dsh-mask
PII masking middleware for DeepSeek Harness — anonymize personal data before it reaches the model, keep it reversible host-side.
Phones, emails, ID cards, bank cards, keys, and more become placeholders at the model boundary; the plaintext never enters your session log.
DSH Market
English · 简体中文 · Español · Português · हिन्दी
dsh plugin --profile <name> add dsh-mask
📖 Ecosystem knowledge base — measured data, not marketing: plugin development guide · plugin-selection data · maintenance criteria.
Install & uninstall
- git channel (latest
main): dsh plugin --profile web add github:PerryLink/dsh-mask (equivalent to installing from git+https://github.com/PerryLink/dsh-mask.git). No build step — index.mjs and lib/ are the shipped artifacts.
- npm channel (published releases):
dsh plugin --profile web add dsh-mask.
- tarball channel:
pnpm pack in this repo, then dsh plugin --profile web add ./dsh-mask-<version>.tgz.
- uninstall:
dsh plugin --profile web remove dsh-mask (or remove the row from the profile patch).
dsh-mask no longer bundles the storage stack. Profiles that already compose it (the web profile does, via @deepseek-ai/dsh-web-app) provide storageDomain, so persistence works out of the box. On a bare profile without storage the plugin still mounts and masks, but the restore table is memory-only (lost on restart) — compose the storage stack in your profile patch, or set persistRestoreTable: false.
⭐ 如果它帮到了你
这个插件是 DSH 插件家族的一员（95 个，全部 Apache-2.0）。如果你在用，给个 star —— 它不会解锁任何功能，但会让下一个人在搜索里更容易找到它。
English: part of a 95 plugin family for DeepSeek Harness. If it is useful, a star helps the next person find it — nothing is gated behind it.
What is dsh-mask?
PII masking middleware for DeepSeek Harness — anonymize personal data before it reaches the model, keep it reversible host-side.
Phones, emails, ID cards, bank cards, keys, and more become placeholders at the model boundary; the plaintext never enters your session log.
Workflow diagram for dsh-mask: call → plugin → result.
Workflow overview: call → plugin → result. The examples below use the real command surface.
Comparison
Measured comparison chart for dsh-mask
108 synthetic samples (60 positive, 48 negative) · from benchmark/RESULTS.md
Compatibility
| Surface | Status |
|---|
| Harness | DeepSeek Harness dsh-v0.2.1-alpha.1 (adapted 2026-09-18): the session envelope keeps its ignorable field for stored-log read compatibility only - Session.append still cannot stamp it, so audit-gate behavior is unchanged. Verified 2026-09-18 against the dsh-v0.1.7-alpha.1 master checkout (full gate chain + profile install smoke). |
| Node | ^22.19.0 || >=24.0.0 |
| Platforms | Anywhere DSH runs (pure host, zero-dependency regex; no browser half) |
| Model | Text models fully supported; no extra model capability required |
What you get
dsh-mask anonymizes personal data at the model boundary — before a message reaches the model — and keeps a restore table host-side so placeholders stay reversible:
- Request-time masking —
agent/pre-step messages are rewritten so phones, emails, ID cards, bank cards, and keys (on by default) and IPs (opt-in) become <PHONE_1>-style placeholders. The masked text is what gets logged and sent to the model.
- Restore table — the
placeholder → original map lives only in memory and a controlled storage domain (dsh_mask); the plaintext never enters the session log.
- Audit, not plaintext — the
mask/applied session event records only "replaced N values + type distribution", never the original text or the mapping.
/mask command — status (counts + distribution), on/off (runtime toggle), restore <text> (unmap placeholders and deliver the plaintext as a file, never as logged text), help.mask_test tool — run a snippet through the detector and see the placeholder result; it never reveals the original values.
user message ──agent/pre-step──▶ placeholders ──model──▶ placeholders ──restore──▶ controlled file
▲ │
└──────── restore table (memory + dsh_mask) ───────────┘
Quick start
dsh plugin --profile web add github:PerryLink/dsh-mask
# 1. install the bundle into your profile
dsh plugin --profile web add github:PerryLink/dsh-mask
# or from npm (published releases)
dsh plugin --profile web add dsh-mask
# 2. verify the row mounts
dsh --profile web --dump-config | grep -A2 'id: mask'
Then tailor the entity list in your profile patch:
- insert:
- id: mask
name: dsh-mask
config:
entities: [phone, email, id-card, bank-card, key]
> /mask status
> /mask restore <PHONE_1>
mask: restored 1 placeholder(s): <PHONE_1>
file: /tmp/dsh-mask-restore/restore-1f0c….txt (mode 0600, one file per session)
Examples
These are the plugin's real entry points; install once, then call one.
