dsh-anchored-monitor

Real-time chain-of-thought anchoring monitor & intervention for DeepSeek Harness. Watch the we / let's / let me fingerprint of every reasoning block, stay in the spec band, and pull the model back automatically when the trajectory drifts.

English | 简体中文

---

Why this exists

DeepSeek V4 Pro conditions heavily on what the first request shows it. The community measured the consequence: the official Minimal preset (46-character persona + bash/str_replace_editor) anchors a collective "we" trajectory and scores 99/96 on Project2, while the full Standard preset anchors an actor-style "let me" trajectory and scores 91. Behavior is path-committed: once anchored, expanding the tool catalog perturbs at most one reasoning block — the mode never flips back on its own.

Anchored presets (dsh-anchored-standard) solve the bootstrap. This plugin solves what happens after anchoring: external factors (imperative hints, oversized injected context) can drag the trajectory out of the spec band, and nothing detects or repairs that drift — until now.

What it does

• Live fingerprinting — every reasoning block is scored for the spec trajectory markers (we, let's, we'll, we need, …) versus the react marker (let me), over a sliding window.
• Three-band model — exactly the quantization measured by dsh-router-standard: persona_ratio < 0.2 → spec, 0.2–0.5 → mixed (the unstable transition band), ≥ 0.5 → react.
• Tiered intervention — automatic, cooldown-gated, with hysteresis recovery:

Level	Trigger	Action
L1 hint	entering the mixed band	injects a suggestive hint (never imperative — commands flip we→let me)
L2 reset	entering the react band	next request gets the 46-char Minimal persona + bash/str_replace_editor only; monitor window/baseline reset
L3 restart	L2 retries exhausted	recommends restarting the session

• Liquid-glass Web UI — a sidebar entry opens a frosted-glass panel floating above the conversation (drag / resize / remembered). Collapsed, it becomes a rheostat-style bar showing thinking intensity as a slider position plus a rolling log ticker. DeepSeek white/gray/blue palette; dark mode uses neutral grays matching the shell. The bar is a fixed-width rounded rectangle (320px, viewport-adaptive) — no size jumping. zh/en language toggle in the panel header and settings.
• Zero-setup auto-start — the host plugin spawns the monitor process when DSH starts (15s watchdog keeps it alive). No manual steps.
• Live streaming ECG — the host subscribes to llm/stream and pushes reasoning-delta chunks (1s throttle), so the charts and the bar move while the model thinks — not just after each turn. Counts are additive, so window aggregates stay exact.
• Intervention master switch — a toggle in the panel header turns interventions on/off at runtime (persisted across restarts); off = monitor-only, no interruptions — handy when switching models.
• Full settings page — every parameter (window, lexicon, scoring weights, band boundaries, thresholds, cooldowns, hint templates, bootstrap pair, log rotation) is editable in the DSH settings page with explanations and a floating save bar; saving restarts the monitor to apply changes. Charts only draw threshold lines for enabled trigger rules.
• In-loop interventions with auto-continue — the host observes every session's reasoning, executes L1/L2/L3 automatically, and then continues the task instead of stopping: L1 injects a suggestive context hint (never imperative); L2 cancels the running turn and soft-restarts the conversation (the next request runs under the 46-char Minimal persona + bash/str_replace_editor); L3 applies the same soft restart plus restart advice.
• Experiment-first — every lexicon entry, weight, threshold, window, band boundary and cooldown lives in YAML (config/*.yaml, validated against config/schema.json). JSONL experiment logs, offline replay and grid-search calibration scripts included.
• Decoupled — the monitor is an independent Node process; it never touches the model's context or compute.

Requirements

• Node.js ≥ 22.19
• DeepSeek Harness 0.1.0-rc.5+ (for the web plugin & preset)

Install

# 1) install the web plugin into your web profile (dsh CLI = pnpm forwarder)
dsh plugin --profile web add @a9i5k4/dsh-anchored-monitor

# 2) start the monitor process
npx anchored-monitor --profile demo

# 3) restart DeepSeek Harness (host bundle) and refresh the web GUI

You should now see 锚定监控 / Anchored Monitor in the left sidebar footer. Click it to open the glass panel; click the floating bar to expand/collapse.

To change the monitor address: edit ~/.dsh/anchored-monitor.json or POST /api/anchored-monitor/config with { "monitorUrl": "http://127.0.0.1:9301" }.

