dsh-shadow-mind

Parallel cognitive runtime for DeepSeek Harness (DSH / Cordis).

This project is a DSH plugin implementation of pi-shadow-mind. The plugin runs multiple "Shadow Mind" agents beside the main agent to provide independent reviews, fact-checking, and parallel cognitive work: after each main-agent turn, a heartbeat scheduler randomly activates configured shadows, each with its own responsibility, tool allowlist, and run timeout.

Status: functional prototype. Core heartbeat scheduling, restricted-tool shadow agents, per-run timeouts, lifecycle cleanup, and management tools work end-to-end. See Known Limitations below.

Relation to pi-shadow-mind

• Original project: https://github.com/liuzhengdongfortest/pi-shadow-mind
• This repository is a DSH plugin (dsh-plugin) implementation of pi-shadow-mind: the shadow-mind concepts are reimplemented on the DSH/Cordis runtime as an independent package — not a fork, and not a line-by-line translation.
• DSH uses different primitives (continuable subagents, native background notices, and the Cordis plugin model), so the implementation adapts the original design to them.

Features

• Heartbeat scheduling: After each main-agent turn, randomly activates configured Shadow Minds.
• Restricted-tool shadows: Each shadow receives a sanitized main-session trajectory and an explicit tool allowlist. The default allowlist is read-only; configuring other tools can broaden that access.
• Per-run timeout: timeout_seconds (or the config default) bounds each shadow run; an expired run is interrupted and its slot is released.
• Lifecycle cleanup: subagent/end removes finished shadow runs, so slots and active counts stay accurate within an epoch.
• Management tools: Create, update, delete, list, enable, and disable shadow definitions via model tools. Persistent writes are gated behind the DSH approval service when one is mounted.
• Config tools: Read and write the global configuration via model tools (writes validate the merged result before persisting).
• Pause/resume/epoch: /shadow pause and /shadow resume; pausing also aborts running shadows. New user input increments the epoch and cancels running shadows from the previous epoch.
• Auto toggle: /shadow auto on|off actually enables/disables heartbeat activation.
• Tool-call argument redaction: Tool-call arguments are redacted before being forwarded to shadows (credentials are not leaked). Tool results are summarized.

Installation

DSH plugins are loaded through a Cordis composition (profile or agent preset).

1. Install the package into a DSH profile

dsh plugin --profile web add @winterchenhuan/dsh-shadow-mind@latest

For local development, build the package before adding the directory so dist/index.js exists:

cd /path/to/dsh-shadow-mind
npm install
npm run build
dsh plugin --profile web add /path/to/dsh-shadow-mind

The profile stores the local package as a path dependency; rerun npm run build after source changes, then restart DSH.

The current repository version is 0.1.7. The @latest suffix makes the add command re-resolve to the newest published version; a bare add <pkg> keeps the range already recorded in the profile (e.g. ^0.1.6) and the lockfile-pinned version, so it will not upgrade an existing installation. To test an unpublished checkout, use the local-directory procedure above.

The package ships a cordis.patch.yml. The profile loader reads the package's dsh.bundle.patch manifest and applies it automatically; no manual profile cordis.patch.yml editing is needed.

2. Restart DSH

Restart DSH and the plugin will be loaded. The package exports a default Cordis plugin factory from dist/index.js.

Configuration

Global configuration is a user-settings namespace: the single source of truth is the shadow-mind: section of ~/.dsh/settings.yaml. It can be edited from three places that stay in sync:

• Settings page: DSH Web 设置页的「可配置插件」里有一张 Shadow Mind 卡片，按字段编辑、保存即写入（支持覆盖标记与重置）。
• Model tools: write_shadow_config / read_shadow_config（写入前需要用户批准）。
• Manual edit: directly edit the shadow-mind: section of ~/.dsh/settings.yaml; changes are picked up live.

Shadow definitions (Markdown) still live in:

$DSH_HOME/agent/shadow-minds/
├── grounded-reviewer.md
├── requirement-keeper.md
└── ...

$DSH_HOME defaults to ~/.dsh (honoring the DSH_HOME environment variable), so the default location is ~/.dsh/agent/shadow-minds/.

