dsh-pii-anonymizer

dsh-pii-anonymizer

A native DeepSeek Harness plugin that replaces detected PII with stable placeholders in what the model is sent: your prompt, the system prompt, runtime context, workspace instructions, recalled sessions, relayed agent messages, and tool results.

Detection uses fixed rules and checksums (never a model). Every detected value is replaced with a keyed placeholder such as [[EMAIL_3f9a2c1d]] before the request goes to the provider. The real values do stay on this computer: the host app commits your message to the local session log exactly as you typed it, and files and tool outputs on disk are not changed.

you type:    email a@x.com or call +31 6 1234 5678
model sees:  email [[EMAIL_6c1f0a9e]] or call [[PHONE_d24b7c03]]
             + a notice: "🔒 PII shield: 2 item(s) in this step's input were
               replaced with placeholders before being sent to the model
               provider (1 EMAIL, 1 PHONE)."
local log:   the message as typed — raw values remain in the local session
             log and transcript.

The notice is a user-role message: it is shown in your transcript and the model receives it (that is how the harness delivers plugin notices). It carries only types and counts — never values or placeholders. Turn it off with notify: false.

Anonymization is one-way: the plugin keeps no value → placeholder table and has no restore path, so nothing it holds can put the real values back.

Where it hooks

Each listener is registered with prepend: true, so it is the outermost listener of its waterfall and rewrites the final value after every other plugin has contributed.

Hook (package)	What is rewritten
system-prompt/assemble (dsh-system-prompt)	prompt section text (e.g. a terminal app's workspace section), runtime-context text, and string prompt variables. The agent loop renders the assembled contexts into its form: "snapshot" user message, so snapshots (e.g. a memory index) are covered here, and the loop's "has the snapshot changed?" check stays stable.
agent/pre-step (dsh-agent-loop)	the text blocks of every message entering the step, whatever its source: the user prompt, runtime-context snapshots, workspace instructions (AGENTS.md), recalled sessions (recall), relayed agent messages (relay), skill catalogs, and contexts attached by tool results.
tools/post-execute (dsh-tools)	the tool result content the model reads (file reads, grep, shell output, memory reads…), a block decision's feedback, and attached contexts.

Why a native plugin, not a hook

The Claude Code / Codex hook bridges cannot do this. HookOutput in @deepseek-ai/dsh-hook-protocol exposes only additionalContext; its updatedInput is documented as "PARSED but NOT honored." There is no prompt-rewrite channel. Only a native Cordis plugin can replace the messages, prompt, and tool results that enter the model request.

Detection scope

Deterministic means structural. The detectors find things with a grammar or a checksum:

Type	Trigger	Validation
API_KEY	known prefixes (sk-, sk-ant-, AKIA…, ghp_…, github_pat_…, xox…, AIza…)	—
IBAN	CC00…, compact or space-grouped	mod-97 checksum
CREDIT_CARD	13–19 digits, spaces/dashes allowed	Luhn checksum
SSN	DDD-DD-DDDD	none
EMAIL	RFC-ish address shape	none
IPV6 / IPV4	address shapes	octet range for IPv4
PHONE	digit run with + ( ) . - space, optional leading (, 7–15 digits	rejects ISO dates and version strings

Free-form names, postal addresses, and other context-dependent PII are not detected — a deterministic rule cannot catch them without unacceptable false positives. Adding an ML/NER stage would break the "deterministic" property.

Overlaps are resolved by priority (checksum types beat the loose phone rule), then span length, then position.

Pseudonymization model

A placeholder is [[TYPE_<hash>]], where <hash> is the first 8 hex digits of HMAC-SHA256(key, type + normalized value). Normalization folds trivial spelling differences (email case, IBAN/card/phone spacing and separators) into one value.

• Stable everywhere. The same value gets the same placeholder in every session, in subagents, and after a restart, so a placeholder saved into a memory or a note keeps meaning the same thing. Different values get different placeholders (barring a 32-bit truncation collision).
• Keyed. The key is a per-install secret: 32 random bytes in $DSH_HOME/pii-anonymizer.key (default ~/.dsh/pii-anonymizer.key), created on first use with mode 0600 (an existing file with looser permissions is tightened). Without the key, a placeholder cannot be matched against guessed values, which matters for low-entropy values such as phone numbers. Keep the file private; deleting it changes every placeholder from then on. If the file cannot be read or created, the plugin logs a warning and uses a process-only key: values are still replaced, but placeholders are not stable across restarts.
• No raw PII on disk. The key file holds only the random key. The plugin keeps no value table and writes no values anywhere.
• Idempotent. Existing placeholders are never re-detected, so text that has already been rewritten passes through unchanged.

What is and is not covered

Covered (model-facing text):

• your prompt and every other message entering a step — snapshots, workspace instructions, recall, relay, catalogs, tool-result contexts;
• system-prompt sections, runtime-context sections, and string prompt variables;
• tool result content, including errors returned by a tool body.

