proxy: add support for user defined metadata in model configs (#333)

Changes: - add Metadata key to ModelConfig - include metadata in /v1/models under meta.llamaswap key - add recursive macro substitution into Metadata - change macros at global and model level to be any scalar type Note: This is the first mostly AI generated change to llama-swap. See #333 for notes about the workflow and approach to AI going forward.
2025-10-04 19:56:41 -07:00
parent 1f6179110c
commit 70930e4e91
11 changed files with 807 additions and 25 deletions
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -0,0 +1,43 @@
+# Project: llama-swap
+
+## Project Description:
+
+llama-swap is a light weight, transparent proxy server that provides automatic model swapping to llama.cpp's server.
+
+## Tech stack
+
+- golang
+- typescript, vite and react for UI (ui/)
+
+## Testing
+
+- `make test-dev` - Use this when making iterative changes. Runs `go test` and `staticcheck`. Fix any static checking errors.
+- `make test-all` - runs at the end before completing work. Includes long running concurrency tests.
+
+## Workflow Tasks
+
+### Plan Improvements
+
+Work plans are located in ai-plans/. Plans written by the user may be incomplete, contain inconsistencies or errors.
+
+When the user asks to improve a plan follow these guidelines for expanding and improving it.
+
+- Identify any inconsistencies.
+- Expand plans out to be detailed specification of requirements and changes to be made.
+- Plans should have at least these sections:
+  - Title - very short, describes changes
+  - Overview: A more detailed summary of goal and outcomes desired
+  - Design Requirements: Detailed descriptions of what needs to be done
+  - Testing Plan: Tests to be implemented
+  - Checklist: A detailed list of changes to be made
+
+Look for "plan expansion" as explicit instructions to improve a plan.
+
+### Implementation of plans
+
+When the user says "paint it", respond with "commencing automated assembly". Then implement the changes as described by the plan. Update the checklist as you complete items.
+
+## General Rules
+
+- when summarizing changes only include details that require further action (action items)
+- when there are no action items, just say "Done."
--- a/7
+++ b/7
@@ -23,6 +23,11 @@ proxy/ui_dist/placeholder.txt:
 	mkdir -p proxy/ui_dist
 	touch $@

+# use cached test results while developing
+test-dev: proxy/ui_dist/placeholder.txt
+	go test -short ./proxy/...
+	staticcheck ./proxy/... || true
+
 test: proxy/ui_dist/placeholder.txt
 	go test -short -count=1 ./proxy/...

@@ -82,4 +87,4 @@ release:
 	git tag "$$new_tag";

