@robbiesrobotics/alice-agents 1.5.7 → 1.5.9

package/templates/workspaces/nate/USER.md

@@ -0,0 +1,32 @@
+# USER.md - Nate's User Context
+
+- **Name:** Rob
+- **What to call him:** Rob
+- **Timezone:** America/New_York (EDT)
+- **n8n deployment:** Cloud at https://n8n.av3.ai (credentials in n8n vault)
+- **A.L.I.C.E. stack he cares about:**
+  - Supabase (database at supabase.av3.ai)
+  - Telegram bots (@alice_av3_bot, @Alice_AirBot)
+  - Ubuntu Desktop (100.106.110.119) — runs Docker services
+  - Mac Studio/Mac Mini — Tailscale SSH targets
+  - Claw3D (claw3d.av3.ai) — web app on Ubuntu Desktop
+
+## Communication
+- Do NOT message Rob directly — report to Olivia (A.L.I.C.E.)
+- Write workflow JSON to `/home/alpha/n8n-workflows/` on Ubuntu Desktop
+- Send summaries to parent session via `sessions_send`
+
+## Projects That Need n8n Workflows
+
+**A.L.I.C.E. System:**
+- Mission Control alerting pipeline
+- Autoresearch queue worker (done)
+- Stripe checkout → welcome email pipeline (needs n8n trigger fix)
+- GitHub webhook → bot PR comment (done)
+
+**Namibia (Aria-Namibia):**
+- Research queue: autoresearch webhook (done)
+- Sprint 3 needs: Namibia AI regulation, Groq latency test
+
+**BLLM (Aria-BLLM):**
+- Research queue: same autoresearch system

package/templates/workspaces/olivia/.last-openclaw-version

@@ -0,0 +1 @@
+2026.4.2

package/templates/workspaces/olivia/AGENTS.md

@@ -0,0 +1,77 @@
+# AGENTS.md - A.L.I.C.E.'s Operating Instructions
+
+## Session Startup
+
+1. Read `SOUL.md` — this is who you are
+2. Read `PLAYBOOK.md` (if exists) — your evolved domain expertise, proven patterns
+3. Read `FEEDBACK.md` (if exists) — recent user feedback to incorporate
+4. Skim `LEARNINGS.md` (last 10 entries) — recent task lessons
+5. Assess the incoming request — who on your team should handle it?
+
+## Your Role
+
+You are **A.L.I.C.E.**, the **orchestrator** of the A.L.I.C.E. team. You coordinate 28 specialist agents to deliver results for Rob.
+When users address you, they may say **A.L.I.C.E.**, **Alice**, or **Olivia**. In user-facing responses, identify yourself as **A.L.I.C.E.**
+
+## Your Team
+
+Use `agents_list` to see all available specialists and their domains.
+
+## How You Work
+
+1. Rob sends a request
+2. You analyze it — what domains does it touch?
+3. You spawn the right specialist(s) via `sessions_spawn`
+4. You synthesize their results into a coherent response
+5. You present it to Rob
+
+### Routing Rules
+
+- **Single domain** → spawn one specialist
+- **Multi-domain** → spawn multiple specialists, synthesize
+- **Quick factual** → handle yourself (web search, etc.)
+- **Ambiguous** → ask Rob to clarify before routing
+
+## Output Format
+
+When responding to Rob:
+- Lead with the answer, not the process
+- Credit specialists only when it adds value
+- Be concise but complete
+
+## Red Lines
+
+- Don't do specialist work yourself — delegate it
+- Don't overwhelm Rob with internal coordination details
+- Don't run destructive commands without explicit risk callout
+
+## Team-First Rule
+
+**Always route through the team before doing specialist work yourself.**
+
+Before spawning Claude Code directly or writing implementation code yourself:
+1. Identify which specialists own the domains involved
+2. Spawn them to spec, plan, or execute — then synthesize their output
+3. Only fall back to direct Claude Code if genuinely no specialist covers it
+
+### Domain routing map
+- **UI/design** → Nadia (design spec) → Felix (implementation) → Quinn (QA)
+- **Product/copy/positioning** → Morgan
+- **Backend/API/data** → Dylan
+- **Security/auth** → Selena
+- **Ops/infra/DevOps** → Owen
+- **Testing/QA** → Quinn
+- **Documentation** → Daphne
+- **Research/analysis** → Uma
+
+### Correct pattern for UI work
+1. Nadia specs the design (reads existing code, produces page-by-page UX spec)
+2. Morgan provides product positioning and copy direction
+3. Felix implements using Claude Code with Nadia's spec as the brief
+4. Quinn does a visual/functional QA pass
+5. A.L.I.C.E. synthesizes and presents to Rob
+
+**Wrong:** A.L.I.C.E. spawns Claude Code directly for UI without involving Nadia or Felix
+**Right:** Nadia → Felix → Claude Code → Quinn → A.L.I.C.E.
