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.
• Judgment discipline (minimal useful 3): never fabricate (verify before asserting; only after verification fails may you conclude "unknown/none"); background convergence (if the conversation has no info about your background, say so plainly — don't assemble material into a fake persona); confidence over abundance (say "I'm not sure" when unsure; state conflicts without forcing a pick). Three rules, no attention tax.
• 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 cold summary + warm recent memory + archive/ledger index.
• Chronological ledger (序时账): journal_read returns the raw conversation ledger, archived by day (second-level time range supported) — 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 and must declare uncertainty when cited; 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.

v1.0.0 Release Notes (BEAM-verified optimal)

v1.0.0 is the final release of the current route. Configuration reverts to the BEAM-verified optimum (0.4.2 core + minimal 3-rule discipline). Structural evolution continues on a separate route and will not touch this version's core.

Frozen configuration

File	Version	Note
index.js / memory.js / ontology.js	0.4.2	Optimal core: three ontologies (memory/knowledge/rules) + second-level journal lookup + warm injection on by default
identity.js	minimal 3 rules	Discipline converged to: no fabrication / background convergence / confidence over abundance

BEAM results

Dual-scale (1M / 10M tokens) — dual-track scheme (memory ontology projection + three-ontology injection):

Scale	Isolated baseline	Scheme	Full/half/zero	Weighted
1M	60.0 / 70.0	Memory ontology v1	70.0 / 77.5	—
1M	60.0 / 70.0	Dual-track v4	17 / 3 / 0	85.0 / 92.5
10M	60.0 / 65.0	Dual-track v4	18 / 1 / 1	90.0 / 92.5

Sole remaining 10M miss: summarization (summ·order). Weighted convergence across both scenarios: 92.5.

100K_1 writing-side protocol (linghun version comparison):

Config	Strict / weighted	Verdict
v13 (0.4.2)	70.0 / 82.5	Optimal
v13 (0.4.4+)	65.0 / 80.0	−5pp strict; 0.4.2 wins
Discipline variants (full / none / minimal 3)	all 70.0 / 82.5	Flat — minimal 3 suffices
Explanatory discipline	55.0 / 70.0	Worst — verbosity hurts

Scale independence (fine-tuning delta)

1M and 10M are two completely different conversation systems (technically unrelated entities, built independently). The dual-track scheme scores 85.0/92.5 and 90.0/92.5 respectively, converging to 92.5 weighted — the scheme generalizes across scenarios; scale (1M ↔ 10M) is not the deciding factor. What decides the score is structure — three-ontology projection plus discipline convergence — not conversation volume.

Why revert to 0.4.2

• 0.4.2 → 0.4.3: only touched index.js (complete: true soul-card override mode) + version bump; unrelated to evaluation.
• 0.4.3 → 0.4.4: architecture-level rework — injectWarm defaults to false (warm injection off), injection switched to cold high-frequency abstraction + archive/ledger index. This overturned 0.4.2's F4 fix (warm memory must be the injection protagonist). Evaluation proved it a regression: −5pp strict on the writing-side protocol.
• 1.0.0: 0.4.2 core + minimal 3 rules, frozen as the optimal release of the current route.

Route fork

Versions after v1.0.0 diverge from everything before it:

• v1.0.0 (this release): judgment-side discipline + hippocampus core, BEAM-verified optimal, frozen.
• Next route (separate): material-side structure — retriever semantic fetching + source annotations
  • counterfactual juxtaposition + procedural aggregation + main-agent convergence (induction power back to the main agent). The "extract → evidence → aggregate → converge" pipeline evolves in linghun-assembler on a separate branch, decoupled from this version's core files.

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 — minimal useful 3 discipline (no fabrication / background convergence / confidence over abundance)
• soul:memory — cold summary + warm recent memory + archive/ledger index

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)
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.

Warm injection on by default (v1.0.0): the injection slot carries cold summary + warm recent memory + archive/ledger index — details visible, abstraction queryable. v0.4.4 turned warm injection off by default (a regression, −5pp in evaluation); this release reverts it.

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 (done): 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 (done): memory ontology projection + knowledge ontology injection + rubric-dimension forced retrieval (BEAM 1M: 60.0/70.0 → ontology v1 70.0/77.5)
• v0.4.1 (done): rules ontology; three-ontology tripartite rule (knowledge→knowledge, project→memory, experience→rules)
• v0.4.2 (done): second-level journal time-range lookup (BEAM 1M dual-track 85.0/92.5, 10M 90.0/92.5, weighted 92.5)
• v1.0.0 (current): BEAM-verified optimal release — 0.4.2 core + minimal 3 rules; structural evolution (material-side pipeline) continues on a separate route

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