Detailed example
Command
mask_test
Input
{"text":"电话 13812345678"}
Output
mask_test: replaced 1 PII value(s) (PHONE=1)
masked:
电话 <PHONE_1>
Observed from plugins/dsh-mask/index.mjs:243-250; test/index.test.mjs:306-315
Failure branch
Input
/mask restore <PHONE_1> with restoreDir=/tmp/dsh-mask-probe-not-a-dir (a regular file)
Output
mask: restore failed: EEXIST: file already exists, mkdir '/tmp/dsh-mask-probe-not-a-dir'
Observed from plugins/dsh-mask/index.mjs:501-506; plugins/dsh-mask/lib/deliver.mjs:89-97; plugins/dsh-mask/test/index.test.mjs:278-290
Brief examples
mask_test/mask restore <text>
Configuration
All tunables are Schemastery Config fields (changeable from cordis.yml). An id-targeted override replaces the whole row — restate every key you need. cordis.patch.yml documents each key inline.
| Key | Default | Meaning |
|---|
enabled | true | Master switch; false unregisters the listener, the /mask command, and the mask_test tool |
mode | regex | Detection mode; only regex is implemented (regex+ner for name/address recognition is reserved and fails loud) |
entities | [phone, email, id-card, bank-card, key] | Which PII types to mask; ip is also regex-capable (opt-in), person/address require NER |
scope | [messages] | Masking surface(s); messages masks agent/pre-step messages, tools masks tool-result text on tools/post-execute. Accepts a string or an array, e.g. [messages, tools] |
registerCommand | true | Register the /mask command |
registerTools | true | Register the mask_test tool when the tools service is present |
persistRestoreTable | true | Persist the restore table to the controlled dsh_mask storage domain (false = memory only) |
maxRestoreEntriesPerSession | 500 | Per-session restore entry cap (oldest evicted first) |
maxSessions | 1000 | In-memory session cap (least-recently-used evicted, mapping reloaded on demand) |
maskClientEnabled | false | Feature flag for the browser half "reveal" bubble (defensive; off by default until the live slot catalog verifies the target slot). The key is schema-declared and validated, but no runtime code reads it yet, so it changes nothing until the browser half ships |
restoreDir | '' | Where /mask restore delivers the restored plaintext. Empty = os.tmpdir()/dsh-mask-restore; a relative value resolves against the host process working directory. The directory is created and the file written , one file per session |
Example override in your profile patch:
- insert:
- id: mask
name: dsh-mask
config:
entities: [phone, email, id-card, bank-card, key, ip]
persistRestoreTable: false
registerCommand: true
Tools & surfaces
| Surface | Reveals plaintext | Notes |
|---|
agent/pre-step masking | never | Rewrites messages to placeholders before they are logged or sent to the model |
tools/post-execute masking | never | Rewrites tool-result text blocks to placeholders before they are logged or fed back to the model (scope: tools) |
/mask status | never | Enabled state, total replaced, type distribution |
/mask on / /mask off | never | Runtime toggle (resets to config.enabled on restart) |
/mask restore <text> | yes (explicit) | Unmaps placeholders back to the values stored for this session and writes them to a controlled file (one per session, mode 0600); the result text and the session log carry only that path and the placeholder names |
mask_test | never | Masks a snippet and reports the placeholder result + counts |
Permissions & data
- Permissions:
dsh-mask performs no network requests and stores no credentials; it reads the session at the agent/pre-step boundary, writes its own dsh_mask storage domain, and writes the /mask restore delivery file (the manifest declares filesystem:write for exactly that). The dshWorkshop manifest also declares network:none and credentials:none.
- Data: the
placeholder → original restore table lives in memory and, when persistRestoreTable: true, in the controlled dsh_mask storage domain — it is never written to the session log. The only other plaintext copy is the delivery file an explicit /mask restore writes under restoreDir (mode 0600, one file per session, overwritten by the next restore); nothing writes plaintext to disk implicitly.
- Session log:
mask/applied is declared in types.d.ts and appended only when the host records the type (see Known limitations). Its payload is counts + type distribution only.
Security boundaries
- Plaintext never enters the session log. The masked (placeholder) form is what gets logged and sent to the model, so model-visible content is reconstructable from the log in placeholder form; the originals stay in the restore table. The same rule governs the reveal direction: a command's result text is logged verbatim by the host and rendered from that same event, so
/mask restore writes the plaintext to a controlled file and returns only its path.
- Sanitize before display/log.
lib/sanitize.mjs redacts PII, secrets, and URL credentials before any text reaches the model or the log; mask_test, /mask status, and every /mask error path never echo originals.
- Controlled restore.
/mask restore is the single explicit reveal surface, and it only reads the mapping for the active session. Failure fails closed: if the delivery file cannot be written, the command reports an error and never falls back to echoing the plaintext.
- Fail closed. Unimplemented
mode (regex+ner), unknown scope values, NER-only entities, and out-of-bounds numbers all fail loudly at load.
- Registrations are effects. The listener, command, tool, and storage-domain close are all Cordis effects — stop/hot-reload removes them.