Quick demo

git clone https://github.com/Aik358/dsh-anchored-monitor.git
cd dsh-anchored-monitor
npm install && npm run build

npm run demo:generate                # 300 synthetic reasoning blocks
npm run dev -- --profile demo        # monitor + dashboard on :9301
npm run demo:feed                    # live-feed the blocks (watch L1→L2 cascade)

Agent-side preset (optional)

Copy preset/ into ~/.dsh/.agent-presets/anchored-monitor to let the harness agent push its own reasoning blocks to the monitor and execute the L1/L2/L3 interventions inside the agent loop (pair it with an anchoring preset such as dsh-anchored-standard for the first-round anchor).

API

The monitor process exposes (default http://127.0.0.1:9301):

Method	Path	Description
GET	/api/overview	sessions + selected snapshot + tail events in one call
GET	/api/sessions	session summaries
GET	/api/sessions/:id	full snapshot (history / interventions / baseline)
GET	/api/events?sessionId=&limit=	tail of the experiment JSONL
POST	/api/push	push a reasoning block {sessionId, text, sequence?, timestamp?}
POST	/api/sessions/:id/ack	acknowledge an intervention
POST	/api/sessions/:id/reset	trigger a manual L2 reset
GET	/api/stream	SSE event stream

The web plugin proxies these through /api/anchored-monitor/* (loopback-only).

Scripts

Command	Description
npm run demo:generate	generate a synthetic session JSONL (labelled, for calibration)
npm run demo:feed	live-feed blocks into a running monitor
npm run replay -- --file x.jsonl	offline replay with a summary report
npm run calibrate -- --file x.jsonl	grid-search window/weights/thresholds
npm run preview:build	snapshot the dashboard into a standalone HTML

Configuration

See docs/experiment-params.md for the full parameter reference. Everything is YAML — no hardcoded tuning.

FAQ

What happens after an intervention — does the conversation stop? No. Every intervention auto-continues the task: L1 injects a hint and the next turn keeps going; L2 stops the running turn (soft restart) and immediately re-enters context — the next request continues the task under the Minimal persona + bootstrap pair; L3 does the same and adds restart advice.

Why did the chart feel slow / frozen? Before 0.2.0, data only arrived once per finished turn. The host now streams reasoning-delta chunks to the monitor every ~1s, so the curve moves while the model thinks.

How do I disable interventions? Use the panel-header switch (persisted), or set intervention.enabled: false in the settings — monitoring continues, interventions stop.

Does L2 truncate the model's context? No. L2 replaces only the next request's persona section with the 46-char Minimal sentence and narrows the visible tool catalog to the bootstrap pair. All conversation history is preserved; the monitor only resets its own fingerprint statistics (invisible to the model). The plan-mode and all other sections are kept, because dropping them causes re-exploration amnesia (measured by dsh-router-standard).

Why is the transition band treated as a warning? The mixed band (0.2–0.5) is the training-distribution gap: measured scores are lower than either stable band. Entering it triggers the L1 hint; entering the react band escalates.

Is it safe to run? The monitor is read-only with respect to the model: it consumes reasoning text and sends intervention signals. All HTTP routes are loopback-only. Reasoning text may be sensitive — the experiment log is local by default; rotate/disable it in experiment_log.

Plugin author notes (read before extending)

Streaming waterfall rule — llm/stream is a stream-passthrough waterfall: listeners earlier in the chain iterate the return value of the listeners after them. Therefore:

1. Producer side: an llm/stream listener MUST be a plain function that returns an async generator. Never declare it async — the generator gets wrapped in a Promise and upstream for await consumers crash with next(...) is not a function or its return value is not async iterable.
2. Consumer side: always for await (const chunk of await next()) — await first, then iterate; safe regardless of what downstream returns.
3. agent/pre-step / system-prompt/assemble are value-passing events; async + await next() is correct there.

A violation took down every model request with zero logged events — if all sessions suddenly fail after a bundle reorder or a new plugin, audit the llm/stream chain first.

Single-intervention-executor rule — L1/L2/L3 must have exactly one executor. The Web plugin (host half) owns interventions and is the default; the agent preset (preset/) ships with handleInterventions: false and only pushes reasoning. Enabling both would double-register agent/pre-step / system-prompt/assemble and fire L2 resets twice.

Credits

Built on the measured results of these community projects (shallow-cloned in ../references for audit):

• xiaobright/dsh-anchored-standard — two-phase anchor preset (bootstrap pair, promotion gate, post-promotion resident set)
• ruler770525/dsh-anchored-flash — fingerprint counting (we/let's/let me) and the E1/E1.5 hint-wording experiments
• yjh051108/dsh-router-standard — dual-attractor paper and the three-band quantization (bandOf)
• KDB-Wind/dsh-minimal-anchored — minimal-tool anchoring alternative
• deepseek-ai/deepseek-harness — the host framework this plugin extends

License

MIT © 2026 Aik358