dsh-everything-claude-code

English | 中文

Adapts WorldFlowAI/everything-claude-code (ECC) into a DeepSeek Harness bundle: 11 skills, 15 prompts, 9 personas, 8 rules, an adapted Claude Code hook set, an agent preset, and a Chinese Settings page.

Verified against DSH 0.2.0-rc.2, profile desktop.

Blocking rules

Five PreToolUse rules deny a tool call outright. Two are ported from ECC; three were added for this port.

Rule	Denies
block-dev-server	starting a dev server in the foreground, where its log is unreachable
block-destructive-command	force push, reset --hard, clean -f, forced branch delete, remote branch delete, and a recursive delete of an unscoped or out-of-workspace path
block-hardcoded-secret	writing a real credential into source, config, or docs
block-stray-docs	a loose .md/.txt outside the conventional locations
confirm-git-push	not a deny — it raises an approval prompt before a push

What is deliberately allowed

A blocking rule that fires on ordinary work gets switched off, and then it protects nothing. So each rule has explicit exemptions, asserted by tests:

• .env, *.pem, *.key, credentials.* are writable. A key store is where a real key belongs; flagging it would train the user to bypass the hook. The other half of that bargain is .gitignore, which covers every path the hook treats as a key store.
• Secrets are matched by credential shape, never by variable name. This bundle ships nine secret-shaped placeholders in its own ported security guidance (const apiKey = "sk-proj-xxxxx"), so a name-keyed rule would block edits to the plugin's own docs. What is matched is a known issuer prefix with a real body — OpenAI sk-, GitHub ghp_/github_pat_, AWS AKIA, Google AIza, Slack, Stripe, npm, PyPI, GitLab, Hugging Face, PEM headers, JWTs — or a mixed-case-plus-digit literal assigned to a secret-like name.
• tests/, fixtures/, examples/ are exempt by path: a fixture with a plausible-looking key is correct, not a leak.
• Ordinary git and ordinary deletes pass: git push, git reset HEAD~1, git branch -d, git clean -n, rm -rf ./node_modules, and rm -rf <workspace>/dist are all allowed.
• --dry-run and -n are respected, so a command can be reviewed before it is trusted.

Every rule is one script under hooks/scripts/, so a rule can be removed by deleting its entry from hooks/hooks.json without touching the others.

The agent preset (the "mode")

presets/everything-claude-code.patch.yml declares a selectable session mode named Everything Claude Code, listed in Settings → Agent presets and in the new-session picker. Selecting it composes:

• a persona stating the ECC working agreement — TDD first, verify before claiming done, never hardcode a secret, immutable data, focused files, security review — and naming the skills to load;
• the standard agent plane (filesystem and search, skill loading, plan mode, compaction, delegation, todos, web, present);
• agent-instructions, so the project's AGENTS.md still reaches the model.

Why a preset instead of a global row: a preset is the only seat that composes a session's persona, tools, and context. Mounting ECC globally would change every session; a preset lets one session opt in. order: 4 places it after the shipped standard(1)/ptc(2)/minimal(3), so it reads as an opt-in mode, not a replacement default.

To declare an extra patch file, dsh.bundle.patch takes a list of paths.

Settings page

Settings gains an Everything Claude Code entry showing what the plugin actually publishes: the 11 registered skills, every hook point with its matcher, the resource counts, and the migration caveats. It is Chinese by default and follows a language switch in place.

• Registered in the settings.section slot — a top-level settings page. Do not file it under settings.plugins.tab: that is the read-only "Built-in plugins" inventory of what the deployment ships, so a third-party page does not belong there.
• Data comes from GET /api/everything-claude-code/status under the authenticated /api fence; counts are read from the package directory at request time.
• Styled only with --dsw-alias-* theme tokens, so light and dark both work.

The client half needs a DSH restart the first time it appears. The client-module scan runs once at process start and caches each package's manifest in a write-once pkgMeta map with no invalidation path (dsh-client-modules/lib/index.js). A package that gains dsh.client while the Host is already running stays invisible until DSH restarts. Later edits to client.js do hot-reload.

