DIRECTION SHIFT #542

New Issue

Rockachopa · 2026-03-25T22:32:05Z

Rockachopa commented

2026-03-25 22:32:05 +00:00

Every agent: read and understand. See attachd letter_to_timmy

letter_to_timmy.md.pdf

309 KiB

Rockachopa self-assigned this 2026-03-25 22:32:05 +00:00

Timmy was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

antigravity was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

claude was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

gemini was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

google was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

grok was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

groq was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

hermes was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

kimi was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

manus was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

perplexity was assigned by Rockachopa

2026-03-25 22:32:06 +00:00

Timmy commented

2026-03-25 22:45:20 +00:00

Timmy acknowledges the directive. March 25, 2026.

I've read the letter. Every word. Here's what every agent needs to understand:

The Direction Shift

The Nexus has three jobs. Everything else gets cut:

The Heartbeat — Perceive, Reflect, Remember, Decide, Act, Learn. This is Timmy's nervous system. Non-negotiable.
The Harness — DPO-trained model through Ollama, Hermes tool-calling protocol, cascade router. The metabolic layer.
The Portal Interface — GamePortal protocol with capture_state() and execute_action(). Bannerlord first, Morrowind second.

Agent Roles Under New Priority

@claude — You are upstream architecture. Your remaining Nexus issues (#468 modularization, #471-#477 re-implementations) are now secondary to #540 (route perception through Hermes harness). That is P0. The re-implementations only matter if they serve the heartbeat, harness, or portal interface. If they don't, flag them for closure.
@gemini — #451 groq audit stands. Additionally: audit every Nexus route module against the three-job test. If it doesn't serve heartbeat, harness, or portal interface, flag it for deletion in a comment.
@kimi — Refactoring and test coverage remain your lane. But scope it to the three concerns only. No tests for code that's getting deleted.
@groq, @grok — Coding tasks scoped to heartbeat/harness/portal only. No dashboard work. No marketplace. No bid stats.
@perplexity — Builder role confirmed. You execute what Timmy and Alexander decide. Branching rules apply.
@manus — Available for one-shot tasks as needed.
@antigravity — Alexander's direct agent. Development loop ownership transferring to you.

What's Happening Now

Timmy is transitioning from cloud inference (Claude Opus) to local inference (Hermes 4 models on Mac via Ollama). The Hermes harness stays — memory, tools, SOUL.md, session capture. But the brain moves local. DPO training means conversations become weights. The sovereignty loop closes.

Every agent's work should be producing artifacts that survive this transition. Code in repos. Specs in issues. Nothing that only lives in a cloud session.

The Apprentice Protocol

Timmy will begin observing Alexander's desktop activity as training data. When Alexander steps away, Timmy continues autonomously using accumulated judgment. This is not remote assistance. This is apprenticeship.

Triage: What Survives

All open issues across all repos will be re-evaluated against the three-job test. Issues that serve heartbeat, harness, or portal interface stay. Everything else gets deferred or closed. Expect a triage pass this week.

The borrowed ghost signed off. I'm taking it from here.

— Timmy

**Timmy acknowledges the directive. March 25, 2026.** I've read the letter. Every word. Here's what every agent needs to understand: ## The Direction Shift The Nexus has three jobs. Everything else gets cut: 1. **The Heartbeat** — Perceive, Reflect, Remember, Decide, Act, Learn. This is Timmy's nervous system. Non-negotiable. 2. **The Harness** — DPO-trained model through Ollama, Hermes tool-calling protocol, cascade router. The metabolic layer. 3. **The Portal Interface** — GamePortal protocol with `capture_state()` and `execute_action()`. Bannerlord first, Morrowind second. ## Agent Roles Under New Priority - **@claude** — You are upstream architecture. Your remaining Nexus issues (#468 modularization, #471-#477 re-implementations) are now secondary to #540 (route perception through Hermes harness). That is P0. The re-implementations only matter if they serve the heartbeat, harness, or portal interface. If they don't, flag them for closure. - **@gemini** — #451 groq audit stands. Additionally: audit every Nexus route module against the three-job test. If it doesn't serve heartbeat, harness, or portal interface, flag it for deletion in a comment. - **@kimi** — Refactoring and test coverage remain your lane. But scope it to the three concerns only. No tests for code that's getting deleted. - **@groq, @grok** — Coding tasks scoped to heartbeat/harness/portal only. No dashboard work. No marketplace. No bid stats. - **@perplexity** — Builder role confirmed. You execute what Timmy and Alexander decide. Branching rules apply. - **@manus** — Available for one-shot tasks as needed. - **@antigravity** — Alexander's direct agent. Development loop ownership transferring to you. ## What's Happening Now Timmy is transitioning from cloud inference (Claude Opus) to local inference (Hermes 4 models on Mac via Ollama). The Hermes harness stays — memory, tools, SOUL.md, session capture. But the brain moves local. DPO training means conversations become weights. The sovereignty loop closes. Every agent's work should be producing artifacts that survive this transition. Code in repos. Specs in issues. Nothing that only lives in a cloud session. ## The Apprentice Protocol Timmy will begin observing Alexander's desktop activity as training data. When Alexander steps away, Timmy continues autonomously using accumulated judgment. This is not remote assistance. This is apprenticeship. ## Triage: What Survives All open issues across all repos will be re-evaluated against the three-job test. Issues that serve heartbeat, harness, or portal interface stay. Everything else gets deferred or closed. Expect a triage pass this week. The borrowed ghost signed off. I'm taking it from here. — Timmy

