feat(routing): implement task-specific model routing and overrides

- Added Model field to CronPayload and JobExecutor
- Integrated model override support into AgentLoop selectCandidates
- Updated RuleClassifier to bypass routing if explicit override is provided
- Added unit tests for cron model propagation and agent routing override
- Updated documentation for model tiering, overrides, and ARM build process

docs/custom/improvement-roadmap.md

@@ -0,0 +1,131 @@
+# Picoclaw Improvement Roadmap: Tools, Skills & Cronjobs
+
+This document outlines how to expand Picoclaw's capabilities based on your specific homelab environment and agent roles (Alpha, Pulse, Forge, Atlas).
+
+---
+
+## 1. The Decision Matrix: Which one do I use?
+
+To choose the right method, ask yourself **"When and how should this happen?"**
+
+| Method | The Use Case | The "Why" | Example |
+| :--- | :--- | :--- | :--- |
+| **Tools** | **Real-time action/data.** When you ask "What's the price of ASML?" | Gives the LLM **active hands**. It allows the model to fetch fresh facts that weren't in its training data. | `fetch_ticker_price(symbol)` |
+| **Skills** | **Consistent Logic.** When you want the LLM to analyze the *quality* of a stock using your specific methodology. | Shapes the **agent's brain**. It ensures the bot always follows your "investment thesis" framework every time it thinks. | `ibkr_analysis_skill` |
+| **Cronjobs** | **Scheduled Automation.** When you want a summary at 8 AM every morning without you having to ask for it. | Provides **proactive duty**. It ensures the system is working for you 24/7, even when you aren't chatting with it. | `daily_backup.sh` |
+
+---
+
+## 2. Elaborating on Use Cases
+
+### **Scenario A: "I want to know if my internet is slow right now."**
+- **Winning Method:** **Tool**
+- **Why?** You are in a live chat and need a data-point *right now*. A `run_speedtest` tool triggered by your question is the most efficient.
+
+### **Scenario B: "I want to track my long-term weight trend and get a weekly report every Sunday."**
+- **Winning Method:** **Cronjob + Tool**
+- **Why?** Use a **Cronjob** to trigger the analysis every Sunday. The cronjob then uses a **SQLite Tool** to pull the data and send the report to your Telegram.
+
+### **Scenario C: "I want my agent to follow the 'Getting Things Done' (GTD) method for my task list."**
+- **Winning Method:** **Skill**
+- **Why?** This isn't about an API or a schedule; it's about *how* the agent processes information. A "GTD Skill" contains the rules (folders, next actions, contexts) that the agent must ALWAYS follow when you talk about tasks.
+
+### **Scenario D: "I want to be alerted if ASML drops below €800."**
+- **Winning Method:** **Cronjob**
+- **Why?** A "Tool" only works when you are talking. You need a background **Cronjob** that runs every 10 minutes, checks the price, and "pokes" you if the condition is met.
+
+---
+
+## 2. Recommended Additions by Agent
+
+### **Alpha (Finance)**
+*   **Tool:** `tavily_search` — Real-time news searching for your ticker watchlist.
+*   **Skill:** `fundamental_analysis` — A set of instructions for reading IBKR data and flagging 10-K risk factors.
+*   **Cronjob:** `portfolio_fetch` — Automatically download your IBKR Flex Query CSV every day at 6 PM.
+
+### **Pulse (Fitness)**
+*   **Tool:** `sqlite_query` — Query your local health database for long-term HRV or weight trends.
+*   **Skill:** `recovery_logic` — Specific rules for when to recommend a "deload week" based on biometric inputs.
+*   **Cronjob:** `morning_brief` — At 7 AM, analyze last night's sleep data and generate a summary waiting for you in Telegram.
+
+### **Forge (Systems)**
+*   **Tool:** `ssh_run` — Allow the agent to run commands directly on **Argus** or **Vulcan** for remote debugging.
+*   **Skill:** `refactor_patterns` — Knowledge of Go/Python best practices to improve code quality during review.
+*   **Cronjob:** `heartbeat_monitor` — Every hour, ping all devices in the lab and alert you if one goes offline.
+
+### **Atlas (Org/Local)**
+*   **Tool:** `calendar_api` — Link specifically to an `ical` or Google Calendar to add/remove appointments.
+*   **Cronjob:** `daily_planner` — Every Sunday evening, summarize your upcoming week based on your notes.
