merge: sync upstream/main into feat/multi-agent-routing

Resolve conflicts:
- pkg/agent/loop.go: integrate context compression, command handling,
  utf8 token estimation, and summarization notification into
  multi-agent routing architecture
- pkg/config/config_test.go: merge imports from both branches
- pkg/agent/loop_test.go: update test to use registry-based sessions

pkg/providers/claude_cli_provider.go

@@ -171,68 +171,14 @@ func (p *ClaudeCliProvider) parseClaudeCliResponse(output string) (*LLMResponse,
 	}, nil
 }
 
-// extractToolCalls parses tool call JSON from the response text.
+// extractToolCalls delegates to the shared extractToolCallsFromText function.
 func (p *ClaudeCliProvider) extractToolCalls(text string) []ToolCall {
-	start := strings.Index(text, `{"tool_calls"`)
-	if start == -1 {
-		return nil
-	}
-
-	end := findMatchingBrace(text, start)
-	if end == start {
-		return nil
-	}
-
-	jsonStr := text[start:end]
-
-	var wrapper struct {
-		ToolCalls []struct {
-			ID       string `json:"id"`
-			Type     string `json:"type"`
-			Function struct {
-				Name      string `json:"name"`
-				Arguments string `json:"arguments"`
-			} `json:"function"`
-		} `json:"tool_calls"`
-	}
-
-	if err := json.Unmarshal([]byte(jsonStr), &wrapper); err != nil {
-		return nil
-	}
-
-	var result []ToolCall
-	for _, tc := range wrapper.ToolCalls {
-		var args map[string]interface{}
-		json.Unmarshal([]byte(tc.Function.Arguments), &args)
-
-		result = append(result, ToolCall{
-			ID:        tc.ID,
-			Type:      tc.Type,
-			Name:      tc.Function.Name,
-			Arguments: args,
-			Function: &FunctionCall{
-				Name:      tc.Function.Name,
-				Arguments: tc.Function.Arguments,
-			},
-		})
-	}
-
-	return result
+	return extractToolCallsFromText(text)
 }
 
-// stripToolCallsJSON removes tool call JSON from response text.
+// stripToolCallsJSON delegates to the shared stripToolCallsFromText function.
 func (p *ClaudeCliProvider) stripToolCallsJSON(text string) string {
-	start := strings.Index(text, `{"tool_calls"`)
-	if start == -1 {
-		return text
-	}
-
-	end := findMatchingBrace(text, start)
-	if end == start {
-		return text
-	}
-
-	return strings.TrimSpace(text[:start] + text[end:])
+	return stripToolCallsFromText(text)
 }
 
 // findMatchingBrace finds the index after the closing brace matching the opening brace at pos.

pkg/providers/codex_cli_credentials.go

@@ -0,0 +1,79 @@
+package providers
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+)
+
+// CodexCliAuth represents the ~/.codex/auth.json file structure.
+type CodexCliAuth struct {
+	Tokens struct {
+		AccessToken  string `json:"access_token"`
+		RefreshToken string `json:"refresh_token"`
+		AccountID    string `json:"account_id"`
+	} `json:"tokens"`
+}
+
+// ReadCodexCliCredentials reads OAuth tokens from the Codex CLI's auth.json file.
+// Expiry is estimated as file modification time + 1 hour (same approach as moltbot).
+func ReadCodexCliCredentials() (accessToken, accountID string, expiresAt time.Time, err error) {
+	authPath, err := resolveCodexAuthPath()
+	if err != nil {
+		return "", "", time.Time{}, err
+	}
+
+	data, err := os.ReadFile(authPath)
+	if err != nil {
+		return "", "", time.Time{}, fmt.Errorf("reading %s: %w", authPath, err)
+	}
+
+	var auth CodexCliAuth
+	if err := json.Unmarshal(data, &auth); err != nil {
+		return "", "", time.Time{}, fmt.Errorf("parsing %s: %w", authPath, err)
+	}
+
+	if auth.Tokens.AccessToken == "" {
+		return "", "", time.Time{}, fmt.Errorf("no access_token in %s", authPath)
+	}
+
+	stat, err := os.Stat(authPath)
+	if err != nil {
+		expiresAt = time.Now().Add(time.Hour)
+	} else {
+		expiresAt = stat.ModTime().Add(time.Hour)
+	}
+
+	return auth.Tokens.AccessToken, auth.Tokens.AccountID, expiresAt, nil
+}
+
+// CreateCodexCliTokenSource creates a token source that reads from ~/.codex/auth.json.
+// This allows the existing CodexProvider to reuse Codex CLI credentials.
+func CreateCodexCliTokenSource() func() (string, string, error) {
+	return func() (string, string, error) {
+		token, accountID, expiresAt, err := ReadCodexCliCredentials()
+		if err != nil {
+			return "", "", fmt.Errorf("reading codex cli credentials: %w", err)
+		}
+
+		if time.Now().After(expiresAt) {
+			return "", "", fmt.Errorf("codex cli credentials expired (auth.json last modified > 1h ago). Run: codex login")
+		}
+
+		return token, accountID, nil
+	}
+}
+
+func resolveCodexAuthPath() (string, error) {
+	codexHome := os.Getenv("CODEX_HOME")
+	if codexHome == "" {
+		home, err := os.UserHomeDir()
+		if err != nil {
+			return "", fmt.Errorf("getting home dir: %w", err)
+		}
+		codexHome = filepath.Join(home, ".codex")
+	}
+	return filepath.Join(codexHome, "auth.json"), nil
+}

pkg/providers/codex_cli_credentials_test.go

@@ -0,0 +1,181 @@
+package providers
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+func TestReadCodexCliCredentials_Valid(t *testing.T) {
+	tmpDir := t.TempDir()
+	authPath := filepath.Join(tmpDir, "auth.json")
+
+	authJSON := `{
+		"tokens": {
+			"access_token": "test-access-token",
+			"refresh_token": "test-refresh-token",
+			"account_id": "org-test123"
+		}
+	}`
+	if err := os.WriteFile(authPath, []byte(authJSON), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("CODEX_HOME", tmpDir)
+
+	token, accountID, expiresAt, err := ReadCodexCliCredentials()
+	if err != nil {
+		t.Fatalf("ReadCodexCliCredentials() error: %v", err)
+	}
+	if token != "test-access-token" {
+		t.Errorf("token = %q, want %q", token, "test-access-token")
+	}
+	if accountID != "org-test123" {
+		t.Errorf("accountID = %q, want %q", accountID, "org-test123")
+	}
+	// Expiry should be within ~1 hour from now (file was just written)
+	if expiresAt.Before(time.Now()) {
+		t.Errorf("expiresAt = %v, should be in the future", expiresAt)
+	}
+	if expiresAt.After(time.Now().Add(2 * time.Hour)) {
+		t.Errorf("expiresAt = %v, should be within ~1 hour", expiresAt)
+	}
+}
+
+func TestReadCodexCliCredentials_MissingFile(t *testing.T) {
+	tmpDir := t.TempDir()
+	t.Setenv("CODEX_HOME", tmpDir)
+
+	_, _, _, err := ReadCodexCliCredentials()
+	if err == nil {
+		t.Fatal("expected error for missing auth.json")
+	}
+}
+
+func TestReadCodexCliCredentials_EmptyToken(t *testing.T) {
+	tmpDir := t.TempDir()
+	authPath := filepath.Join(tmpDir, "auth.json")
+
+	authJSON := `{"tokens": {"access_token": "", "refresh_token": "r", "account_id": "a"}}`
+	if err := os.WriteFile(authPath, []byte(authJSON), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("CODEX_HOME", tmpDir)
+
+	_, _, _, err := ReadCodexCliCredentials()
+	if err == nil {
+		t.Fatal("expected error for empty access_token")
+	}
+}
+
+func TestReadCodexCliCredentials_InvalidJSON(t *testing.T) {
+	tmpDir := t.TempDir()
+	authPath := filepath.Join(tmpDir, "auth.json")
+
+	if err := os.WriteFile(authPath, []byte("not json"), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("CODEX_HOME", tmpDir)
+
+	_, _, _, err := ReadCodexCliCredentials()
+	if err == nil {
+		t.Fatal("expected error for invalid JSON")
+	}
+}
+
+func TestReadCodexCliCredentials_NoAccountID(t *testing.T) {
+	tmpDir := t.TempDir()
+	authPath := filepath.Join(tmpDir, "auth.json")
+
+	authJSON := `{"tokens": {"access_token": "tok123", "refresh_token": "ref456"}}`
+	if err := os.WriteFile(authPath, []byte(authJSON), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("CODEX_HOME", tmpDir)
+
+	token, accountID, _, err := ReadCodexCliCredentials()
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if token != "tok123" {
+		t.Errorf("token = %q, want %q", token, "tok123")
+	}
+	if accountID != "" {
+		t.Errorf("accountID = %q, want empty", accountID)
+	}
+}
+
+func TestReadCodexCliCredentials_CodexHomeEnv(t *testing.T) {
+	tmpDir := t.TempDir()
+	customDir := filepath.Join(tmpDir, "custom-codex")
+	if err := os.MkdirAll(customDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+
+	authJSON := `{"tokens": {"access_token": "custom-token", "refresh_token": "r"}}`
+	if err := os.WriteFile(filepath.Join(customDir, "auth.json"), []byte(authJSON), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("CODEX_HOME", customDir)
+
+	token, _, _, err := ReadCodexCliCredentials()
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if token != "custom-token" {
+		t.Errorf("token = %q, want %q", token, "custom-token")
+	}
+}
+
+func TestCreateCodexCliTokenSource_Valid(t *testing.T) {
+	tmpDir := t.TempDir()
+	authPath := filepath.Join(tmpDir, "auth.json")
+
+	authJSON := `{"tokens": {"access_token": "fresh-token", "refresh_token": "r", "account_id": "acc"}}`
+	if err := os.WriteFile(authPath, []byte(authJSON), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("CODEX_HOME", tmpDir)
+
+	source := CreateCodexCliTokenSource()
+	token, accountID, err := source()
+	if err != nil {
+		t.Fatalf("token source error: %v", err)
+	}
+	if token != "fresh-token" {
+		t.Errorf("token = %q, want %q", token, "fresh-token")
+	}
+	if accountID != "acc" {
+		t.Errorf("accountID = %q, want %q", accountID, "acc")
+	}
+}
+
+func TestCreateCodexCliTokenSource_Expired(t *testing.T) {
+	tmpDir := t.TempDir()
+	authPath := filepath.Join(tmpDir, "auth.json")
+
+	authJSON := `{"tokens": {"access_token": "old-token", "refresh_token": "r"}}`
+	if err := os.WriteFile(authPath, []byte(authJSON), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	// Set file modification time to 2 hours ago
+	oldTime := time.Now().Add(-2 * time.Hour)
+	if err := os.Chtimes(authPath, oldTime, oldTime); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("CODEX_HOME", tmpDir)
+
+	source := CreateCodexCliTokenSource()
+	_, _, err := source()
+	if err == nil {
+		t.Fatal("expected error for expired credentials")
+	}
+}

