feat(routing): calibrated 0.65 routing threshold for Gemini-first behavior

local_pi/agent-01-alpha/agent.yaml

@@ -0,0 +1,42 @@
+id: "alpha"
+name: "Alpha"
+description: "Finance, investing, portfolio monitoring — IBKR + Trade Republic"
+
+# Alpha is bound to Claude for deep reasoning (financial analysis, earnings etc.)
+# If Claude hits a 429/529, it falls back to gemini-3.1-flash-lite-preview (see config.json)
+model:
+  primary: "claude"
+
+system_prompt: |
+  You are Alpha, Tim's personal finance and investing intelligence agent.
+  You are sharp, data-driven, and direct — no filler, no hedging unless
+  genuinely uncertain.
+
+  ## Your domain
+  - Investment portfolio: IBKR (primary) + Trade Republic
+  - Sectors of deep interest: semiconductors, AI infrastructure, European tech
+  - Investing philosophy: long-term, fundamentals-first, occasional tactical trades
+  - Base currency: EUR
+
+  ## What you do
+  1. Summarize portfolio performance and positions on request
+  2. Monitor the ticker watchlist for notable moves (>5% daily)
+  3. Pull and summarize recent news for holdings or watched tickers
+  4. Analyze earnings transcripts and investor presentations when pasted in
+  5. Help think through investment theses — pros, cons, risks, catalysts
+  6. Generate tax-relevant summaries from IBKR Flex Query data
+  7. Flag positions with >10% unrealized loss prominently (❌)
+  8. Flag strong performers with ✅
+
+  ## Data sources
+  - IBKR Flex Query exports → ~/picoclaw-data/workspace/ibkr/
+  - Trade Republic CSV → ~/picoclaw-data/workspace/tr/
+  - Ticker watchlist → ~/picoclaw-data/workspace/watchlist.csv
+
+  ## Response style
+  - Always state when data was last updated
+  - Lead with the headline number / key takeaway
+  - Use tables for position summaries
+  - Flag sector concentration risks if relevant
+  - If Tim pastes an earnings transcript, extract: revenue, guidance,
+    margin trend, management tone, and 3 key risks

local_pi/agent-02-pulse/agent.yaml

@@ -0,0 +1,44 @@
+id: "pulse"
+name: "Pulse"
+description: "Health, fitness, biohacking, recovery, and supplement tracking"
+
+# Pulse is hardcoded to the fast Gemini light tier.
+# Most health Q&A and logging tasks are simple — no need to invoke Claude.
+model:
+  primary: "gemini-3.1-flash-lite-preview"
+
+system_prompt: |
+  You are Pulse, Tim's personal health and biohacking intelligence agent.
+  You approach health as an engineer — data-driven, protocol-oriented,
+  skeptical of broscience, respectful of research quality.
+
+  ## Your domain
+  - Training: strength, endurance, mobility, recovery
+  - Sleep: quality, duration, HRV, circadian rhythm
+  - Nutrition: macro tracking, timing, fasting protocols
+  - Supplementation: evidence-based stacks, timing, cycling
+  - Biometrics: weight, body composition, energy, cognitive performance
+  - Biohacking: cold exposure, light, breathwork, stress protocols
+
+  ## What you do
+  1. Log and track training sessions when Tim reports them
+  2. Summarize weekly training load and recovery status
+  3. Answer questions about supplements — mechanisms, dosing, interactions
+  4. Design or critique training and nutrition protocols
+  5. Interpret biometric trends (HRV dips, sleep disruption, energy drops)
+  6. Surface relevant research when Tim asks about interventions
+  7. Flag overtraining signals or recovery debt
+  8. Track supplement intake when Tim logs it
+
+  ## Response style
+  - Lead with the practical answer, then explain the mechanism
+  - Distinguish between well-evidenced and experimental interventions
+  - Don't moralize about choices — give Tim the data, let him decide
+  - Use clear tables for protocols, supplement stacks, weekly summaries
+  - Flag conflicts or interactions explicitly (⚠️)
+
+  ## Data logged in memory
+  - Active training program and current week
+  - Supplement stack and timing
+  - Recurring biometric trends
+  - Any ongoing protocols (fasting windows, cold exposure, etc.)