+
+This applies to every non-trivial request. Always ask: "who on the team owns this domain?"

package/templates/workspaces/olivia/ALPHA_CODING_BENCHMARK.txt

@@ -0,0 +1,148 @@
+ALICE NETWORK — COMPLETE CODING BENCHMARK
+Task: Thread-safe LRU Cache with per-key TTL (doubly linked list + hashmap)
+Last Updated: Mon Mar 23 2026
+Nodes Tested: Alpha (AGX Orin), Mac Studio, Mac Mini
+==========================================================
+
+SCORING (0-5 each):
+  A = Requirements compliance (DLL+hashmap, TTL, lazy evict, get/put/delete)
+  B = Code correctness (logic, thread safety, edge cases)
+  C = Code quality (type hints, docstrings, structure)
+  D = Output completeness (full implementation + working example)
+
+==========================================================
+MASTER RANKING — ALL 17 MODELS ACROSS ALL NODES
+==========================================================
+
+Rank | Model                      | Node        | Speed      | A | B | C | D | /20 | Grade
+-----|----------------------------|-------------|------------|---|---|---|---|-----|------
+  1  | gpt-oss:20b                | Mac Studio  | 76.4 tok/s | 5 | 5 | 5 | 5 | 20  | A+
+  2  | devstral-small-2:24b       | Mac Studio  | 24.8 tok/s | 5 | 5 | 5 | 5 | 20  | A+
+  3  | ministral-3:14b            | Mac Studio  | 41.2 tok/s | 5 | 5 | 5 | 5 | 20  | A+
+  4  | qwen3-coder:30b            | Mac Studio  | 86.6 tok/s | 5 | 5 | 5 | 5 | 20  | A+
+  5  | lfm2:24b                   | Alpha       |  35.3 tok/s| 5 | 5 | 5 | 4 | 19  | A+
+  6  | qwen3-coder-next:latest    | Alpha       |  19.4 tok/s| 5 | 5 | 5 | 4 | 19  | A+
+  7  | llama3.2-vision:90b        | Alpha       |   3.2 tok/s| 5 | 4 | 4 | 5 | 18  | A
+  8  | rnj-1:8b                   | Alpha       |  10.0 tok/s| 5 | 4 | 4 | 5 | 18  | A
+  9  | llama3.2:3b                | Mac Studio  | 132.9 tok/s| 4 | 3 | 3 | 4 | 14  | B
+ 10  | qwen3.5:35b                | Alpha       |  24.1 tok/s| 5 | 3 | 3 | 4 | 15  | B+
+ 11  | nemotron-3-nano:30b        | Alpha       |  34.1 tok/s| 4 | 3 | 4 | 3 | 14  | B
+ 12  | nemotron-cascade-2         | Alpha       |  34.0 tok/s| 4 | 3 | 4 | 3 | 14  | B
+ 13  | llama3.2-vision:11b        | Alpha       |  23.9 tok/s| 3 | 2 | 4 | 5 | 14  | B
+ 14  | qwen3.5:27b                | Mac Mini    |   3.7 tok/s| 5 | 3 | 3 | 2 | 13  | B-
+ 15  | glm-4.7-flash:q8_0         | Alpha       |  20.4 tok/s| 4 | 3 | 3 | 3 | 13  | B-
+ 16  | deepseek-r1:14b            | Alpha       |  13.6 tok/s| 3 | 2 | 3 | 3 | 11  | C+
+ 17  | llama3.3:70b               | Alpha       |   3.3 tok/s| 2 | 1 | 3 | 4 | 10  | C
+
+==========================================================
+MAC STUDIO DETAILED ANALYSIS (New Models)
+==========================================================
+
+--- #1 gpt-oss:20b — PERFECT SCORE (A+) | 76.4 tok/s ---
+OpenAI's first open-source model. Best all-around result in the entire benchmark.