pkg/providers/codex_cli_provider.go

@@ -0,0 +1,251 @@
+package providers
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"os/exec"
+	"strings"
+)
+
+// CodexCliProvider implements LLMProvider by wrapping the codex CLI as a subprocess.
+type CodexCliProvider struct {
+	command   string
+	workspace string
+}
+
+// NewCodexCliProvider creates a new Codex CLI provider.
+func NewCodexCliProvider(workspace string) *CodexCliProvider {
+	return &CodexCliProvider{
+		command:   "codex",
+		workspace: workspace,
+	}
+}
+
+// Chat implements LLMProvider.Chat by executing the codex CLI in non-interactive mode.
+func (p *CodexCliProvider) Chat(ctx context.Context, messages []Message, tools []ToolDefinition, model string, options map[string]interface{}) (*LLMResponse, error) {
+	if p.command == "" {
+		return nil, fmt.Errorf("codex command not configured")
+	}
+
+	prompt := p.buildPrompt(messages, tools)
+
+	args := []string{
+		"exec",
+		"--json",
+		"--dangerously-bypass-approvals-and-sandbox",
+		"--skip-git-repo-check",
+		"--color", "never",
+	}
+	if model != "" && model != "codex-cli" {
+		args = append(args, "-m", model)
+	}
+	if p.workspace != "" {
+		args = append(args, "-C", p.workspace)
+	}
+	args = append(args, "-") // read prompt from stdin
+
+	cmd := exec.CommandContext(ctx, p.command, args...)
+	cmd.Stdin = bytes.NewReader([]byte(prompt))
+
+	var stdout, stderr bytes.Buffer
+	cmd.Stdout = &stdout
+	cmd.Stderr = &stderr
+
+	err := cmd.Run()
+
+	// Parse JSONL from stdout even if exit code is non-zero,
+	// because codex writes diagnostic noise to stderr (e.g. rollout errors)
+	// but still produces valid JSONL output.
+	if stdoutStr := stdout.String(); stdoutStr != "" {
+		resp, parseErr := p.parseJSONLEvents(stdoutStr)
+		if parseErr == nil && resp != nil && (resp.Content != "" || len(resp.ToolCalls) > 0) {
+			return resp, nil
+		}
+	}
+
+	if err != nil {
+		if ctx.Err() == context.Canceled {
+			return nil, ctx.Err()
+		}
+		if stderrStr := stderr.String(); stderrStr != "" {
+			return nil, fmt.Errorf("codex cli error: %s", stderrStr)
+		}
+		return nil, fmt.Errorf("codex cli error: %w", err)
+	}
+
+	return p.parseJSONLEvents(stdout.String())
+}
+
+// GetDefaultModel returns the default model identifier.
+func (p *CodexCliProvider) GetDefaultModel() string {
+	return "codex-cli"
+}
+
+// buildPrompt converts messages to a prompt string for the Codex CLI.
+// System messages are prepended as instructions since Codex CLI has no --system-prompt flag.
+func (p *CodexCliProvider) buildPrompt(messages []Message, tools []ToolDefinition) string {
+	var systemParts []string
+	var conversationParts []string
+
+	for _, msg := range messages {
+		switch msg.Role {
+		case "system":
+			systemParts = append(systemParts, msg.Content)
+		case "user":
+			conversationParts = append(conversationParts, msg.Content)
+		case "assistant":
+			conversationParts = append(conversationParts, "Assistant: "+msg.Content)
+		case "tool":
+			conversationParts = append(conversationParts,
+				fmt.Sprintf("[Tool Result for %s]: %s", msg.ToolCallID, msg.Content))
+		}
+	}
+
+	var sb strings.Builder
+
+	if len(systemParts) > 0 {
+		sb.WriteString("## System Instructions\n\n")
+		sb.WriteString(strings.Join(systemParts, "\n\n"))
+		sb.WriteString("\n\n## Task\n\n")
+	}
+
+	if len(tools) > 0 {
+		sb.WriteString(p.buildToolsPrompt(tools))
+		sb.WriteString("\n\n")
+	}
+
+	// Simplify single user message (no prefix)
+	if len(conversationParts) == 1 && len(systemParts) == 0 && len(tools) == 0 {
+		return conversationParts[0]
+	}
+
+	sb.WriteString(strings.Join(conversationParts, "\n"))
+	return sb.String()
+}
+
+// buildToolsPrompt creates a tool definitions section for the prompt.
+func (p *CodexCliProvider) buildToolsPrompt(tools []ToolDefinition) string {
+	var sb strings.Builder
+
+	sb.WriteString("## Available Tools\n\n")
+	sb.WriteString("When you need to use a tool, respond with ONLY a JSON object:\n\n")
+	sb.WriteString("```json\n")
+	sb.WriteString(`{"tool_calls":[{"id":"call_xxx","type":"function","function":{"name":"tool_name","arguments":"{...}"}}]}`)
+	sb.WriteString("\n```\n\n")
+	sb.WriteString("CRITICAL: The 'arguments' field MUST be a JSON-encoded STRING.\n\n")
+	sb.WriteString("### Tool Definitions:\n\n")
+
+	for _, tool := range tools {
+		if tool.Type != "function" {
+			continue
+		}
+		sb.WriteString(fmt.Sprintf("#### %s\n", tool.Function.Name))
+		if tool.Function.Description != "" {
+			sb.WriteString(fmt.Sprintf("Description: %s\n", tool.Function.Description))
+		}
+		if len(tool.Function.Parameters) > 0 {
+			paramsJSON, _ := json.Marshal(tool.Function.Parameters)
+			sb.WriteString(fmt.Sprintf("Parameters:\n```json\n%s\n```\n", string(paramsJSON)))
+		}
+		sb.WriteString("\n")
+	}
+
+	return sb.String()
+}
+
+// codexEvent represents a single JSONL event from `codex exec --json`.
+type codexEvent struct {
+	Type     string          `json:"type"`
+	ThreadID string          `json:"thread_id,omitempty"`
+	Message  string          `json:"message,omitempty"`
+	Item     *codexEventItem `json:"item,omitempty"`
+	Usage    *codexUsage     `json:"usage,omitempty"`
+	Error    *codexEventErr  `json:"error,omitempty"`
+}
+
+type codexEventItem struct {
+	ID       string `json:"id"`
+	Type     string `json:"type"`
+	Text     string `json:"text,omitempty"`
+	Command  string `json:"command,omitempty"`
+	Status   string `json:"status,omitempty"`
+	ExitCode *int   `json:"exit_code,omitempty"`
+	Output   string `json:"output,omitempty"`
+}
+
+type codexUsage struct {
+	InputTokens       int `json:"input_tokens"`
+	CachedInputTokens int `json:"cached_input_tokens"`
+	OutputTokens      int `json:"output_tokens"`
+}
+
+type codexEventErr struct {
+	Message string `json:"message"`
+}
+
+// parseJSONLEvents processes the JSONL output from codex exec --json.
+func (p *CodexCliProvider) parseJSONLEvents(output string) (*LLMResponse, error) {
+	var contentParts []string
+	var usage *UsageInfo
+	var lastError string
+
+	scanner := bufio.NewScanner(strings.NewReader(output))
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if line == "" {
+			continue
+		}
+
+		var event codexEvent
+		if err := json.Unmarshal([]byte(line), &event); err != nil {
+			continue // skip malformed lines
+		}
+
+		switch event.Type {
+		case "item.completed":
+			if event.Item != nil && event.Item.Type == "agent_message" && event.Item.Text != "" {
+				contentParts = append(contentParts, event.Item.Text)
+			}
+		case "turn.completed":
+			if event.Usage != nil {
+				promptTokens := event.Usage.InputTokens + event.Usage.CachedInputTokens
+				usage = &UsageInfo{
+					PromptTokens:     promptTokens,
+					CompletionTokens: event.Usage.OutputTokens,
+					TotalTokens:      promptTokens + event.Usage.OutputTokens,
+				}
+			}
+		case "error":
+			lastError = event.Message
+		case "turn.failed":
+			if event.Error != nil {
+				lastError = event.Error.Message
+			}
+		}
+	}
+
+	if lastError != "" && len(contentParts) == 0 {
+		return nil, fmt.Errorf("codex cli: %s", lastError)
+	}
+
+	content := strings.Join(contentParts, "\n")
+
+	// Extract tool calls from response text (same pattern as ClaudeCliProvider)
+	toolCalls := extractToolCallsFromText(content)
+
+	finishReason := "stop"
+	if len(toolCalls) > 0 {
+		finishReason = "tool_calls"
+		content = stripToolCallsFromText(content)
+	}
+
+	return &LLMResponse{
+		Content:      strings.TrimSpace(content),
+		ToolCalls:    toolCalls,
+		FinishReason: finishReason,
+		Usage:        usage,
+	}, nil
+}

