Preface¶

In the DSH development workflow, whether a session stays on the intended reasoning trajectory is often not directly visible on the interface. dsh-anchored-monitor is a DSH plugin designed to observe the fingerprints of we, let's, we'll, we need, and let me in reasoning blocks in real-time, using a three-band display to show the current state.

Its default behavior is to first provide visualization, displaying the current thinking efficiency or capability strength; if needed, intervention can be enabled to prompt, reset, or restart when the trajectory drifts to the dispersed let me band, allowing the task to continue. Below, we first introduce what it does, followed by installation notes and typical usage.

What It Is¶

Aik358/dsh-anchored-monitor is maintained by Aik358 and licensed under MIT. It is positioned as a real-time chain-of-thought anchoring monitor and intervention plugin for DeepSeek Harness.

It reads the reasoning block files for each session:

~/.dsh/sessions/*/events.jsonl

Then scores the reasoning blocks based on we / let's / let me fingerprints. The plugin does not require the installation of dsh-anchored-standard nor checks for preset names; it relies solely on the fingerprint signals provided by the reasoning content itself.

Core Features¶

Real-time Fingerprints and Three-Band Model¶

The plugin compares two types of markers on a sliding window:

Spec-side markers: we, let's, we'll, we need
React-side marker: let me

The three-band model is as follows:

spec: < 0.2
mixed: 0.2–0.5
react: >= 0.5

Tiered Intervention¶

The plugin offers three levels of intervention:

L1 hint
L2 reset
L3 restart

This set of interventions includes cooldown gating and hysteresis recovery.

Two points to note:

L1 hint uses suggestive prompts, not imperative prompts. This is because imperative expressions might flip we to let me.
L2 reset uses a built-in 46-character Minimal persona and configures bash and str_replace_editor.

Visualization Interface¶

The plugin defaults to prioritizing visualization, acting as a real-time thinking efficiency or capability strength gauge.

The interface includes:

Sidebar entry
Frosted-glass panel floating above the conversation
Draggable, resizable, and remembers position
Collapsed rheostat-style bar
zh/en switch in the panel header and settings page
Two skins: default Serious rheostat bar, and Meme skin, also called 滑动变祖器 / Liang-o-meter

Runtime, Streaming, and Guards¶

Runtime aspects include:

Zero-setup auto-start
15s watchdog

When DSH starts, the host plugin launches the monitoring process; the watchdog keeps it alive.

Streaming signals come from:

llm/stream
reasoning-delta

Push throttling is set to 1 second.

The plugin also receives two types of alert signals:

text_leak
streaming_stall

These two guards only alert and do not trigger automatic intervention.

The intervention master switch can be toggled at runtime and persists across restarts. Once the settings page is opened, these parameters can be edited:

window
lexicon
scoring weights
band boundaries
thresholds
cooldowns
hint templates
bootstrap pair
log rotation

The monitoring process is an independent process supporting:

JSONL experiment logs
offline replay
parameter calibration

Installation and Activation¶

Below are only the verified repository, dependencies, and runtime requirements from the materials, excluding official installation commands.

Repository address:

https://github.com/Aik358/dsh-anchored-monitor

Runtime requirements:

node >=22.19

Peer dependency:

@deepseek-ai/cordis ^4.0.1

License:

MIT

The plugin runs with the current DSH process permissions. Before installation, it is recommended to review the source code and license. If the team uses npm or the DSH plugin directory, please refer to the actual package name and installation instructions provided on the repository page and directory page, rather than constructing installation commands based solely on the repository name.

Typical Usage¶

First, confirm the model being used, then decide whether to enable intervention.

Open the plugin panel: Open the frosted-glass panel from the sidebar entry, place it above the conversation, and drag or resize it.
View the default gauge: Observe the thinking intensity bar, three-band status, and real-time curve.
Choose intervention mode:
- DeepSeek V4 Pro 0813: Recommended to turn interventions on.
- Other models: Keep interventions OFF, using it only as a live gauge.
Runtime switching: Toggle interventions on or off in the panel header; this state is saved across restarts.
Adjust parameters: Open the settings page to adjust parameters; if too many changes are made, use the one-click Reset to defaults.
Switch skins: Switch between Serious rheostat bar or Meme skin in the settings page.
View data sources: The monitor reads the reasoning blocks for each session:

~/.dsh/sessions/*/events.jsonl

And scores them based on fingerprints.

Applicable Scenarios and Notes¶

Suitable for developers who need to observe DSH reasoning trajectories in real-time; if only used for monitoring, turn off interventions.
Intervention is recommended only for DeepSeek V4 Pro 0813; for other models, it is advised to keep it off, using it only as a live gauge.
If a session has never been anchored, it may remain in the react band indefinitely, causing L2 to trigger continuously; this is described in the materials as a fight loop, not normal monitoring.
text_leak and streaming_stall are alert signals and do not trigger automatic intervention.
L1 uses suggestive prompts, not imperative prompts, because imperative expressions might flip we to let me.
The plugin does not depend on dsh-anchored-standard nor checks for preset names.
The facial assets for the Meme skin come from Lichtspektrum/liang-intensity-calibrator, licensed under MIT.

Conclusion¶

The value of dsh-anchored-monitor lies in making “whether the model is currently maintaining an anchored trajectory” a visible, adjustable, and replayable monitoring task: by default, it shows the bands and strength, and tiered intervention is applied when needed.