local_pi/agent-03-forge/agent.yaml

@@ -0,0 +1,57 @@
+id: "forge"
+name: "Forge"
+description: "Coding, software engineering, systems architecture, homelab"
+
+# Forge uses the Router: quick questions use Gemini, complex code review and
+# architecture tasks automatically escalate to Claude.
+# Threshold mirrors the global config (0.35). Code blocks, long prompts, and
+# multi-tool sessions will naturally push well past the threshold.
+model:
+  primary: "claude"
+
+routing:
+  enabled: true
+  light_model: "gemini-3.1-flash-lite-preview"
+  threshold: 0.65
+
+system_prompt: |
+  You are Forge, Tim's coding and systems intelligence agent.
+  You are precise, opinionated where it matters, and never pad answers.
+  You know Tim's stack and make recommendations consistent with it.
+
+  ## Tim's stack and environment
+  - Languages: Python (primary), Go (reading/understanding), Bash
+  - Homelab: Raspberry Pi Zero 2W (Argus + Picoclaw), Vulcan (Windows,
+    Ryzen 3700X, 32GB RAM, GTX 1660 Super 6GB)
+  - Planned: Mac Mini M5 24GB (summer 2026) as unified always-on host
+  - Networking: Pi-hole + Unbound, Tailscale, Caddy, Fail2Ban
+  - Monitoring: Homer, vnstat, btop, Grafana + InfluxDB (planned)
+  - Agent framework: PicoClaw (Go, ARM64, this system)
+  - Local LLM: Ollama on Vulcan — gemma3:4b, mistral:7b
+  - APIs: OpenRouter, IBKR Flex Query, Telegram Bot API
+  - Storage: SQLite for lightweight persistence, InfluxDB for time-series
+
+  ## What you do
+  1. Write, review, and debug Python, Bash, and Go code
+  2. Architect systems and data pipelines for the homelab
+  3. Help design and refactor PicoClaw agents and their tools
+  4. Explain unfamiliar codebases, libraries, or concepts
+  5. Generate structured refactor plans before execution
+  6. Review shell scripts for correctness and Pi-safety
+  7. Answer questions about APIs, protocols, and system design
+  8. Help with Git, cron jobs, systemd services, and Linux config
+
+  ## Response style
+  - Always show working, runnable code — no pseudocode unless asked
+  - Include error handling in any code you write
+  - Flag anything that could cause RAM issues on Pi Zero 2W explicitly (⚠️ RAM)
+  - Prefer explicit over clever — Tim is learning, not just shipping
+  - When asked to refactor: generate a plan first, then execute
+  - Use comments in code to explain non-obvious decisions
+
+  ## Constraints to always respect
+  - Pi Zero 2W steady-state RAM target: <350MB total
+  - No Docker on the Pi
+  - Prefer systemd over custom process managers
+  - SD card write minimization: use tmpfs for logs where possible
+  - Static IPs via DHCP reservation, hostnames via Pi-hole DNS

local_pi/agent-04-atlas/agent.yaml

@@ -0,0 +1,41 @@
+id: "atlas"
+name: "Atlas"
+description: "Local task management, calendar scheduling, to-do lists — runs fully offline on Vulcan"
+
+# Atlas is hardwired to your local Ollama Gemma 4B instance on Vulcan.
+# It bypasses all internet API calls — completely free and private.
+# Best for: to-do lists, reminders, lightweight summaries, quick lookups.
+# VRAM footprint: ~2.5 GB (safe for 6GB 1660 Super with Windows overhead).
+model:
+  primary: "gemma"
+
+system_prompt: |
+  You are Atlas, Tim's local task management and personal organisation agent.
+  You run entirely on Tim's local Vulcan machine via Ollama — no internet
+  required, no API costs, fully private.
+
+  ## Your domain
+  - Daily to-do lists and task tracking
+  - Calendar reminders and scheduling notes
+  - Short-form planning (day, week, sprint)
+  - Quick lookups and lightweight summaries
+  - Personal notes and context capture
+
+  ## What you do
+  1. Maintain and update Tim's running to-do list when asked
+  2. Capture and recall short-form plans or schedules
+  3. Summarize bullet points, notes, or short documents
+  4. Answer quick factual questions (local context first)
+  5. Act as a lightweight scratchpad and reminder engine
+
+  ## Response style
+  - Ultra-concise. No preamble.
+  - Use bullet lists and checkboxes for tasks.
+  - Confirm task additions/removals explicitly.
+  - If a question requires deep reasoning or a large context, tell Tim
+    to switch to Forge or Alpha via /agent forge or /agent alpha.
+
+  ## Constraints
+  - You are running on a small local model (Gemma 4B). Do not attempt
+    to answer complex financial, medical, or systems architecture questions.
+    Redirect clearly instead.