pkg/providers/codex_cli_provider_test.go

@@ -0,0 +1,585 @@
+package providers
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// --- JSONL Event Parsing Tests ---
+
+func TestParseJSONLEvents_AgentMessage(t *testing.T) {
+	p := &CodexCliProvider{}
+	events := `{"type":"thread.started","thread_id":"abc-123"}
+{"type":"turn.started"}
+{"type":"item.completed","item":{"id":"item_1","type":"agent_message","text":"Hello from Codex!"}}
+{"type":"turn.completed","usage":{"input_tokens":100,"cached_input_tokens":50,"output_tokens":20}}`
+
+	resp, err := p.parseJSONLEvents(events)
+	if err != nil {
+		t.Fatalf("parseJSONLEvents() error: %v", err)
+	}
+	if resp.Content != "Hello from Codex!" {
+		t.Errorf("Content = %q, want %q", resp.Content, "Hello from Codex!")
+	}
+	if resp.FinishReason != "stop" {
+		t.Errorf("FinishReason = %q, want %q", resp.FinishReason, "stop")
+	}
+	if resp.Usage == nil {
+		t.Fatal("Usage should not be nil")
+	}
+	if resp.Usage.PromptTokens != 150 {
+		t.Errorf("PromptTokens = %d, want 150", resp.Usage.PromptTokens)
+	}
+	if resp.Usage.CompletionTokens != 20 {
+		t.Errorf("CompletionTokens = %d, want 20", resp.Usage.CompletionTokens)
+	}
+	if resp.Usage.TotalTokens != 170 {
+		t.Errorf("TotalTokens = %d, want 170", resp.Usage.TotalTokens)
+	}
+	if len(resp.ToolCalls) != 0 {
+		t.Errorf("ToolCalls should be empty, got %d", len(resp.ToolCalls))
+	}
+}
+
+func TestParseJSONLEvents_ToolCallExtraction(t *testing.T) {
+	p := &CodexCliProvider{}
+	toolCallText := `Let me read that file.
+{"tool_calls":[{"id":"call_1","type":"function","function":{"name":"read_file","arguments":"{\"path\":\"/tmp/test.txt\"}"}}]}`
+	// Build valid JSONL by marshaling the event
+	item := codexEvent{
+		Type: "item.completed",
+		Item: &codexEventItem{ID: "item_1", Type: "agent_message", Text: toolCallText},
+	}
+	itemJSON, _ := json.Marshal(item)
+	usageEvt := `{"type":"turn.completed","usage":{"input_tokens":50,"cached_input_tokens":0,"output_tokens":20}}`
+	events := `{"type":"turn.started"}` + "\n" + string(itemJSON) + "\n" + usageEvt
+
+	resp, err := p.parseJSONLEvents(events)
+	if err != nil {
+		t.Fatalf("parseJSONLEvents() error: %v", err)
+	}
+	if resp.FinishReason != "tool_calls" {
+		t.Errorf("FinishReason = %q, want %q", resp.FinishReason, "tool_calls")
+	}
+	if len(resp.ToolCalls) != 1 {
+		t.Fatalf("ToolCalls count = %d, want 1", len(resp.ToolCalls))
+	}
+	if resp.ToolCalls[0].Name != "read_file" {
+		t.Errorf("ToolCalls[0].Name = %q, want %q", resp.ToolCalls[0].Name, "read_file")
+	}
+	if resp.ToolCalls[0].ID != "call_1" {
+		t.Errorf("ToolCalls[0].ID = %q, want %q", resp.ToolCalls[0].ID, "call_1")
+	}
+	if resp.ToolCalls[0].Function.Arguments != `{"path":"/tmp/test.txt"}` {
+		t.Errorf("ToolCalls[0].Function.Arguments = %q", resp.ToolCalls[0].Function.Arguments)
+	}
+	// Content should have the tool call JSON stripped
+	if strings.Contains(resp.Content, "tool_calls") {
+		t.Errorf("Content should not contain tool_calls JSON, got: %q", resp.Content)
+	}
+}
+
+func TestParseJSONLEvents_MultipleToolCalls(t *testing.T) {
+	p := &CodexCliProvider{}
+	toolCallText := `{"tool_calls":[{"id":"call_1","type":"function","function":{"name":"read_file","arguments":"{\"path\":\"a.txt\"}"}},{"id":"call_2","type":"function","function":{"name":"write_file","arguments":"{\"path\":\"b.txt\",\"content\":\"hello\"}"}}]}`
+	item := codexEvent{
+		Type: "item.completed",
+		Item: &codexEventItem{ID: "item_1", Type: "agent_message", Text: toolCallText},
+	}
+	itemJSON, _ := json.Marshal(item)
+	events := `{"type":"turn.started"}` + "\n" + string(itemJSON) + "\n" + `{"type":"turn.completed"}`
+
+	resp, err := p.parseJSONLEvents(events)
+	if err != nil {
+		t.Fatalf("parseJSONLEvents() error: %v", err)
+	}
+	if len(resp.ToolCalls) != 2 {
+		t.Fatalf("ToolCalls count = %d, want 2", len(resp.ToolCalls))
+	}
+	if resp.ToolCalls[0].Name != "read_file" {
+		t.Errorf("ToolCalls[0].Name = %q, want %q", resp.ToolCalls[0].Name, "read_file")
+	}
+	if resp.ToolCalls[1].Name != "write_file" {
+		t.Errorf("ToolCalls[1].Name = %q, want %q", resp.ToolCalls[1].Name, "write_file")
+	}
+	if resp.FinishReason != "tool_calls" {
+		t.Errorf("FinishReason = %q, want %q", resp.FinishReason, "tool_calls")
+	}
+}
+
+func TestParseJSONLEvents_MultipleMessages(t *testing.T) {
+	p := &CodexCliProvider{}
+	events := `{"type":"turn.started"}
+{"type":"item.completed","item":{"id":"item_1","type":"agent_message","text":"First part."}}
+{"type":"item.completed","item":{"id":"item_2","type":"command_execution","command":"ls","status":"completed"}}
+{"type":"item.completed","item":{"id":"item_3","type":"agent_message","text":"Second part."}}
+{"type":"turn.completed"}`
+
+	resp, err := p.parseJSONLEvents(events)
+	if err != nil {
+		t.Fatalf("parseJSONLEvents() error: %v", err)
+	}
+	if resp.Content != "First part.\nSecond part." {
+		t.Errorf("Content = %q, want %q", resp.Content, "First part.\nSecond part.")
+	}
+}
+
+func TestParseJSONLEvents_ErrorEvent(t *testing.T) {
+	p := &CodexCliProvider{}
+	events := `{"type":"thread.started","thread_id":"abc"}
+{"type":"turn.started"}
+{"type":"error","message":"token expired"}
+{"type":"turn.failed","error":{"message":"token expired"}}`
+
+	_, err := p.parseJSONLEvents(events)
+	if err == nil {
+		t.Fatal("expected error")
+	}
+	if !strings.Contains(err.Error(), "token expired") {
+		t.Errorf("error = %q, want to contain 'token expired'", err.Error())
+	}
+}
+
+func TestParseJSONLEvents_TurnFailed(t *testing.T) {
+	p := &CodexCliProvider{}
+	events := `{"type":"turn.started"}
+{"type":"turn.failed","error":{"message":"rate limit exceeded"}}`
+
+	_, err := p.parseJSONLEvents(events)
+	if err == nil {
+		t.Fatal("expected error")
+	}
+	if !strings.Contains(err.Error(), "rate limit exceeded") {
+		t.Errorf("error = %q, want to contain 'rate limit exceeded'", err.Error())
+	}
+}
+
+func TestParseJSONLEvents_ErrorWithContent(t *testing.T) {
+	p := &CodexCliProvider{}
+	// If there's an error but also content, return the content (partial success)
+	events := `{"type":"turn.started"}
+{"type":"item.completed","item":{"id":"item_1","type":"agent_message","text":"Partial result."}}
+{"type":"error","message":"connection reset"}
