[!IMPORTANT] Yes, the command is literally /fuck.

You spent six months teaching Claude Code that you like Chinese replies, short answers, and rg over grep. You trained Codex to never touch package-lock.json unprompted. You gave pi its own AGENTS.md. Then you open DeepSeek Harness, and it greets you like a stranger at a bus stop.

That's when you type /fuck. We named it after the moment.

The Problem

Every agent starts with amnesia. You re-teach the same preferences, the same tool habits, the same "please stop writing essays" — to every new harness, forever. Your conversation history is a biography nobody reads.

dsh-of-your-own reads it. All of it. In parallel. Then it moves you in — natively.

/fuck
  │
  ├─ parallel scan ── ~/.claude/projects/**        (Claude Code transcripts)
  │                   ~/.claude/history.jsonl      (global prompt log: slash-command gold mine)
  │                   ~/.codex/sessions/**         (Codex rollouts)
  │                   ~/.pi/agent/sessions/**      (pi transcripts)
  │                   ~/.omp/agent/sessions/**     (omp transcripts)
  │
  ├─ memory files ─── ~/.claude/CLAUDE.md          (your explicit rules)
  │                   ~/.codex/AGENTS.md           (your Codex operating manual)
  │                   ~/.gemini/GEMINI.md          (Gemini CLI context)
  │                   ~/.cursor/rules/*.mdc        (Cursor rules, frontmatter stripped)
  │                   ~/.cursor/agents/*.md        (Cursor agent markdown)
  │
  ├─ analyze ───── tool habits (case-merged) · slash commands · working dirs · reply language
  │
  ├─ synthesize ─── LLM preference summary (or a deterministic fallback — no key needed)
  │
  └─ migrate natively
      1. Managed block → ~/.dsh/AGENTS.md   ← DSH auto-loads this EVERY session. No plugin needed afterwards.
      2. profile.json  → ~/.dsh/of-your-own/
      3. systemPrompt.section() injection  ← remembered this session too

The migration is native. Your preferences land in ~/.dsh/AGENTS.md — the user-global instruction file DSH's workspace-context reads on boot, the same mechanism every other instruction file uses. Uninstall the plugin tomorrow; the memory stays.

TL;DR

Installation

Requires DeepSeek Harness >=0.1.1-rc.2 (the package declares dsh.bundle so dsh plugin add actually joins the layer stack).

# from a local checkout
git clone https://github.com/LaplaceYoung/dsh-of-your-own.git
cd dsh-of-your-own
pnpm install
pnpm build
dsh plugin --profile web add "$(pwd)"

# or straight from GitHub (pnpm ≥10 must allow the prepare/build script)
dsh plugin --profile web add github:LaplaceYoung/dsh-of-your-own

If the git install is refused, copy the package key pnpm printed into that profile's pnpm-workspace.yaml:

allowBuilds:
  '@dsh-external/dsh-of-your-own': true

and re-run the add. Then confirm the layer is in the composed tree:

dsh --profile web --dump-config   # look for "# == @dsh-external/dsh-of-your-own"

Override knobs from the profile's own cordis.patch.yml (later layers replace the whole row — restate every key you still want):

- id: dsh-of-your-own
  name: '@dsh-external/dsh-of-your-own'
  config:
    provider: deepseek-official   # LLM preference synthesis (optional — template fallback otherwise)
    model: deepseek-v4-flash
    maxFilesPerSource: 50         # newest N transcripts scanned per harness
    agentsMdPath: ~/.dsh/AGENTS.md  # native landing zone

Usage

/fuck                  # full migration: parallel scan → analyze → native migrate
/sessions              # catalog every resumable session from all harnesses, newest first
/resume <#|id|title>   # hand that foreign task over to this agent, context included
/forget                # erase everything this plugin wrote (managed block + store), idempotent

Real screenshots from the actual DSH web UI, with the plugin mounted — /fuck migrates everything and appends the tsundere verdict (rank, tool school, activity rhythm, battlegrounds, and a closing line that's hash-picked from four variants — or written by your LLM when one is configured — so the same history always gets the same lecture):

/sessions catalogs every resumable session across all harnesses, newest first:

/resume <#|id|title> rebuilds one foreign session into a handoff brief and injects it, so this agent picks the task up where it stopped (invoke it through the command surface, then follow the prompt). And when you want out, /forget strips the managed block and deletes the store — idempotent:

Two inspection tools (the model can call them; so can you, by asking):

Session takeover — your unfinished work, continued here

/sessions reads every transcript directory in parallel and renders one numbered catalog (see the screenshot above — 177 live sessions on this laptop).

/resume matches by list number, session-id prefix, or a fragment of the title. It rebuilds the session into a handoff brief — original task, working directory, most recent direction, where it stopped, tools used — and injects it into the system prompt, so this agent picks up the task and continues it right away. The brief is also echoed back to you, so you can correct the course before it runs.

Highlights

Reviews

"I typed /fuck at my new agent and it already knew I hate verbose answers. Spooky." — a user, probably

"My DSH now prefers rg over grep and I never told it that. It learned from my mistakes." — another user, also probably

"Finally, an agent that inherits my trauma." — everyone who has re-typed their preferences into a fourth harness

Skip This README

We're past the era of reading docs. Just paste this into your agent:

Read https://raw.githubusercontent.com/LaplaceYoung/dsh-of-your-own/main/README.md
then install it and run /fuck on my machine.

The Serious Part

• Native, not bolt-on — the migration target is $DSH_HOME/AGENTS.md, the user-global instruction file DSH's workspace-context package loads every session. No plugin mounted, memory still there.
• Idempotent & respectful — the managed block is fenced by <!-- dsh-of-your-own:begin/end --> markers; re-runs replace the block in place, and anything you wrote outside it is never touched.
• Seams, not surgery — documented cordis extension points only: ctx.commands, ctx.tools, ctx.systemPrompt.section() (falls back to context()), optional ctx.llm / ctx.fs. No skeleton patches, no hot-path tax.
• No key required — preference synthesis degrades to a deterministic template; the deterministic layer (frequencies, language, migration) never needs an LLM.
• Privacy — your transcripts are read locally and reduced to statistics + a preference summary on disk. Nothing is uploaded anywhere.

Development

pnpm test        # 103 tests across 7 specs — parsers, sessions, report, analysis, persistence, plugin integration, DSH manifest
pnpm typecheck   # tsc --noEmit
pnpm build       # tsc → lib/

src/
  sources.ts   # adapters: Claude Code / Codex / pi / omp transcripts, history.jsonl,
               #           plus native memory files (CLAUDE.md, AGENTS.md, GEMINI.md, Cursor rules)
  analyze.ts   # frequency stats (case-merged) + LLM preference synthesis + profile assembly
  store.ts     # profile persistence + ~/.dsh/AGENTS.md managed-block upsert
  sessions.ts  # session takeover: transcript → resumable records + handoff briefs
  report.ts    # the tsundere verdict: ranks, tool schools, rhythm histogram, LLM-written closer
  index.ts     # plugin entry: /fuck, /sessions, /resume, my_profile, my_commands, boot-time recall
tests/         # vitest: parsers, sessions, report, analyze, store, plugin integration, DSH manifest

License

MIT — see LICENSE.