@whutzefengxie-ops/dsh-shadow-mind
面向 DeepSeek Harness 的概率性后台 Shadow agent 审查
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:whutzefengxie-ops/dsh-shadow-mind#fcf53b2e5c8b9b302034dd62124bf04077543f7bShadow Mind 以 @whutzefengxie-ops/dsh-shadow-mind 发布，当前版本为 0.2.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
中文 | English
An independently versioned DeepSeek Harness bundle that starts fresh background Shadow agents after eligible tool-using root turns, validates their structured findings, relays accepted reports into the root session, and exposes configuration and evidence in the Web UI.
The core design idea for this project comes from pi-shadow-mind. This repository is an independent implementation for the DeepSeek Harness plugin system, not an official fork of that project. Its runtime, Sessions, subagents, permissions, persistence, and Web UI use DeepSeek Harness extension mechanisms.
The plugin requires DeepSeek Harness 0.1.2-alpha.3 or newer (peer range >=0.1.2-alpha.3 <0.2.0). Pin a reviewed commit when installing from GitHub:
dsh plugin --profile web add github:whutzefengxie-ops/dsh-shadow-mind#<commit-sha>
Install a local checkout during development:
dsh plugin --profile web add /path/to/dsh-shadow-mind
Restart the selected profile after installation. Verify the composed rows without starting the Web server:
dsh --profile web --dump-config
The output must contain shadow-mind-runtime and tool-shadow-mind. Update by running add with a new reviewed SHA. Remove the bundle with:
dsh plugin --profile web remove @whutzefengxie-ops/dsh-shadow-mind
GitHub installation consumes the committed lib/ directory. This package has no prepare script and does not require pnpm allowBuilds authorization.
Open Settings → Plugins → Shadow Mind. The page owns:
Definitions are stored in $DSH_HOME/shadow-minds/*.md. A minimal deterministic acceptance definition is:
---
id: acceptance-reviewer
name: Acceptance Reviewer
enabled: true
debug: false
activation_probability: 1
active_for_models:
- '*'
tools: []
capture: since-compaction
context: minimal
think_first: true
---
Review the completed task. If there is a concrete defect or missing requirement, return a concise report with verdict `challenge` or `gap` and only rendered sequence numbers in `refs`. Return `silent` when the review applies but adds nothing actionable, or `not_relevant` when it does not apply.
Set the definition's activation_probability to 1 for deterministic acceptance. If run_with_model is omitted, the child inherits the root route; set a complete provider/model route to use another model. The default Shadow tools are read, grep, and glob; definition tools extend that allowlist and may carry write authority if the inherited sandbox permits it. The disabled starter library under examples/shadow-minds/ demonstrates the anchored probe vocabulary and is never installed into $DSH_HOME automatically.
Every Shadow child can be bound to the DSH deployment's configured providers, models, and reasoning efforts. The Settings page offers linked provider/model/effort dropdowns populated from the live DSH LLM directory, including each model's adapter-advertised reasoning efforts. Leaving a field empty inherits the default: the root route for provider/model, and the selected model's default effort for reasoning. The stored wire format remains the provider/model route string, so model-facing management tools are unchanged.
Shadow scheduling requires a completed root turn containing at least one durable tool result. In a new session, ask the root agent to read a repository file and analyze it. One human command covers the case automatic scheduling misses: /shadow new forces an immediate review while this session has not admitted any Shadow run yet. After that, the per-turn conversation card owns the run controls: its Retry button re-runs a specific failed or aborted run, and the Pause/Resume button in the card header stops or restarts scheduling for the session (pausing also cancels admitted work). Active runs, total admitted runs, and the last outcome stay visible on the settings page and the per-turn conversation cards.
When Shadow scheduling starts, a running placeholder appears immediately below the reviewed root response and warns that sending a new message cancels the review. The same card updates in place to report, silent, not-relevant, aborted, or failed. Repeated reviews remain at their actual turn positions. Report bodies use DSH's Markdown renderer for GFM, tables, code blocks, and TeX with its unsafe-content filtering.
An accepted report becomes a durable root user message and triggers a follow-up, but that relay updates the existing anchored card instead of adding a tail card. silent, not_relevant, aborted, and failed remain visible without entering the root agent, so presentation state cannot create a Shadow loop. A silent or not_relevant output that carries an explanatory body is tolerated: its body text is never relayed, the run settles as the declared status instead of failing validation (only a report carries body text), the discarded body is reported through a runtime warning that names the Shadow and run id, and a non-report-body-discarded debug record captures its presence, length, and content hash when debug: true is set.