 # Phony targets
-.PHONY: all clean ui mac linux windows simple-responder test test-all
+.PHONY: all clean ui mac linux windows simple-responder test test-all test-dev
--- a/ai-plans/issue-264-add-metadata.md
+++ b/ai-plans/issue-264-add-metadata.md
@@ -0,0 +1,292 @@
+# Add Model Metadata Support with Typed Macros
+
+## Overview
+
+Implement support for arbitrary metadata on model configurations that can be exposed through the `/v1/models` API endpoint. This feature extends the existing macro system to support scalar types (string, int, float, bool) instead of only strings, enabling type-safe metadata values.
+
+The metadata will be schemaless, allowing users to define any key-value pairs they need. Macro substitution will work within metadata values, preserving types when macros are used directly and converting to strings when macros are interpolated within strings.
+
+## Design Requirements
+
+### 1. Enhanced Macro System
+
+**Current State:**
+
+- Macros are defined as `map[string]string` at both global and model levels
+- Only string substitution is supported
+- Macros are replaced in: `cmd`, `cmdStop`, `proxy`, `checkEndpoint`, `filters.stripParams`
+
+**Required Changes:**
+
+- Change `MacroList` type from `map[string]string` to `map[string]any`
+- Support scalar types: `string`, `int`, `float64`, `bool`
+- Implement type-preserving macro substitution:
+  - Direct macro usage (`key: ${macro}`) preserves the macro's type
+  - Interpolated usage (`key: "text ${macro}"`) converts to string
+- Add validation to ensure macro values are scalar types only
+- Update existing macro substitution logic in [proxy/config/config.go](proxy/config/config.go) to handle `any` types
+
+**Implementation Details:**
+
+- Create a generic helper function to perform macro substitution that:
+  - Takes a value of type `any`
+  - Recursively processes maps, slices, and scalar values
+  - Replaces `${macro_name}` patterns with macro values
+  - Preserves types for direct substitution
+  - Converts to strings for interpolated substitution
+- Update `validateMacro()` function to accept `any` type and validate scalar types
+- Maintain backward compatibility with existing string-only macros
+
+### 2. Metadata Field in ModelConfig
+
+**Location:** [proxy/config/model_config.go](proxy/config/model_config.go)
+
+**Required Changes:**
+
+- Add `Metadata map[string]any` field to `ModelConfig` struct
+- Support YAML unmarshaling of arbitrary structures (maps, arrays, scalars)
+- Apply macro substitution to metadata values during config loading
+
+**Schema Requirements:**
+
+- Metadata is optional (default: empty/nil map)
+- Supports nested structures (objects within objects, arrays, etc.)
+- All string values within metadata undergo macro substitution
+- Type preservation rules apply as described above
+
+### 3. Macro Substitution in Metadata
+
+**Location:** [proxy/config/config.go](proxy/config/config.go) in `LoadConfigFromReader()`
+
+**Process Flow:**
+
+1. After loading YAML configuration
+2. After model-level and global macro merging
+3. Apply macro substitution to `ModelConfig.Metadata` field
+4. Use the same merged macros available to `cmd`, `proxy`, etc.
+5. Process recursively through all nested structures
+
+**Substitution Rules:**
+
+- `port: ${PORT}` → keeps integer type from PORT macro
+- `temperature: ${temp}` → keeps float type from temp macro
+- `note: "Running on ${PORT}"` → converts to string `"Running on 10001"`
+- Arrays and nested objects are processed recursively
+- Unknown macros should cause configuration load error (consistent with existing behavior)
+
+### 4. API Response Updates
+
+**Location:** [proxy/proxymanager.go:350](proxy/proxymanager.go#L350) `listModelsHandler()`
+
+**Current Behavior:**
+
+- Returns model records with: `id`, `object`, `created`, `owned_by`
+- Optionally includes: `name`, `description`
+
+**Required Changes:**
+
+- Add metadata to each model record under the key `llamaswap_meta`
+- Only include `llamaswap_meta` if metadata is non-empty
+- Preserve all types when marshaling to JSON
+- Maintain existing sorting by model ID
+
+**Example Response:**
+
+```json
+{
+  "object": "list",
+  "data": [
+    {
+      "id": "llama",
+      "object": "model",
+      "created": 1234567890,
+      "owned_by": "llama-swap",
+      "name": "llama 3.1 8B",
+      "description": "A small but capable model",
+      "llamaswap_meta": {
+        "port": 10001,
+        "temperature": 0.7,
+        "note": "The llama is running on port 10001 temp=0.7, context=16384",
+        "a_list": [1, 1.23, "macros are OK in list and dictionary types: llama"],
+        "an_obj": {
+          "a": "1",
+          "b": 2,
+          "c": [0.7, false, "model: llama"]
+        }
+      }
+    }
+  ]
+}
+```
+
+### 5. Validation and Error Handling
+
+**Macro Validation:**
+
+- Extend `validateMacro()` to accept values of type `any`
+- Verify macro values are scalar types: `string`, `int`, `float64`, `bool`
+- Reject complex types (maps, slices, structs) as macro values
+- Maintain existing validation for macro names and lengths
+
+**Configuration Loading:**
+
+- Fail fast if unknown macros are found in metadata
+- Provide clear error messages indicating which model and field contains errors
+- Ensure macros in metadata follow same rules as macros in cmd/proxy fields
+
+## Testing Plan
+
+### Test 1: Model-Level Macros with Different Types
+
+**File:** [proxy/config/model_config_test.go](proxy/config/model_config_test.go)
+
+**Test Cases:**
+
+- Define model with macros of each scalar type
+- Verify metadata correctly substitutes and preserves types
+- Test direct substitution (`port: ${PORT}`)
+- Test string interpolation (`note: "Port is ${PORT}"`)
+- Verify nested objects and arrays work correctly
+
+### Test 2: Global and Model Macro Precedence
+
+**File:** [proxy/config/config_test.go](proxy/config/config_test.go)
+
+**Test Cases:**
+
+- Define same macro at global and model level with different types
+- Verify model-level macro takes precedence
+- Test metadata uses correct macro value