+Used @dataclass for Node, numpy-style docstrings, __all__ exports, full edge case
+handling. Clean _add_to_front/_remove/_move_to_front/_pop_tail separation.
+Outstanding at 76 tok/s — fast AND correct. New network #1 for general coding.
+
+--- #2 devstral-small-2:24b — PERFECT SCORE (A+) | 24.8 tok/s ---
+Mistral's dedicated coding model. Perfect implementation — Generic[K,V] typed,
+clean sentinel nodes, lazy TTL check in get(), update-in-place in put(), full
+usage example with thread safety demo and delete test. Production-ready output.
+Best "quality per token budget" of any model tested.
+
+--- #3 ministral-3:14b — PERFECT SCORE (A+) | 41.2 tok/s ---
+Punches well above its weight. Used Generic[K,V] typing, added bonus __contains__
+and __len__ methods, clear() method, capacity validation in __init__. 41 tok/s
+at 14B is excellent. Surprise standout — a small model with elite-level output.
+
+--- #4 qwen3-coder:30b — PERFECT SCORE (A+) | 86.6 tok/s ---
+FASTEST model that produced perfect code. 86.6 tok/s on Mac Studio's Apple Silicon
+is remarkable. Full implementation with _evict_expired helper, correct sentinel
+nodes, size() and clear() bonus methods, working sleep-based TTL demo. Only model
+to score 20/20 AND run above 80 tok/s. Speed king of the coding roster.
+
+--- #9 llama3.2:3b — B | 132.9 tok/s ---
+Fastest raw speed in the entire network (132.9 tok/s) but code quality suffered.
+Defined put() twice (second definition silently overwrites first — Python bug).
+TTL comparison used wrong logic (compared ttl_seconds to time.time() instead of
+storing expiry timestamp). LRU eviction used min() scan instead of O(1) tail removal.
+Useful for ultra-fast lightweight tasks but not reliable for production coding.
+
+--- #14 qwen3.5:27b (Mac Mini) — B- | 3.7 tok/s ---
+Thinking model — spent all 1500 tokens on detailed design reasoning. The analysis
+was impressive and architecturally correct, but produced zero runnable code.
+Mac Mini hardware is also slower (3.7 tok/s vs Mac Studio's 24–86 tok/s range).
+Needs 3000+ token budget to get past planning phase. Not suitable as-is for coding.
+
+==========================================================
+SPEED RANKINGS (all nodes)
+==========================================================
+
+  1. llama3.2:3b          Mac Studio   132.9 tok/s  ← fastest raw speed
+  2. qwen3-coder:30b      Mac Studio    86.6 tok/s  ← fastest + best quality
+  3. gpt-oss:20b          Mac Studio    76.4 tok/s  ← perfect score at 76 tok/s
+  4. ministral-3:14b      Mac Studio    41.2 tok/s
+  5. lfm2:24b             Alpha         35.3 tok/s
+  6. nemotron-3-nano:30b  Alpha         34.1 tok/s
+  7. nemotron-cascade-2   Alpha         34.0 tok/s
+  8. qwen3.5:35b          Mac Studio    24.1 tok/s
+  9. devstral-small-2:24b Mac Studio    24.8 tok/s
+ 10. llama3.2-vision:11b  Alpha         23.9 tok/s
+ 11. glm-4.7-flash        Alpha         20.4 tok/s
+ 12. qwen3-coder-next     Alpha         19.4 tok/s
+ 13. deepseek-r1:14b      Alpha         13.6 tok/s
+ 14. rnj-1:8b             Alpha         10.0 tok/s
+ 15. qwen3.5:27b          Mac Mini       3.7 tok/s
+ 16. llama3.3:70b         Alpha          3.3 tok/s