Known limitations
- Regex only. Name (
person) and address (address) recognition needs an external NER recognizer, which the pure-host zero-dependency form does not bundle; mode: regex+ner and those entities fail loudly at load. The PII types covered out of the box are phone, email, ID card, bank card, key, and (opt-in) IP.
- Region-specific patterns. The
phone and id-card detectors match mainland-China formats only: phone is 1[3-9] followed by nine digits, and id-card is an 18-character Chinese resident ID (17 digits plus a digit or X). Phone numbers and national identifiers from other countries are not detected. email, ip, and key are region-agnostic; bank-card accepts any 16-19 digit run at a lower confidence score.
- Display-layer restore needs a client half. Masking is fully host-side, but transparently un-masking the assistant bubbles in the client UI is a browser-half feature this pure-host form does not ship. The host side keeps the restore table and the exported
RestoreStore seam (its methods take a session id), so a future client half would reach them through a host remote rather than directly; today the unmasking surface is the /mask restore <text> command, which delivers the plaintext as a file because a command result is logged verbatim (see Security boundaries), and the maskClientEnabled key is validated but read by no runtime code yet.
- No no-argument
/mask restore. Restoring every placeholder of the most recent assistant message is not offered, and the command says so when invoked without text. Two host facts block it: the restore table is a flat per-session {placeholder: original} map with no per-message index, so that message has to be located in the session event log; and the host deprecated Session.snapshotEvents() / eventAt() / ownEvents() with an explicit "new calls are prohibited", leaving an incremental projection as the only sanctioned route — which this plugin's declared host corridor (>=0.1.2-rc.1) cannot rely on. Placeholders keep working one at a time instead: /mask restore <PHONE_1>.
- Session events on
0.1.7-rc.2. The harness records 59 session event types and none of them is mask/*; its Session.append also cannot stamp the ignorable envelope, so the session-log audit appends are skipped and sessions keep loading. The gate is not silent about it: the first refusal per session logs one visible warning naming the skipped type and this documented limitation. The plugin enables the append automatically once a host records the type or supports the envelope.
Development
pnpm install # node ^22.19 || >=24
pnpm run typecheck && pnpm run typecheck:ci # two rulers: checkout face (guarded; unverifiable without the checkout) + published-line face
pnpm test # node --test
pnpm run verify:self-contained # dependency specs resolve from the registry
pnpm run verify:artifacts # shipped files present + index.mjs importable
pnpm run check:readmes # five-language README consistency
pnpm pack # the published tarball
There is no build step: pure ESM, index.mjs and lib/ are the shipped artifacts.
Benchmark
The PII benchmark (per-type P/R/F1 over 108 synthetic samples) is published in benchmark/RESULTS.md; regenerate it with node benchmark/run.mjs (no build step, zero new dependencies).
Interoperability with other DSH plugins
Verified against DSH 0.2.0-rc.2 (the runtime this README ships for) and the high-star plugin set surveyed on 2026-10-05.
This plugin does not interfere with other plugins, including the widely installed high-star ones:
- No tool-name collision. Every tool is namespaced; no bare name owned by a shipped tool or another plugin is registered.
- No service-key collision. It provides no service key at all, so it cannot collide on one.
- No slot collision. It registers no client slot key, so it cannot contend for a
shadows-shipped-ui seat.
- No HTTP route collision. It registers no
webServer prefix.
- No patch-layer collision. The bundle patch only
inserts its own row; it never overrides a built-in row's config.
- No global mutation. It does not patch prototypes, rewrite
process.env, or replace the global fetch dispatcher.
Shared event listeners are non-interfering by construction. It observes the ordering-sensitive events agent/pre-step, tools/post-execute with ctx.on() — Cordis's broadcast registration, where every listener runs and none can starve another. Every listener here delegates through next(), so the chain is never short-circuited, and a mutation is applied to the value next() produced rather than returned in its place:
agent/pre-step — also used by dsh-routing-suite (7000★, 7 listeners), modlens (4122★), dsh-purge (3317★), dsh-agent-teams (1923★), dsh-context (1849★).tools/post-execute — also used by cc-safety-net (1576★).
Static evidence: dsh-plugin-doctor K10–K13 report pass for every check on this repository.
Topics
dsh, dsh-plugin, deepseek-harness, deepseek, cordis, pii, mask, privacy, anonymization, security
Contributors
- @PerryLink — creator and maintainer: the regex PII detector ported from Pii-Stripper-Middleware, the
agent/pre-step masking seam, the restore table, the /mask command and mask_test tool, and the five-language docs.
PerryLink DSH Plugin Family
This project is part of the DeepSeek Harness plugin family from PerryLink. The roster stays collapsed below so every README keeps a short first screen; expand it to see every row and its status.
Expand all plugins
License
LICENSE (Apache License 2.0) © 2026 dsh-mask contributors
