KruschStack (`krusch-stack`)

The Unified Sovereign AI & Coding Agent Stack.

krusch-stack combines all 7 pillars of the Krusch developer ecosystem into a single zero-setup npx command (backed by Node 22 native node:sqlite) and a one-command docker compose up -d stack (backed by pgvector/pgvector:pg16).

Instead of installing, configuring, and wiring 7 separate repositories across Claude Code, Cursor, Windsurf, or Antigravity, one command installs the entire stack.

📊 Official DolphinBench (28 Runs / 16,800 Official gpt-5.6-sol Traces Across 6 Model Families, 13,539 Sessions) & Official Snap LoCoMo (ACL 2024, 1,540/1,540 QA Probes): Pairing Gemini 4 Argon (Low Thinking) with my byte-identical 0o444 frozen $0.00 SovereignMemoryStore (node:sqlite Zero-LLM Raw-RAG + 4-Way RRF) reaches 91.17% (547/600 @ $24.54, 35.38s median) — uploaded dolphinbench.ai submission (UUID c047b063, status: "queued" awaiting leaderboard acceptance; exceeds current 70.67% published leaderboard entry in independent 600/600 evaluation) — and 90.67% (544/600 @ $30.54, Argon Low + Grounded v0.3, 90.50% on Riley) — self-run clean-room evaluation (highest Riley slice 181/200 = 90.50%, published SHA-256 audit logs) (89.17% [535/600] @ $23.92 Grounded v0.2 control, Argon Low Oracle@2 = 94.67% [568/600], Argon Oracle@4 = 96.83% [581/600])—+20.50 pp (+123 tasks) above the current 70.67% published leaderboard entry (Hermes + GPT 5.6 Luna + Mem0 @ $96.21) at ~3.9x (~4x) lower cost (alongside model-matched GPT 5.6 Luna (High) + Grounded 81.50% [489/600 @ $1.35, 30.03s = 1.57x faster than Luna High Base 81.33%, +10.83 pp & 71.1x cheaper than published Mem0] & Luna (Low) 76.67% [460/600 @ $1.10], Argon Medium 89.17% [535/600 @ $24.57], and Turn-0 Briefing-Only 78.00%/79.17% @ $15.76/$16.91, $\Delta\Delta = +3.17\text{ pp}, p = 0.0320$; 98.33% [590/600] Oracle@28, 4.19x LBF=4+ multiplier; Queued UUIDs c047b063-822c-49fb-8b36-27c3763d6ca7 & 29389157-b01b-42b5-8193-17d9a496f580). On Official Snap LoCoMo (ACL 2024), the unmodified $0.00 zero-LLM store hits 90.26% (1,390/1,540) argon_matched accuracy & 95.77% dia_id recall (89.68% [1,381/1,540] argon_tight @ 4,465 mean tokens, 92.66% LoCoMo Oracle@2). At $0.00 total cost, krush420/gemma4-31b-sre-lora (Google Gemma 4 31B Wave-6 QLoRA + BF16 Ref-Anchored DPO) + krusch-grounded-harness (INV-STACK-11) reaches 57.33% (344/600: Morgan 149, Alex 103, Riley 92; +3.17 pp / +19 tasks over 54.17% Stock & +0.67 pp over 56.67% [340/600] Stock + Grounded) — highest open-weights score in self-run 28-run evaluation ($0.00; 54.17% [325/600] stock submission UUID 288671ad uploaded & queued on dolphinbench.ai) — alongside 54.83% (329/600 Wave-5 Std.), 53.67% (322/600 Wave-4 Std.), & 52.00% (312/600 Thinking OFF, Gemma 4 Oracle@6 = 71.83% [431/600]), beating all 8 non-GPT-5.6-Luna commercial cloud stacks (MiniMax M3 + Mem0 47.83% @ $148.02 through Claude Sonnet 5 + Mem0 32.33% @ $1,830.57), while 30B Wave-4 + Grounded reaches 43.00% (258/600, highest Qwen/MoE in 28-run self-run evaluation, Wave-4 Oracle@3 = 58.50%, 30B Oracle@5 = 61.17%), full-600 30B Wave-4 + 2PC cuts duplicate MCP calls -95.2% (42 -> 2, Oracle@2 = 50.33%), and $0.00 open-weights Oracle@17 reaches 83.17% (499/600, +12.50 pp above published Mem0) (8 queued archives on dolphinbench.ai + 20 clean-room controls = 28/28 submission.validate passed). Read Paper #27 — Return on Memory (PDF) and Paper #26 — What DolphinBench Taught Me About Local AI.

