refactor(ai-tools): use cheaper worker models

Route routine discovery and helper agents to lower-cost models so
subagent delegation stays cheap by default.

Remove gpt-5.4-nano references because current Codex model guidance
lists gpt-5.4-mini and Spark instead.

Author: khaneliman

modules/common/ai-tools/agents.nix

@@ -14,8 +14,8 @@ let
         "Glob"
       ];
       model = {
-        claude = "opus";
-        copilot = "claude-opus-4.8";
+        claude = "sonnet";
+        copilot = "claude-sonnet-4.6";
         antigravity = "gemini-3.1-pro-preview";
         opencode = "openai/gpt-5.5";
       };

modules/common/ai-tools/base.md

@@ -42,9 +42,10 @@ Respond like smart caveman.
   subagents when tool support exists or context would pollute main thread.
 - Give subagents only task, paths, constraints, and exit criteria. Do not copy
   full conversation.
-- Prefer Claude Haiku/Sonnet or Codex/OpenCode `gpt-5.4-mini` /
-  `gpt-5.3-codex-spark` for routine workers. Use stronger models only when
-  ambiguity, risk, or reasoning depth justifies cost.
+- Prefer `gpt-5.4-mini` for robust unsupervised Codex/OpenCode exploration and
+  test workers. Use `gpt-5.3-codex-spark` only for narrow deterministic tool
+  runs, active latency-sensitive pairing, or quota backup. Use stronger models
+  when ambiguity, risk, or reasoning depth justifies cost.
 - Use skills for durable workflows, MCP/live tools for external state, and
   one-shot commands for atomic prompts.
 

modules/common/ai-tools/commands.nix

@@ -3,7 +3,16 @@ let
   importedCommands = lib.importSubdirs ./commands { };
   aiAgents = import ./agents.nix { inherit lib; };
 
-  agentModels = lib.mapAttrs (_name: agent: agent.model) aiAgents.agents;
+  agentModelDefaults = {
+    explore = {
+      claude = "haiku";
+      copilot = "claude-haiku-4.5";
+      antigravity = "gemini-3.1-flash-lite-preview";
+      opencode = "openai/gpt-5.4-mini";
+    };
+  };
+
+  agentModels = agentModelDefaults // lib.mapAttrs (_name: agent: agent.model) aiAgents.agents;
 
   commandAgents = {
     analyze-git-history = "explore";

modules/home/programs/terminal/tools/codex/default.nix

@@ -43,7 +43,7 @@ let
 
     # Cheapest local utility profile for triage and simple transforms.
     nano = {
-      model = "gpt-5.4-nano";
+      model = "gpt-5.4-mini";
       model_reasoning_effort = "none";
       model_verbosity = "low";
       plan_mode_reasoning_effort = "low";

modules/home/programs/terminal/tools/opencode/default.nix

@@ -37,10 +37,11 @@ in
         refactorerModel = aiTools.agents.refactorer.model.opencode;
       in
       {
-        opencode-coding = "opencode --model openai/gpt-5.3-codex-spark";
+        opencode-coding = "opencode --model openai/gpt-5.4-mini";
         opencode-deep = "opencode --model ${refactorerModel}";
-        opencode-nano = "opencode --model openai/gpt-5.4-nano";
+        opencode-nano = "opencode --model openai/gpt-5.4-mini";
         opencode-research = "opencode --agent refactorer";
+        opencode-spark = "opencode --model openai/gpt-5.3-codex-spark";
       }
       // lib.optionalAttrs config.services.exo.enable {
         opencode-exo = ''f(){ model="$1"; shift; opencode --model "exo/$model" "$@"; }; f'';
@@ -75,7 +76,7 @@ in
           model = refactorerModel;
           share = "manual";
           autoupdate = false;
-          small_model = "openai/gpt-5.3-codex-spark";
+          small_model = "openai/gpt-5.4-mini";
           default_agent = "refactorer";
           compaction = {
             auto = true;
@@ -96,7 +97,7 @@ in
             };
             nano = {
               template = "Keep each action minimal and targeted for small-surface modifications.";
-              model = "openai/gpt-5.4-nano";
+              model = "openai/gpt-5.4-mini";
               agent = "refactorer";
               subtask = true;
             };

modules/home/programs/terminal/tools/opencode/oh-my-openagent.nix

@@ -32,21 +32,20 @@ in
         debuggerModel = aiTools.agents.debugger.model.opencode;
         refactorerModel = aiTools.agents.refactorer.model.opencode;
         testRunnerModel = aiTools.agents.test-runner.model.opencode;
-        quickModel = "openai/gpt-5.3-codex-spark";
-        nanoModel = "openai/gpt-5.4-nano";
+        sparkModel = "openai/gpt-5.3-codex-spark";
+        miniModel = "openai/gpt-5.4-mini";
         opencodeSkillsPath = toString config.programs.opencode.skills;
 
         deliberateFallbackModels = [
           {
-            model = "openai/gpt-5.3-codex";
+            model = "openai/gpt-5.4";
             reasoningEffort = "high";
           }
           testRunnerModel
         ];
 
-        fastFallbackModels = [
-          quickModel
-          nanoModel
+        explorationFallbackModels = [
+          sparkModel
         ];
 
         defaultConfig = {
@@ -74,20 +73,19 @@ in
             };
             explore = {
               model = testRunnerModel;
-              fallback_models = fastFallbackModels;
+              fallback_models = explorationFallbackModels;
             };
             librarian = {
               model = testRunnerModel;
-              fallback_models = fastFallbackModels;
+              fallback_models = explorationFallbackModels;
             };
           };
 
           categories = {
             quick = {
-              model = quickModel;
+              model = sparkModel;
               fallback_models = [
-                nanoModel
-                testRunnerModel
+                miniModel
               ];
               description = "Fast, minimal edits and low-surface changes.";
             };
@@ -103,14 +101,13 @@ in
             };
             unspecified-low = {
               model = testRunnerModel;
-              fallback_models = fastFallbackModels;
+              fallback_models = explorationFallbackModels;
               description = "Cheaper general subtasks and verification work.";
             };
             writing = {
               model = testRunnerModel;
               fallback_models = [
-                quickModel
-                nanoModel
+                sparkModel
               ];
               description = "Documentation and prose-heavy tasks.";
             };

modules/home/programs/terminal/tools/pi-coding-agent/default.nix

@@ -56,7 +56,7 @@ in
         ".pi/agent/settings.json".text = builtins.toJSON (
           lib.recursiveUpdate {
             defaultProvider = "openai-codex";
-            defaultModel = "gpt-5.3-codex-spark";
+            defaultModel = "gpt-5.4-mini";
             defaultThinkingLevel = "high";
             enableInstallTelemetry = false;
             collapseChangelog = true;

modules/home/programs/terminal/tools/t3code/default.nix

@@ -294,6 +294,10 @@ in
                 provider = "codex";
                 model = "gpt-5.4";
               }
+              {
+                provider = "codex";
+                model = "gpt-5.4-mini";
+              }
               {
                 provider = "codex";
                 model = "gpt-5.3-codex-spark";