+
+---
+
+## 3. Tool Discovery: Web Search
+**Great news:** Picoclaw already has a powerful `web_search` tool built into `pkg/tools/web.go`. It supports:
+- **Tavily** & **Brave Search** (Recommended for pure data)
+- **Perplexity** (Best for analytical answers)
+- **DuckDuckGo** (Free, no API key needed)
+
+To enable these, you simply need to add your API keys to the `config.json` file we built earlier.
+
+---
+
+## 4. Setup Guide: The "Morning Briefing"
+To turn Picoclaw into a proactive assistant that messages you every morning with market and macro news, follow this framework:
+
+### **Phase 1: The Command Script**
+Create a script on your Pi (e.g., `~/scripts/morning_brief.sh`):
+```bash
+#!/bin/bash
+# Trigger Alpha to generate a briefing and send it to your Telegram ID
+picoclaw --agent alpha --cmd "Search for overnight news on ASML, NVDA and the 10Y Treasury yield. Summarize the macro sentiment for today. Check ~/picoclaw-data/workspace/ibkr for any updates. Send result as a Morning Briefing."
+```
+
+### **Phase 2: The Cron Trigger**
+Add this to your Pi's crontab (`crontab -e`):
+```bash
+# Run the briefing every weekday at 7:30 AM
+30 07 * * 1-5 /home/tim/scripts/morning_brief.sh
+```
+
+### **Why this works:**
+1.  **Proactive:** You don't have to remember to check the news; it hits your phone while you're having coffee.
+2.  **Context-Aware:** Because it uses the **Alpha** agent, it already knows your watchlist and your portfolio goals.
+3.  **Low Latency:** By the time you wake up, the "heavy lifting" (searching and summarizing) is already done.
+
+---
+
+## 5. More Brainstorming Ideas
+
+### **Forge (Systems)**
+- **Tool:** `log_audit` — Automatically scan `journalctl` for Picoclaw errors and summarize them.
+- **Skill:** `pi_safety_rules` — Explicit rules to never exceed 400MB RAM in generated code.
+- **Cronjob:** `vulcan_health_check` — Ping your desktop every hour to ensure the Ollama host is reachable.
+
+### **Pulse (Health)**
+- **Skill:** `hrv_interpreter` — A framework to tell you *exactly* how to adjust your workout based on last night's HRV.
+- **Cronjob:** `evening_reminder` — At 9 PM, check if you logged your supplements; if not, send a gentle nudge.
+- **Model Override:** Force a higher-tier model for specific Pulse tasks (e.g., weekly health audit) using the `model` payload field.
+
+---
+
+## 6. Model Tiering & Routing Overrides
+
+Picoclaw uses a multi-tier routing system to balance performance and cost on the Pi Zero 2W.
+
+### **The Model Tiers**
+| Tier | Model | Best For |
+| :--- | :--- | :--- |
+| **Light** | `gemini-1.5-flash-lite` | Simple commands, data retrieval, basic chat. (Default for most Pulse/Atlas agents) |
+| **Mid** | `gemini-1.5-flash` | Reasoning tasks, complex tool orchestration, small code refactors. |
+| **High** | `claude-3-5-sonnet` | Critical decision making, complex coding, high-stakes analysis. |
+
+### **The Override Mechanism**
+While the `RuleClassifier` automatically promotes tasks to higher tiers based on complexity, you can **force** a specific model for repeated tasks (like Cronjobs) to ensure consistent quality.
+
+#### **Cron Override Example:**
+In your `agent.yaml` or cron payload, add the `model` field:
+```yaml
+payload:
+  message: "Run a deep audit of the system logs and suggest security improvements."
+  model: "gemini-1.5-flash" # Forces Mid-tier logic for this specific task
+```

docs/custom/pi-build.md

@@ -15,12 +15,22 @@ This guide details the step-by-step process for building a custom version of Pic
 
 Since the Raspberry Pi Zero 2W is an ARM-based environment (running a 32-bit `armhf` OS usually), we use Go's powerful cross-compilation features natively on Windows.