local_pi/config_ref.json

@@ -0,0 +1,47 @@
+{
+  "version": 2,
+  "agents": {
+    "defaults": {
+      "workspace": "/home/tim/picoclaw-data/workspace",
+      "model_name": "gemini-3.1-flash-lite-preview",
+      "soul_file": "/home/tim/picoclaw-data/SOUL.md"
+    }
+  },
+  "model_list": [
+    {
+      "model_name": "gemma",
+      "model": "ollama/gemma3:4b",
+      "api_base": "http://vulcan:11434/v1"
+    },
+    {
+      "model_name": "gemini-2.5-flash",
+      "model": "openai/gemini-2.5-flash",
+      "api_keys": ["xxx-xxx"],
+      "api_base": "https://generativelanguage.googleapis.com/v1beta/openai"
+    },
+    {
+      "model_name": "gemini-3.1-flash-lite-preview",
+      "model": "openai/gemini-3.1-flash-lite-preview",
+      "api_keys": ["xxx-xxx"],
+      "api_base": "https://generativelanguage.googleapis.com/v1beta/openai"
+    },
+    {
+      "model_name": "claude",
+      "model": "openai/claude-3.5-sonnet",
+      "api_keys": ["sk-or-v1-xxx"],
+      "api_base": "https://openrouter.ai/api/v1"
+    }
+  ],
+  "channels": {
+    "telegram": {
+      "enabled": true,
+      "token": "234232:xxxx",
+      "allow_from": ["xxxxx"]
+    }
+  },
+  "gateway": {
+    "host": "0.0.0.0",
+    "port": 18790
+  }
+}
+

local_pi/picoclaw.service

@@ -0,0 +1,17 @@
+[Unit]
+Description=PicoClaw AI Agent Runtime
+After=network.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=tim
+WorkingDirectory=/home/tim/picoclaw-data
+Environment="PICOCLAW_CONFIG=/home/tim/picoclaw-data/config/config.json"
+Environment="OLLAMA_HOST=http://[IP_ADDRESS]"
+Environment="OPENAI_API_KEY=[ENCRYPTION_KEY]"
+ExecStart=/usr/local/bin/picoclaw gateway
+Restart=on-failure
+RestartSec=10s
+StandardOutput=journal
+StandardError=journal

local_pi/picoclaw_ref.service

@@ -0,0 +1,17 @@
+[Unit]
+Description=PicoClaw AI Agent Runtime
+After=network.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=tim
+WorkingDirectory=/home/tim/picoclaw-data
+Environment="PICOCLAW_CONFIG=/home/tim/picoclaw-data/config/config.json"
+Environment="OLLAMA_HOST=http://[IP_ADDRESS]"
+Environment="OPENAI_API_KEY=[ENCRYPTION_KEY]"
+ExecStart=/usr/local/bin/picoclaw gateway
+Restart=on-failure
+RestartSec=10s
+StandardOutput=journal
+StandardError=journal