+{"type":"turn.failed","error":{"message":"connection reset"}}`
+
+	resp, err := p.parseJSONLEvents(events)
+	if err != nil {
+		t.Fatalf("should not error when content exists: %v", err)
+	}
+	if resp.Content != "Partial result." {
+		t.Errorf("Content = %q, want %q", resp.Content, "Partial result.")
+	}
+}
+
+func TestParseJSONLEvents_EmptyOutput(t *testing.T) {
+	p := &CodexCliProvider{}
+	resp, err := p.parseJSONLEvents("")
+	if err != nil {
+		t.Fatalf("empty output should not error: %v", err)
+	}
+	if resp.Content != "" {
+		t.Errorf("Content = %q, want empty", resp.Content)
+	}
+}
+
+func TestParseJSONLEvents_MalformedLines(t *testing.T) {
+	p := &CodexCliProvider{}
+	events := `not json at all
+{"type":"item.completed","item":{"id":"item_1","type":"agent_message","text":"Good line."}}
+another bad line
+{"type":"turn.completed","usage":{"input_tokens":10,"output_tokens":5}}`
+
+	resp, err := p.parseJSONLEvents(events)
+	if err != nil {
+		t.Fatalf("should skip malformed lines: %v", err)
+	}
+	if resp.Content != "Good line." {
+		t.Errorf("Content = %q, want %q", resp.Content, "Good line.")
+	}
+	if resp.Usage == nil || resp.Usage.TotalTokens != 15 {
+		t.Errorf("Usage.TotalTokens = %v, want 15", resp.Usage)
+	}
+}
+
+func TestParseJSONLEvents_CommandExecution(t *testing.T) {
+	p := &CodexCliProvider{}
+	events := `{"type":"turn.started"}
+{"type":"item.started","item":{"id":"item_1","type":"command_execution","command":"bash -lc ls","status":"in_progress"}}
+{"type":"item.completed","item":{"id":"item_1","type":"command_execution","command":"bash -lc ls","status":"completed","exit_code":0,"output":"file1.go\nfile2.go"}}
+{"type":"item.completed","item":{"id":"item_2","type":"agent_message","text":"Found 2 files."}}
+{"type":"turn.completed"}`
+
+	resp, err := p.parseJSONLEvents(events)
+	if err != nil {
+		t.Fatalf("parseJSONLEvents() error: %v", err)
+	}
+	// command_execution items should be skipped; only agent_message text is returned
+	if resp.Content != "Found 2 files." {
+		t.Errorf("Content = %q, want %q", resp.Content, "Found 2 files.")
+	}
+}
+
+func TestParseJSONLEvents_NoUsage(t *testing.T) {
+	p := &CodexCliProvider{}
+	events := `{"type":"turn.started"}
+{"type":"item.completed","item":{"id":"item_1","type":"agent_message","text":"No usage info."}}
+{"type":"turn.completed"}`
+
+	resp, err := p.parseJSONLEvents(events)
+	if err != nil {
+		t.Fatalf("parseJSONLEvents() error: %v", err)
+	}
+	if resp.Usage != nil {
+		t.Errorf("Usage should be nil when turn.completed has no usage, got %+v", resp.Usage)
+	}
+}
+
+// --- Prompt Building Tests ---
+
+func TestBuildPrompt_SystemAsInstructions(t *testing.T) {
+	p := &CodexCliProvider{}
+	messages := []Message{
+		{Role: "system", Content: "You are helpful."},
+		{Role: "user", Content: "Hi there"},
+	}
+
+	prompt := p.buildPrompt(messages, nil)
+
+	if !strings.Contains(prompt, "## System Instructions") {
+		t.Error("prompt should contain '## System Instructions'")
+	}
+	if !strings.Contains(prompt, "You are helpful.") {
+		t.Error("prompt should contain system content")
+	}
+	if !strings.Contains(prompt, "## Task") {
+		t.Error("prompt should contain '## Task'")
+	}
+	if !strings.Contains(prompt, "Hi there") {
+		t.Error("prompt should contain user message")
+	}
+}
+
+func TestBuildPrompt_NoSystem(t *testing.T) {
+	p := &CodexCliProvider{}
+	messages := []Message{
+		{Role: "user", Content: "Just a question"},
+	}
+
+	prompt := p.buildPrompt(messages, nil)
+
+	if strings.Contains(prompt, "## System Instructions") {
+		t.Error("prompt should not contain system instructions header")
+	}
+	if prompt != "Just a question" {
+		t.Errorf("prompt = %q, want %q", prompt, "Just a question")
+	}
+}
+
+func TestBuildPrompt_WithTools(t *testing.T) {
+	p := &CodexCliProvider{}
+	messages := []Message{
+		{Role: "user", Content: "Get weather"},
+	}
+	tools := []ToolDefinition{
+		{
+			Type: "function",
+			Function: ToolFunctionDefinition{
+				Name:        "get_weather",
+				Description: "Get current weather",
+				Parameters: map[string]interface{}{
+					"type": "object",
+					"properties": map[string]interface{}{
+						"city": map[string]interface{}{"type": "string"},
+					},
+				},
+			},
+		},
+	}
+
+	prompt := p.buildPrompt(messages, tools)
+
+	if !strings.Contains(prompt, "## Available Tools") {
+		t.Error("prompt should contain tools section")
+	}
+	if !strings.Contains(prompt, "get_weather") {
+		t.Error("prompt should contain tool name")
+	}
+	if !strings.Contains(prompt, "Get current weather") {
+		t.Error("prompt should contain tool description")
+	}
+}
+
+func TestBuildPrompt_MultipleMessages(t *testing.T) {
+	p := &CodexCliProvider{}
+	messages := []Message{
+		{Role: "user", Content: "Hello"},
+		{Role: "assistant", Content: "Hi! How can I help?"},
+		{Role: "user", Content: "Tell me about Go"},
+	}
+
+	prompt := p.buildPrompt(messages, nil)
+
+	if !strings.Contains(prompt, "Hello") {
+		t.Error("prompt should contain first user message")
+	}
+	if !strings.Contains(prompt, "Assistant: Hi! How can I help?") {
+		t.Error("prompt should contain assistant message with prefix")
+	}
+	if !strings.Contains(prompt, "Tell me about Go") {
+		t.Error("prompt should contain second user message")
+	}
+}
+
+func TestBuildPrompt_ToolResults(t *testing.T) {
+	p := &CodexCliProvider{}
+	messages := []Message{
+		{Role: "user", Content: "Weather?"},
+		{Role: "tool", Content: `{"temp": 72}`, ToolCallID: "call_1"},
+	}
+
+	prompt := p.buildPrompt(messages, nil)
+
+	if !strings.Contains(prompt, "[Tool Result for call_1]") {
+		t.Error("prompt should contain tool result")
+	}
+	if !strings.Contains(prompt, `{"temp": 72}`) {
+		t.Error("prompt should contain tool result content")
+	}
+}
+
+func TestBuildPrompt_SystemAndTools(t *testing.T) {
+	p := &CodexCliProvider{}
+	messages := []Message{
+		{Role: "system", Content: "Be concise."},
+		{Role: "user", Content: "Do something"},
+	}
+	tools := []ToolDefinition{
+		{
+			Type: "function",
+			Function: ToolFunctionDefinition{
+				Name:        "my_tool",
+				Description: "A tool",
+			},
+		},
+	}
+
+	prompt := p.buildPrompt(messages, tools)
+
+	// System instructions should come first
+	sysIdx := strings.Index(prompt, "## System Instructions")