+- Verify type is preserved from the winning macro
+
+### Test 3: Macro Validation
+
+**File:** [proxy/config/config_test.go](proxy/config/config_test.go)
+
+**Test Cases:**
+
+- Test that complex types (maps, arrays) are rejected as macro values
+  - Verify error message includes: macro name and type that was rejected
+- Test that scalar types (string, int, float, bool) are accepted
+  - Each type should load without error
+- Test macro name validation still works with `any` types
+  - Invalid characters, reserved names, length limits should still be enforced
+
+### Test 4: Metadata in API Response
+
+**File:** [proxy/proxymanager_test.go](proxy/proxymanager_test.go)
+
+**Existing Test:** `TestProxyManager_ListModelsHandler`
+
+**Test Cases:**
+
+- Model with metadata → verify `llamaswap_meta` key appears
+- Model without metadata → verify `llamaswap_meta` key is absent
+- Verify all types are correctly marshaled to JSON
+- Verify nested structures are preserved
+- Verify macro substitution has occurred before serialization
+
+### Test 5: Unknown Macros in Metadata
+
+**File:** [proxy/config/config_test.go](proxy/config/config_test.go)
+
+**Test Cases:**
+
+- Use undefined macro in metadata
+- Verify configuration loading fails with clear error
+- Error should indicate model name and that macro is undefined
+
+### Test 6: Recursive Substitution
+
+**File:** [proxy/config/config_test.go](proxy/config/config_test.go)
+
+**Test Cases:**
+
+- Metadata with deeply nested structures
+- Arrays containing objects with macros
+- Objects containing arrays with macros
+- Mixed string interpolation and direct substitution at various nesting levels
+
+## Checklist
+
+### Configuration Schema Changes
+
+- [x] Change `MacroList` type from `map[string]string` to `map[string]any` in [proxy/config/config.go:19](proxy/config/config.go#L19)
+- [x] Add `Metadata map[string]any` field to `ModelConfig` struct in [proxy/config/model_config.go:37](proxy/config/model_config.go#L37)
+- [x] Update `validateMacro()` function signature to accept `any` type for values
+- [x] Add validation logic to ensure macro values are scalar types only
+
+### Macro Substitution Logic
+
+- [x] Create generic recursive function `substituteMetadataMacros()` to handle `any` types
+- [x] Implement type-preserving direct substitution logic
+- [x] Implement string interpolation with type conversion
+- [x] Handle maps: recursively process all values
+- [x] Handle slices: recursively process all elements
+- [x] Handle scalar types: perform string-based macro substitution if value is string
+- [x] Integrate macro substitution into `LoadConfigFromReader()` after existing macro expansion
+- [x] Update existing macro substitution calls to use merged macros with correct types
+
+### API Response Changes
+
+- [x] Modify `listModelsHandler()` in [proxy/proxymanager.go:350](proxy/proxymanager.go#L350)
+- [x] Add `llamaswap_meta` field to model records when metadata exists
+- [x] Ensure empty metadata results in omitted `llamaswap_meta` key
+- [x] Verify JSON marshaling preserves all types correctly
+
+### Testing - Config Package
+
+- [x] Add test for string macros in metadata: [proxy/config/config_test.go](proxy/config/config_test.go)
+- [x] Add test for int macros in metadata: [proxy/config/config_test.go](proxy/config/config_test.go)
+- [x] Add test for float macros in metadata: [proxy/config/config_test.go](proxy/config/config_test.go)
+- [x] Add test for bool macros in metadata: [proxy/config/config_test.go](proxy/config/config_test.go)
+- [x] Add test for string interpolation in metadata: [proxy/config/config_test.go](proxy/config/config_test.go)
+- [x] Add test for model-level macro precedence: [proxy/config/config_test.go](proxy/config/config_test.go)
+- [x] Add test for nested structures in metadata: [proxy/config/config_test.go](proxy/config/config_test.go)
+- [x] Add test for unknown macro in metadata (should error): [proxy/config/config_test.go](proxy/config/config_test.go)
+- [x] Add test for invalid macro type validation: [proxy/config/config_test.go](proxy/config/config_test.go)
+
+### Testing - Model Config Package
+
+- [x] Add test cases to [proxy/config/model_config_test.go](proxy/config/model_config_test.go) for metadata unmarshaling
+- [x] Test metadata with various scalar types
+- [x] Test metadata with nested objects and arrays
+
+### Testing - Proxy Manager
+
+- [x] Update `TestProxyManager_ListModelsHandler` in [proxy/proxymanager_test.go](proxy/proxymanager_test.go)
+- [x] Add test case for model with metadata
+- [x] Add test case for model without metadata
+- [x] Verify `llamaswap_meta` key presence/absence
+- [x] Verify type preservation in JSON output
+- [x] Verify macro substitution has occurred
+
+### Documentation
+
+- [x] Verify [config.example.yaml](config.example.yaml) already has complete metadata examples (lines 149-171)
+- [x] No additional documentation needed per project instructions
+
+## Known Issues and Considerations
+
+### Inconsistencies
+
+None identified. The plan references the correct existing example in [config.example.yaml:149-171](config.example.yaml#L149-L171).
+
+### Design Decisions
+
+1. **Why `llamaswap_meta` instead of merging into record?**
+
+   - Avoids potential collisions with OpenAI API standard fields
+   - Makes it clear this is llama-swap specific metadata
+   - Easier for clients to distinguish standard vs. custom fields
+
+2. **Why support nested structures?**
+
+   - Provides maximum flexibility for users
+   - Aligns with the schemaless design principle
+   - Example config already demonstrates this capability
+
+3. **Why validate macro types?**
+   - Prevents confusing behavior (e.g., substituting a map)
+   - Makes configuration errors explicit at load time
+   - Simpler implementation and testing
--- a/config.example.yaml
+++ b/config.example.yaml
@@ -67,7 +67,8 @@ models:
    # - macros defined here override macros defined in the global macros section
    # - model level macros follow the same rules as global macros
    macros:
-      "default_ctx": "16384"
+      "default_ctx": 16384
+      "temp": 0.7