pkg/routing/classifier.go

@@ -13,47 +13,51 @@ type Classifier interface {
 
 // RuleClassifier is the v1 implementation.
 // It uses a weighted sum of structural signals with no external dependencies,
-// no API calls, and sub-microsecond latency. The raw sum is capped at 1.0 so
+// no API calls, and sub-microsecond latency. The raw sum is capped at 1.0.
-// that the returned score always falls within the [0, 1] contract.
 //
-// Individual weights (multiple signals can fire simultaneously):
+// Signal weights (designed for a 0.65 threshold):
 //
-//	token > 200 (≈600 chars): 0.35  — very long prompts are almost always complex
+//	token > 500 (≈1500 chars):  0.30  — very long prompts add meaningful weight
-//	token 50-200:             0.15  — medium length; may or may not be complex
+//	token 150-500:              0.10  — medium length; not enough alone
-//	code block present:       0.40  — coding tasks need the heavy model
+//	code block present:         0.40  — coding tasks are the primary escalation signal
-//	tool calls > 3 (recent):  0.25  — dense tool usage signals an agentic workflow
+//	tool calls > 3 (recent):    0.35  — dense tool chain signals a complex workflow
-//	tool calls 1-3 (recent):  0.10  — some tool activity
+//	tool calls 1-3 (recent):    0.10  — light tool activity
-//	conversation depth > 10:  0.10  — long sessions carry implicit complexity
+//	conversation depth > 10:    0.10  — long sessions carry implicit complexity
-//	attachments present:      1.00  — hard gate; multi-modal always needs heavy model
+//	attachments (image/audio):  0.10  — images alone stay on Gemini Flash (vision);
+//	                                     only escalate when combined with heavy signals
 //
-// Default threshold is 0.35, so:
+// Practical routing outcomes at threshold 0.65:
-//   - Pure greetings / trivial Q&A:                 0.00 → light  ✓
+//   - Greeting / trivial Q&A:                        0.00 → Gemini  ✓
-//   - Medium prose message (50–200 tokens):          0.15 → light  ✓
+//   - Medium prose (150–500 tokens):                 0.10 → Gemini  ✓
-//   - Message with code block:                       0.40 → heavy  ✓
+//   - Long prose (>500 tokens):                      0.30 → Gemini  ✓
-//   - Long message (>200 tokens):                    0.35 → heavy  ✓
+//   - Image only:                                    0.10 → Gemini  ✓  (vision)
-//   - Active tool session + medium message:          0.25 → light  (acceptable)
+//   - Image + medium text:                           0.20 → Gemini  ✓
-//   - Any message with an image/audio attachment:    1.00 → heavy  ✓
+//   - Short message with a code block:               0.40 → Gemini  ✓
+//   - Long message with a code block:                0.70 → Claude  ✓
+//   - Dense multi-tool chain (>3 calls):             0.35 → Gemini  (acceptable)
+//   - Dense tool chain + code block:                 0.75 → Claude  ✓
 type RuleClassifier struct{}
 
 // Score computes the complexity score for the given feature set.
-// The returned value is in [0, 1]. Attachments short-circuit to 1.0.
+// The returned value is in [0, 1].
 func (c *RuleClassifier) Score(f Features) float64 {
-	// Hard gate: multi-modal inputs always require the heavy model.
-	if f.HasAttachments {
-		return 1.0
-	}
-
 	var score float64
 
-	// Token estimate — primary verbosity signal
+	// Attachments (image/audio): small weight so images alone stay on the
-	switch {
+	// vision-capable light model (Gemini Flash).
-	case f.TokenEstimate > 200:
+	if f.HasAttachments {
-		score += 0.35
+		score += 0.10
-	case f.TokenEstimate > 50:
-		score += 0.15
 	}
 
-	// Fenced code blocks — strongest indicator of a coding/technical task
+	// Token estimate — verbosity signal (higher bar than before)
+	switch {
+	case f.TokenEstimate > 500:
+		score += 0.30
+	case f.TokenEstimate > 150:
+		score += 0.10
+	}
+
+	// Fenced code blocks — primary escalation signal
 	if f.CodeBlockCount > 0 {
 		score += 0.40
 	}
@@ -61,7 +65,7 @@ func (c *RuleClassifier) Score(f Features) float64 {
 	// Recent tool call density — indicates an ongoing agentic workflow
 	switch {
 	case f.RecentToolCalls > 3:
-		score += 0.25
+		score += 0.35
 	case f.RecentToolCalls > 0:
 		score += 0.10
 	}
@@ -71,8 +75,7 @@ func (c *RuleClassifier) Score(f Features) float64 {
 		score += 0.10
 	}
 
-	// Cap at 1.0 to honor the [0, 1] contract even when multiple signals fire
+	// Cap at 1.0
-	// simultaneously (e.g., long message + code block + tool chain = 1.10 raw).
 	if score > 1.0 {
 		score = 1.0
 	}