dsh-superpowers

English | 中文

Superpowers for DeepSeek Harness (DSH): Jesse Vincent's skill library for disciplined agentic development (brainstorming, planning, test-driven development, systematic debugging, verification, code review), packaged as one DSH bundle.

Upstream	obra/superpowers v6.4.2, commit 8ca22dba (pin)
Skills	all 15, with every upstream file under skills/ (74), each verified by git blob hash. 73 are byte-identical to upstream
Tested on	DeepSeek Harness Desktop 0.2.0-rc.2, Windows 11

What you get

The bundle adds three plugin rows to your DSH profile. Each one can be switched on or off on its own.

Row	Default	What it does
dsh-superpowers-skills	on	Serves the skills to every session, in every workspace.
dsh-superpowers-bootstrap	on	Puts upstream's session-start text, which wraps using-superpowers, in every session's system prompt, followed by the DSH tool mapping. It replaces upstream's SessionStart hook, which DSH does not have.
dsh-superpowers-gate	off	Optional enforcement. It denies write/edit to source files until the session has loaded a skill. This row is not part of upstream.

How it differs from upstream: hardly at all, on purpose. Upstream's porting guide says to never edit skill bodies to fit a harness, and this port follows it:

• The skills ship verbatim. The one exception is a single line in using-superpowers's "Platform Adaptation" list, which points to the DSH mapping. That pointer is the only edit the guide allows.
• One DSH-only file: using-superpowers/references/dsh-tools.md. It maps the skills' actions to DSH tools. It explains that superpowers:<name> loads as <name>, that subagents are limited to one level, and how to run the bash helper scripts on Windows. The bootstrap inlines it, so every session has it.
• Every change is a reviewable file in overlays/. Nothing in skills/ is maintained by hand.

Install

From the DSH Desktop Plugins page, choose Install, then enter:

@blugart-dev/dsh-superpowers

This installs the latest release from npm, and DSH's plugin manager checks compatibility before downloading anything. To install straight from GitHub instead, use github:blugart-dev/dsh-superpowers#v1.0.1 (a release tag) or github:blugart-dev/dsh-superpowers (tracks main).

You can also ask an agent in Creator mode: plugin_manager { action: install_bundle, spec: "@blugart-dev/dsh-superpowers" }.

Restart DSH after installing or updating. A replaced package needs a fresh module generation.

Check that it works

Open a new session in any workspace, send a message, and then run:

npm run inspect-session

This needs a clone of this repository. It reads DSH's own session log, so it reports what the model actually received:

Bootstrap  present once in the system prompt
Skills     2 loaded
  14:04:26  brainstorming
  14:10:21  writing-plans

See docs/VERIFYING.md for what each line means and for the manual equivalent.

Configure

Toggle a row on the Plugins page, or with:

plugin_manager { action: set_plugin, target: dsh-superpowers-bootstrap, enabled: false }

To arm the gate, enable its row and override its config with gate: true in your profile patch. A config override replaces the whole config block, so restate every field you need:

- id: dsh-superpowers-gate
  disabled: false
  config:
    gate: true
    artifactsWritable: true
    artifactPrefixes: [docs/superpowers/plans/, docs/superpowers/specs/, research/, notes/]
    escapeSkill: { enabled: true, name: superpowers-workflow }
    announceInPrompt: true
    requiredSkills: []   # e.g. [test-driven-development, systematic-debugging]

• requiredSkills narrows which skills unlock writes. The default empty list means any skill unlocks them. The escape hatch always unlocks, so a configured gate can never lock a session out.
• Recovery. A blocked session is told exactly how to recover: load skill {"name": "superpowers-workflow"}. Only the denial message names that skill. The announcement and the skill catalog deliberately don't, because when they did, agents loaded it as a ritual instead of a real skill.
• Arming. The gate refuses to arm if its escape hatch does not resolve.
• Limits. The gate cannot see shell writes, and its behaviour in forked subagents depends on what the fork inherits (see docs/VERIFYING.md).

Every row also accepts diagnosticsLog: <absolute path>. Each activation decision is then appended to that file, which is useful when a row seems inactive.

Skill names, overrides and collisions

DSH has no skill namespaces. Upstream's skills are addressed as superpowers:<name> on Claude Code, but here they use bare names (brainstorming, test-driven-development), and the skills refer to each other by those names.

• To replace one of these skills, put a skill with the same name in a project's .dsh/skills/ or in ~/.dsh/skills/. Those roots outrank this bundle (rank 600), so your version wins, and the rest of the methodology keeps pointing at it.
• Another skill pack can collide. If it ships a skill with the same name at a higher priority, it shadows ours silently; DSH logs only a warning. Check npm run inspect-session if a skill behaves unexpectedly.

Subagents

Upstream's session-start hook never reaches subagents. To match that, the bootstrap section is empty in delegated sessions (delegationDepth > 0). Set subagents: true on the dsh-superpowers-bootstrap row to include it there too.

Known limitations

• Tested platforms. Live sessions have run on Windows 11 with DSH Desktop only. The compat workflow checks that the bundle composes on Ubuntu and Windows, but makes no model calls. macOS is untested.
• Models. Tested only with DeepSeek's models, mostly deepseek-flash.
• No effectiveness numbers. The tests check that the methodology fires: the right skill loads before code, tests come first, and verification is fresh. They do not measure whether projects turn out better than without it, and no such figure is claimed.
• Model per subagent. Upstream's subagent-driven development picks a model for each subagent. On DSH that needs the subagent model selection setting, which is off by default; otherwise every subagent runs on the session's model.
• The mapping is still young. So far, every long field test has found something in the DSH tool mapping to correct. Expect fixes in patch releases.
• Wrap-up verification. In about a third of wrap-up turns the model restates an earlier test result instead of re-running the tests, even though verification-before-completion forbids it (see CHANGELOG.md).
• The gate is opt-in and cannot see writes made through the shell.

Uninstall

Remove the bundle from the Plugins page, or run plugin_manager { action: remove_bundle, target: "@blugart-dev/dsh-superpowers" }, then restart DSH.

Develop

Requires Node ≥ 23.6 and git.

npm run verify        # sync:check + contract checks + tests
npm run compat        # does the bundle compose on the installed DSH? (no model calls)
npm run eval:harness  # behavioural tests in real headless sessions (spends tokens)

• CONTRIBUTING.md: how to edit a skill, upgrade upstream, and add a test.
• docs/ARCHITECTURE.md: how the pieces fit, and why.
• evals/: behavioural scenarios for checking that DSH actually works the Superpowers way.
• docs/RELEASING.md: the release checklist.
• SECURITY.md: what runs where, and how to report a vulnerability.

License

MIT. The skills are derived from obra/superpowers, Copyright (c) 2025 Jesse Vincent, under the MIT license in LICENSE.superpowers. See LICENSE for this package.