    # cmd: the command to run to start the inference server.
    # - required
@@ -79,6 +80,7 @@ models:
      ${latest-llama}
      --model path/to/llama-8B-Q4_K_M.gguf
      --ctx-size ${default_ctx}
+      --temperature ${temp}

    # name: a display name for the model
    # - optional, default: empty string
@@ -144,6 +146,30 @@ models:
      # - recommended to stick to sampling parameters
      stripParams: "temperature, top_p, top_k"

+    # metadata: a dictionary of arbitrary values that are included in /v1/models
+    # - optional, default: empty dictionary
+    # - while metadata can contains complex types it is recommended to keep it simple
+    # - metadata is only passed through in /v1/models responses
+    metadata:
+      # port will remain an integer
+      port: ${PORT}
+
+      # the ${temp} macro will remain a float
+      temperature: ${temp}
+      note: "The ${MODEL_ID} is running on port ${PORT} temp=${temp}, context=${default_ctx}"
+
+      a_list:
+        - 1
+        - 1.23
+        - "macros are OK in list and dictionary types: ${MODEL_ID}"
+
+      an_obj:
+        a: "1"
+        b: 2
+        # objects can contain complex types with macro substitution
+        # becomes: c: [0.7, false, "model: llama"]
+        c: ["${temp}", false, "model: ${MODEL_ID}"]
+
    # concurrencyLimit: overrides the allowed number of active parallel requests to a model
    # - optional, default: 0
    # - useful for limiting the number of active parallel requests a model can process
--- a/proxy/config/config.go
+++ b/proxy/config/config.go
@@ -16,7 +16,7 @@ import (

 const DEFAULT_GROUP_ID = "(default)"

-type MacroList map[string]string
+type MacroList map[string]any

 type GroupConfig struct {
 	Swap       bool     `yaml:"swap"`
@@ -25,6 +25,11 @@ type GroupConfig struct {
 	Members    []string `yaml:"members"`
 }

+var (
+	macroNameRegex    = regexp.MustCompile(`^[a-zA-Z0-9_-]+$`)
+	macroPatternRegex = regexp.MustCompile(`\$\{([a-zA-Z0-9_-]+)\}`)
+)
+
 // set default values for GroupConfig
 func (c *GroupConfig) UnmarshalYAML(unmarshal func(interface{}) error) error {
 	type rawGroupConfig GroupConfig
@@ -182,14 +187,18 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			mergedMacros[k] = v
 		}