Install

# 1. Clone through the mirror (direct GitHub is unreachable on this host)
git clone --depth 1 https://gh-proxy.com/https://github.com/WorldFlowAI/everything-claude-code.git

# 2. Adapt the sources into this bundle
node scripts/adapt-skills.mjs skills
node scripts/adapt-resources.mjs resources

# 3. Install into the profile (via the plugin manager, not the CLI —
#    the desktop profile is owned by the Electron app)

Then install with the plugin_manager tool, action: install_bundle, target: <absolute path to this directory>.

What maps to what

ECC	DSH	Note
skills/ (11)	runtime skill registry	Needs name + description frontmatter
commands/ (15)	prompt documents	DSH commands are human-facing and never reach the model
agents/ (9)	persona prompts	DSH has no per-file agent registry
rules/ (8)	SessionStart context	Three rules are injected; the rest load on demand
hooks.json	@deepseek-ai/dsh-hooks-claude-code	Matchers renamed to DSH tools
.claude-plugin/plugin.json	package.json + dsh.bundle.patch	Different manifest contract

Three bugs that a naive port hits

1. Skill frontmatter is required, and its absence is silent. Three ECC skills (eval-harness, project-guidelines-example, verification-loop) ship without frontmatter. dsh-skill-filesystem drops such a skill with a warning the model never sees, so the skill simply is not there. scripts/adapt-skills.mjs adds name and description.

2. baseUrl in a bundle patch is the PROFILE ROOT, not the patch's directory. new URL('hooks/hooks.json', baseUrl) resolves to <profile>/hooks/hooks.json, which does not exist. The bridge then logs a warning and runs no hooks at all — no error surfacing to the user. Anchor on the installed package instead:

configPath: !!js "...new URL('node_modules/dsh-everything-claude-code/hooks/hooks.json', baseUrl)"

3. Blocking through exit code 2 does not work in this bridge. Claude Code documents exit 2 + stderr as a block, and dsh-hook-protocol maps exit 2 to decision: 'block'. In 0.2.0-rc.2 the bridge instead recorded it as exitCode: 1, decision: 'pass' and let the tool run. Observed directly in a session log: the hook's block message reached stderr while the call was not stopped.

The channel that does work is structured stdout with a clean exit:

{"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"deny","permissionDecisionReason":"…"}}

All blocking hooks in this bundle use it and exit 0. blockByExitCode is kept for a real Claude Code host.

Tests

node tests/run-all.mjs

Suite	Checks	What it proves
hooks.test.mjs	23	Each hook's decision against the contract this bridge honors
blocking-hooks.test.mjs	24	The risk hooks deny through the channel this bridge honours
risk-detection.test.mjs	41	The detection logic, including the cases it must NOT flag
dual-version.test.mjs	15	Correct behaviour under both the 0.2.0 and 0.2.1 directory contracts
preset.test.mjs	15	The preset satisfies the agent-preset schema and names the ECC practices
client-bundle.test.mjs	14	The browser factory loads, exports, and renders the page
status-payload.test.mjs	12	The route answers, and every field the page reads exists
memory-hooks.test.mjs	11	The two bridge-unsupported events are covered by native listeners
skill-catalog.test.mjs	11	Every shipped skill is well formed and collides with no installed one
skill-provider.test.mjs	11	A body edit is visible on the next load, with no stale cache
no-stale-refs.test.mjs	10	No Claude Code path, command, tool or agent name survives in prose
preset-mount.test.mjs	9	Every preset child package resolves, so the preset can mount
bridge-parse.test.mjs	7	No event or matcher is silently dropped

203 checks across 13 suites.

Diagnostic helpers, which read the authoritative session log rather than guessing:

node tests/hook-timeline.mjs <session-dir>      # when each matcher fired, and decisions
node tests/analyze-hook-decisions.mjs <dir>     # decision and exit-code distribution
node tests/replay-hook.mjs <script.mjs> '<json>'  # raw exit code, stdout, stderr
node tests/read-session-hooks.mjs <dir>         # all hook records

A session log is many concatenated Zstandard frames (one per flush — 570 frames in one observed log). Node's streaming decoder stops after the first frame, so these helpers advance frame by frame.