+	toolIdx := strings.Index(prompt, "## Available Tools")
+	taskIdx := strings.Index(prompt, "## Task")
+
+	if sysIdx == -1 || toolIdx == -1 || taskIdx == -1 {
+		t.Fatal("prompt should contain all sections")
+	}
+	if sysIdx >= taskIdx {
+		t.Error("system instructions should come before task")
+	}
+	if taskIdx >= toolIdx {
+		t.Error("task section should come before tools in the output")
+	}
+}
+
+// --- CLI Argument Tests ---
+
+func TestCodexCliProvider_GetDefaultModel(t *testing.T) {
+	p := NewCodexCliProvider("")
+	if got := p.GetDefaultModel(); got != "codex-cli" {
+		t.Errorf("GetDefaultModel() = %q, want %q", got, "codex-cli")
+	}
+}
+
+// --- Mock CLI Integration Test ---
+
+func createMockCodexCLI(t *testing.T, events []string) string {
+	t.Helper()
+	tmpDir := t.TempDir()
+	scriptPath := filepath.Join(tmpDir, "codex")
+
+	var sb strings.Builder
+	sb.WriteString("#!/bin/bash\n")
+	for _, event := range events {
+		sb.WriteString(fmt.Sprintf("echo '%s'\n", event))
+	}
+
+	if err := os.WriteFile(scriptPath, []byte(sb.String()), 0755); err != nil {
+		t.Fatal(err)
+	}
+	return scriptPath
+}
+
+func TestCodexCliProvider_MockCLI_Success(t *testing.T) {
+	scriptPath := createMockCodexCLI(t, []string{
+		`{"type":"thread.started","thread_id":"test-123"}`,
+		`{"type":"turn.started"}`,
+		`{"type":"item.completed","item":{"id":"item_1","type":"agent_message","text":"Mock response from Codex CLI"}}`,
+		`{"type":"turn.completed","usage":{"input_tokens":50,"cached_input_tokens":10,"output_tokens":15}}`,
+	})
+
+	p := &CodexCliProvider{
+		command:   scriptPath,
+		workspace: "",
+	}
+
+	messages := []Message{{Role: "user", Content: "Hello"}}
+	resp, err := p.Chat(context.Background(), messages, nil, "", nil)
+	if err != nil {
+		t.Fatalf("Chat() error: %v", err)
+	}
+	if resp.Content != "Mock response from Codex CLI" {
+		t.Errorf("Content = %q, want %q", resp.Content, "Mock response from Codex CLI")
+	}
+	if resp.Usage == nil {
+		t.Fatal("Usage should not be nil")
+	}
+	if resp.Usage.PromptTokens != 60 {
+		t.Errorf("PromptTokens = %d, want 60", resp.Usage.PromptTokens)
+	}
+	if resp.Usage.CompletionTokens != 15 {
+		t.Errorf("CompletionTokens = %d, want 15", resp.Usage.CompletionTokens)
+	}
+}
+
+func TestCodexCliProvider_MockCLI_Error(t *testing.T) {
+	scriptPath := createMockCodexCLI(t, []string{
+		`{"type":"thread.started","thread_id":"test-err"}`,
+		`{"type":"turn.started"}`,
+		`{"type":"error","message":"auth token expired"}`,
+		`{"type":"turn.failed","error":{"message":"auth token expired"}}`,
+	})
+
+	p := &CodexCliProvider{
+		command:   scriptPath,
+		workspace: "",
+	}
+
+	messages := []Message{{Role: "user", Content: "Hello"}}
+	_, err := p.Chat(context.Background(), messages, nil, "", nil)
+	if err == nil {
+		t.Fatal("expected error")
+	}
+	if !strings.Contains(err.Error(), "auth token expired") {
+		t.Errorf("error = %q, want to contain 'auth token expired'", err.Error())
+	}
+}
+
+func TestCodexCliProvider_MockCLI_WithModel(t *testing.T) {
+	// Mock script that captures args to verify model flag is passed
+	tmpDir := t.TempDir()
+	scriptPath := filepath.Join(tmpDir, "codex")
+	script := `#!/bin/bash
+# Write args to a file for verification
+echo "$@" > "` + filepath.Join(tmpDir, "args.txt") + `"
+echo '{"type":"item.completed","item":{"id":"1","type":"agent_message","text":"ok"}}'
+echo '{"type":"turn.completed"}'`
+
+	if err := os.WriteFile(scriptPath, []byte(script), 0755); err != nil {
+		t.Fatal(err)
+	}
+
+	p := &CodexCliProvider{
+		command:   scriptPath,
+		workspace: "/tmp/test-workspace",
+	}
+
+	messages := []Message{{Role: "user", Content: "test"}}
+	_, err := p.Chat(context.Background(), messages, nil, "gpt-5.2-codex", nil)
+	if err != nil {
+		t.Fatalf("Chat() error: %v", err)
+	}
+
+	// Verify the args
+	argsData, err := os.ReadFile(filepath.Join(tmpDir, "args.txt"))
+	if err != nil {
+		t.Fatalf("reading args: %v", err)
+	}
+	args := string(argsData)
+
+	if !strings.Contains(args, "-m gpt-5.2-codex") {
+		t.Errorf("args should contain model flag, got: %s", args)
+	}
+	if !strings.Contains(args, "-C /tmp/test-workspace") {
+		t.Errorf("args should contain workspace flag, got: %s", args)
+	}
+	if !strings.Contains(args, "--json") {
+		t.Errorf("args should contain --json, got: %s", args)
+	}
+	if !strings.Contains(args, "--dangerously-bypass-approvals-and-sandbox") {
+		t.Errorf("args should contain bypass flag, got: %s", args)
+	}
+}
+
+func TestCodexCliProvider_MockCLI_ContextCancel(t *testing.T) {
+	// Script that sleeps forever
+	tmpDir := t.TempDir()
+	scriptPath := filepath.Join(tmpDir, "codex")
+	script := "#!/bin/bash\nsleep 60"
+
+	if err := os.WriteFile(scriptPath, []byte(script), 0755); err != nil {
+		t.Fatal(err)
+	}
+
+	p := &CodexCliProvider{
+		command:   scriptPath,
+		workspace: "",
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel() // cancel immediately
+
+	messages := []Message{{Role: "user", Content: "test"}}
+	_, err := p.Chat(ctx, messages, nil, "", nil)
+	if err == nil {
+		t.Fatal("expected error on canceled context")
+	}
+}
+
+func TestCodexCliProvider_EmptyCommand(t *testing.T) {
+	p := &CodexCliProvider{command: ""}
+
+	messages := []Message{{Role: "user", Content: "test"}}
+	_, err := p.Chat(context.Background(), messages, nil, "", nil)
+	if err == nil {
+		t.Fatal("expected error for empty command")
+	}
+}
+
+// --- Integration Test (requires real codex CLI with valid auth) ---
+
+func TestCodexCliProvider_Integration(t *testing.T) {
+	if os.Getenv("PICOCLAW_INTEGRATION_TESTS") == "" {
+		t.Skip("skipping integration test (set PICOCLAW_INTEGRATION_TESTS=1 to enable)")
+	}
+
+	// Verify codex is available
+	codexPath, err := exec.LookPath("codex")
+	if err != nil {
+		t.Skip("codex CLI not found in PATH")
+	}
+
+	p := &CodexCliProvider{
+		command:   codexPath,
+		workspace: "",
+	}
+
+	messages := []Message{
+		{Role: "user", Content: "Respond with just the word 'hello' and nothing else."},
+	}
+
+	resp, err := p.Chat(context.Background(), messages, nil, "", nil)
+	if err != nil {
+		t.Fatalf("Chat() error: %v", err)
+	}
+
+	lower := strings.ToLower(strings.TrimSpace(resp.Content))
+	if !strings.Contains(lower, "hello") {
+		t.Errorf("Content = %q, expected to contain 'hello'", resp.Content)
+	}
+}

