jinguanghai/deepseek-harness-forge-plugins#forge-memory

Everything is a Plugin. — Bring the Forge (铸剑炉) arsenal into DeepSeek Harness.

12 production-grade tools · 3 plugins · zero npm dependencies · pure Cordis — the fastest way to give your DeepSeek Harness agent a real compute brain: symbolic math, formal logic, theorem proving, regex verification, model checking, code repair, TCM diagnosis, and memory recall.

---

✨ Why Forge Plugins?

🧮 Real computation	Symbolic math and logic that never hallucinate — every tool runs a compiled gate binary and returns verified results
🧠 Formal reasoning	TPTP first-order theorem proving (EPROVER) + state-machine model checking, not just text generation
🀄 TCM diagnosis	Unique 八极八势 (Eight Trigrams) pattern diagnosis and herb-pair retrieval across an ancient-formula corpus — the only TCM plugin for DeepSeek Harness
🪶 Zero deps	Pure Cordis plugins, no npm install, no framework bootstrap code — drop-in and go
🔒 Sandboxed	All file access goes through ctx.fs (host-plane discipline) — no bare fs calls
🎁 Prebuilt binaries	Windows .exe included in Releases — Linux/macOS build from src/go in one command

---

🔌 Plugins and Tools

Plugin	Tool	What it does	Example result
forge-gates	forge_math	Symbolic simplification / evaluation	3*(x+1)-3*x → 3
	forge_logic	Formal proof / equivalence check	x>0 and x<10 -> x>0 → proved
	forge_regex	Fullmatch regex validation	[A-Z]\d{3} vs B456 → pass
	forge_eprover	TPTP first-order theorem proving	CNF axioms → theorem / counter-sat
	forge_system	State-machine model checking (deadlock/invariant)	6 states → passed
	forge_repair	Code defect fix suggestions	bare except: → except Exception:
forge-tcm	tcm_diagnose	八极八势 TCM pattern diagnosis	text/vector → trigram + strategy
	tcm_herb_pair	Herb-pair co-occurrence retrieval	附子+干姜 → 916 formulas
forge-memory	memory_recall	BM25 experience recall	query → ranked snippets
	memory_fold_list	Folded-memory index (π/φ)	list folded sections
	memory_fold_preview	Preview a folded section	content preview
	memory_fold_deep	Deep-read a folded section	full content

12 tools · 3 plugins · MIT · works with dsh ≥ 0.1.0-rc.5

---

🚀 Quick Start

# 1. Put this repo anywhere, e.g.:
#    ~/.dsh/.agent-presets/forge-plus/   (repo root = preset dir)

# 2. Register plugins in your agent preset (see examples/agent.cordis.yml.snippet)

# 3. Start DeepSeek Harness:
npx @deepseek-ai/dsh web

agent.cordis.yml snippet:

# inside your preset directory
plugins:
  - path: ./plugins/forge-gates
  - path: ./plugins/forge-tcm
  - path: ./plugins/forge-memory

Then just ask your agent:

"Simplify 3(x+1)-3x" → 3 ✅ "Prove that x>0 follows from x>0 and x<10" → proved ✅ "Check regex [A-Z]\d{3} against B456 and order B456 ok" → pass/fail ✅ "附子 配 干姜 有哪些方剂?" → ranked formulas from your corpus ✅

---

🔧 Environment

Variable	Default	Purpose
FORGE_GATE_BIN	<repo>/bin/windows	gate binary dir (Linux/macOS: go build from src/go, see below)
FORGE_HOME	process.cwd()	memory data dir (memory.json / .forge-temp/memory_current.json / data/memory_current.json)
FORGE_DATA	cwd / data/	TCM formula dir (must contain formula_db.json)

🔨 Build Gates from Source (Linux / macOS / Windows)

cd src/go
for g in math logic regex eprover system repair tcm; do
  go build -o ${g}_gate${EXT} ${g}_gate.go
done   # Windows: set EXT=.exe

💡 Windows users: prebuilt .exe binaries are attached to the v0.1.0 Release — no build needed.

---

🏗️ Architecture

DeepSeek Harness (dsh) agent
        │  inject: ['tools']
        ▼
   ┌─ forge-plus preset ──────────────────┐
   │ forge-gates ── forge-memory ── forge-tcm │
   │        └──────┐         └────────┐      │
   │        subprocess               │      │
   │               ▼                 ▼      │
   │     ┌─ gate binaries ──┐  ┌─ data ─────┐│
   │     │ math / logic /   │  │ memory.json││
   │     │ regex / eprover /│  │ formula_db ││
   │     │ system / repair  │  │ (yours)    ││
   │     │ tcm              │  └────────────┘│
   │     └──────────────────┘                │
   └─────────────────────────────────────────┘
Gate protocol: tools invoke gate binaries, each writes ONE JSON object to stdout.

---

📊 Data Notes

• forge-tcm herb-pair retrieval requires formula_db.json — a private ancient-formula corpus (家传古籍数据) that is NOT distributed with this repo. Provide your own and point FORGE_DATA at it.
• forge-memory reads memory.json / .forge-temp/memory_current.json / data/memory_current.json — bring your own memory files (empty templates in data/).
• The gates themselves are pure logic (no private data baked in).

---

🎯 Project Lineage

Forge (铸剑炉) is a general-purpose digital agent: a multi-language compiler sandbox with 18 tool gates (Python / Go / Node / Rust / EPROVER / TCM / …), an LLM brain over a code body, and self-evolving architecture. This repo is its dsh-native plugin family — same gates, same discipline, now Everything is a Plugin.

• Main project: github.com/jinguanghai/forge
• DeepSeek Harness: deepseek-ai/deepseek-harness

📜 License

MIT © jinguanghai

INSTALL

Three plugins are published as Cordis bundles -- install any of them into a dsh profile:

# 1 dsh CLI (local dir / tarball)
dsh plugin add ./plugins/forge-gates
# 2 pnpm (npm registry / local tarball)
pnpm add dsh-forge-gates
# 3 Manual: symlink/copy the plugin dir into your profile's node_modules,
#    then add the bundle entry to package.json's dsh.profile.bundles.

Prebuilt Windows .exe binaries ship inside each bundle (bin/windows/). Latest bundle tarballs: https://github.com/jinguanghai/deepseek-harness-forge-plugins/releases/latest

Linux/macOS: build the 7 gate binaries from src/go/ in one command, see BUILDING.