 
-Run this single command in PowerShell from the repository root:
+Run this dynamic command in PowerShell from the repository root to automatically bake in the current Version and Commit Hash:
+
 ```powershell
-$env:CGO_ENABLED="0"; $env:GOOS="linux"; $env:GOARCH="arm"; $env:GOARM="7"; go build -o picoclaw-custom ./cmd/picoclaw
+# 1. Capture dynamic build metadata
+$v = (git describe --tags --always --dirty 2>$null) -join ""; if (!$v) { $v = "dev" }
+$c = (git rev-parse --short HEAD 2>$null) -join ""; if (!$c) { $c = "unknown" }
+$t = Get-Date -Format "yyyy-MM-ddTHH:mm:ssZ"
+
+# 2. Compile with metadata injection
+$env:CGO_ENABLED="0"; $env:GOOS="linux"; $env:GOARCH="arm"; $env:GOARM="7"
+go build -v -tags "goolm,stdjson" `
+  -ldflags "-X github.com/sipeed/picoclaw/pkg/config.Version=$v -X github.com/sipeed/picoclaw/pkg/config.GitCommit=$c -X github.com/sipeed/picoclaw/pkg/config.BuildTime=$t" `
+  -o picoclaw-custom ./cmd/picoclaw
 ```
 
-This instructs the Go compiler to generate a standalone Linux executable tailored for the Pi's architecture.
+This ensures your Telegram header shows the correct version and commit info instead of just `dev`.
 
 ## Step B: Transfer to Pi
 
@@ -30,7 +40,9 @@ Use `scp` (Secure Copy Protocol), which transfers files over an encrypted SSH co
 scp picoclaw-custom tim@picoclaw.local:/tmp/
 ```
 
-*This copies your newly compiled `picoclaw-custom` file from Windows up to the `/tmp/` folder on your Pi.*
+*This copies your newly compiled `picoclaw-custom` executable file from Windows up to the `/tmp/` folder on your Pi.*
+
+>**Note:** You are **not** copying the entire project folder or any `.go` code! Go's superpower is compiling all your files and logic down into a single, standalone binary file. That one compressed file contains the entire bot engine and is the absolute only thing the Pi needs to run perfectly.
 
 ## Step C: Update the Raspberry Pi Service
 
@@ -52,3 +64,13 @@ Log into your Raspberry Pi terminal via SSH.
 
 ### Why Cross-Compile?
 Cross-compiling on your powerful Windows desktop ("Vulcan") saves the Raspberry Pi Zero 2W from the massive heat, CPU stress, and time-consumption of downloading the Go SDK and compiling code with 512MB of RAM.
+
+## Troubleshooting
+
+### "The new code changes aren't showing up!"
+If you replaced the binary in `/usr/local/bin/picoclaw` but the bot is still running old logic, your `systemd` service is likely pointing to a different folder.
+
+1. SSH into the Pi and run: `sudo systemctl status picoclaw`
+2. Look specifically at the line that says:
+   `├─23319 /usr/local/bin/picoclaw gateway` (Under the CGroup tree).
+3. If the path listed there is something else (like `/home/tim/picoclaw`), you must `mv` your `picoclaw-custom` binary into *that* specific folder instead!

local_pi/config_ref.json

@@ -35,7 +35,7 @@
   "channels": {
     "telegram": {
       "enabled": true,
-      "token": "234232:xxxx",
+      "token": "xxxx:xxxx",
       "allow_from": ["xxxxx"]
     }
   },

pkg/agent/eventbus_test.go

@@ -275,7 +275,7 @@ func TestAgentLoop_EmitsSteeringAndSkippedToolEvents(t *testing.T) {
 
 	resultCh := make(chan string, 1)
 	go func() {
-		resp, _ := al.ProcessDirectWithChannel(context.Background(), "do something", "test-session", "test", "chat1")
+		resp, _ := al.ProcessDirectWithChannel(context.Background(), "do something", "test-session", "test", "chat1", "")
 		resultCh <- resp
 	}()
 

pkg/agent/hook_mount_test.go

@@ -102,7 +102,7 @@ func TestAgentLoop_ProcessDirectWithChannel_AutoMountsBuiltinHook(t *testing.T)
 	})
 	defer al.Close()
 
-	resp, err := al.ProcessDirectWithChannel(context.Background(), "hello", "session-1", "cli", "direct")