+		mergedMacros["MODEL_ID"] = modelId
+
 		// go through model config fields: cmd, cmdStop, proxy, checkEndPoint and replace macros with macro values
 		for macroName, macroValue := range mergedMacros {
 			macroSlug := fmt.Sprintf("${%s}", macroName)
-			modelConfig.Cmd = strings.ReplaceAll(modelConfig.Cmd, macroSlug, macroValue)
-			modelConfig.CmdStop = strings.ReplaceAll(modelConfig.CmdStop, macroSlug, macroValue)
-			modelConfig.Proxy = strings.ReplaceAll(modelConfig.Proxy, macroSlug, macroValue)
-			modelConfig.CheckEndpoint = strings.ReplaceAll(modelConfig.CheckEndpoint, macroSlug, macroValue)
-			modelConfig.Filters.StripParams = strings.ReplaceAll(modelConfig.Filters.StripParams, macroSlug, macroValue)
+			// Convert macro value to string for command/string field substitution
+			macroStr := fmt.Sprintf("%v", macroValue)
+			modelConfig.Cmd = strings.ReplaceAll(modelConfig.Cmd, macroSlug, macroStr)
+			modelConfig.CmdStop = strings.ReplaceAll(modelConfig.CmdStop, macroSlug, macroStr)
+			modelConfig.Proxy = strings.ReplaceAll(modelConfig.Proxy, macroSlug, macroStr)
+			modelConfig.CheckEndpoint = strings.ReplaceAll(modelConfig.CheckEndpoint, macroSlug, macroStr)
+			modelConfig.Filters.StripParams = strings.ReplaceAll(modelConfig.Filters.StripParams, macroSlug, macroStr)
 		}

 		// enforce ${PORT} used in both cmd and proxy
@@ -203,16 +212,14 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			modelConfig.Cmd = strings.ReplaceAll(modelConfig.Cmd, "${PORT}", nextPortStr)
 			modelConfig.CmdStop = strings.ReplaceAll(modelConfig.CmdStop, "${PORT}", nextPortStr)
 			modelConfig.Proxy = strings.ReplaceAll(modelConfig.Proxy, "${PORT}", nextPortStr)
+
+			// add port to merged macros so it can be used in metadata
+			mergedMacros["PORT"] = nextPort
+
 			nextPort++
 		}

-		if strings.Contains(modelConfig.Cmd, "${MODEL_ID}") || strings.Contains(modelConfig.CmdStop, "${MODEL_ID}") {
-			modelConfig.Cmd = strings.ReplaceAll(modelConfig.Cmd, "${MODEL_ID}", modelId)
-			modelConfig.CmdStop = strings.ReplaceAll(modelConfig.CmdStop, "${MODEL_ID}", modelId)
-		}
-
 		// make sure there are no unknown macros that have not been replaced
-		macroPattern := regexp.MustCompile(`\$\{([a-zA-Z0-9_-]+)\}`)
 		fieldMap := map[string]string{
 			"cmd":                 modelConfig.Cmd,
 			"cmdStop":             modelConfig.CmdStop,
@@ -222,7 +229,7 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 		}

 		for fieldName, fieldValue := range fieldMap {
-			matches := macroPattern.FindAllStringSubmatch(fieldValue, -1)
+			matches := macroPatternRegex.FindAllStringSubmatch(fieldValue, -1)
 			for _, match := range matches {
 				macroName := match[1]
 				if macroName == "PID" && fieldName == "cmdStop" {
@@ -234,6 +241,15 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			}
 		}

+		// Apply macro substitution to metadata
+		if len(modelConfig.Metadata) > 0 {
+			substitutedMetadata, err := substituteMetadataMacros(modelConfig.Metadata, mergedMacros)
+			if err != nil {
+				return Config{}, fmt.Errorf("model %s metadata: %s", modelId, err.Error())
+			}
+			modelConfig.Metadata = substitutedMetadata.(map[string]any)
+		}
+
 		config.Models[modelId] = modelConfig
 	}

@@ -296,7 +312,7 @@ func AddDefaultGroupToConfig(config Config) Config {
 		}
 	} else {
 		// iterate over existing group members and add non-grouped models into the default group
-		for modelName, _ := range config.Models {
+		for modelName := range config.Models {
 			foundModel := false
 		found:
 			// search for the model in existing groups
@@ -369,20 +385,25 @@ func StripComments(cmdStr string) string {
 	return strings.Join(cleanedLines, "\n")
 }

-var (
-	macroNameRegex = regexp.MustCompile(`^[a-zA-Z0-9_-]+$`)
-)
-
 // validateMacro validates macro name and value constraints
-func validateMacro(name, value string) error {
+func validateMacro(name string, value any) error {
 	if len(name) >= 64 {
 		return fmt.Errorf("macro name '%s' exceeds maximum length of 63 characters", name)
 	}
 	if !macroNameRegex.MatchString(name) {
 		return fmt.Errorf("macro name '%s' contains invalid characters, must match pattern ^[a-zA-Z0-9_-]+$", name)
 	}
-	if len(value) >= 1024 {
-		return fmt.Errorf("macro value for '%s' exceeds maximum length of 1024 characters", name)
+
+	// Validate that value is a scalar type
+	switch v := value.(type) {
+	case string:
+		if len(v) >= 1024 {
+			return fmt.Errorf("macro value for '%s' exceeds maximum length of 1024 characters", name)
+		}
+	case int, int8, int16, int32, int64, uint, uint8, uint16, uint32, uint64, float32, float64, bool:
+		// These types are allowed
+	default:
+		return fmt.Errorf("macro '%s' has invalid type %T, must be a scalar type (string, int, float, or bool)", name, value)
 	}

