields, PRFAQ keeps the graded verdict and distillate, and retrospective keeps per-finding dispositions separate from the epic verdict. Unselected upstream capabilities (build/code-review/customize/bmod runtime) are out of scope, not silently dropped.
• using-superpowers is the always-on cost — its body is injected into every non-subagent turn; the packaged body is about 2.4× obra's and now carries one merged red-flag/rationalization table instead of two, plus the dsh priority and tiny-edit rules. The table was merged and trimmed to stay under the 8,000-byte bootstrap budget. Frontmatter is stripped before injection. More than twenty pinned phrases in tests/superpowers.spec.ts gate the edit, so trimming further needs its own measured change rather than a drive-by cut.
• SDD scripts resolve the repository from the working directory — run them from the target repository root, or pass -C DIR. A skill-directory cwd writes ledgers and briefs into the wrong repository.
• Reference packs are on-demand disk payload — they add no catalog entries and no always-on prompt cost, but they ship in the tarball and are read only when the owner links them at the relevant step. A pack linked from the wrong step is dead weight; the link-integrity gate catches broken paths, not bad routing.
• Peers come from DeepSeek Harness — this bundle does not vendor @deepseek-ai/dsh-*; install into a profile that already has the CLI/base stack.
• Fork product bases may mount the publish scope — some fork dsh-base compositions depend on @firefly0621/dsh-superpowers so the product default and npm identity stay one package; this monorepo mounts it only via overlay patch or dsh plugin add. Splitting a private workspace name from the publish name remains deferred.
• Description collision checks warn first — packaged catalog pairwise Jaccard ≥ 0.55 logs a warning; promoting that check to fail is a follow-up after the fused catalog stabilizes.

Dev Note

The package owns skill discovery, bundle filtering, invocation metadata, and the system-prompt bootstrap. @deepseek-ai/dsh-skill, dsh-skill-filesystem, and dsh-tool-skill remain upstream and unmodified.