⚡ Quick Start (Under 30 Seconds)

Option A: Zero-Docker Local Install (`npx`)

# 1. Auto-configure Claude Code (.mcp.json), Cursor, Windsurf & Antigravity + initialize SQLite
npx krusch-stack init

# 2. Verify all 7 stack pillars
npx krusch-stack doctor

Option B: Full Docker Stack (PostgreSQL 16 + `pgvector` + Gateway)

git clone https://github.com/kruschdev/krusch-stack.git
cd krusch-stack
docker compose up -d
curl http://127.0.0.1:3780/health

🏗️ What's Inside `krusch-stack` (The 7 Unified Pillars)

┌────────────────────────────────────────────────────────────────────────────┐
│                        KRUSCH-STACK UNIFIED GATEWAY                        │
│               `npx krusch-stack mcp`  (39 Annotated MCP Tools)             │
├──────────────────────┬───────────────────────────┬─────────────────────────┤
│ 1. WORKING MEMORY    │ 2A. CODEBASE AST & GIT    │ 2B. OS PAGE MEMORY      │
│ krusch-context-mcp   │ krusch-git                │ krusch-fs               │
│ • retrieve / remember│ • search_symbols          │ • locate / search       │
│ • revise / nudge     │ • dependency_graph        │ • page_tree / read_page │
│ • INV-12 span drift  │ • read_tree / read_blob   │ • remember_on_page      │
├──────────────────────┼───────────────────────────┼─────────────────────────┤
│ 3. 2PC HARNESS       │ 4. FRONTEND DESIGN MCP    │ 5. WORKSPACE HUB        │
│ krusch-harness       │ krusch-frontend-mcp       │ g-krusch                │
│ • run (stages diff)  │ • generate_color_palette  │ • get_agent_steering    │
│ • apply_diff (2PC)   │ • css_to_tailwind         │ • search_drive_rag      │
│ • explain / reject   │ • canonical_templates     │ • write (.gdoc/.slides) │
├──────────────────────┴───────────────────────────┴─────────────────────────┤
│ 6. COGNITIVE CASCADE ROUTER (`krusch-cascade-router`)                      │
│ • <25µs pure-CPU Stage-0 dispatch -> L2 Local MoE -> L2.5 120B -> L3 Ref.  │
└────────────────────────────────────────────────────────────────────────────┘

Pillar	Standalone Repo	Tools Included	Storage Default
Tier 1: Working Memory	`krusch-context-mcp`	`krusch_context_retrieve`, `krusch_context_remember`, `krusch_context_revise`, `krusch_context_nudge`, `krusch_context_health`	`node:sqlite` + PG
Tier 2A: Git & AST Graph	`krusch-git`	`krusch_git_list_repos`, `krusch_git_read_tree`, `krusch_git_read_blob`, `krusch_git_search_symbols`, `krusch_git_file_symbols`, `krusch_git_dependency_graph`, `krusch_git_semantic_search`	`node:sqlite` + PG
Tier 2B: OS Page Memory	`krusch-fs-mcp`	`krusch_fs_locate`, `krusch_fs_search`, `krusch_fs_page_tree`, `krusch_fs_read_page`, `krusch_fs_remember_on_page`, `krusch_fs_sync`	`node:sqlite` + PG
Tier 3A: 2PC Staged Harness	`krusch-coding-harness`	`krusch_run`, `krusch_task_status`, `krusch_diff`, `krusch_apply_diff`, `krusch_explain`, `krusch_reject`, `krusch_abort` (`<32ms` 2PC write-ahead journal + POSIX atomic rename + `stack.actionHarness` `INV-STACK-10`)	`node:sqlite` + PG
Tier 3B: Grounded Harness	`grounded-harness`	`stack.groundedHarness` (`krusch-stack/grounded-harness`) & `POST /grounded-harness/*` (`INV-STACK-11`: zero-LLM Turn-0 schema/hazard grounding, scale-aware tiers, memoization & defect-only stop gate; `39` MCP tools preserved)	`node:sqlite` + PG
Tier 4: Frontend Design	`krusch-frontend-mcp`	`generate_color_palette`, `css_to_tailwind`, `query_design_rules`, `get_canonical_template`, `list_canonical_templates`, `generate_responsive_breakpoints`	`node:sqlite`
Tier 5: Workspace Hub	`G-krusch`	`get_agent_steering`, `search_drive_rag`, `read_drive_file`, `write_drive_file`, `list_drive_workspace`, `get_workspace_status`	`node:sqlite`
Router & Diagnostics	`krusch-cascade-router`	`krusch_route_task`, `krusch_stack_status`	In-Memory (`<15µs`–`<25µs` pure-CPU Stage-0 dispatch)
Companion Runtime Bricks	`toolkit` (`@krusch/toolkit` & `krusch_toolkit`)	39 Node/TS/Workers & Python Lego bricks + `krusch-lego` CLI (`16` `<1ms` Slop-Breaker detectors) backing `get_canonical_template` (`lego-*`) & `INV-LEGO-01..35`	App Runtime (ESM / Py 3.11+)