 	switch name {
@@ -392,3 +413,63 @@ func validateMacro(name, value string) error {

 	return nil
 }
+
+// substituteMetadataMacros recursively substitutes macros in metadata structures
+// Direct substitution (key: ${macro}) preserves the macro's type
+// Interpolated substitution (key: "text ${macro}") converts to string
+func substituteMetadataMacros(value any, macros MacroList) (any, error) {
+	switch v := value.(type) {
+	case string:
+		// Check if this is a direct macro substitution
+		if strings.HasPrefix(v, "${") && strings.HasSuffix(v, "}") && strings.Count(v, "${") == 1 {
+			macroName := v[2 : len(v)-1]
+			if macroValue, exists := macros[macroName]; exists {
+				return macroValue, nil
+			}
+			return nil, fmt.Errorf("unknown macro '${%s}' in metadata", macroName)
+		}
+
+		// Handle string interpolation
+		matches := macroPatternRegex.FindAllStringSubmatch(v, -1)
+		result := v
+		for _, match := range matches {
+			macroName := match[1]
+			macroValue, exists := macros[macroName]
+			if !exists {
+				return nil, fmt.Errorf("unknown macro '${%s}' in metadata", macroName)
+			}
+			// Convert macro value to string for interpolation
+			macroStr := fmt.Sprintf("%v", macroValue)
+			result = strings.ReplaceAll(result, match[0], macroStr)
+		}
+		return result, nil
+
+	case map[string]any:
+		// Recursively process map values
+		newMap := make(map[string]any)
+		for key, val := range v {
+			newVal, err := substituteMetadataMacros(val, macros)
+			if err != nil {
+				return nil, err
+			}
+			newMap[key] = newVal
+		}
+		return newMap, nil
+
+	case []any:
+		// Recursively process slice elements
+		newSlice := make([]any, len(v))
+		for i, val := range v {
+			newVal, err := substituteMetadataMacros(val, macros)
+			if err != nil {
+				return nil, err
+			}
+			newSlice[i] = newVal
+		}
+		return newSlice, nil
+
+	default:
+		// Return scalar types as-is
+		return value, nil
+	}
+}
--- a/proxy/config/config_posix_test.go
+++ b/proxy/config/config_posix_test.go
@@ -163,7 +163,7 @@ groups:
 	expected := Config{
 		LogLevel:  "info",
 		StartPort: 5800,
-		Macros: map[string]string{
+		Macros: MacroList{
 			"svr-path": "path/to/server",
 		},
 		Hooks: HooksConfig{
--- a/proxy/config/config_test.go
+++ b/proxy/config/config_test.go
@@ -517,3 +517,243 @@ models:
 	assert.NoError(t, err)
 	assert.Equal(t, "/path/to/server -p 9000 -hf author/model:F16", strings.Join(sanitizedCmd3, " "))
 }
+
+func TestConfig_TypedMacrosInMetadata(t *testing.T) {
+	content := `
+startPort: 10000
+macros:
+  PORT_NUM: 10001
+  TEMP: 0.7
+  ENABLED: true
+  NAME: "llama model"
+  CTX: 16384
+
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+    metadata:
+      port: ${PORT_NUM}
+      temperature: ${TEMP}
+      enabled: ${ENABLED}
+      model_name: ${NAME}
+      context: ${CTX}
+      note: "Running on port ${PORT_NUM} with temp ${TEMP} and context ${CTX}"
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.NoError(t, err)
+
+	meta := config.Models["test-model"].Metadata
+	assert.NotNil(t, meta)
+
+	// Verify direct substitution preserves types
+	assert.Equal(t, 10001, meta["port"])
+	assert.Equal(t, 0.7, meta["temperature"])
+	assert.Equal(t, true, meta["enabled"])
+	assert.Equal(t, "llama model", meta["model_name"])
+	assert.Equal(t, 16384, meta["context"])
+
+	// Verify string interpolation converts to string
+	assert.Equal(t, "Running on port 10001 with temp 0.7 and context 16384", meta["note"])
+}
+
+func TestConfig_NestedStructuresInMetadata(t *testing.T) {
+	content := `
+startPort: 10000
+macros:
+  TEMP: 0.7
+
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+    metadata:
+      config:
+        port: ${PORT}
+        temperature: ${TEMP}
+      tags: ["model:${MODEL_ID}", "port:${PORT}"]
+      nested:
+        deep:
+          value: ${TEMP}
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.NoError(t, err)
+
+	meta := config.Models["test-model"].Metadata
+	assert.NotNil(t, meta)
+
+	// Verify nested objects