Example settings.yaml section (fields absent from the section fall back to schema defaults):

shadow-mind:
  heartbeatProbability: 0.33
  maxParallelShadows: 2
  defaultShadowTimeoutSeconds: 120
  defaultThinkingLevel: low

Field reference:

Field	Default	Meaning
heartbeatProbability	0.333…	Per-turn activation probability per eligible shadow, 0..1
maxParallelShadows	2	Concurrently running shadows, 1..8
defaultShadowTimeoutSeconds	300	Per-shadow run timeout when the definition sets none
headlessDrainTimeoutSeconds	120	Grace window for settling children at shutdown
resultBatchWindowMs	400	Window coalescing multiple reports into one wakeup
defaultShadowModel	(unset → main model)	Model used when a shadow declares none
defaultThinkingLevel	low	minimal / low / medium / high
randomSeed	(unset → random)	Deterministic heartbeat RNG seed; applies to the next session

Legacy config.json migration

An old $DSH_HOME/agent/shadow-minds/config.json is imported into the settings section automatically on the first startup after upgrading, then renamed to config.json.migrated so it cannot resurrect stale values later. A file that fails validation is renamed aside too — the plugin keeps running on defaults and the values survive inside the .migrated copy.

The timeout and heartbeat fields are active. headless_drain_timeout_seconds, result_batch_window_ms, and default_thinking_level are accepted for configuration compatibility but are not active yet; see Known Limitations.

Example shadow definition grounded-reviewer.md:

---
id: grounded-reviewer
name: Project Grounding Checker
enabled: true
activation_probability: 0.6
run_with_model: openai/gpt-5-mini
thinking_level: low
tools:
  - read
  - grep
  - glob
---

Check whether the main agent's claims are supported by the current workspace. If nothing is worth reporting, reply exactly: NOT_RELEVANT.

Usage

After installation, restart the target profile and start DSH Web normally:

dsh web

(dsh --profile web is equivalent.)

In the Web UI, continue using the main agent. Shadow activations happen after main-agent turns according to the configured probability. Use /shadow status first to confirm the plugin loaded and to see registry/config diagnostics.

Typical flow:

1. /shadow status — confirm the plugin loaded and see the current state.
2. /shadow list — see which shadows are defined and enabled.
3. /shadow probe grounded-reviewer — manually run one shadow once to check it works.
4. Shadows then activate automatically after main-agent turns (heartbeat), as long as auto mode is ON.

Commands

Use the single /shadow umbrella command:

/shadow status
/shadow probe <id> [tools]
/shadow list
/shadow clean
/shadow auto <on|off>
/shadow pause
/shadow resume

Command	Purpose
/shadow status	Show the current state: active/paused, auto mode, epoch, running shadow count, event count, plus config/registry diagnostics.
/shadow probe <id> [tools]	Manually run one shadow once without waiting for a heartbeat. <id> selects the shadow (omit it to pick a random enabled one). Optional comma-separated tools overrides that shadow's tool allowlist.
/shadow list	List all shadow definitions: enabled/disabled, activation probability, and tool allowlist.
/shadow clean	Abort all currently running shadows and release their slots.
/shadow auto <on|off>	Enable/disable heartbeat activation. off stops random activations; already-running shadows are not affected.
/shadow pause / /shadow resume	Pause/resume the whole shadow system for the current session. Pausing also aborts running shadows. A new user message (new epoch) also stops shadows left over from the previous epoch.

Management Tools

These are registered as model-callable tools:

• list_shadows
• create_shadow
• update_shadow
• delete_shadow
• enable_shadow
• disable_shadow
• trigger_shadow
• read_shadow_config
• write_shadow_config

Known Limitations

The following areas are not fully implemented yet:

Area	Status
Independent Shadow AgentSession	Partial: DSH continuable subagents are used instead of a separate agent session.
report_to_main tool	Missing: DSH native background notices are used instead. Report batching and steer/followUp are not yet replicated.
Tool allowlist resolution	Simplified: missing-tool reporting and report_to_main injection are not implemented.
Per-shadow model auth check	Missing: run_with_model is passed as agentOptions, but auth validation is not performed.
Per-shadow timeout_seconds	Enforced: each run is bounded by timeout_seconds or default_shadow_timeout_seconds; expired runs are interrupted and their slots released.
Per-shadow thinking_level	Not applied: DSH uses reasoning effort, which is not yet mapped.
Shutdown drain / headless mode	Missing: no headless drain on process shutdown.
UI status panel / message renderer	Partial: a Client indicator exists in the dynamic prototype; the real package currently only exposes /shadow.
Debug session logs	Not needed: shadow runs are DSH continuable subagents, and DSH Web already shows their execution (trajectory, tool calls, results) live. The debug frontmatter field was removed accordingly.
Test suite	Minimal: vitest suite covers the pure scheduling, parsing, serialization, config, and drain logic; no harness integration tests yet.

See DESIGN.md for the project's design goals; some of them are not fully implemented yet.

Development

npm install
npm run typecheck
npm run verify   # typecheck + unit tests
npm run build

npm pack runs prepack (i.e. npm run build) automatically, so a release tarball always contains dist/index.js.

License

MIT