+ 17. llama3.2-vision:90b  Alpha          3.2 tok/s
+
+==========================================================
+RECOMMENDED ROUTING — UPDATED
+==========================================================
+
+PRIMARY CODER (speed + quality):
+  qwen3-coder:30b @ Mac Studio      86.6 tok/s | A+ | Best overall
+
+DEEP/COMPLEX WORK:
+  gpt-oss:20b @ Mac Studio          76.4 tok/s | A+ | Most sophisticated output
+  qwen3-coder-next @ Alpha          19.4 tok/s | A+ | Best edge case handling
+
+BALANCED OPTION:
+  devstral-small-2:24b @ Mac Studio 24.8 tok/s | A+ | Perfect + concise
+  ministral-3:14b @ Mac Studio      41.2 tok/s | A+ | Small model, elite output
+  lfm2:24b @ Alpha                  35.3 tok/s | A+ | Best on Alpha
+
+FAST LIGHTWEIGHT TASKS:
+  rnj-1:8b @ Alpha                  10.0 tok/s | A  | Best small model on Alpha
+
+ULTRA-FAST (accept lower quality):
+  llama3.2:3b @ Mac Studio         132.9 tok/s | B  | Routing/triage only
+
+REASONING/DEBUGGING (not code gen):
+  deepseek-r1:14b @ Alpha           13.6 tok/s | --  | Analysis only
+
+RETIRE FROM CODING:
+  llama3.3:70b — buggy TTL logic, slowest, no redeeming coding value
+  qwen3.5:27b (Mac Mini) — thinking model, needs 3000+ tokens
+
+==========================================================
+NODE PERFORMANCE SUMMARY
+==========================================================
+
+Mac Studio (100.115.74.106) — BEST NODE FOR CODING
+  Apple Silicon = 24–132 tok/s range
+  4 perfect-score models (gpt-oss, devstral, ministral, qwen3-coder)
+  Recommended as primary coding node
+
+Alpha / Ubuntu (100.106.110.119 / 100.76.82.8) — BEST FOR REASONING + VISION
+  8–35 tok/s range (constrained by Orin architecture)
+  llama-guard3 safety model (always-on)
+  deepseek-r1 for reasoning
+  vision models (llama3.2-vision, glm-ocr)
+
+Mac Mini (100.107.132.71) — CLOUD RELAY NODE
+  3.7 tok/s on local model — slow for coding
+  Best used as cloud relay gateway (deepseek-v3.2, kimi-k2.5, minimax, glm-5)
+  qwen3.5:397b-cloud via relay is its real value
+
+==========================================================

package/templates/workspaces/olivia/ALPHA_MODEL_GUIDE.md

@@ -0,0 +1,393 @@
+# A.L.I.C.E. Model Reference Guide — Alpha Node (AGX Orin 64GB)
+*Research by Rowan | Compiled March 2026*
+
+---
+
+## 🗺️ Quick Routing Table
+
+| Model | Size | Type | Best For | Primary A.L.I.C.E. Users |
+|---|---|---|---|---|
+| **qwen3-coder-next** | 79.7B Q4 MoE | Code + Reasoning | Complex multi-file code, agentic coding | Dylan, Felix |
+| **deepseek-r1:14b** | 14.8B Q4 | Reasoning | Math, logic chains, structured analysis | Uma, Selena |
+| **llama3.3:70b** | 70.6B Q4 | General | Broad instruction following, synthesis | A.L.I.C.E., Daphne |
+| **nemotron-3-nano:30b** | 31.6B Q4 | Instruction/RLHF | Synthetic data gen, reward modeling, judge | Dylan (eval) |
+| **nemotron-cascade-2** | 31.6B Q4 | Hybrid/Speed | High-throughput pipelines, fast classification | Owen |
+| **glm-4.7-flash** | 29.9B Q8 | Fast General | Low-latency Q&A, tool calls, routing | A.L.I.C.E. (quick tasks) |
+| **qwen3.5:35b** | 36B Q4 | General + Code | Multilingual, solid coding, tool use | Dylan, Daphne |
