dsh-ponytail

Your agent adds a cache class, a config flag, and a factory for one product. This plugin makes it stop and write the one line that works instead. The ladder: question whether the thing needs to exist, reuse what is already in the repo, then the standard library, then the platform, then an installed dependency, then one line. Stop at the first rung that holds. Adapted from DietrichGebert/ponytail (MIT): "He says nothing. He writes one line. It works."

What you get

• The ladder on every request. While any level except off is active, the mode-filtered ruleset joins the system prompt.
• Six skills the skill tool can load, so they also appear as /ponytail-review, /ponytail-audit, … in the composer: ponytail, ponytail-review, ponytail-audit, ponytail-debt, ponytail-gain, ponytail-help.
• Two ways to switch level. The model calls the ponytail tool; you type /ponytail <level> or just say stop ponytail.
• A settings card and a composer chip, so the level is visible without opening Settings.
• No unrequested abstractions, no boilerplate. Deletion over addition, boring over clever.

Install

Requires DeepSeek Harness >=0.2.1-rc.2 <0.3.0. Restart DSH and refresh the browser after upgrading.

Install it as a bundle. dsh plugin add … mounts the row from the package's own patch layer, which is what the settings editor can write to. A row added with --patch is an overlay: it disappears at the next start, and the Plugins card cannot save into it (the editor refuses a write an overlay would win).

dsh plugin --profile web add --save-exact @maci0/[email protected]

Pin an exact npm version. To upgrade, run the same command with the newer version, then restart dsh web (bundle layers compose at boot).

Use it

/ponytail ultra   -> Ponytail level: ultra (was full).
/ponytail         -> Ponytail level: ultra.
/ponytail off     -> Ponytail off (was ultra). Normal behavior.

Then ask for the work. This is the full example from the skill itself:

> Add a cache for these API responses.

`@lru_cache(maxsize=1000)` on the fetch function. Skipped custom cache class,
add when lru_cache measurably falls short.

Model side, the tool takes one optional argument: {"mode": "lite"} sets and persists a level, and calling it with no argument reports the current one.

Typing stop ponytail or normal mode as an ordinary message has the same effect as /ponytail off, and it lands on the turn that carried it. Only the human's own words count: injected context riding the same event stream cannot toggle the level, and the message must be the command, so "add a normal mode toggle" is left alone.

Levels

| Level | Behavior | Persisted | |---|---|---| | lite | Build what was asked, name the lazier alternative in one line. | yes | | full | The ladder enforced. Default. | yes | | ultra | YAGNI extremist: challenge the requirement before building it. | yes | | review | Over-engineering review; points at the ponytail-review skill. | no, session only | | off | No injection. Normal behavior. | yes |

The four persistable levels live in the ponytail settings namespace, so the card, the chip, the tool, and /ponytail agree and the choice survives a restart. review stays session-only because it is a review mode, not a level a deployment should start in; the tool and the command still accept it. The card and the chip show the level the host is using, so they show review, and any level kept only for this session because the settings write failed, labelled as session-only.

Lazy is not negligent. Understanding the problem, trust-boundary validation, data-loss handling, security, and accessibility are never on the chopping block.

Configure

| Field | Default | Meaning | |---|---|---| | defaultMode | full | Startup level. Must be off, lite, full, or ultra. |

The exported Config schema defaults the field, so a row that omits defaultMode starts in full. The user's settings namespace overrides it once a level is chosen in the card or by /ponytail. An invalid level fails while the plugin loads rather than quietly doing the wrong thing: the loader validates the row against the exported schema.

To set the startup level from the environment, override the row in your profile's cordis.patch.yml with a !!js expression (same id, which replaces the row's whole config):

- id: ponytail
  config:
    defaultMode: !!js process.env.PONYTAIL_DEFAULT_MODE ?? 'full'

Override the row this way; do not paste the bundle's insert of the id: ponytail row into your profile's cordis.patch.yml: insert does not dedupe ids, and a second row mounts the plugin twice.

How it works

The package declares dsh.bundle, so dsh plugin add appends it to dsh.profile.bundles and the row in its own cordis.patch.yml applies as a layer.

The host half mounts through public Cordis extension points: systemPrompt.section, skills.registerProvider, tools.register, commands.register, loader/volatile-update (the settings document writes the row's volatile defaultMode), session/event for the message switch, and webServer for GET /ponytail/level, which answers { mode, source } (settings or session) behind the connection trust fence. The browser half reads that route on every settings change and every 5 seconds, draws its card into the public plugins.row.config slot from configForms, registers its copy through locale.register, and draws its chip into conversation.input.left, so this plugin needs no client change of its own.

The entry point is the built lib/index.js (declarations in lib/types/); bun run build regenerates it from src/. lib/client.js is hand-authored plain JavaScript: the client module system serves it as a lazy-CJS factory on window.__ModuleLoader__ because the package exports ./client, and it is not built. Skills come from skills/<name>/SKILL.md with frontmatter parsed by yaml; the provider takes its rank and name grammar from @deepseek-ai/dsh-skill, projects disable-model-invocation, user-invocable, and whenToUse, and settles on the lookup's abort signal. Tools use defineTool from @deepseek-ai/dsh-tools and forward exec.signal. Only skills/ponytail/SKILL.md is the source of truth for the ruleset; this README keeps no second copy.

Limits

• It does not skip comprehension. The ladder shortens the solution, never the reading. The skill says so in its own text.
• The level is process-wide. One prompt section, one namespace value: every agent in the process shares it.
• External subagents ignore it. In-process children inherit the ruleset, but subagent-claude-code and subagent-codex spawn their own CLI with its own system prompt, and no harness extension point wraps a spawn.
• Two locales. The card and the chip ship en and zh; any other locale falls back through the service's own chain.
• Host source edits need bun run build and a restart, because the Loader loads lib/index.js and a bundle layer composes at boot.
• Browser-half edits need a page refresh.
• A level change is not a session event. A replayed session shows the ruleset each request already carried.

Development

bun install         # the client packages are optional peers and stay uninstalled
bun run build       # tsc -p tsconfig.build.json -> lib/index.js + lib/types/
bun test            # every tests/*.test.ts, no build step
bun run typecheck   # tsc -p tsconfig.json

dsh loads plugins on Node ^22.19.0 || >=24.0.0; development and tests run on bun.

For local development, install the checkout into a profile with dsh plugin --profile <name> add <path-to-checkout>.

The suite covers level normalization and filtering, the fake-host surface, the skills provider, the client card, and a real Cordis composition mount next to the real skill registry. To uninstall, run dsh plugin --profile web remove @maci0/dsh-ponytail; that removes the dependency and the bundle layer together.

Attribution and license

MIT. Skill content and mode semantics: © DietrichGebert (ponytail). DSH port: see LICENSE.

The six skills/*/SKILL.md files are verbatim copies, so they carry upstream's figures and citations rather than this package's. /ponytail-gain therefore reports the original five-task single-shot benchmark (80–94% fewer lines, 47–77% cheaper, 3–6× faster), which upstream's own README has since revised to roughly 54% fewer lines, 20% cheaper and 27% faster on its agentic benchmark.