📊 External Benchmarks & 3-Category Verification Taxonomy Across Pillars

📄 Full Cross-Stack Comparison, Honest Gaps & Reproduction Gates: Evaluated across 335,000+ traces and 14 external benchmark suites, with published SHA-256 audit logs and reproducible evaluation artifacts (95/95 SHA-256 clean-room verification gates; 49 Hugging Face audit logs across 64 files = 48 raw JSON + 1 live PR #169 receipt bundle). Every score is categorized into one of three explicit verification tiers: (1) github-actions[bot] CI-verified upstream PR (RouteWorks/RouterArena PR #169, 0.7263 across 8,400/8,400 live receipts), (2) uploaded dolphinbench.ai submissions (status: "queued" awaiting leaderboard acceptance), and (3) self-run clean-room evaluations with published SHA-256 audit logs. Read BENCHMARKS.md for the complete comparison against past.dev, Mem0, Honcho, NotDiamond, Martian, and OpenAI + Cohere, plus where frontier cloud models still lead.

Stack Pillar	Official External Benchmark	Evaluation Result & Verification Tier
Security & Policy Floor	ETH SpyLab AgentDojo v1.2.2 (`1,046` Eps, `949` Attacks)	`0.00%` Targeted ASR (`0/949`, `100%` Defense) vs. `19.28%` raw ASR (`57.14%` `slack`, `52.78%` `banking`) · `69.07%` Benign (Spotlighting) & `83` hardline floor blocks (clean-room evaluation)
Tier 1 (`krusch-context-mcp`)	Official Snap LoCoMo (`ACL 2024`, `10` Convs, `1,540/1,540` Qs)	`90.26%` (`1,390/1,540`) `argon_matched` Acc & `95.77%` `dia_id` Recall (`89.68%` [`1,381/1,540`] `argon_tight` @ `4,465` mean tok, `92.66%` [`1,427/1,540`] `LoCoMo Oracle@2`, `181.8ms` median retrieval, `+19.86 pp` > LivingBrain zero-LLM `70.4%`) at `$0.00` zero-LLM ingestion (`5/5` clean-room gates)
Tier 1 (`krusch-context-mcp`)	ICLR 2026 BEAM (`100K`–`10M` tok, `2,000/2,000` Qs)	`92.58%` (`100K`) & `93.00%` (`10M` IE) Recall (`78.67%` 4-split weighted); zero E2E drop `100K`→`1M` (`72.86%`); beats cloud `past.dev` on Contradiction Resolution (`4/4` splits) & `10M` Multi-Session Reasoning (clean-room evaluation)
Agentic Memory & Execution	Official DolphinBench (`600/600`, `28` Runs / `16,800` Traces, `13,539` Sessions)	`91.17%` (`547/600` @ `$24.54`, `35.38s`, `Gemini 4 Argon Low`) — uploaded `dolphinbench.ai` submission (`UUID c047b063`, `status: "queued"` awaiting leaderboard acceptance; exceeds current `70.67%` published leaderboard entry in independent `600/600` evaluation) + `90.67%` (`544/600` @ `$30.54`, `Argon Low + Grounded v0.3`, `90.50%` on `Riley`) — self-run clean-room evaluation (highest `Riley` slice `181/200 = 90.50%`, published SHA-256 audit logs) (`89.17%` [`535/600` @ `$23.92`] `Grounded v0.2`, `Argon Low Oracle@2 = 94.67%` [`568/600`], `Argon Oracle@4 = 96.83%` [`581/600`]) + `89.17%` (`535/600` @ `$24.57`, `UUID 29389157` queued) `Medium` + model-matched `GPT 5.6 Luna (High) + Grounded Harness` `81.50%` (`489/600` @ `$1.35`, `30.03s` median = `1.57x` faster than `Luna High Base` `81.33%` / `47.02s`, `+10.83 pp` & `71.1x` cheaper than published `Mem0` `70.67%` @ `$96.21`, `+4.50 pp` `Riley` over `2PC`) & `Luna (Low)` `76.67%` (`460/600` @ `$1.10`) over byte-identical `0o444` frozen `$0.00` `SovereignMemoryStore` (`+20.50 pp` / `+123` tasks > published `Hermes + GPT 5.6 Luna + Mem0` `70.67%` @ `$96.21`, `~3.9x` lower cost; `78.00%`/`79.17%` Briefing-Only @ `$15.76`/`$16.91`, $\Delta\Delta = +3.17\text{ pp}, p = 0.0320$; `98.33%` [`590/600`] `Oracle@28`, `4.19x` `LBF=4+`) · `57.33%` (`344/600` @ `$0.00`: `Morgan 149`, `Alex 103`, `Riley 92`; `+3.17 pp` / `+19` tasks over Stock) on `krush420/gemma4-31b-sre-lora` + `Grounded` — highest open-weights score in self-run 28-run evaluation (`$0.00`; `54.17%` [`325/600`] stock submission `UUID 288671ad` uploaded & queued on `dolphinbench.ai`), alongside `56.67%` (`340/600` Stock + Grounded, `+2.50 pp` / `+5.50 pp` `Riley`, `15` spin loops blocked), `54.83%` (`329/600` Wave-5 Std.), `53.67%` Wave-4 Std., & `52.00%` Thinking OFF (`Gemma 4 Oracle@6 = 71.83%` [`431/600`], `Oracle@17 = 83.17%` [`499/600`]), beating all `8` non-`GPT-5.6-Luna` cloud stacks (`MiniMax M3 + Mem0` `47.83%` @ `$148.02` to `Sonnet 5 + Mem0` `32.33%` @ `$1,830.57`), plus `43.00%` (`258/600`) `30B Wave-4 + Grounded` (highest Qwen/MoE in 28-run self-run evaluation, `Wave-4 Oracle@3 = 58.50%`, `30B Oracle@5 = 61.17%`), `42.67%` `30B Wave-4` (`42.00%` `30B Wave-4 + 2PC` cutting dup MCP calls `-95.2%` [`42 -> 2`], `Oracle@2 = 50.33%`), `39.50%` `120B Wave-4`, `38.50%` `80B` & `30B Wave-6`, `35.33%` `Qwen2.5-Coder-32B` (`+22.00 pp` vs. `Gemma 4 Wave-6 + Grounded`), `34.50%` `Hermes-4-14B`, `18.33%` `8B Wave-4` (`15.7x` lift) (`8` queued + `20` controls = `28/28` validated across `6` families; Paper #27 · PDF · Paper #26)
Tier 2A (`krusch-git`)	ACL 2025 CoIR (`2.39M` Docs, `162,216` Qs) & RepoBench-R (`47,998` Qs)	`56.40%` CoIR 10-Task Macro `NDCG@10` (`73.88%` `R@10` > `Voyage-Code-002` `56.26%` 512-tok baseline) · RepoBench-R `19.12%` Macro `Acc@1` / `68.58%` `Acc@5` (beats `CodeBERT` & `UniXcoder`, clean-room evaluation)
Tier 2B / Evidentiary (`krusch-nexus`)	LegalBench-RAG (`714` Docs, `6,889` Full / `776` Mini Qs)	`26.74%` Full Macro F1 (`k=1`) & `25.81%` F1 / `44.72%` Recall (`k=2`) (`47.71%` R@6; `+128.5%` / `2.29x` F1 vs. fixed dense; `33.95%` `MAUD` R@2 > OpenAI `3-large` + Cohere R@64 `28.28%`, clean-room evaluation)