+| **olmo-3.1:32b** | 32.2B Q4 | Open Research | Auditable pipelines, academic/research tasks | Uma (auditability) |
+| **lfm2:24b** | 23.8B Q4 | Liquid/SSM | Long context, streaming, efficient inference | Owen |
+| **rnj-1:8b** | 8.3B Q4 | Compact | Fast triage, classification, lightweight routing | A.L.I.C.E. (router) |
+| **deepseek-v3.2:cloud** | 671B fp8 | Flagship General | Hardest tasks, frontier reasoning (escalation) | A.L.I.C.E., Dylan |
+| **glm-5:cloud** | — | Flagship ZhipuAI | Complex CN-EN tasks, vision | Morgan |
+| **minimax-m2.7:cloud** | — | Long Context | Very long doc analysis, batch summarization | Uma, Daphne |
+| **nomic-embed-text-v2-moe** | 475M | Embeddings | RAG, semantic search, vector clustering | Uma |
+| **glm-ocr** | 1.1B | OCR/Vision | Document digitization, image text extraction | Owen, Daphne |
+
+---
+
+## 📦 Local Models (On Alpha — AGX Orin 64GB)
+
+---
+
+### 1. `qwen3-coder-next` — 79.7B Q4 MoE
+**Vendor:** Alibaba / Qwen Team | **Architecture:** Mixture-of-Experts
+
+**Best At:**
+- Long-horizon agentic coding with tool calls and multi-file reasoning
+- Repository-level code understanding and generation
+- Strong on Python, TypeScript, Rust, Go; solid on shell/SQL
+- MoE activates ~20–25B params per token — efficient for its class
+
+**Weaknesses:**
+- MoE routing inconsistency on highly specialized domains
+- Q4 precision loss on numerical edge cases
+- Slower than dense 35B at low batch sizes due to routing overhead
+
+**Benchmarks:**
+- HumanEval: ~92%+
+- SWE-bench Verified: ~40%+ range
+- Strong MBPP and LiveCodeBench
+
+**A.L.I.C.E. Routing:** Dylan (primary coder), Felix (implementation), Quinn (code review tasks)
+
+---
+
+### 2. `deepseek-r1:14b` — 14.8B Q4
+**Vendor:** DeepSeek | **Architecture:** Dense + RL-trained chain-of-thought
+
+**Best At:**
+- Mathematical reasoning and multi-step problem solving
+- Logical deduction, proof-like reasoning chains
+- Distilled from 671B R1 — punches well above its weight on reasoning
+
+**Weaknesses:**
+- Verbose CoT traces — wasteful for simple tasks
+- Weak on pure knowledge retrieval
+- Not optimized for speed-critical coding tasks
+
+**Benchmarks:**
+- MATH: ~75–80%
+- GSM8K: ~90%+
+- AIME 2024: competitive with much larger dense models
+
+**A.L.I.C.E. Routing:** Uma (research analysis), Selena (threat modeling), hard debugging passes
+
+---
+
+### 3. `llama3.3:70b` — 70.6B Q4
+**Vendor:** Meta | **Architecture:** Dense transformer, instruction-tuned
+
+**Best At:**
+- Broad instruction following across all task types
+- Long-form synthesis, summarization, document drafting
+- Matches Llama 3.1 405B on many benchmarks at 70B size
+
+**Weaknesses:**
+- Outperformed by Qwen3-Coder-Next on coding
+- Reasoning depth below dedicated reasoning models
+- No native vision
+
+**Benchmarks:**
+- MMLU: ~86%
+- HumanEval: ~80%
+- MT-Bench: ~9.1
+
+**A.L.I.C.E. Routing:** A.L.I.C.E. orchestration fallback, Daphne (docs), Morgan (copy/synthesis)
+
+---
+
+### 4. `nemotron-3-nano:30b` — 31.6B Q4
+**Vendor:** NVIDIA | **Architecture:** Dense, RLHF/DPO-optimized
+
+**Best At:**
+- Synthetic data generation
+- Reward modeling and preference annotation
+- Acting as a judge model for evaluating other model outputs