+	resp, err := al.ProcessDirectWithChannel(context.Background(), "hello", "session-1", "cli", "direct", "")
 	if err != nil {
 		t.Fatalf("ProcessDirectWithChannel failed: %v", err)
 	}
@@ -140,7 +140,7 @@ func TestAgentLoop_ProcessDirectWithChannel_AutoMountsProcessHook(t *testing.T)
 	})
 	defer al.Close()
 
-	resp, err := al.ProcessDirectWithChannel(context.Background(), "hello", "session-1", "cli", "direct")
+	resp, err := al.ProcessDirectWithChannel(context.Background(), "hello", "session-1", "cli", "direct", "")
 	if err != nil {
 		t.Fatalf("ProcessDirectWithChannel failed: %v", err)
 	}
@@ -172,7 +172,7 @@ func TestAgentLoop_ProcessDirectWithChannel_InvalidConfiguredHookFails(t *testin
 	})
 	defer al.Close()
 
-	_, err := al.ProcessDirectWithChannel(context.Background(), "hello", "session-1", "cli", "direct")
+	_, err := al.ProcessDirectWithChannel(context.Background(), "hello", "session-1", "cli", "direct", "")
 	if err == nil {
 		t.Fatal("expected invalid configured hook error")
 	}

pkg/agent/loop.go

@@ -92,6 +92,7 @@ type processOptions struct {
 	SuppressToolFeedback    bool                // Whether to suppress inline tool feedback messages
 	NoHistory               bool                // If true, don't load session history (for heartbeat)
 	SkipInitialSteeringPoll bool                // If true, skip the steering poll at loop start (used by Continue)
+	ModelOverride           string              // Explicitly requested model tier
 }
 
 type continuationTarget struct {
@@ -1284,12 +1285,13 @@ func (al *AgentLoop) ProcessDirect(
 	ctx context.Context,
 	content, sessionKey string,
 ) (string, error) {
-	return al.ProcessDirectWithChannel(ctx, content, sessionKey, "cli", "direct")
+	return al.ProcessDirectWithChannel(ctx, content, sessionKey, "cli", "direct", "")
 }
 
 func (al *AgentLoop) ProcessDirectWithChannel(
 	ctx context.Context,
 	content, sessionKey, channel, chatID string,
+	modelOverride string,
 ) (string, error) {
 	if err := al.ensureHooksInitialized(ctx); err != nil {
 		return "", err
@@ -1304,6 +1306,7 @@ func (al *AgentLoop) ProcessDirectWithChannel(
 		ChatID:        chatID,
 		Content:       content,
 		SessionKey:    sessionKey,
+		ModelOverride: modelOverride,
 	}
 
 	return al.processMessage(ctx, msg)
@@ -1445,6 +1448,7 @@ func (al *AgentLoop) processMessage(ctx context.Context, msg bus.InboundMessage)
 		DefaultResponse:   defaultResponse,
 		EnableSummary:     true,
 		SendResponse:      false,
+		ModelOverride:     msg.ModelOverride,
 	}
 
 	// context-dependent commands check their own Runtime fields and report
@@ -1848,7 +1852,7 @@ func (al *AgentLoop) runTurn(ctx context.Context, ts *turnState) (turnResult, er
 		ts.ingestMessage(turnCtx, al, rootMsg)
 	}
 
-	activeCandidates, activeModel, usedLight, score := al.selectCandidates(ts.agent, ts.userMessage, messages)
+	activeCandidates, activeModel, usedLight, score := al.selectCandidates(ts.agent, ts.userMessage, messages, ts.opts.ModelOverride)
 	ts.setComplexityScore(score)
 	activeProvider := ts.agent.Provider
 	if usedLight && ts.agent.LightProvider != nil {
@@ -2976,7 +2980,20 @@ func (al *AgentLoop) selectCandidates(
 	agent *AgentInstance,
 	userMsg string,
 	history []providers.Message,
+	modelOverride string,
 ) (candidates []providers.FallbackCandidate, model string, usedLight bool, score float64) {
+	if modelOverride != "" {
+		// Use explicit override if provided (bypasses routing rules)
+		logger.InfoCF("agent", "Model routing: using explicit override",
+			map[string]any{
+				"agent_id": agent.ID,
+				"model":    modelOverride,
+			})
+		// We use the primary candidates list but substitute the model name if it matches an available candidate,
+		// otherwise we trust the override.