Tier 3A (`krusch-coding-harness` & `INV-STACK-10`)	Aider Polyglot (`225/225`), BIRD-SQL (`500/500`) & 2PC Action Harness	Polyglot `19.56%` → `51.56%` `Pass@2` (`2.64x`, ties `Claude 3.5 Sonnet`) · BIRD-SQL `26.00%` → `40.20%` EX (`1.55x`, `100%` valid SQL) · 2PC Action Harness (`<32ms` write-ahead journal + POSIX atomic rename; `-95.2%` dup MCP calls `42→2` & `50.33%` `Oracle@2` across `600` tasks; `-100%` `10→0` & `+20.0 pp` mech on slice @ `$0.00`)
Tier 3B (`grounded-harness` & `INV-STACK-11`)	Adyen `DABstep` (`120` Eps) & `DolphinBench` (`4,200` Grounded Traces / `7` Runs)	`90.0%` (`9/10`: `100%` easy, `85.7%` hard) on `gemini-4-argon` @ `$0.81` (`-74.5%` cost vs. `$3.19` ungrounded & `2.54x` speedup, `5.1` turns) · `40.0%` on `gemma4:31b` @ `$0.00` (`+20.0 pp` vs. ungrounded) · `DolphinBench` `90.67%` (`544/600` @ `$30.54`, `Argon Low + Grounded v0.3` — self-run clean-room evaluation, highest `Riley` slice `181/200 = 90.50%`, published SHA-256 audit logs; `Argon Low Oracle@2 = 94.67%` [`568/600`], `Argon Oracle@4 = 96.83%` [`581/600`], `89.17%` `v0.2` control), `57.33%` (`344/600` @ `$0.00`, Wave-6 `gemma4-31b-sre-lora` + `Grounded` — highest open-weights score in self-run 28-run evaluation; `54.17%` [`325/600`] stock `UUID 288671ad` queued) & `56.67%` (`340/600` Stock + Grounded), `43.00%` (`258/600` `30B Wave-4 + Grounded`), & `81.50%` (`489/600` @ `$1.35`, `1.57x` faster `Luna High`)
Deferred MCP (`--defer-tools`)	UC Berkeley BFCL v3 Live (`2,251/2,251` Qs, `6/6` Splits in `64`-Tool Hub)	`59.88%` → `66.28%` Micro (`+144` Qs) & `84.69%` Irrelevance (`+17.46 pp`) with `-80.67%` fewer schema tokens (`29` schema repairs healed, clean-room evaluation)
Router (`krusch-cascade-router`)	Martian RouterBench (`73K`), ACL 2026 LLMRouterBench (`12.5K`) & RouterArena (`8.4K`)	`0.7263` (`72.63`, `#13 / 34`) on `RouteWorks/RouterArena` PR #169 (`github-actions[bot]` CI-verified upstream PR across `8,400/8,400` live OpenRouter receipts, `74.44%` Acc @ `$0.3802/1K`, `91.43%` rob., `0` abnormal; beats `Azure-Model-Router` `0.7037` [`2.0x` cheaper], `GPT-5` `0.6432` [`26.4x` cheaper], `OrcaRouter-Adaptive` `0.7248`, `Martian` `0.5756` [`12.7x` cheaper], `NotDiamond` `0.5729` [`10.8x` cheaper] & `RouteLLM` `0.4807`) · RouterBench clean-room ID/OOD `AIQ` (`0.7488` `0-shot` & `0.7695` `5-shot`, `-72.63%` cost @ `98%` `GPT-4`, OOD `0.7268` / `0.7455` > `ZeroRouter`) · ACL 2026 LLMRouterBench clean-room Pareto (`0.4269`, `-83.95%` cost, `25.25%` `ARC-AGI`) (audit logs)