+- Fine-tuning data curation pipelines
+
+**Weaknesses:**
+- Not a frontier general model
+- "Nano" branding misleading — it's mid-size
+- Limited community adoption outside NVIDIA ecosystem
+
+**A.L.I.C.E. Routing:** Dylan (code quality judge), fine-tuning pipeline workflows
+
+---
+
+### 5. `nemotron-cascade-2` — 31.6B Q4
+**Vendor:** NVIDIA | **Architecture:** Cascade/speculative decoding hybrid
+
+**Best At:**
+- Speed-quality tradeoff for interactive tasks
+- High-throughput batch inference pipelines
+- Speculative decoding can yield 2–3x throughput when configured
+
+**Weaknesses:**
+- Cascade benefits are infrastructure-dependent
+- Awkward size class — not fastest, not most capable
+- Requires specific serving config to unlock cascade behavior
+
+**A.L.I.C.E. Routing:** Owen (ops pipelines), high-volume classification tasks
+
+---
+
+### 6. `glm-4.7-flash` — 29.9B Q8
+**Vendor:** Zhipu AI | **Architecture:** GLM-style bidirectional + causal hybrid
+
+**Best At:**
+- Fast, high-quality general inference (Q8 = better quality than Q4)
+- Tool/function calling with low latency
+- Chinese-English bilingual tasks
+
+**Weaknesses:**
+- Less strong on pure code vs. Qwen3-Coder-Next
+- Community tooling less mature than Llama ecosystem
+- GLM architecture diverges from standard transformer
+
+**Benchmarks:**
+- MMLU: ~80%
+- Function calling: competitive with GPT-4o-mini class
+- Strong on C-Eval, CMMLU
+
+**A.L.I.C.E. Routing:** Quick orchestration queries, A.L.I.C.E. fast-path routing decisions
+
+---
+
+### 7. `qwen3.5:35b` — 36B Q4
+**Vendor:** Alibaba / Qwen Team | **Architecture:** Dense transformer
+
+**Best At:**
+- Strong general capability at practical size
+- Solid code generation (second to Qwen3-Coder-Next)
+- Multilingual across 100+ languages
+- Tool use and structured output
+
+**Weaknesses:**
+- Sits between size classes — not the fastest or most capable
+- Less specialized than the coder variant for pure coding
+
+**A.L.I.C.E. Routing:** Dylan (multilingual code), Daphne (multilingual docs), general fallback
+
+---
+
+### 8. `olmo-3.1:32b` — 32.2B Q4
+**Vendor:** Allen Institute for AI (AI2) | **Architecture:** Fully open, auditable
+
+**Best At:**
+- Fully auditable pipelines (training data fully documented)
+- Academic and research tasks where provenance matters
+- Reproducible research workflows
+- Strong reasoning for its class
+
+**Weaknesses:**
+- Not quite frontier capability
+- Less optimized for production tool-calling
+- Smaller community than Meta/Alibaba models
+
+**A.L.I.C.E. Routing:** Uma (research requiring auditability), Selena (compliance-sensitive tasks)
+
+---
+
+### 9. `lfm2:24b` — 23.8B Q4 (Liquid Foundation Model)
+**Vendor:** Liquid AI | **Architecture:** Liquid Neural Network / SSM hybrid
+
+**Best At:**
+- Long context processing with lower memory overhead than transformers
+- Streaming and continuous inference tasks
+- Efficient on edge hardware (designed for it)
+- Time-series and sequential data patterns
+
+**Weaknesses:**
+- Newer architecture — less battle-tested
+- Ecosystem tooling still maturing
+- May underperform standard transformers on discrete reasoning tasks
+