Not covered:

• Local storage. The session log, transcript, memory files, and any file on disk keep the raw values; this shield changes only what is sent to the provider. The raw values are protected by the file permissions of $DSH_HOME and nothing else.
• Undetectable PII. Free-form names, postal addresses, and other context-dependent PII are not detected (see Detection scope).
• A complete: true system-prompt section. dsh-system-prompt restores an effective complete section from its pre-waterfall text after the system-prompt/assemble waterfall, so a plugin cannot rewrite it.
• Tool schemas (tool names, descriptions, parameter schemas) are sent as is.
• Tool results that skip tools/post-execute. The registry finalizes a few outcomes without post-execute — cancellations and pipeline failures (for example a throwing tools/pre-execute or tools/execute listener) — and a tool's own finalizeContent callback runs after post-execute. Their text reaches the model unrewritten. Denials, approval rejections, tool errors and successes all pass through post-execute and are covered. A downstream post-execute listener that replaces the result's value is left alone (the registry forbids replacing both value and content).
• Model output. The assistant's replies are not rewritten; placeholders are not restored in them.
• Requests made outside the agent loop by other plugins (for example a session-title generator) only see what the loop already rewrote — the title request built from your messages receives the placeholders, but a plugin that reads raw files itself and calls a model directly is outside this shield.

Rewriting injected context can occasionally hide a value the agent genuinely needs (for example an IP address in a workspace instruction). Disable the system-prompt or tool-result rewrite with systemPrompt: false or toolResults: false if that matters more to you than the coverage.

Configuration

- id: pii-anonymizer
  name: dsh-pii-anonymizer
  config:
    enabled: true        # master switch
    notify: true         # add the notice message when a step's input had PII
    debug: false         # log load/shield events (counts only, never values)
    systemPrompt: true   # rewrite system prompt + runtime context
    toolResults: true    # rewrite tool result content
    keyFile: ''          # empty = $DSH_HOME/pii-anonymizer.key

Installation (as a bundle)

This package is a DSH bundle: package.json declares dsh.bundle.patch, and cordis.patch.yml inserts the pii-anonymizer row. A profile lists it in both dependencies and dsh.profile.bundles — which is what the moqi-tui /plugins picker reads (it reads the manifest, not the Cordis loader).

Install it from the registry into a profile:

npm install dsh-pii-anonymizer          # inside ~/.dsh/profiles/<name>
# or, from the terminal app:
/plugins add dsh-pii-anonymizer

tools/register-profile.mjs <profileDir> makes the manifest edit for a local checkout, and writes a package.json.bak beside it:

node tools/register-profile.mjs ~/.dsh/profiles/tui

The package must also be resolvable from the profile, for example by symlinking the checkout into the shared resolution root:

ln -s "$PWD" ~/.dsh/profiles/node_modules/dsh-pii-anonymizer

Restart the app to load it. Rollback per profile: restore its package.json.bak.

Development

npm test             # node --test: detectors, keyed pseudonymizer, key file, and
                     # the pre-step / assemble / post-execute rewrites

The unit tests import only src/detectors.mjs, src/anonymizer.mjs and src/hooks.mjs, which have no harness dependencies, so CI needs no install. src/index.mjs imports the harness peers; for local module-load smoke tests, symlink node_modules/@deepseek-ai to the @deepseek-ai directory of your harness install. An installed plugin resolves them from the profile instead.

End-to-end test (no credentials, no network)

tools/mock-provider.mjs is a tiny OpenAI-compatible SSE server; tools/e2e.patch.yml routes the deepseek adapter at it and loads this plugin. A real harness loop then runs against the mock, and tools/requests.jsonl records exactly what the provider received. With MOCK_READ_FILE set, the mock first asks the agent to read that file, so the tool-result path is exercised too.

# 1. throwaway harness home (never ~/.dsh), with this checkout resolvable
export DSH_HOME=$(mktemp -d)
dsh --profile pitest --from-default-profile headless --dump-config >/dev/null
ln -s "$PWD" "$DSH_HOME/profiles/node_modules/dsh-pii-anonymizer"

# 2. a scratch workspace with PII in an instruction file and a data file
WORK=$(mktemp -d) && git -C "$WORK" init -q
echo 'Owner: owner@corp.example'                 > "$WORK/AGENTS.md"
echo 'customer: c@corp.example +31 6 9876 5432'  > "$WORK/notes.txt"

# 3. mock provider
MOCK_READ_FILE="$WORK/notes.txt" node tools/mock-provider.mjs &

# 4. one real turn
(cd "$WORK" && DSH_E2E_KEY=dummy dsh --profile pitest \
  --patch "$OLDPWD/tools/e2e.patch.yml" "My email is jane.doe@corp.example; read notes.txt")

# 5. confirm the provider only ever saw placeholders
grep -c 'corp.example' tools/requests.jsonl      # expect 0
grep -o '\[\[[A-Z_]*_[0-9a-f]*\]\]' tools/requests.jsonl | sort | uniq -c

Verified: across the agent requests and the session-title request, the provider received placeholders for the prompt email, the AGENTS.md email (workspace instructions), and the email and phone in the read tool result; no corp.example value appeared. A second run with the same DSH_HOME produced the identical placeholders, and the runtime-context snapshot was sent once, not re-sent every step. The raw values do remain in the local session log, as documented above.

Status

• Deterministic detectors with checksum validation
• Keyed, cross-session stable placeholders (one-way, no reverse map)
• User prompt and all injected messages (snapshot, instructions, recall, relay, catalogs, tool contexts) via agent/pre-step
• System prompt, runtime context and variables via system-prompt/assemble
• Tool result content via tools/post-execute
• Run through a real headless harness against a mock provider