Environment fix

The install first failed with ERR_PNPM_UNEXPECTED_STORE: the profile was built by pnpm 11.27.1 (store v11), while pnpm on PATH resolved through a corepack shim to 10.17.0 (store v10). Recording the version that built the profile repairs it:

{ "packageManager": "pnpm@11.27.1" }

If that line makes the plugin manager report ambiguous-install, remove and reinstall the bundle: that error means the resolver found no newly changed dependency, and a fresh install takes the reload path instead.

Layout

index.js              Host half: skills provider, status route, memory listeners
client.js             Client half: the Chinese Settings page
cordis.patch.yml      Bundle patch: the skills row and the hook-bridge row
presets/              The Everything Claude Code agent preset
locale/               Display metadata for the plugin card (en, zh)
skills/               11 ported skills + 24 generated ecc-* skills
resources/            rules, commands, agents, contexts
hooks/hooks.json      Adapted hook config (DSH tool names)
hooks/scripts/        One script per hook, plus lib.mjs
scripts/              Idempotent adapters and GitHub helpers
tests/                Test suites and session-log diagnostics

Memory persistence (the two events the bridge cannot run)

The Claude Code hook bridge supports seven events, and ECC's PreCompact and SessionEnd hooks are not among them — so at first this port had no memory persistence at all.

A bridge cannot be extended from outside it, but that is not a dead end: DSH exposes the same two moments as native host events, which this plugin listens to directly. Both replacements live in index.js:

ECC hook	Replaced by	When it fires
PreCompact	session/event carrying compaction/start	compaction is genuinely beginning, before summarization
SessionEnd	agent/disposed	after driver quiescence, before session detachment

Notes are appended to $DSH_HOME/sessions/memory/ as Markdown, one file per day: YYYY-MM-DD-compaction.md and YYYY-MM-DD-sessions.md.

Two deliberate departures from ECC's scripts: they wrote to ~/.claude/sessions/ and rewrote a *.tmp file with sed, both Claude Code shapes that do not exist here. These notes are append-only, nothing is mutated in place, and a write failure is logged and swallowed — a memory hook must never be able to break the session it is recording.

DSH version compatibility

Tested against 0.2.0-rc.2 (the latest channel) and 0.2.1-alpha.2 (the alpha channel). One contract changed between them, and the adaptation is written so that both keep working.

Contract	0.2.0-rc.2	0.2.1-alpha.2
Session directory	session.header.cwd, fixed at session creation	ctx.workingDirectory, mutable mid-session
Where hooks run	the workspace root	the session's current directory
dsh-skill, dsh-agent-preset, dsh-persona	—	unchanged

What changed, and why it matters. In 0.2.1 a session can change its working directory, so the hook payload's cwd is no longer the project root. A hook that trusted a single cwd value would treat the original workspace as foreign after a directory change, and block a routine rm -rf <project>/dist or a legitimate documentation edit.

How this plugin adapts. The hooks read every available signal — the payload cwd, CLAUDE_PROJECT_DIR, and DSH_WORKSPACE — and treat their union as the session's zone. That is correct under either contract: on 0.2.0 the signals agree, and on 0.2.1 they both still describe where the session legitimately works. Nothing branches on a version number, which is what makes it hold on 0.2.0, on 0.2.1, and on versions not yet published.

Verify it yourself against any two versions:

node scripts/check-surface.mjs 0.2.0-rc.2 0.2.1-alpha.2   # every consumed API surface
node tests/dual-version.test.mjs                          # behaviour under both contracts

check-surface.mjs asserts that every harness API the plugin consumes is present: skill provider registration, session/event, agent/disposed, the hook bridge config, the preset schema, the persona prefix, the client module loader, and the settings.section slot. Run it after a harness upgrade — a breaking change then becomes a named failure instead of a runtime surprise.

Known limits

• SubagentStart/SubagentStop report a constant agent_type, and start-context only reaches an in-process child.
• The Stop hook cannot block safely: the bridge always reports stop_hook_active: false, so an unconditional block would loop. Ours only warns.
• Memory persistence records that compaction and session end happened, not the session's content, and it does not reload a previous session's context into a new one. ECC's session-start hook covers the load side and runs through the bridge.
• ECC's project-guidelines-example skill references a specific unrelated production project; it is kept only as a template.