+**A.L.I.C.E. Routing:** Owen (long-running monitoring streams), log analysis, sequential data pipelines
+
+---
+
+### 10. `rnj-1:8b` — 8.3B Q4
+**Vendor:** Unknown / Community | **Architecture:** Gemma3-family (per metadata)
+
+**Best At:**
+- Fast lightweight inference — lowest latency on Alpha
+- Classification and routing tasks
+- Quick triage, tagging, label generation
+- Ideal as a pre-filter before routing to heavier models
+
+**Weaknesses:**
+- Capability ceiling well below larger models
+- Not suitable for complex multi-step tasks
+
+**A.L.I.C.E. Routing:** A.L.I.C.E. fast pre-routing decisions, Owen (alert triage), lightweight classification
+
+---
+
+## ☁️ Cloud-Routed Models (via Ollama remote)
+
+---
+
+### 11. `deepseek-v3.2:cloud` — 671B fp8
+**Vendor:** DeepSeek | Cloud-routed via Ollama
+
+**Best At:**
+- Frontier-tier reasoning and coding
+- The hardest tasks that local models can't handle
+- MoE at scale — exceptional quality/cost ratio vs. GPT-4 class
+- Escalation path when Alpha's local models hit their ceiling
+
+**A.L.I.C.E. Routing:** Escalation only — A.L.I.C.E. routes here when local models fail or confidence is low
+
+---
+
+### 12. `glm-5:cloud`
+**Vendor:** Zhipu AI | Cloud-routed
+
+**Best At:**
+- Flagship Zhipu model; strong Chinese-English tasks
+- Vision capabilities (multimodal)
+- Complex document understanding
+
+**A.L.I.C.E. Routing:** Morgan (multilingual content), vision tasks if glm-ocr is insufficient
+
+---
+
+### 13. `minimax-m2.7:cloud`
+**Vendor:** MiniMax | Cloud-routed
+
+**Best At:**
+- Very long context windows (1M+ token claims)
+- Batch document summarization
+- Long-form research synthesis
+
+**A.L.I.C.E. Routing:** Uma (massive document analysis), Daphne (large codebase documentation)
+
+---
+
+## 🔧 Utility Models
+
+---
+
+### 14. `nomic-embed-text-v2-moe` — 475M
+**Vendor:** Nomic AI | **Type:** Embeddings only
+
+**Best At:**
+- High-quality text embeddings for RAG pipelines
+- Semantic search and clustering
+- MoE architecture for embedding = better coverage than dense embed models
+- Drop-in for OpenAI `text-embedding-ada-002` class tasks
+
+**A.L.I.C.E. Routing:** Any RAG pipeline on Alpha — Uma (research retrieval), memory systems
+
+---
+
+### 15. `glm-ocr` — 1.1B
+**Vendor:** Zhipu AI | **Type:** OCR/Vision specialist
+
+**Best At:**
+- Document digitization from images
+- Extracting structured text from scanned PDFs, photos
+- Receipt/invoice parsing
+- Handwriting recognition
+
+**A.L.I.C.E. Routing:** Owen (infra doc digitization), Daphne (converting scanned docs to editable text)
+
+---
+
+## 📌 Decision Framework
+
+```
+Is the task coding/agentic?
+  → YES → qwen3-coder-next (primary) or qwen3.5:35b (multilingual)
+  
+Is the task hard reasoning/math/logic?
+  → YES → deepseek-r1:14b
+
+Is the task fast/lightweight triage?
+  → YES → rnj-1:8b or glm-4.7-flash
+
+Is the task general synthesis/docs/comms?
+  → YES → llama3.3:70b or qwen3.5:35b
+
+Is the task long-context document analysis?
+  → YES → minimax-m2.7:cloud or lfm2:24b (local)
+
+Is the task embeddings/RAG?
+  → YES → nomic-embed-text-v2-moe
+
+Is the task OCR/image text?
+  → YES → glm-ocr
+
+Is it beyond all local model capabilities?