Set debug: true on a definition when diagnosing production behavior. $DSH_HOME/shadow-minds/logs/<shadow-id>.jsonl records admission, child start, cancellation request, terminal outcome, and report delivery with lifecycle stage, stable reason code, cancellation source, and provider stop reason. It excludes prompts, report bodies, tool arguments, credentials, absolute paths, and stacks; a discarded non-report body is represented only by its length and SHA-256 hash. For example, new user input is USER_MESSAGE_RECEIVED, a Shadow deadline is SHADOW_TIMEOUT, and an abort not attributed to the plugin is PROVIDER_ABORTED.
A failure report usually arrives as only a subagent error message or a child subagent session id. The zero-dependency tools/shadow-debug.mjs reconstructs the full run context from either fragment:
node tools/shadow-debug.mjs trace <childSessionId|runId|rootSessionId> # one run: timeline + inputs + child-session evidence
node tools/shadow-debug.mjs find <error text|reason code> # locate runs by error content
node tools/shadow-debug.mjs runs [--failed] [--shadow <id>] # list recent runs
node tools/shadow-debug.mjs health # definition debug flags and log health
It resolves $DSH_HOME/shadow-minds/logs/*.jsonl (run timeline and input metadata), $DSH_HOME/shadow-minds/<id>.md (definition and review prompt), and $DSH_HOME/sessions/*/<childSessionId>/session.jsonl.zstd (the full child event stream: prompt, LLM request headers, tool errors, turn termination). When run inside a DSH session it infers paths from environment variables. The Chinese debugging guide documents the workflow, reason-code/stage tables, and a symptom-to-evidence map; an equivalent agent skill ships at .agents/skills/shadow-debug/SKILL.md.
The default trajectory projection omits reasoning, raw tool-result text, and tool arguments. Prompt injection remains possible in projected user and assistant text, so tool allowlists, inherited sandbox policy, fixed child approval policy, and disclosure limits remain security controls.
Definitions apply to one Harness home rather than one profile or workspace. Child Sessions follow the Harness persistence policy. Concurrent Shadows do not share a transaction, and write-capable tools can race with the root agent or other Shadows.
This plugin is developed and released from its own repository and is installed into a Harness profile as a bundle. An earlier in-repo copy of the plugin (packages/shadow-mind in the Harness source tree) predates this standalone repository and still uses the older {status, content} structured-output contract without verdict/severity/refs; it is not part of this repository's build or install path and is not updated here. Fixes in this repository apply to the released bundle only.
Accepted reports and their anchored cards survive a service restart. Runtime counters, the latest-run diagnostic, non-report lifecycle cards, and pause state are current-process data. The current release also resets spentChars on restart, so a restart can reopen a configured soft or hard budget before the next real user message.
Follow the installation and runtime validation plan for a reproducible source-host deployment, real-model acceptance run, restart checks, and the budget-persistence remediation. See the target architecture, review conditioning, and review-quality directions for current runtime contracts. The settings field tiers document which Web form fields are required, common, advanced, or intentionally not exposed. The Chinese technical design retains the Pi reference analysis and independent release topology. See SECURITY.md before publishing logs or changing an installation pin.
The devDependencies link against the DeepSeek Harness source (../deepseek-harness, a sibling of this checkout) because the dsh-0.1.2 client surface is not published to npm yet. Prepare it at the pinned release commit first (see .github/workflows/ci.yml):
corepack enable
git clone https://github.com/whutzefengxie-ops/deepseek-harness.git ../deepseek-harness
git -C ../deepseek-harness checkout dd6322d604e00eec1ba5e0c8541159906a21094a
pnpm --dir ../deepseek-harness install --frozen-lockfile
pnpm --dir ../deepseek-harness run build:lib
Then install and check this repository (CI=1 skips pnpm's interactive confirmation when the modules directory must be rebuilt from scratch):
CI=1 pnpm install --frozen-lockfile
pnpm run check
lib/ is a reviewed release artifact and must be committed whenever source behavior changes. Never commit local Harness state, credentials, sessions, logs, or acceptance exports.