hoyyang/dsh-best-ponytail

Ponytail (MIT) packaged for DeepSeek Harness: your AI agent now thinks like the laziest senior dev in the room — the best code is the code you never write.

中文 · Releases · Changelog

Install

dsh plugin add dsh-best-ponytail   # npm channel (the exclusive install name)
dsh plugin add github:hoyyang/dsh-best-ponytail   # or via GitHub

Zero config, zero behavior change on install: the always-on injection defaults to off, and all six skills are usable immediately. Tested on dsh 0.1.5. Dual-channel distribution: npm as dsh-best-ponytail (the original name dsh-ponytail is taken by an unrelated package) + GitHub.

What you get

Six skills (passive + explicit invocation)

• ponytail — the 7-rung ladder before any code: does it need to exist → already in the codebase → stdlib → native platform → installed dependency → one line → the minimum that works
• ponytail-review — over-engineering-only review: reinvented stdlib, needless dependencies, speculative abstractions, dead flexibility; one line per finding: where / what to cut / what replaces it
• ponytail-audit — whole-repo bloat audit: a ranked list of what to delete, simplify, or replace with stdlib; report only, no fixes applied
• ponytail-debt — harvests every ponytail: marker comment into a debt ledger, so "later" never silently becomes "never"
• ponytail-gain — the upstream benchmark scoreboard: less code, less cost, more speed (medians)
• ponytail-help — quick-reference card for all modes, skills, and commands

One command, always in reach

• /ponytail — switch the always-on level per session (lite / full / ultra / off), effective next turn
• /ponytail status — current level and its source (config default or session override)
• /ponytail reset — clear the session override, back to config default

Optional always-on injection (the upstream hook, translated)

• systemPrompt section — the ladder is welded into every turn's system prompt; 100% of coding happens under the rules, quick one-line edits included
• four config levels — off / lite / full / ultra; full is the level the upstream benchmark measured
• session override — /ponytail off silences it for this session only; restart falls back to config

Packaging

• upstream untouched — all 193 upstream files (skills, hooks, benchmarks, docs, ...) vendored verbatim under upstream/, pinned to commit e3ba2aa
• zero tools registered — nothing enters the tools schema; uninstall leaves no residue
• MIT with attribution — upstream © DietrichGebert; the adapter layer is MIT too

30-second start

1. Run dsh plugin add dsh-best-ponytail
2. Configure nothing — six skills are already in the catalog
3. Verify: say "ponytail, build me a small X" — the reply should end with skipped: X, add when Y
4. Check the level: type /ponytail status
5. Want always-on: set mode to full in the profile config (see below)
6. Hunt over-engineering: say "audit this codebase"
7. Review debt: say "ponytail debt"
8. Silence for this session: /ponytail off
9. Back to default: /ponytail reset
10. Uninstall: dsh plugin remove dsh-best-ponytail — nothing left behind

Everyday use

• Install and do nothing — the ladder is available in coding tasks, skills visible in the catalog
• For 100% coding coverage: set config mode to full, every turn carries the ladder
• Want minimalist now: say "ponytail ultra" — the agent questions whether the requirement should exist at all
• Want gentle now: say "ponytail lite" — normal build plus one lazier alternative
• Rework after the fact: say "review for over-engineering" and decide from the deletion list
• Inheriting an unfamiliar repo: run "audit this codebase" first, slim top-down by rank
• Long-lived projects: periodic "ponytail debt" keeps verbal promises from evaporating
• Reporting savings: say "ponytail gain" for the upstream benchmark scoreboard
• Forgot the commands: say "ponytail help" — the quick-reference card lists everything
• Parallel sessions: levels are per-session; /ponytail off affects only the current one

Use cases

• Rapid prototyping — "build me X" no longer returns five dependencies and three layers of abstraction
• Legacy refactors — review mode hunts only what can be deleted: reinvented stdlib, dead options, scaffolding for a future that never came
• Review second pass — beside correctness review, a dedicated complexity hunter, one line per finding
• Debt inventory — run the ledger periodically and turn "later" into a visible list
• Team consistency — everyone runs the same ruleset; output style stays uniform across agents
• Cost & speed — less code means fewer tokens and faster delivery; upstream measured -20% cost, -27% time
• Learning taste — watch the agent ask "does this need to exist" first; train your own trade-off instinct
• Monorepo slimming — the audit's ranked list doubles as a refactor backlog
• Dependency hygiene — ladder rung 5 forces the installed-dependency check before any new package
• Model migration — rules are model-agnostic; switch models without switching discipline