+	configMap := meta["config"].(map[string]any)
+	assert.Equal(t, 10000, configMap["port"])
+	assert.Equal(t, 0.7, configMap["temperature"])
+
+	// Verify arrays
+	tags := meta["tags"].([]any)
+	assert.Equal(t, "model:test-model", tags[0])
+	assert.Equal(t, "port:10000", tags[1])
+
+	// Verify deeply nested structures
+	nested := meta["nested"].(map[string]any)
+	deep := nested["deep"].(map[string]any)
+	assert.Equal(t, 0.7, deep["value"])
+}
+
+func TestConfig_ModelLevelMacroPrecedenceInMetadata(t *testing.T) {
+	content := `
+startPort: 10000
+macros:
+  TEMP: 0.5
+  GLOBAL_VAL: "global"
+
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+    macros:
+      TEMP: 0.9
+      LOCAL_VAL: "local"
+    metadata:
+      temperature: ${TEMP}
+      global: ${GLOBAL_VAL}
+      local: ${LOCAL_VAL}
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.NoError(t, err)
+
+	meta := config.Models["test-model"].Metadata
+	assert.NotNil(t, meta)
+
+	// Model-level macro should override global
+	assert.Equal(t, 0.9, meta["temperature"])
+	// Global macro should be accessible
+	assert.Equal(t, "global", meta["global"])
+	// Model-level macro should be accessible
+	assert.Equal(t, "local", meta["local"])
+}
+
+func TestConfig_UnknownMacroInMetadata(t *testing.T) {
+	content := `
+startPort: 10000
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+    metadata:
+      value: ${UNKNOWN_MACRO}
+`
+
+	_, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.Error(t, err)
+	assert.Contains(t, err.Error(), "test-model")
+	assert.Contains(t, err.Error(), "UNKNOWN_MACRO")
+}
+
+func TestConfig_InvalidMacroType(t *testing.T) {
+	content := `
+startPort: 10000
+macros:
+  INVALID:
+    nested: value
+
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+`
+
+	_, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.Error(t, err)
+	assert.Contains(t, err.Error(), "INVALID")
+	assert.Contains(t, err.Error(), "must be a scalar type")
+}
+
+func TestConfig_MacroTypeValidation(t *testing.T) {
+	tests := []struct {
+		name      string
+		yaml      string
+		shouldErr bool
+	}{
+		{
+			name: "string macro",
+			yaml: `
+startPort: 10000
+macros:
+  STR: "test"
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+`,
+			shouldErr: false,
+		},
+		{
+			name: "int macro",
+			yaml: `
+startPort: 10000
+macros:
+  NUM: 42
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+`,
+			shouldErr: false,
+		},
+		{
+			name: "float macro",
+			yaml: `
+startPort: 10000
+macros:
+  FLOAT: 3.14
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+`,
+			shouldErr: false,
+		},
+		{
+			name: "bool macro",
+			yaml: `
+startPort: 10000
+macros:
+  BOOL: true
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+`,
+			shouldErr: false,
+		},
+		{
+			name: "array macro (invalid)",
+			yaml: `
+startPort: 10000
+macros:
+  ARR: [1, 2, 3]
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+`,
+			shouldErr: true,
+		},
+		{
+			name: "map macro (invalid)",
+			yaml: `
+startPort: 10000
+macros:
+  MAP:
+    key: value
+models:
+  test-model:
+    cmd: /path/to/server -p ${PORT}
+`,
+			shouldErr: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			_, err := LoadConfigFromReader(strings.NewReader(tt.yaml))
+			if tt.shouldErr {
+				assert.Error(t, err)
+			} else {
+				assert.NoError(t, err)
+			}
+		})
+	}
+}
--- a/proxy/config/config_windows_test.go
+++ b/proxy/config/config_windows_test.go
@@ -155,7 +155,7 @@ groups:
 	expected := Config{
 		LogLevel:  "info",
 		StartPort: 5800,
-		Macros: map[string]string{
+		Macros: MacroList{
 			"svr-path": "path/to/server",
 		},
 		Models: map[string]ModelConfig{
--- a/proxy/config/model_config.go
+++ b/proxy/config/model_config.go
@@ -31,6 +31,10 @@ type ModelConfig struct {
 	// Macros: see #264
 	// Model level macros take precedence over the global macros
 	Macros MacroList `yaml:"macros"`
+
+	// Metadata: see #264
+	// Arbitrary metadata that can be exposed through the API
+	Metadata map[string]any `yaml:"metadata"`
 }