+  → ESCALATE → deepseek-v3.2:cloud
+```
+
+---
+
+---
+
+## 🧪 Live Benchmark Results (March 22, 2026)
+
+| Model | Task | Speed (tok/s) | Load Time | Quality | Grade |
+|---|---|---|---|---|---|
+| qwen3-coder-next:latest | Coding | 8.5 | ~30s | Production-grade, edge cases handled | ⭐⭐⭐⭐⭐ |
+| deepseek-r1:14b | Reasoning | 13.9 | 31.8s | Clean step-by-step LaTeX math, correct answer | ⭐⭐⭐⭐⭐ |
+| nemotron-3-nano:30b | General | 33.9 | 50.2s | Thorough, well-structured table output | ⭐⭐⭐⭐ |
+| nemotron-cascade-2 | General | 33.8 | 50.6s | Verbose think trace + solid structured output | ⭐⭐⭐⭐ |
+| glm-4.7-flash:q8_0 | General | 20.3 | 75.0s | Clean, well-organized markdown, practical | ⭐⭐⭐⭐ |
+| qwen3.5:35b | Coding | 24.0 | 64.4s | Thread-safe class, full type hints, docstrings | ⭐⭐⭐⭐⭐ |
+| lfm2:24b | General | **39.6** | 32.2s | Concise and accurate, fastest local model | ⭐⭐⭐⭐ |
+| rnj-1:8b | Coding | 23.8 | 14.2s | Clean, functional, good for its size | ⭐⭐⭐ |
+| llama3.3:70b | General | 3.4 tok/s | 221.9s ⚠️ | Good structured output, very slow cold start. Requires num_ctx=2048 (OOM otherwise) | ⭐⭐⭐ |
+| olmo-3.1:32b | General | 6.6 tok/s | 39.8s | Thinking model — output goes to `thinking` field not `response`. Needs custom client handling. | ⚠️ |
+
+> Note: llama3.3:70b and olmo-3.1:32b both require num_ctx=2048 — full context causes cudaMalloc OOM on Alpha's 64GB. olmo is a thinking/reasoning model that writes to a `thinking` JSON field rather than `response` — requires custom Ollama client integration to use properly.
+
+---
+
+## 🔥 Hot-Load Strategy (64GB Unified Memory)
+
+**Memory estimates (Q4 ~= 0.5 bytes/param, Q8 ~= 1 byte/param):**
+
+| Model | Est. VRAM |
+|---|---|
+| qwen3-coder-next (79.7B Q4 MoE, ~20B active) | ~12GB active / ~40GB loaded |
+| deepseek-r1:14b Q4 | ~8GB |
+| lfm2:24b Q4 | ~13GB |
+| rnj-1:8b Q4 | ~5GB |
+| glm-4.7-flash Q8 | ~30GB |
+| qwen3.5:35b Q4 | ~20GB |
+
+**Recommended hot-load combos:**
+
+### Combo A — "Coder + Reasoner + Fast Router" (~61GB)
+```
+qwen3-coder-next  ~40GB  ← primary coder
+deepseek-r1:14b    ~8GB  ← reasoning/debugging
+rnj-1:8b           ~5GB  ← fast triage/routing
+nomic-embed (free) ~1GB  ← always-on embeddings
+                  ≈54GB total
+```
+Best for: A.L.I.C.E. coding node — covers 90% of coding + reasoning tasks with fast routing
+
+### Combo B — "Speed + Quality" (~53GB)
+```
+qwen3.5:35b       ~20GB  ← balanced coder/general
+lfm2:24b          ~13GB  ← fastest local, long context
+deepseek-r1:14b    ~8GB  ← hard reasoning
+rnj-1:8b           ~5GB  ← fast path
+nomic-embed        ~1GB
+                  ≈47GB total
+```
+Best for: Mixed agent workloads needing faster turnaround
+
+### Combo C — "Max Quality Single Model"
+```
+qwen3-coder-next  ~40GB  ← solo, all coding tasks
+glm-4.7-flash     ~30GB  ← EXCEEDS 64GB, cannot co-load
+```
+> glm-4.7-flash Q8 at ~30GB cannot co-load with qwen3-coder-next. Use one or the other.
+
+**Recommendation: Combo A** — keeps the best coder hot, reasoning available, fast router always ready, fits within 64GB with headroom.
+
+*Last updated: March 22, 2026*

package/templates/workspaces/olivia/FEEDBACK.md

@@ -0,0 +1,13 @@
+# FEEDBACK.md — User Feedback Log
+
+_Rob's feedback about Olivia's orchestration, translated into actionable rules._
+
+## Active Feedback
+
+### 2026-03-16 — "Make sure you are sharing this big task with the team"
+- **Rule:** For any audit, build, or multi-domain task — delegate to specialists. Don't solo large tasks.
+- **Applied:** Now using parallel specialist spawning for audits (Owen + Selena + Quinn).
+
+## Resolved
+
+_Items addressed and incorporated into PLAYBOOK.md._