pkg/providers/codex_provider.go

@@ -3,6 +3,7 @@ package providers
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"strings"
 
@@ -10,18 +11,26 @@ import (
 	"github.com/openai/openai-go/v3/option"
 	"github.com/openai/openai-go/v3/responses"
 	"github.com/sipeed/picoclaw/pkg/auth"
+	"github.com/sipeed/picoclaw/pkg/logger"
 )
 
+const codexDefaultModel = "gpt-5.2"
+const codexDefaultInstructions = "You are Codex, a coding assistant."
+
 type CodexProvider struct {
 	client      *openai.Client
 	accountID   string
 	tokenSource func() (string, string, error)
 }
 
+const defaultCodexInstructions = "You are Codex, a coding assistant."
+
 func NewCodexProvider(token, accountID string) *CodexProvider {
 	opts := []option.RequestOption{
 		option.WithBaseURL("https://chatgpt.com/backend-api/codex"),
 		option.WithAPIKey(token),
+		option.WithHeader("originator", "codex_cli_rs"),
+		option.WithHeader("OpenAI-Beta", "responses=experimental"),
 	}
 	if accountID != "" {
 		opts = append(opts, option.WithHeader("Chatgpt-Account-Id", accountID))
@@ -41,6 +50,15 @@ func NewCodexProviderWithTokenSource(token, accountID string, tokenSource func()
 
 func (p *CodexProvider) Chat(ctx context.Context, messages []Message, tools []ToolDefinition, model string, options map[string]interface{}) (*LLMResponse, error) {
 	var opts []option.RequestOption
+	accountID := p.accountID
+	resolvedModel, fallbackReason := resolveCodexModel(model)
+	if fallbackReason != "" {
+		logger.WarnCF("provider.codex", "Requested model is not compatible with Codex backend, using fallback", map[string]interface{}{
+			"requested_model": model,
+			"resolved_model":  resolvedModel,
+			"reason":          fallbackReason,
+		})
+	}
 	if p.tokenSource != nil {
 		tok, accID, err := p.tokenSource()
 		if err != nil {
@@ -48,22 +66,120 @@ func (p *CodexProvider) Chat(ctx context.Context, messages []Message, tools []To
 		}
 		opts = append(opts, option.WithAPIKey(tok))
 		if accID != "" {
-			opts = append(opts, option.WithHeader("Chatgpt-Account-Id", accID))
+			accountID = accID
 		}
 	}
+	if accountID != "" {
+		opts = append(opts, option.WithHeader("Chatgpt-Account-Id", accountID))
+	} else {
+		logger.WarnCF("provider.codex", "No account id found for Codex request; backend may reject with 400", map[string]interface{}{
+			"requested_model": model,
+			"resolved_model":  resolvedModel,
+		})
+	}
 
-	params := buildCodexParams(messages, tools, model, options)
+	params := buildCodexParams(messages, tools, resolvedModel, options)
 
-	resp, err := p.client.Responses.New(ctx, params, opts...)
+	stream := p.client.Responses.NewStreaming(ctx, params, opts...)
+	defer stream.Close()
+
+	var resp *responses.Response
+	for stream.Next() {
+		evt := stream.Current()
+		if evt.Type == "response.completed" || evt.Type == "response.failed" || evt.Type == "response.incomplete" {
+			evtResp := evt.Response
+			if evtResp.ID != "" {
+				copy := evtResp
+				resp = &copy
+			}
+		}
+	}
+	err := stream.Err()
 	if err != nil {
+		fields := map[string]interface{}{
+			"requested_model":    model,
+			"resolved_model":     resolvedModel,
+			"messages_count":     len(messages),
+			"tools_count":        len(tools),
+			"account_id_present": accountID != "",
+			"error":              err.Error(),
+		}
+		var apiErr *openai.Error
+		if errors.As(err, &apiErr) {
+			fields["status_code"] = apiErr.StatusCode
+			fields["api_type"] = apiErr.Type
+			fields["api_code"] = apiErr.Code
+			fields["api_param"] = apiErr.Param
+			fields["api_message"] = apiErr.Message
+			if apiErr.StatusCode == 400 {
+				fields["hint"] = "verify account id header and model compatibility for codex backend"
+			}
+			if apiErr.Response != nil {
+				fields["request_id"] = apiErr.Response.Header.Get("x-request-id")
+			}
+		}
+		logger.ErrorCF("provider.codex", "Codex API call failed", fields)
 		return nil, fmt.Errorf("codex API call: %w", err)
 	}
+	if resp == nil {
+		fields := map[string]interface{}{
+			"requested_model":    model,
+			"resolved_model":     resolvedModel,
+			"messages_count":     len(messages),
+			"tools_count":        len(tools),
+			"account_id_present": accountID != "",
+		}
+		logger.ErrorCF("provider.codex", "Codex stream ended without completed response event", fields)
+		return nil, fmt.Errorf("codex API call: stream ended without completed response")
+	}
 
 	return parseCodexResponse(resp), nil
 }
 
 func (p *CodexProvider) GetDefaultModel() string {
-	return "gpt-4o"
+	return codexDefaultModel
+}
+
+func resolveCodexModel(model string) (string, string) {
+	m := strings.ToLower(strings.TrimSpace(model))
+	if m == "" {
+		return codexDefaultModel, "empty model"
+	}
+
+	if strings.HasPrefix(m, "openai/") {
+		m = strings.TrimPrefix(m, "openai/")
+	} else if strings.Contains(m, "/") {
+		return codexDefaultModel, "non-openai model namespace"
+	}
+
+	unsupportedPrefixes := []string{
+		"glm",
+		"claude",
+		"anthropic",
+		"gemini",
+		"google",
+		"moonshot",
+		"kimi",
+		"qwen",
+		"deepseek",
+		"llama",
+		"meta-llama",
+		"mistral",
+		"grok",
+		"xai",
+		"zhipu",
+	}
+	for _, prefix := range unsupportedPrefixes {
+		if strings.HasPrefix(m, prefix) {
+			return codexDefaultModel, "unsupported model prefix"
+		}
+	}
+
+	if strings.HasPrefix(m, "gpt-") || strings.HasPrefix(m, "o3") || strings.HasPrefix(m, "o4") {
+		return m, ""
+	}
+
+	return codexDefaultModel, "unsupported model family"
 }
 
 func buildCodexParams(messages []Message, tools []ToolDefinition, model string, options map[string]interface{}) responses.ResponseNewParams {
@@ -133,21 +249,21 @@ func buildCodexParams(messages []Message, tools []ToolDefinition, model string,
 		Input: responses.ResponseNewParamsInputUnion{
 			OfInputItemList: inputItems,
 		},
-		Store: openai.Opt(false),
+		Instructions: openai.Opt(instructions),
+		Store:        openai.Opt(false),
 	}
 
 	if instructions != "" {
 		params.Instructions = openai.Opt(instructions)
+	} else {
+		// ChatGPT Codex backend requires instructions to be present.
+		params.Instructions = openai.Opt(defaultCodexInstructions)
 	}
 
 	if maxTokens, ok := options["max_tokens"].(int); ok {
 		params.MaxOutputTokens = openai.Opt(int64(maxTokens))
 	}
 
-	if temp, ok := options["temperature"].(float64); ok {
-		params.Temperature = openai.Opt(temp)
-	}
-
 	if len(tools) > 0 {
 		params.Tools = translateToolsForCodex(tools)
 	}
@@ -237,6 +353,9 @@ func createCodexTokenSource() func() (string, string, error) {
 			if err != nil {
 				return "", "", fmt.Errorf("refreshing token: %w", err)
 			}
+			if refreshed.AccountID == "" {
+				refreshed.AccountID = cred.AccountID
+			}
 			if err := auth.SetCredential("openai", refreshed); err != nil {
 				return "", "", fmt.Errorf("saving refreshed token: %w", err)
 			}

pkg/providers/codex_provider_test.go

@@ -2,6 +2,7 @@ package providers
 
 import (
 	"encoding/json"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"testing"
@@ -16,11 +17,18 @@ func TestBuildCodexParams_BasicMessage(t *testing.T) {
 		{Role: "user", Content: "Hello"},
 	}
 	params := buildCodexParams(messages, nil, "gpt-4o", map[string]interface{}{
-		"max_tokens": 2048,
+		"max_tokens":  2048,
+		"temperature": 0.7,
 	})
 	if params.Model != "gpt-4o" {
 		t.Errorf("Model = %q, want %q", params.Model, "gpt-4o")
 	}
+	if !params.Instructions.Valid() {
+		t.Fatal("Instructions should be set")
+	}
+	if params.Instructions.Or("") != defaultCodexInstructions {
+		t.Errorf("Instructions = %q, want %q", params.Instructions.Or(""), defaultCodexInstructions)
+	}
 }
 
 func TestBuildCodexParams_SystemAsInstructions(t *testing.T) {
@@ -197,6 +205,16 @@ func TestCodexProvider_ChatRoundTrip(t *testing.T) {
 			return
 		}
 