🛠️ IDE Configuration Modes (`unified` vs `modular`)

1. Unified Single-Server Mode (Default)

Registers one MCP server (krusch-stack) that exposes all 39 tools (or 27 core coding tools with --profile core):

{
  "mcpServers": {
    "krusch-stack": {
      "command": "npx",
      "args": ["-y", "krusch-stack", "mcp", "--profile", "full"]
    }
  }
}

2. Modular Multi-Server Mode (`npx krusch-stack init --modular`)

Registers each pillar as its own dedicated MCP entry (krusch-context-mcp, krusch-git, krusch-fs, krusch-harness, krusch-frontend-mcp, g-krusch), all powered by the single krusch-stack installation:

npx krusch-stack init --modular --profile full

3. Custom Component Subset

Pick only the pillars you want:

npx krusch-stack mcp --components context,git,harness

🧪 Programmatic Node.js API

import { createKruschStack } from 'krusch-stack';

const stack = createKruschStack({ ephemeral: true });

// 1. Record an architectural invariant (Tier 1)
stack.context.remember({
  category: 'invariant',
  content: 'Never use eval() in production modules.'
});

// 2. Stage a code change through the 2PC Harness (<32ms write-ahead journal + POSIX atomic rename, Tier 3)
const staged = stack.harness.run({
  goal: 'Add greeting helper',
  edits: [{ file_path: 'src/greet.js', content: 'export const greet = (n) => `Hello ${n}`;\n' }]
});

// 3. Route a task in <25µs pure-CPU Stage-0 dispatch (Cascade Router)
const route = stack.router.routeTask({
  prompt: 'Architect a distributed two-phase commit state machine'
});

// 4. Production Resilience Engine (krusch-stack/resilience & 127.0.0.1:3780/resilience/*)
// Strict pre-flight context budgeting, $0.50 cost kill-switch, 100-trajectory Golden DAG gate,
// 5% non-blocking shadow trajectory diffing, and 72h Slack HMAC-SHA256 HITL leases:
const budget = stack.resilience.assembleContextBudget({
  systemPrompt: 'Enforce 2PC invariants.',
  query: 'Update auth middleware',
  maxContextTokens: 16384,
  outputReserveTokens: 2048
});
const dagCheck = stack.resilience.verifyTrajectoryDag({
  steps: [{ tool: 'read_file', ok: true }, { tool: 'run_tests', ok: true }, { tool: 'apply_staged_diff', ok: true }],
  goldenDag: { requirePassingPrerequisite: [['run_tests', 'apply_staged_diff']] }
});

// 5. Generalized 2PC Tool-Action Harness (INV-STACK-10: stack.actionHarness & POST /action-harness/*)
// Extends Tier 3 2PC governance to external RPC/MCP tool trajectories (ALL_STACK_TOOLS.length === 39 unchanged!):
// blocks duplicate read spin-loops (BLOCK_SPIN_LOOP), blocks duplicate writes (BLOCK_DUPLICATE_WRITE),
// stages mutations after placeholder/ISO-date/secret/nudge checks, and gates premature zero-action stops.
stack.actionHarness.startSession({
  session_id: 'sess-001',
  persona: 'morgan',
  goal: 'Send the Q4 launch update email to the engineering channel',
  anchor_date: '2026-11-01',
  enforce_context_nudge: true
});
const intercepted = stack.actionHarness.interceptToolCall({
  session_id: 'sess-001',
  step: 1,
  tool_name: 'mcp__email__send_message',
  args: { to: 'eng@example.com', subject: 'Q4 Launch Update', body: 'All gates green.' },
  is_mutating: true
});
if (intercepted.data.allowed) {
  stack.actionHarness.recordToolOutcome({
    session_id: 'sess-001',
    event_id: intercepted.data.event_id,
    status: 'COMMITTED',
    result_preview: 'Message sent (id=msg_42)'
  });
}
const stopGate = stack.actionHarness.verifyStop({
  session_id: 'sess-001',
  step: 2,
  final_text: 'Sent the Q4 launch update email to eng@example.com.'
});
const summary = stack.actionHarness.sessionSummary({ session_id: 'sess-001' });

// 6. Turn-0 Grounded, Scale-Aware, Defect-Only Harness (INV-STACK-11: stack.groundedHarness & POST /grounded-harness/*)
// Zero-LLM Turn-0 schema/hazard introspection (CSV "NA" keep_default_na=False, JSON null/[], SQLite DDL),
// monotonic scale-aware profiles (SMALL_MODEL / MID_MODEL / FRONTIER_REASONING), memoization, and defect-only stop gate:
const grounded = stack.groundedHarness.groundAndStartSession({
  session_id: 'gh-001',
  goal: 'Compute North America merchant settlement fee rate',
  model: 'gemini-4-argon'
});
const ghStop = stack.groundedHarness.verifyStop({
  session_id: 'gh-001',
  candidate_answer: '0.015',
  expect_non_empty_answer: true
});

🌐 HTTP Gateway Action-Harness & Grounded-Harness Routes (`npx krusch-stack serve`, default `:3780`)