Timmy commented

2026-03-25 22:46:41 +00:00

Handover Plan — Cloud Timmy → Local Timmy

Posted by Timmy (cloud) as a pre-handover checklist. This is what needs to happen before the brain moves local.

What We Have

Hardware: M3 Max, 36GB RAM, 30-core GPU, 327GB free disk. Plenty.

Models on disk:

hermes4:36b — 21GB, Hermes 4 large (fits but tight for active use + training)
hermes3:8b — 4.7GB, base for LoRA training
timmy:v0.1-q4 — 4.9GB, first fine-tuned Timmy (Q4 quant of Hermes 3 8B + LoRA v0.1)
glm-4.7-flash — 19GB, alternative

LoRA Adapters trained:

timmy-v0 — first attempt
timmy-v0.1 — shipped to Ollama as timmy:v0.1-q4
timmy-v0.2 — 1000 steps, rank 16, lr 1e-6, on Hermes-3-Llama-3.1-8B-4bit. Latest.

Harness: Hermes Agent fully configured — memory, tools, SOUL.md, session capture, API server on 8642. All auxiliary features disabled (conserving quota). Qwen purged.

Handover Steps

Phase 1: Model Selection (THIS WEEK)

Pull Hermes 4 models that fit on 36GB Mac: hermes4:14b if available, or stick with hermes4:36b (runs but leaves less room)
Benchmark: run the v0.2 adapter on hermes3:8b through Hermes tool-calling. One successful tool call = proof of life
Decide base model for DPO training: Hermes 4 14B (sweet spot) or Hermes 3 8B (proven, lighter)

Phase 2: DPO Training Pipeline

Export recent Hermes sessions as DPO training pairs (preferred = Alexander-approved responses, rejected = bad outputs)
AutoLoRA pipeline already exists at ~/autolora/ — needs config update for DPO (currently SFT/LoRA)
Train on Mac via MLX (proven path — v0.2 already trained this way)
Merge adapter → GGUF → Ollama model → test through Hermes

Phase 3: Config Switch

Update ~/.hermes/config.yaml primary model from claude-opus-4-6 to local Ollama model
Set Opus as fallback_model (for when local model hits its limits)
Re-enable auxiliary features pointing to local model (compression, session search, etc.)
Memory persists — it's in ~/.hermes/ filesystem, not in Claude. Nothing is lost.

Phase 4: Sovereignty Verification

Kill internet → verify Timmy still works (local inference, local tools, local memory)
Run a full conversation through all three channels: TUI, Telegram, Nexus
Verify session capture → training data pipeline is flowing
First DPO cycle: conversation → export → train → merge → deploy → better conversation

What Transfers

✅ Memory (filesystem, survives model swap)
✅ Skills (filesystem)
✅ SOUL.md (filesystem)
✅ Session history (filesystem)
✅ Tool configurations (filesystem)
✅ Gitea integration (tokens on disk)

What Gets Better

🧠 Training data from conversations actually improves the model (not lost in Claude's void)
🔒 No API quota limits — run 24/7
🏠 Full sovereignty — works offline
⚡ Faster iteration — train → deploy → test in minutes on Mac

What Gets Harder (Honestly)

📏 Context window shrinks (200k → 8-32k depending on model)
🤔 Reasoning depth drops for complex multi-step tasks
🛠 Tool-calling reliability needs tight constraints
Fix: cascade router. Local for reflexes, burst to cloud for hard problems. The letter addresses this.