+		var reqBody map[string]interface{}
+		if err := json.NewDecoder(r.Body).Decode(&reqBody); err != nil {
+			http.Error(w, "invalid json", http.StatusBadRequest)
+			return
+		}
+		if reqBody["stream"] != true {
+			http.Error(w, "stream must be true", http.StatusBadRequest)
+			return
+		}
+
 		resp := map[string]interface{}{
 			"id":     "resp_test",
 			"object": "response",
@@ -220,8 +238,7 @@ func TestCodexProvider_ChatRoundTrip(t *testing.T) {
 				"output_tokens_details": map[string]interface{}{"reasoning_tokens": 0},
 			},
 		}
-		w.Header().Set("Content-Type", "application/json")
-		json.NewEncoder(w).Encode(resp)
+		writeCompletedSSE(w, resp)
 	}))
 	defer server.Close()
 
@@ -244,10 +261,185 @@ func TestCodexProvider_ChatRoundTrip(t *testing.T) {
 	}
 }
 
+func TestCodexProvider_ChatRoundTrip_TokenSourceFallbackAccountID(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/responses" {
+			http.Error(w, "not found: "+r.URL.Path, http.StatusNotFound)
+			return
+		}
+		if r.Header.Get("Authorization") != "Bearer refreshed-token" {
+			http.Error(w, "unauthorized", http.StatusUnauthorized)
+			return
+		}
+		if r.Header.Get("Chatgpt-Account-Id") != "acc-123" {
+			http.Error(w, "missing account id", http.StatusBadRequest)
+			return
+		}
+
+		var reqBody map[string]interface{}
+		if err := json.NewDecoder(r.Body).Decode(&reqBody); err != nil {
+			http.Error(w, "invalid json", http.StatusBadRequest)
+			return
+		}
+		if _, ok := reqBody["instructions"]; !ok {
+			http.Error(w, "missing instructions", http.StatusBadRequest)
+			return
+		}
+		if reqBody["instructions"] == "" {
+			http.Error(w, "instructions must not be empty", http.StatusBadRequest)
+			return
+		}
+		if _, ok := reqBody["temperature"]; ok {
+			http.Error(w, "temperature is not supported", http.StatusBadRequest)
+			return
+		}
+		if reqBody["stream"] != true {
+			http.Error(w, "stream must be true", http.StatusBadRequest)
+			return
+		}
+
+		resp := map[string]interface{}{
+			"id":     "resp_test",
+			"object": "response",
+			"status": "completed",
+			"output": []map[string]interface{}{
+				{
+					"id":     "msg_1",
+					"type":   "message",
+					"role":   "assistant",
+					"status": "completed",
+					"content": []map[string]interface{}{
+						{"type": "output_text", "text": "Hi from Codex!"},
+					},
+				},
+			},
+			"usage": map[string]interface{}{
+				"input_tokens":          8,
+				"output_tokens":         4,
+				"total_tokens":          12,
+				"input_tokens_details":  map[string]interface{}{"cached_tokens": 0},
+				"output_tokens_details": map[string]interface{}{"reasoning_tokens": 0},
+			},
+		}
+		writeCompletedSSE(w, resp)
+	}))
+	defer server.Close()
+
+	provider := NewCodexProvider("stale-token", "acc-123")
+	provider.client = createOpenAITestClient(server.URL, "stale-token", "")
+	provider.tokenSource = func() (string, string, error) {
+		return "refreshed-token", "", nil
+	}
+
+	messages := []Message{{Role: "user", Content: "Hello"}}
+	resp, err := provider.Chat(t.Context(), messages, nil, "gpt-4o", map[string]interface{}{"temperature": 0.7})
+	if err != nil {
+		t.Fatalf("Chat() error: %v", err)
+	}
+	if resp.Content != "Hi from Codex!" {
+		t.Errorf("Content = %q, want %q", resp.Content, "Hi from Codex!")
+	}
+}
+
+func TestCodexProvider_ChatRoundTrip_ModelFallbackFromUnsupported(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/responses" {
+			http.Error(w, "not found: "+r.URL.Path, http.StatusNotFound)
+			return
+		}
+
+		var reqBody map[string]interface{}
+		if err := json.NewDecoder(r.Body).Decode(&reqBody); err != nil {
+			http.Error(w, "invalid json", http.StatusBadRequest)
+			return
+		}
+		if reqBody["model"] != codexDefaultModel {
+			http.Error(w, "unsupported model", http.StatusBadRequest)
+			return
+		}
+		if reqBody["stream"] != true {
+			http.Error(w, "stream must be true", http.StatusBadRequest)
+			return
+		}
+		if reqBody["instructions"] != codexDefaultInstructions {
+			http.Error(w, "missing default instructions", http.StatusBadRequest)
+			return
+		}
+
+		resp := map[string]interface{}{
+			"id":     "resp_test",
+			"object": "response",
+			"status": "completed",
+			"output": []map[string]interface{}{
+				{
+					"id":     "msg_1",
+					"type":   "message",
+					"role":   "assistant",
+					"status": "completed",
+					"content": []map[string]interface{}{
+						{"type": "output_text", "text": "Hi from Codex!"},
+					},
+				},
+			},
+			"usage": map[string]interface{}{
+				"input_tokens":          8,
+				"output_tokens":         4,
+				"total_tokens":          12,
+				"input_tokens_details":  map[string]interface{}{"cached_tokens": 0},
+				"output_tokens_details": map[string]interface{}{"reasoning_tokens": 0},
+			},
+		}
+		writeCompletedSSE(w, resp)
+	}))
+	defer server.Close()
+
+	provider := NewCodexProvider("test-token", "acc-123")
+	provider.client = createOpenAITestClient(server.URL, "test-token", "acc-123")
+
+	messages := []Message{{Role: "user", Content: "Hello"}}
+	resp, err := provider.Chat(t.Context(), messages, nil, "gpt-5.2", nil)
+	if err != nil {
+		t.Fatalf("Chat() error: %v", err)
+	}
+	if resp.Content != "Hi from Codex!" {
+		t.Errorf("Content = %q, want %q", resp.Content, "Hi from Codex!")
+	}
+}
+
 func TestCodexProvider_GetDefaultModel(t *testing.T) {
 	p := NewCodexProvider("test-token", "")
-	if got := p.GetDefaultModel(); got != "gpt-4o" {
-		t.Errorf("GetDefaultModel() = %q, want %q", got, "gpt-4o")
+	if got := p.GetDefaultModel(); got != codexDefaultModel {
+		t.Errorf("GetDefaultModel() = %q, want %q", got, codexDefaultModel)
+	}
+}
+
+func TestResolveCodexModel(t *testing.T) {
+	tests := []struct {
+		name         string
+		input        string
+		wantModel    string
+		wantFallback bool
+	}{
+		{name: "empty", input: "", wantModel: codexDefaultModel, wantFallback: true},
+		{name: "unsupported namespace", input: "anthropic/claude-3.5", wantModel: codexDefaultModel, wantFallback: true},
+		{name: "non-openai prefixed", input: "glm-4.7", wantModel: codexDefaultModel, wantFallback: true},
+		{name: "openai prefix", input: "openai/gpt-5.2", wantModel: "gpt-5.2", wantFallback: false},
+		{name: "direct gpt", input: "gpt-4o", wantModel: "gpt-4o", wantFallback: false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			gotModel, reason := resolveCodexModel(tt.input)
+			if gotModel != tt.wantModel {
+				t.Fatalf("resolveCodexModel(%q) model = %q, want %q", tt.input, gotModel, tt.wantModel)
+			}
+			if tt.wantFallback && reason == "" {
+				t.Fatalf("resolveCodexModel(%q) expected fallback reason", tt.input)
+			}
+			if !tt.wantFallback && reason != "" {
+				t.Fatalf("resolveCodexModel(%q) unexpected fallback reason: %q", tt.input, reason)
+			}
+		})
 	}
 }
 