Input / output

Live echo samples (captured in-session):

$ /ponytail status
ponytail always-on injection: FULL (config default) — injected into the system prompt, effective next turn
Skills (passive + explicit) are available regardless of this setting.
Switch: /ponytail lite|full|ultra|off · /ponytail reset (back to config default: full)

$ /ponytail ultra
ponytail always-on injection set to ULTRA for THIS session — effective next turn.
ponytail always-on injection: ULTRA (session override) — injected into the system prompt, effective next turn
Skills (passive + explicit) are available regardless of this setting.
Switch: /ponytail lite|full|ultra|off · /ponytail reset (back to config default: full)

You say / type	You get
/ponytail status	current level + source + how to switch
/ponytail ultra	YAGNI-extremist level for this session, next turn
"ponytail, add a cache"	the minimal ladder-passed implementation + one skipped line
"review for over-engineering"	one-line-per-finding deletion list
"audit this codebase"	ranked repo-wide bloat list, no fixes applied
"ponytail debt"	ledger of every ponytail: marker
"ponytail gain"	upstream benchmark scoreboard (-54% LOC / -20% cost / -27% time)
"stop ponytail"	natural style for this session

Outputs

• Minimal diff — the shortest working change the ladder allows, no unrequested abstraction layers
• One skipped line — every lazy cut is traceable: what was skipped, when to add it back
• Review deletion list — one line per finding: location / what to cut / what replaces it
• Audit ranked report — repo-wide over-engineering priorities, usable directly as a refactor backlog
• Debt ledger — every ponytail: marker collected, each with its upgrade path
• Gain scoreboard — upstream benchmark medians: -54% LOC / -20% cost / -27% time
• Status text — instant echo of current level, source, and switch instructions
• Installable tgz — prebuilt GitHub Release asset, one dsh plugin add away

Advanced (config)

The always-on level lives in the profile config (default off):

- id: dsh-best-ponytail
  config:
    mode: full   # off | lite | full | ultra

How it works

A bundle plugin mapping upstream concepts onto DSH: skills/ are served through a ctx.skills provider (passive trigger by description match); the upstream SessionStart hook's per-turn ruleset injection becomes a ctx.systemPrompt section controlled by config.mode with in-memory per-session overrides; the upstream commands/ponytail.toml level semantics become the /ponytail host command. Everything else is vendored verbatim for provenance.

Reliability & verification

• Smoke test — the real apply() against a mock ctx: 6 skills registered, ladder body served, level state machine, unknown-level usage message (npm test)
• Cold-start check — boot-check six static classes (junction / bundle manifest / disabled conflicts / runtime imports / client id consistency / duplicate insert ids) all green
• Lifecycle — two measured inject/uninstall cycles: catalog appears/disappears live, junction removed, residue check PASS
• Fail loud — missing skill tree, unreadable ruleset, or invalid config throws with the failing path named; no silent fallbacks
• Security — no credentials, zero new tools, one dependency (schemastery ^3.18.0), no install scripts
• Content fidelity — 193 upstream files byte-identical, pinned commit, auditable
• Compatibility — tested on dsh 0.1.5 web profile; degrades with a warning on profiles missing services
• Upgrade safety — profile edits are backed up (.bak-timestamp) before writing
• Clear attribution — upstream MIT declared separately; adapter code is a single lib/index.js
• Known limits — gains concentrate where agents over-build; near zero on already-minimal code; the npm name dsh-ponytail is taken by an unrelated package, hence renamed to dsh-best-ponytail

FAQ

Why nothing changed after install? — mode defaults to off: the skills are active (passive + explicit), the always-on injection needs opting in. Minimal surprise by design.

Related to the npm package dsh-ponytail? — No. That is an independent third-party implementation (ccll, 2026-08); this package renamed to dsh-best-ponytail for npm and ships via both channels.

Does passive triggering miss? — Sometimes; that is the physical ceiling of on-demand skill loading. Set mode to full for 100% coverage while coding.

Are upstream benchmarks/tests really in the package? — Yes, vendored under upstream/ as provenance reference; they do not run inside DSH.

Build locally

npm install && npm run build && npm test   # pure JS, no transpilation: syntax check + smoke test

License

MIT. Upstream ponytail © DietrichGebert (MIT), unmodified — this package only adapts it for DSH; the adapter code is MIT too. See LICENSE.