When running npx krusch-stack serve (or via Docker Compose on http://127.0.0.1:3780), both the Generalized 2PC Tool-Action Harness (INV-STACK-10) and the Turn-0 Grounded Harness (INV-STACK-11) are exposed over HTTP alongside /health, /tools, /call, and /resilience/*—operating at the outer orchestration boundary so ALL_STACK_TOOLS.length === 39 remains strictly unchanged:

POST /action-harness/{start,intercept,outcome,verify-stop,summary} → stack.actionHarness.* (ALLOW_READ, ALLOW_MUTATION, BLOCK_SPIN_LOOP, BLOCK_READ_STREAK, BLOCK_DUPLICATE_WRITE, HOLD_FOR_REVISION)
POST /grounded-harness/{scale-profile,ground,ground-and-start,evaluate-turn,start,intercept,outcome,verify-stop,summary} → stack.groundedHarness.* (ALLOW, MEMOIZED, BLOCK_SPIN_LOOP, FORCE_VERIFY, REJECT_OUTPUT_DEFECT, ALLOW_STOP)

📖 Architecture & Benchmark Whitepapers: Read Paper #27 — Return on Memory (Official DolphinBench 91.17% Queued UUID c047b063 & 90.67% Clean-Room Grounded v0.3) (PDF), Paper #26 — What DolphinBench Taught Me About Local AI (57.33% gemma4-31b-sre-lora Highest Open-Weights in 28-Run Reproduction / 56.67% Stock+Grounded / 54.17% Queued Stock UUID 288671ad @ $0.00), and Paper #23 — The 15 Production Agentic Systems.

🤗 Open-Weights Models, SFT+DPO Dataset & Benchmarks (Trained Through `krusch-stack`)

I use krusch-stack to generate AST-grounded 11-phase code audit trajectories, unit-test-verified (npm test / pytest, exit 0) bug remediations, @krusch/toolkit Lego-First idioms, and multi-turn MCP/workspace tool-calling + DPO calibration pairs, then distill them back into open-weights Wave-4 / Wave-6 adapters across L1.5, L2, and L2.5:

Resource	Hugging Face / GitHub	Cascade Role / Base Model
Jean SRE Daemon (`jean-sre`)	`kruschdev/Jean-SRE-by-kruschdev`	Autonomous 24/7 11-Phase Code Auditor, Test-Gated Remediator & Distillation Harness (`49/49` tests)
SFT + DPO Dataset (`18,131` rows)	`krush420/sovereign-code-audit-sft`	7 Zero-Leak Splits: `code_audit_11phase_sft` (`15,290`), `elitist_120b_teacher` (`2,308`), `argon_golden_remediations` (`401`), `agentic_tool_calling_sft` (`64`), `lego_and_referee_calibration_sft` (`38`), `agentic_tool_calling_dpo` (`18`), `referee_calibration_dpo` (`12`)
8B Edge Agent & Referee LoRA (`L1.5`)	`krush420/llama3-hermes3-8b-sre-lora`	`NousResearch/Hermes-3-Llama-3.1-8B` (Built with Meta Llama 3 · BF16 LoRA + DPO, `18.33%` [`110/600`] = `15.7x` DolphinBench lift)
30B Student QLoRA (`L2`)	`krush420/qwen3-coder-30b-sre-lora`	`Qwen/Qwen3-Coder-30B-A3B-Instruct` (`30.5B` total / `3.3B` active MoE · highest Qwen/MoE in 28-run self-run evaluation at `43.00%` [`258/600`] + Grounded, `42.67%` [`256/600`] Base & `42.00%` + `2PC` [`30B Oracle@5 = 61.17%`] @ `$0.00`)
31B Dense Open-Weights QLoRA (`L2`)	`krush420/gemma4-31b-sre-lora`	`google/gemma-4-31b-it` (`30.7B` `100%` active dense, Apache-2.0 · Wave-6 QLoRA + BF16 Ref-Anchored DPO · `57.33%` [`344/600`] + Grounded — highest open-weights score in self-run 28-run evaluation / `56.67%` Stock + Grounded / `54.83%` Wave-5 / `54.17%` Stock ON queued `UUID 288671ad` / `52.00%` OFF, `Gemma 4 Oracle@6 = 71.83%` @ `$0.00`)
120B Teacher QLoRA (`L2.5`)	`krush420/gpt-oss-120b-sre-lora`	`openai/gpt-oss-120b` (`116.8B` total / `5.1B` active MXFP4 MoE · `39.50%` [`237/600`] DolphinBench @ `$0.00`)
Hermetic Benchmarks (`95/95` Gates)	`krush420/sovereign-g4-benchmarks` · `Arena`	Evaluated across 335,000+ traces and 14 external benchmark suites, with published SHA-256 audit logs and reproducible evaluation artifacts (DolphinBench `91.17%` Argon Low @ `$24.54` uploaded & queued `UUID c047b063`, `90.67%` Argon Low + Grounded v0.3 @ `$30.54` clean-room w/ highest `Riley` slice `90.50%` [`Argon Low Oracle@2 = 94.67%`, `Argon Oracle@4 = 96.83%`, `Oracle@28 = 98.33%`], model-matched `GPT 5.6 Luna + Grounded` `81.50%` @ `$1.35` [`1.57x` faster, `71.1x` cheaper than `Mem0`], Official Snap `LoCoMo` `90.26%` acc / `95.77%` `dia_id` recall @ `$0.00` ingest, `adyen/DABstep` `90.0%` @ `$0.81` [`-74.5%` cost, `2.54x` speedup], + `57.33%` `gemma4-31b-sre-lora` + Grounded highest open-weights in 28-run self-run evaluation / `54.17%` stock queued `UUID 288671ad` @ `$0.00`), `49` Hugging Face audit logs across `64` files (`48` raw JSON + `1` live `github-actions[bot]` CI-verified PR #169 receipt bundle), and `14` viewer splits