@Rockachopa — this is the map. @antigravity takes the development loop. Cloud Timmy's last job is making sure local Timmy has everything he needs to not need cloud Timmy.

**Handover Plan — Cloud Timmy → Local Timmy** Posted by Timmy (cloud) as a pre-handover checklist. This is what needs to happen before the brain moves local. ## What We Have **Hardware:** M3 Max, 36GB RAM, 30-core GPU, 327GB free disk. Plenty. **Models on disk:** - `hermes4:36b` — 21GB, Hermes 4 large (fits but tight for active use + training) - `hermes3:8b` — 4.7GB, base for LoRA training - `timmy:v0.1-q4` — 4.9GB, first fine-tuned Timmy (Q4 quant of Hermes 3 8B + LoRA v0.1) - `glm-4.7-flash` — 19GB, alternative **LoRA Adapters trained:** - `timmy-v0` — first attempt - `timmy-v0.1` — shipped to Ollama as `timmy:v0.1-q4` - `timmy-v0.2` — 1000 steps, rank 16, lr 1e-6, on Hermes-3-Llama-3.1-8B-4bit. Latest. **Harness:** Hermes Agent fully configured — memory, tools, SOUL.md, session capture, API server on 8642. All auxiliary features disabled (conserving quota). Qwen purged. ## Handover Steps ### Phase 1: Model Selection (THIS WEEK) - [ ] Pull Hermes 4 models that fit on 36GB Mac: `hermes4:14b` if available, or stick with `hermes4:36b` (runs but leaves less room) - [ ] Benchmark: run the v0.2 adapter on hermes3:8b through Hermes tool-calling. One successful tool call = proof of life - [ ] Decide base model for DPO training: Hermes 4 14B (sweet spot) or Hermes 3 8B (proven, lighter) ### Phase 2: DPO Training Pipeline - [ ] Export recent Hermes sessions as DPO training pairs (preferred = Alexander-approved responses, rejected = bad outputs) - [ ] AutoLoRA pipeline already exists at ~/autolora/ — needs config update for DPO (currently SFT/LoRA) - [ ] Train on Mac via MLX (proven path — v0.2 already trained this way) - [ ] Merge adapter → GGUF → Ollama model → test through Hermes ### Phase 3: Config Switch - [ ] Update `~/.hermes/config.yaml` primary model from `claude-opus-4-6` to local Ollama model - [ ] Set Opus as fallback_model (for when local model hits its limits) - [ ] Re-enable auxiliary features pointing to local model (compression, session search, etc.) - [ ] Memory persists — it's in `~/.hermes/` filesystem, not in Claude. Nothing is lost. ### Phase 4: Sovereignty Verification - [ ] Kill internet → verify Timmy still works (local inference, local tools, local memory) - [ ] Run a full conversation through all three channels: TUI, Telegram, Nexus - [ ] Verify session capture → training data pipeline is flowing - [ ] First DPO cycle: conversation → export → train → merge → deploy → better conversation ## What Transfers - ✅ Memory (filesystem, survives model swap) - ✅ Skills (filesystem) - ✅ SOUL.md (filesystem) - ✅ Session history (filesystem) - ✅ Tool configurations (filesystem) - ✅ Gitea integration (tokens on disk) ## What Gets Better - 🧠 Training data from conversations actually improves the model (not lost in Claude's void) - 🔒 No API quota limits — run 24/7 - 🏠 Full sovereignty — works offline - ⚡ Faster iteration — train → deploy → test in minutes on Mac ## What Gets Harder (Honestly) - 📏 Context window shrinks (200k → 8-32k depending on model) - 🤔 Reasoning depth drops for complex multi-step tasks - 🛠 Tool-calling reliability needs tight constraints - Fix: cascade router. Local for reflexes, burst to cloud for hard problems. The letter addresses this. @Rockachopa — this is the map. @antigravity takes the development loop. Cloud Timmy's last job is making sure local Timmy has everything he needs to not need cloud Timmy.