插件目录 / Knowledge / dsh-telemetry-redactor
dsh-telemetry-redactor
已验证 · 实测可装 030611
功能简介
dsh-telemetry-redactor — DSH 插件(会话)
可用 — 实测通过,早期项目
dsh-telemetry-redactor — DSH 插件(会话) 实测能干净安装、正常启动。早期项目,但功能可用。
「已验证」表示我们的自动化 CI 在干净 profile 里实际执行了 dsh plugin add 并启动成功——仅此而已。功能描述与版本兼容性均为作者声明。这不是安全审计,也不代表对第三方代码的背书。
README
DSH Telemetry Redactor
English | 中文

Redact supported credential patterns from outbound telemetry copies before configured backends receive them—without rewriting canonical session logs.
dsh plugin --profile web add dsh-telemetry-redactor
Community-maintained and not an official DeepSeek project. Related trust-layer plugins: Verification Receipt, Evidence Audit, and Context Provenance.
dsh-telemetry-redactor is a minimal DeepSeek Harness Profile Bundle that redacts sensitive values from session telemetry before a backend receives them. It mounts on the official session-telemetry/record waterfall, calls next() so other deployment rules still compose, and returns a new recursively redacted record.
The official telemetry coordinator deep-copies canonical session events before this waterfall and contains thrown rules per record. Therefore this plugin changes only the outbound copy: it never rewrites the canonical session log. In this document, fail-closed means only that this coordinator withholds one failing export copy and continues the agent loop. It is not a claim that every telemetry path or listener order is impossible to bypass.
What it redacts
- Values under high-risk key names such as
authorization,cookie,credential,password,secret,token,apiKey,access_token,clientSecret, andprivateKey; recognized credential patterns in key names are replaced too. - Bearer and Basic authorization values embedded in strings.
- Common credential forms including
sk-..., GitHub tokens, Slack tokens, JWT-like triples, andtoken=.../api_key: ...assignments.
Keys are tokenized before matching, so telemetry counters such as inputTokens, output_tokens, tokenUsage, tokenCount, and contextTokenCount, and ordinary fields such as tokenizer, remain intact. Redaction is a safety filter, not a proof that an arbitrary unknown secret format is absent. The exact supported and unsupported cases are in SECRET-MATRIX.md; see SECURITY.md for the trust boundary.
The listener is prepended so it normally wraps deployment rules mounted before or after it and redacts their final output. A deliberately prepended outer listener can still add content after this plugin; review the complete waterfall listener set in security-sensitive deployments.
Install
Install the public package into the selected DSH profile, then inspect the resolved configuration:
dsh plugin --profile web add dsh-telemetry-redactor
dsh --profile web --dump-config
The dump must contain the inserted telemetry-redactor row. The bundle does not add, replace, or enable a telemetry backend; it only protects records handled by whatever backend the deployment already selected.
Configuration
The only option is replacement, which defaults to [REDACTED]. It must contain 1 to 128 characters and must not itself match a supported credential pattern; invalid values fail loudly when the Cordis plugin fiber is awaited.
- id: telemetry-redactor
config:
replacement: '[TELEMETRY-REDACTED]'
The key and pattern rules are fixed security behavior and cannot be disabled through configuration.
Verification
pnpm run typecheck
pnpm test
pnpm run build
pnpm run test:smoke
pnpm run test:performance
pnpm run test:official-head
pnpm run test:official-patch
pnpm run test:packed:clean-env
The tests include a real Cordis Loader composition with the SessionTelemetryCoordinator: the backend receives the replacement while the authoritative session event retains the fixture secret, and an over-deep record is withheld without escaping the capture handler. Release smokes use the packaged frozen fixture for official commit 47f943859bef60e4160492346772ded9b24f765a: it pins the reviewed source hashes, exact published runtime versions and runtime-file hashes, then verifies the required Cordis/session/LLM/coordinator API surface before composing it with the built or tarball-installed plugin. test:official-patch similarly pins the reviewed Include source hash before applying cordis.patch.yml. No environment variable or installed official checkout is required.
Model Experience
None. The plugin observes an outbound telemetry copy after session logging and contributes no prompt text, tool schema, message, or provider request.
KV Cache effect
None; no model request changes.
Known limitations
- Unknown secret formats that have neither a sensitive key nor a recognized string pattern may pass through. Add deterministic coverage before extending the fixed patterns.
- Accessor properties and non-plain objects are rejected rather than read or silently converted. Enumerable JSON data in arrays, ordinary objects, and null-prototype objects is supported; non-enumerable fields are outside the telemetry contract.
- A key-name match redacts the complete value. This favors safety over retaining structure beneath fields named as credentials.
- Redaction occurs synchronously on the telemetry capture path. The implementation is recursive and bounded to 64 nested containers; very large but shallow records still cost linear CPU time.
- A Proxy can run or throw from reflection traps before the plugin can inspect its contents. The official coordinator's own
structuredClonenormally rejects a Proxy before this waterfall; direct third-party dispatch is not a supported hostile-object sandbox. dsh.plugin.jsonis supplemental community metadata. DSH installation is controlled bypackage.json'sdsh.bundlefield andcordis.patch.yml.
License
MIT
安装
装一次目录插件,之后本站所有插件都能让 DeepSeek Harness 自动找、自动装:
dsh plugin add dshbase-catalog 然后对 agent 说「帮我装 dsh-telemetry-redactor」,它会在目录里找到并自动安装。文档:dshbase-catalog · 已验证场景包。
Web profile:
dsh plugin --profile web add dsh-telemetry-redactor Headless(CLI)profile:
dsh plugin --profile headless add dsh-telemetry-redactor 包信息
npm:dsh-telemetry-redactor · 版本 — · 实测环境 dsh 0.1.0-rc.6
实测报告
端到端验证通过:dsh 0.1.0-rc.6 上 L1 安装 + L2 加载 + L3 运行问答。
使用场景
给 agent 一套记忆、知识库或检索层,让它不再跨会话丢上下文。
适合谁
跑长项目、想让 agent 记住决策、文档和偏好而不用每次重讲的人。
二次开发建议
记忆/检索后端是缝——插新存储、调蒸馏策略,或加引用与审计轨迹。