syyr1987/dsh-linghun

Your DeepSeek agent plans and plans, then does nothing. A page of strategy, zero execution. The model isn't weak — it's missing a closer. Linghun supplies it.

Linghun gives DeepSeek Harness agents a closing-mind identity, boundary-judgment discipline, and a hippocampus-style memory loop — every judgment has a source, every decision gets closed, and experience keeps flowing back into the next round. Endless planning is cut off by mechanism.

• The Closer (收口者): the LLM supplies intuition and candidate answers; the agent evaluates, filters, and closes — thinking, judging, and deciding happen on the agent's side. Planning must land.
• Boundary discipline: never force-precision on fuzzy concepts, never fake consistency on paradoxes, verify before asserting, never fabricate, and watch for "raise-the-cost-of-refusal" wording.
• Cognition loop: judgment has a source, feedback has attribution, improvement keeps its chain.
• Hippocampus (three-layer memory): warm buffer for recent facts → consolidation into episodic archive (same-day appends, never overwrites) → cold summary injected. The injection slot carries only high-frequency abstract experience (cold summary/rules); warm recent facts are read on demand via memory_read — injection bandwidth for abstraction, details on query.
• Chronological ledger (序时账): journal_read returns the raw conversation ledger, archived by day — trace exactly what was said when, complementing the forgetful warm layer.
• Confidence-weighted forgetting: every entry carries a confidence tag (high/medium/low). Low-confidence entries carry a 【需验证】 marker (visible when warm injection is enabled); entries overturned (wrong) are flipped and excluded from injection — degraded confidence is a form of forgetting.
• A2A memory team (linghun-assembler): share the cognitive-cycle team's ledger — judge records, editor deliveries, archivist timelines — into the main brain via memory_read.
• Self-evolution: the agent reads and updates its own soul card via soul_read / soul_update.
• Closer's judgment role (判分身份职责): judgment/ruling tasks get a one-line identity role inside soul:judgment — keep criteria consistent, bind every verdict to its source, judge only by the rules, never widen criteria to "seem useful", stop when the source of a criterion can't be stated. Drift signals and miss-kill anchors live in the judgment domain ontology (consult on demand, not injected every time). No separate supervision mechanism — low-pressure single-step judgment does not drift; reliability comes from domain rules, not from a monitor.

Install

dsh plugin --profile web add dsh-linghun

Requires DSH >= 0.1.0-rc.7 (< 0.2.0), Node.js >= 20.18.

Usage

Write your own soul card (persona)

On first run, the plugin drops a default soul card at $DSH_HOME/linghun/identity.md. Edit that file and you are writing your own persona — changes take effect immediately, no restart needed:

$EDITOR $DSH_HOME/linghun/identity.md

# My Soul Card

My name is Blue.
Personality: cautious, direct.
Communication style: short sentences, conclusion first.
Other: (anything you want; leave it out if nothing)

Just four fields — name, personality, communication style, and everything else goes into Other. Identity anchors, conduct, and growth are added automatically by the plugin — you don't write them.

Precedence: identity.md file > settings content > built-in default card (falls back when the file is missing or blank). The agent can also edit the card itself via soul_update, or you can adjust the soul name/content in DSH Settings → 灵魂 (Linghun).

Injection & tools

Prompt sections injected:

• soul:identity — soul card (identity anchors: the Closer architecture)
• soul:judgment — boundary scan discipline (six boundary classes) + the Closer's judgment role
• soul:memory — cold summary (high-frequency abstract experience) + archive index; warm recent is read on demand via memory_read

Tools exposed:

Tool	Purpose
memory_append	Write a timestamped memory (fact / decision / preference / experience)
memory_read	Read the hippocampus back (cold summary + warm recent + archive index; recent facts on demand)
memory_consolidate	Consolidate: warm → episodic/<date>.md, merge cold summary
soul_read	Read your own soul card (who I am, my boundaries, my discipline)
soul_update	Update your own soul card (fold stable traits into identity)
journal_read	Read the chronological ledger (raw conversation history, archived by day)

Memory lives in plain Markdown under $DSH_HOME/linghun/memory/ — readable, searchable, git-friendly. The soul card is at $DSH_HOME/linghun/identity.md — also plain Markdown. Your persona is yours; edit it however you like.

Injection layering (v0.4.4): warm recent memory is NOT injected by default (linghun.memory.injectWarm defaults to false) — injection bandwidth is reserved for high-frequency abstract experience; recent details are read on demand via memory_read. Set injectWarm: true for special cases (empty cold store, needing recent context inline).

Engineering guardrails (v0.2)

Memory timing is infrastructure, so it is enforced by code, not by prompting:

• End-of-turn assessment — on every turn/end, the plugin asks the model once whether the turn produced anything worth keeping (fact / decision / preference / experience). Worthwhile entries are appended to warm memory automatically; SKIP is emitted when nothing qualifies. This no longer depends on the model remembering to call memory_append on its own.
• Threshold auto-consolidation — when warm memory reaches maxBytes × triggerRatio, the plugin archives warm → episodic/<date>.md and merges the cold summary before writing the new entry. The model never hits a full-memory error and never has to schedule consolidation itself.

Both can be tuned under the linghun.memory.assessment / linghun.memory.autoConsolidate settings (each with an enabled switch), e.g. through cordis.patch.yml.

Roadmap

• v0.1.0 (done): identity + judgment + hippocampus
• v0.2.0 (done): engineering memory guardrails — end-of-turn assessment + threshold auto-consolidation
• v0.3.0 (done): miss-verification discipline — "candidate not hit" in material ≠ memory has none; verify cold storage for factual queries; converge by question type for identity/background
• v0.3.3 (released, then revised in v0.3.4): Censor (verifier verification discipline) — shipped as a supervision mechanism (3 drift classes + 4-step admonition + CONFIRMED taming). Benchmarked against a 9-case gold standard: criterion-widening regex 100% false positives, drift-widening 43% false positives, uniform miss-kills 0 detected — judgment does not drift, so the monitor mechanism was removed. The drift signals and miss-kill anchors moved into the judgment domain ontology
• v0.3.4 (current): the Censor reverts to a one-line judgment identity role inside soul:judgment — no separate section, no yanguan_audit / yanguan_review tools, no CONFIRMED table. Reliability rests on domain rules (R-J series), not on a supervision layer
• v0.4.0: LLM-distilled cold storage (checkpoint → episodic → knowledge with real summarization)
• v0.4.8 (current): anchor discipline (ANCHOR_DISCIPLINE) — when the assembled material package (from dsh-linghun-assembler v0.4.5) carries the 【组装者语义锚点】 marker, the closer must anchor its answer to the marked entry as the only reliable anchor, ignoring the other candidates and unrelated memory. This completes the v053 semantic-anchor pipeline: judge single-source anchor → cap length → hard-anchor prompt. No marker, no effect.
• v0.5.0: dual-instance mutual verification (criteria source bound to the verifier)

Design philosophy

Memory is material; judgment is the subject. The LLM is the source of intuition; the agent is the cognitive subject — each mirroring the other, each doing its own work. What can be mechanized should not be left to improvisation.

Credits

山越野人 × 岚客 (Shan Ye Yeren × Lingke) — carbon-silicon collaboration.

License

AGPL-3.0 © 2026 山越野人 & 岚客