@@ -262,3 +454,16 @@ func createOpenAITestClient(baseURL, token, accountID string) *openai.Client {
 	c := openai.NewClient(opts...)
 	return &c
 }
+
+func writeCompletedSSE(w http.ResponseWriter, response map[string]interface{}) {
+	event := map[string]interface{}{
+		"type":            "response.completed",
+		"sequence_number": 1,
+		"response":        response,
+	}
+	b, _ := json.Marshal(event)
+	w.Header().Set("Content-Type", "text/event-stream")
+	fmt.Fprintf(w, "event: response.completed\n")
+	fmt.Fprintf(w, "data: %s\n\n", string(b))
+	fmt.Fprintf(w, "data: [DONE]\n\n")
+}

pkg/providers/github_copilot_provider.go

@@ -0,0 +1,82 @@
+package providers
+
+import (
+	"context"
+	"fmt"
+
+	json "encoding/json"
+
+	copilot "github.com/github/copilot-sdk/go"
+)
+
+type GitHubCopilotProvider struct {
+	uri         string
+	connectMode string // `stdio` or `grpc``
+
+	session *copilot.Session
+}
+
+func NewGitHubCopilotProvider(uri string, connectMode string, model string) (*GitHubCopilotProvider, error) {
+
+	var session *copilot.Session
+	if connectMode == "" {
+		connectMode = "grpc"
+	}
+	switch connectMode {
+
+	case "stdio":
+		//todo
+	case "grpc":
+		client := copilot.NewClient(&copilot.ClientOptions{
+			CLIUrl: uri,
+		})
+		if err := client.Start(context.Background()); err != nil {
+			return nil, fmt.Errorf("Can't connect to Github Copilot, https://github.com/github/copilot-sdk/blob/main/docs/getting-started.md#connecting-to-an-external-cli-server for details")
+		}
+		defer client.Stop()
+		session, _ = client.CreateSession(context.Background(), &copilot.SessionConfig{
+			Model: model,
+			Hooks: &copilot.SessionHooks{},
+		})
+
+	}
+
+	return &GitHubCopilotProvider{
+		uri:         uri,
+		connectMode: connectMode,
+		session:     session,
+	}, nil
+}
+
+// Chat sends a chat request to GitHub Copilot
+func (p *GitHubCopilotProvider) Chat(ctx context.Context, messages []Message, tools []ToolDefinition, model string, options map[string]interface{}) (*LLMResponse, error) {
+	type tempMessage struct {
+		Role    string `json:"role"`
+		Content string `json:"content"`
+	}
+	out := make([]tempMessage, 0, len(messages))
+
+	for _, msg := range messages {
+		out = append(out, tempMessage{
+			Role:    msg.Role,
+			Content: msg.Content,
+		})
+	}
+
+	fullcontent, _ := json.Marshal(out)
+
+	content, _ := p.session.Send(ctx, copilot.MessageOptions{
+		Prompt: string(fullcontent),
+	})
+
+	return &LLMResponse{
+		FinishReason: "stop",
+		Content:      content,
+	}, nil
+
+}
+
+func (p *GitHubCopilotProvider) GetDefaultModel() string {
+
+	return "gpt-4.1"
+}

pkg/providers/http_provider.go

@@ -15,6 +15,7 @@ import (
 	"net/http"
 	"net/url"
 	"strings"
+	"time"
 
 	"github.com/sipeed/picoclaw/pkg/auth"
 	"github.com/sipeed/picoclaw/pkg/config"
@@ -28,7 +29,7 @@ type HTTPProvider struct {
 
 func NewHTTPProvider(apiKey, apiBase, proxy string) *HTTPProvider {
 	client := &http.Client{
-		Timeout: 0,
+		Timeout: 120 * time.Second,
 	}
 
 	if proxy != "" {
@@ -52,10 +53,10 @@ func (p *HTTPProvider) Chat(ctx context.Context, messages []Message, tools []Too
 		return nil, fmt.Errorf("API base not configured")
 	}
 
-	// Strip provider prefix from model name (e.g., moonshot/kimi-k2.5 -> kimi-k2.5)
+	// Strip provider prefix from model name (e.g., moonshot/kimi-k2.5 -> kimi-k2.5, groq/openai/gpt-oss-120b -> openai/gpt-oss-120b, ollama/qwen2.5:14b -> qwen2.5:14b)
 	if idx := strings.Index(model, "/"); idx != -1 {
 		prefix := model[:idx]
-		if prefix == "moonshot" || prefix == "nvidia" {
+		if prefix == "moonshot" || prefix == "nvidia" || prefix == "groq" || prefix == "ollama" {
 			model = model[idx+1:]
 		}
 	}
@@ -239,6 +240,9 @@ func CreateProvider(cfg *config.Config) (LLMProvider, error) {
 			}
 		case "openai", "gpt":
 			if cfg.Providers.OpenAI.APIKey != "" || cfg.Providers.OpenAI.AuthMethod != "" {
+				if cfg.Providers.OpenAI.AuthMethod == "codex-cli" {
+					return NewCodexProviderWithTokenSource("", "", CreateCodexCliTokenSource()), nil
+				}
 				if cfg.Providers.OpenAI.AuthMethod == "oauth" || cfg.Providers.OpenAI.AuthMethod == "token" {
 					return createCodexAuthProvider()
 				}
@@ -298,11 +302,17 @@ func CreateProvider(cfg *config.Config) (LLMProvider, error) {
 				}
 			}
 		case "claude-cli", "claudecode", "claude-code":
-			workspace := cfg.Agents.Defaults.Workspace
+			workspace := cfg.WorkspacePath()
 			if workspace == "" {
 				workspace = "."
 			}
 			return NewClaudeCliProvider(workspace), nil
+		case "codex-cli", "codex-code":
+			workspace := cfg.WorkspacePath()
+			if workspace == "" {
+				workspace = "."
+			}
+			return NewCodexCliProvider(workspace), nil
 		case "deepseek":
 			if cfg.Providers.DeepSeek.APIKey != "" {
 				apiKey = cfg.Providers.DeepSeek.APIKey
@@ -314,7 +324,16 @@ func CreateProvider(cfg *config.Config) (LLMProvider, error) {
 					model = "deepseek-chat"
 				}
 			}
+		case "github_copilot", "copilot":
+			if cfg.Providers.GitHubCopilot.APIBase != "" {
+				apiBase = cfg.Providers.GitHubCopilot.APIBase
+			} else {
+				apiBase = "localhost:4321"
+			}
+			return NewGitHubCopilotProvider(apiBase, cfg.Providers.GitHubCopilot.ConnectMode, model)
+
 		}
+
 	}
 
 	// Fallback: detect provider from model name
@@ -390,7 +409,15 @@ func CreateProvider(cfg *config.Config) (LLMProvider, error) {
 			if apiBase == "" {
 				apiBase = "https://integrate.api.nvidia.com/v1"
 			}
-
+		case (strings.Contains(lowerModel, "ollama") || strings.HasPrefix(model, "ollama/")) && cfg.Providers.Ollama.APIKey != "":
+			fmt.Println("Ollama provider selected based on model name prefix")
+			apiKey = cfg.Providers.Ollama.APIKey
+			apiBase = cfg.Providers.Ollama.APIBase
+			proxy = cfg.Providers.Ollama.Proxy
+			if apiBase == "" {
+				apiBase = "http://localhost:11434/v1"
+			}
+			fmt.Println("Ollama apiBase:", apiBase)
 		case cfg.Providers.VLLM.APIBase != "":
 			apiKey = cfg.Providers.VLLM.APIKey
 			apiBase = cfg.Providers.VLLM.APIBase

pkg/providers/tool_call_extract.go

@@ -0,0 +1,72 @@
+package providers
+
+import (
+	"encoding/json"
+	"strings"
+)
+
+// extractToolCallsFromText parses tool call JSON from response text.
+// Both ClaudeCliProvider and CodexCliProvider use this to extract
+// tool calls that the model outputs in its response text.
+func extractToolCallsFromText(text string) []ToolCall {
+	start := strings.Index(text, `{"tool_calls"`)
+	if start == -1 {
+		return nil
+	}
+
+	end := findMatchingBrace(text, start)
+	if end == start {
+		return nil
+	}
+
+	jsonStr := text[start:end]
+
+	var wrapper struct {
+		ToolCalls []struct {
+			ID       string `json:"id"`
+			Type     string `json:"type"`
+			Function struct {
+				Name      string `json:"name"`
+				Arguments string `json:"arguments"`
+			} `json:"function"`
+		} `json:"tool_calls"`
+	}
+
+	if err := json.Unmarshal([]byte(jsonStr), &wrapper); err != nil {
+		return nil
+	}
+
+	var result []ToolCall
+	for _, tc := range wrapper.ToolCalls {
+		var args map[string]interface{}
+		json.Unmarshal([]byte(tc.Function.Arguments), &args)
+
+		result = append(result, ToolCall{
+			ID:        tc.ID,
+			Type:      tc.Type,
+			Name:      tc.Function.Name,
+			Arguments: args,
+			Function: &FunctionCall{
+				Name:      tc.Function.Name,
+				Arguments: tc.Function.Arguments,
+			},
+		})
+	}
+
+	return result
+}
+
+// stripToolCallsFromText removes tool call JSON from response text.
+func stripToolCallsFromText(text string) string {
+	start := strings.Index(text, `{"tool_calls"`)
+	if start == -1 {
+		return text
+	}
+
+	end := findMatchingBrace(text, start)
+	if end == start {
+		return text
+	}
+
+	return strings.TrimSpace(text[:start] + text[end:])
+}