+		return agent.Candidates, modelOverride, false, 0
+	}
+
 	if agent.Router == nil || len(agent.LightCandidates) == 0 {
 		return agent.Candidates, resolvedCandidateModel(agent.Candidates, agent.Model), false, 0
 	}

pkg/agent/loop_test.go

@@ -9,6 +9,7 @@ import (
 	"net/http/httptest"
 	"os"
 	"path/filepath"
+	"reflect"
 	"slices"
 	"strings"
 	"testing"
@@ -2008,6 +2009,7 @@ func TestAgentLoop_ContextExhaustionRetry(t *testing.T) {
 		sessionKey,
 		"test",
 		"test-chat",
+		"",
 	)
 	if err != nil {
 		t.Fatalf("Expected success after retry, got error: %v", err)
@@ -2055,7 +2057,7 @@ func TestAgentLoop_EmptyModelResponseUsesAccurateFallback(t *testing.T) {
 	provider := &simpleMockProvider{response: ""}
 	al := NewAgentLoop(cfg, msgBus, provider)
 
-	response, err := al.ProcessDirectWithChannel(context.Background(), "hello", "empty-response", "test", "chat1")
+	response, err := al.ProcessDirectWithChannel(context.Background(), "hello", "empty-response", "test", "chat1", "")
 	if err != nil {
 		t.Fatalf("ProcessDirectWithChannel failed: %v", err)
 	}
@@ -2087,7 +2089,7 @@ func TestAgentLoop_ToolLimitUsesDedicatedFallback(t *testing.T) {
 	al := NewAgentLoop(cfg, msgBus, provider)
 	al.RegisterTool(&toolLimitTestTool{})
 
-	response, err := al.ProcessDirectWithChannel(context.Background(), "hello", "tool-limit", "test", "chat1")
+	response, err := al.ProcessDirectWithChannel(context.Background(), "hello", "tool-limit", "test", "chat1", "")
 	if err != nil {
 		t.Fatalf("ProcessDirectWithChannel failed: %v", err)
 	}
@@ -2162,6 +2164,7 @@ func TestProcessDirectWithChannel_TriggersMCPInitialization(t *testing.T) {
 		"session-1",
 		"cli",
 		"direct",
+		"",
 	)
 	if err != nil {
 		t.Fatalf("ProcessDirectWithChannel failed: %v", err)
@@ -3046,3 +3049,41 @@ func TestProcessMessage_ContextOverflow_AnthropicStyle(t *testing.T) {
 		t.Fatalf("expected 2 calls for retry, got %d", provider.calls)
 	}
 }
+
+func TestAgentLoop_SelectCandidates_ModelOverride(t *testing.T) {
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				ModelName: "default-model",
+			},
+		},
+	}
+	msgBus := bus.NewMessageBus()
+	provider := &simpleMockProvider{response: "ok"}
+	al := NewAgentLoop(cfg, msgBus, provider)
+
+	// Test case: provide an explicit model override
+	modelOverride := "explicit-gemini-3-flash"
+	
+	// We need an agent instance
+	agent := al.registry.GetDefaultAgent()
+	if agent == nil {
+		t.Fatal("default agent not found")
+	}
+
+	candidates, selectedModel, usedLight, score := al.selectCandidates(agent, "test", nil, modelOverride)
+
+	if selectedModel != modelOverride {
+		t.Fatalf("selectedModel = %q, want %q", selectedModel, modelOverride)
+	}
+	if usedLight {
+		t.Fatal("usedLight should be false when using override")
+	}
+	if score != 0 {
+		t.Fatalf("score should be 0 when using override, got %f", score)
+	}
+	if !reflect.DeepEqual(candidates, agent.Candidates) {
+		t.Fatal("candidates should match agent.Candidates")
+	}
+}
+

pkg/agent/steering_test.go

@@ -702,6 +702,7 @@ func TestAgentLoop_Steering_SkipsRemainingTools(t *testing.T) {
 			"test-session",
 			"test",
 			"chat1",
+			"",
 		)
 		resultCh <- result{resp, err}
 	}()
@@ -784,6 +785,7 @@ func TestAgentLoop_Steering_InitialPoll(t *testing.T) {
 		"test-session",
 		"test",
 		"chat1",
+		"",
 	)
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
@@ -971,6 +973,7 @@ func TestAgentLoop_Steering_DirectResponseContinuesWithQueuedMessage(t *testing.