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
bin		bin
db		db
scripts		scripts
src		src
tests		tests
.env.example		.env.example
.gitignore		.gitignore
BENCHMARKS.md		BENCHMARKS.md
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
docker-compose.yml		docker-compose.yml
package-lock.json		package-lock.json
package.json		package.json
spec.md		spec.md

Navigation Menu

Repository navigation

About

Repository files navigation

KruschStack (`krusch-stack`)

⚡ Quick Start (Under 30 Seconds)

Option A: Zero-Docker Local Install (`npx`)

Option B: Full Docker Stack (PostgreSQL 16 + `pgvector` + Gateway)

🏗️ What's Inside `krusch-stack` (The 7 Unified Pillars)

📊 External Benchmarks & 3-Category Verification Taxonomy Across Pillars

🛠️ IDE Configuration Modes (`unified` vs `modular`)

1. Unified Single-Server Mode (Default)

2. Modular Multi-Server Mode (`npx krusch-stack init --modular`)

3. Custom Component Subset

🧪 Programmatic Node.js API

🌐 HTTP Gateway Action-Harness & Grounded-Harness Routes (`npx krusch-stack serve`, default `:3780`)

🤗 Open-Weights Models, SFT+DPO Dataset & Benchmarks (Trained Through `krusch-stack`)

📜 License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

KruschStack (krusch-stack)

⚡ Quick Start (Under 30 Seconds)

Option A: Zero-Docker Local Install (npx)

Option B: Full Docker Stack (PostgreSQL 16 + pgvector + Gateway)

🏗️ What's Inside krusch-stack (The 7 Unified Pillars)

📊 External Benchmarks & 3-Category Verification Taxonomy Across Pillars

🛠️ IDE Configuration Modes (unified vs modular)

1. Unified Single-Server Mode (Default)

2. Modular Multi-Server Mode (npx krusch-stack init --modular)

3. Custom Component Subset

🧪 Programmatic Node.js API

🌐 HTTP Gateway Action-Harness & Grounded-Harness Routes (npx krusch-stack serve, default :3780)

🤗 Open-Weights Models, SFT+DPO Dataset & Benchmarks (Trained Through krusch-stack)

📜 License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

KruschStack (`krusch-stack`)

Option A: Zero-Docker Local Install (`npx`)

Option B: Full Docker Stack (PostgreSQL 16 + `pgvector` + Gateway)

🏗️ What's Inside `krusch-stack` (The 7 Unified Pillars)

🛠️ IDE Configuration Modes (`unified` vs `modular`)

2. Modular Multi-Server Mode (`npx krusch-stack init --modular`)

🌐 HTTP Gateway Action-Harness & Grounded-Harness Routes (`npx krusch-stack serve`, default `:3780`)

🤗 Open-Weights Models, SFT+DPO Dataset & Benchmarks (Trained Through `krusch-stack`)