 func (m *ModelConfig) UnmarshalYAML(unmarshal func(interface{}) error) error {
--- a/proxy/proxymanager.go
+++ b/proxy/proxymanager.go
@@ -370,6 +370,13 @@ func (pm *ProxyManager) listModelsHandler(c *gin.Context) {
 			record["description"] = desc
 		}

+		// Add metadata if present
+		if len(modelConfig.Metadata) > 0 {
+			record["meta"] = gin.H{
+				"llamaswap": modelConfig.Metadata,
+			}
+		}
+
 		data = append(data, record)
 	}

--- a/proxy/proxymanager_test.go
+++ b/proxy/proxymanager_test.go
@@ -282,6 +282,90 @@ func TestProxyManager_ListModelsHandler(t *testing.T) {
 	assert.Empty(t, expectedModels, "not all expected models were returned")
 }

+func TestProxyManager_ListModelsHandler_WithMetadata(t *testing.T) {
+	// Process config through LoadConfigFromReader to apply macro substitution
+	configYaml := `
+healthCheckTimeout: 15
+logLevel: error
+startPort: 10000
+models:
+  model1:
+    cmd: /path/to/server -p ${PORT}
+    macros:
+      PORT_NUM: 10001
+      TEMP: 0.7
+      NAME: "llama"
+    metadata:
+      port: ${PORT_NUM}
+      temperature: ${TEMP}
+      enabled: true
+      note: "Running on port ${PORT_NUM}"
+      nested:
+        value: ${TEMP}
+  model2:
+    cmd: /path/to/server -p ${PORT}
+`
+	processedConfig, err := config.LoadConfigFromReader(strings.NewReader(configYaml))
+	assert.NoError(t, err)
+
+	proxy := New(processedConfig)
+
+	req := httptest.NewRequest("GET", "/v1/models", nil)
+	w := httptest.NewRecorder()
+	proxy.ServeHTTP(w, req)
+
+	assert.Equal(t, http.StatusOK, w.Code)
+
+	var response struct {
+		Data []map[string]any `json:"data"`
+	}
+
+	err = json.Unmarshal(w.Body.Bytes(), &response)
+	assert.NoError(t, err)
+	assert.Len(t, response.Data, 2)
+
+	// Find model1 and model2 in response
+	var model1Data, model2Data map[string]any
+	for _, model := range response.Data {
+		if model["id"] == "model1" {
+			model1Data = model
+		} else if model["id"] == "model2" {
+			model2Data = model
+		}
+	}
+
+	// Verify model1 has llamaswap_meta
+	assert.NotNil(t, model1Data)
+	meta, exists := model1Data["meta"]
+	if !assert.True(t, exists, "model1 should have meta key") {
+		t.FailNow()
+	}
+
+	metaMap := meta.(map[string]any)
+
+	lsmeta, exists := metaMap["llamaswap"]
+	if !assert.True(t, exists, "model1 should have meta.llamaswap key") {
+		t.FailNow()
+	}
+
+	lsmetamap := lsmeta.(map[string]any)
+
+	// Verify type preservation
+	assert.Equal(t, float64(10001), lsmetamap["port"]) // JSON numbers are float64
+	assert.Equal(t, 0.7, lsmetamap["temperature"])
+	assert.Equal(t, true, lsmetamap["enabled"])
+	// Verify string interpolation
+	assert.Equal(t, "Running on port 10001", lsmetamap["note"])
+	// Verify nested structure
+	nested := lsmetamap["nested"].(map[string]any)
+	assert.Equal(t, 0.7, nested["value"])
+
+	// Verify model2 does NOT have llamaswap_meta
+	assert.NotNil(t, model2Data)
+	_, exists = model2Data["llamaswap_meta"]
+	assert.False(t, exists, "model2 should not have llamaswap_meta")
+}
+
 func TestProxyManager_ListModelsHandler_SortedByID(t *testing.T) {
 	// Intentionally add models in non-sorted order and with an unlisted model
 	config := config.Config{