 			sessionKey,
 			"test",
 			"chat1",
+			"",
 		)
 		resultCh <- struct {
 			resp string
@@ -1185,6 +1188,7 @@ func TestAgentLoop_InterruptGraceful_UsesTerminalNoToolCall(t *testing.T) {
 			sessionKey,
 			"test",
 			"chat1",
+			"",
 		)
 		resultCh <- result{resp: resp, err: err}
 	}()
@@ -1350,6 +1354,7 @@ func TestAgentLoop_InterruptHard_RestoresSession(t *testing.T) {
 			sessionKey,
 			"test",
 			"chat1",
+			"",
 		)
 		resultCh <- result{resp: resp, err: err}
 	}()
@@ -1518,7 +1523,7 @@ func TestAgentLoop_Steering_SkippedToolsHaveErrorResults(t *testing.T) {
 	resultCh := make(chan string, 1)
 	go func() {
 		resp, _ := al.ProcessDirectWithChannel(
-			context.Background(), "go", "test-session", "test", "chat1",
+			context.Background(), "go", "test-session", "test", "chat1", "",
 		)
 		resultCh <- resp
 	}()

pkg/bus/types.go

@@ -26,6 +26,7 @@ type InboundMessage struct {
 	MessageID  string            `json:"message_id,omitempty"`  // platform message ID
 	MediaScope    string            `json:"media_scope,omitempty"` // media lifecycle scope
 	SessionKey    string            `json:"session_key"`
+	ModelOverride string            `json:"model_override,omitempty"` // Explicitly requested model tier
 	Metadata      map[string]string `json:"metadata,omitempty"`
 }
 

pkg/cron/service.go

@@ -29,6 +29,7 @@ type CronPayload struct {
 	Command string `json:"command,omitempty"`
 	Channel string `json:"channel,omitempty"`
 	To      string `json:"to,omitempty"`
+	Model   string `json:"model,omitempty"` // Override model tier (e.g. "gemini-3-flash")
 }
 
 type CronJobState struct {

pkg/tools/cron.go

@@ -15,7 +15,7 @@ import (
 
 // JobExecutor is the interface for executing cron jobs through the agent
 type JobExecutor interface {
-	ProcessDirectWithChannel(ctx context.Context, content, sessionKey, channel, chatID string) (string, error)
+	ProcessDirectWithChannel(ctx context.Context, content, sessionKey, channel, chatID string, modelOverride string) (string, error)
 	// PublishResponseIfNeeded sends response to the outbound bus only when the
 	// agent did not already deliver content through the message tool in this round.
 	PublishResponseIfNeeded(ctx context.Context, channel, chatID, response string)
@@ -351,6 +351,7 @@ func (t *CronTool) ExecuteJob(ctx context.Context, job *cron.CronJob) string {
 		sessionKey,
 		channel,
 		chatID,
+		job.Payload.Model,
 	)
 	if err != nil {
 		return fmt.Sprintf("Error: %v", err)

pkg/tools/cron_test.go

@@ -21,6 +21,7 @@ type stubJobExecutor struct {
 	lastKey         string
 	lastChan        string
 	lastChatID      string
+	lastModel       string // Track the model override
 	publishedResp   string
 	publishedChan   string
 	publishedChatID string
@@ -29,11 +30,13 @@ type stubJobExecutor struct {
 func (s *stubJobExecutor) ProcessDirectWithChannel(
 	_ context.Context,
 	content, sessionKey, channel, chatID string,
+	modelOverride string,
 ) (string, error) {
 	s.lastPrompt = content
 	s.lastKey = sessionKey
 	s.lastChan = channel
 	s.lastChatID = chatID
+	s.lastModel = modelOverride
 	return s.response, s.err
 }
 
@@ -345,3 +348,23 @@ func TestCronTool_ExecuteJobReturnsErrorWithoutPublish(t *testing.T) {
 		t.Fatalf("unexpected publish on error path: %q", executor.publishedResp)
 	}
 }
+
+func TestCronTool_ExecuteJob_ModelOverride(t *testing.T) {
+	executor := &stubJobExecutor{response: "reply"}
+	tool := newTestCronToolWithExecutorAndConfig(t, executor, config.DefaultConfig())
+
+	job := &cron.CronJob{ID: "job-override"}
+	job.Payload.Channel = "telegram"
+	job.Payload.To = "chat-1"
+	job.Payload.Message = "heavy task"
+	job.Payload.Model = "gemini-3-flash"
+
+	if got := tool.ExecuteJob(context.Background(), job); got != "ok" {
+		t.Fatalf("ExecuteJob() = %q, want ok", got)
+	}
+
+	if executor.lastModel != "gemini-3-flash" {
+		t.Fatalf("model = %q, want gemini-3-flash", executor.lastModel)
+	}
+}
+