diff --git a/cmd/olliesrv/internal/agent/agent.go b/cmd/olliesrv/internal/agent/agent.go
index 580de54..815191c 100644
--- a/cmd/olliesrv/internal/agent/agent.go
+++ b/cmd/olliesrv/internal/agent/agent.go
@@ -4,6 +4,7 @@ import (
 	"context"
 	"fmt"
 	"os"
+	"runtime"
 	"slices"
 	"sync"
 	"sync/atomic"
@@ -362,7 +363,7 @@ func (ag *Agent) SwitchProfile(name string) error {
 //   - "tool"      → logs result
 //   - "state"     → calls ag.SetState()
 //   - "limitretry"→ sets state to "limitretry"
-//   - "usage"     → parses token counts, updates history, checks spending cap
+//   - "usage"     → parses token counts, updates history
 //   - "error"     → logs
 type Event struct {
 	Role         string
@@ -423,11 +424,8 @@ func (ag *Agent) SetCWD(dir string) {
 	ag.stateMu.Unlock()
 	// Rebuild the environment section with the new cwd
 	if ag.runtime != nil && ag.runtime.Preamble != nil {
-		ag.runtime.Preamble.Set(SectionEnv, ag.envBlock)
-		// Update envBlock to reflect new cwd for future agent reloads
-		// (envBlock is set at construction; for cwd changes we just rebuild)
 		isGitRepo := paths.IsGitRepo(dir)
-		ag.envBlock = EnvironmentBlock(dir, "linux", isGitRepo, "")
+		ag.envBlock = EnvironmentBlock(dir, runtime.GOOS, isGitRepo, "")
 		ag.runtime.Preamble.Set(SectionEnv, ag.envBlock)
 	}
 	if ag.runtime.ToolServer != nil {
@@ -548,7 +546,6 @@ func (ag *Agent) SetGenParams(params backend.GenerationParams) {
 	ag.runtime.GenParams = params
 }
 
-// CompactionModel returns the configured compaction model name.
 // ListModels returns available models from the backend.
 func (ag *Agent) ListModels() []string {
 	return ag.runtime.Backend.Models(context.Background())
diff --git a/cmd/olliesrv/internal/agent/agent_config.go b/cmd/olliesrv/internal/agent/agent_config.go
index e362c1c..c598d6f 100644
--- a/cmd/olliesrv/internal/agent/agent_config.go
+++ b/cmd/olliesrv/internal/agent/agent_config.go
@@ -22,13 +22,6 @@ func (p *Prompt) UnmarshalJSON(data []byte) error {
 	return json.Unmarshal(data, &p.Value)
 }
 
-func (p Prompt) MarshalJSON() ([]byte, error) {
-	if len(p.Value) == 1 {
-		return json.Marshal(p.Value[0])
-	}
-	return json.Marshal(p.Value)
-}
-
 type AgentConfig struct {
 	Prompt      Prompt   `json:"prompt,omitempty"`
 	Backend     string   `json:"backend,omitempty"`
diff --git a/cmd/olliesrv/internal/agent/commands.go b/cmd/olliesrv/internal/agent/commands.go
index 0cd44ae..8fbe1b5 100644
--- a/cmd/olliesrv/internal/agent/commands.go
+++ b/cmd/olliesrv/internal/agent/commands.go
@@ -2,12 +2,13 @@ package agent
 
 import (
 	"context"
-	"os/exec"
 	"strings"
+
+	lib9p "ollie/lib9p/client"
 )
 
 // HandleCommand intercepts /commands before they reach the LLM.
-// / prefix routes to ctl via o: /compact → o session/agent ctl compact.
+// Each slash command is a rdwr to the agent's ctl file: /compact → write "compact" to ctl.
 // Returns true if the input was consumed as a command.
 func (ag *Agent) HandleCommand(_ context.Context, input string) bool {
 	if !strings.HasPrefix(input, "/") {
@@ -17,16 +18,20 @@ func (ag *Agent) HandleCommand(_ context.Context, input string) bool {
 	if rest == "" {
 		return false
 	}
-	// Use session/agent context syntax
-	ctx := ag.sessionID + "/" + ag.id
-	args := append([]string{ctx, "ctl"}, strings.Fields(rest)...)
-	cmd := exec.Command("o", args...)
-	out, err := cmd.CombinedOutput()
-	if len(out) > 0 {
-		ag.emit(infoEvent(strings.TrimRight(string(out), "\n")))
-	}
+	fsys, err := lib9p.DialDefault()
 	if err != nil {
 		ag.emit(infoEvent("error: " + err.Error()))
+		return true
+	}
+	defer fsys.Close()
+	ctlPath := "session/" + ag.sessionID + "/agent/" + ag.id + "/ctl"
+	out, err := fsys.Rdwr(ctlPath, []byte(rest))
+	if err != nil {
+		ag.emit(infoEvent("error: " + err.Error()))
+		return true
+	}
+	if len(out) > 0 {
+		ag.emit(infoEvent(strings.TrimRight(string(out), "\n")))
 	}
 	return true
 }
diff --git a/cmd/olliesrv/internal/agent/config_paths.go b/cmd/olliesrv/internal/agent/config_paths.go
index 93fadd0..5d7fde3 100644
--- a/cmd/olliesrv/internal/agent/config_paths.go
+++ b/cmd/olliesrv/internal/agent/config_paths.go
@@ -13,12 +13,6 @@ func AgentsDirs() []string {
 
 // AgentConfigPath resolves the config file path for a named agent.
 func AgentConfigPath(agentsDir, name string) string {
-	for _, dir := range AgentsDirs() {
-		p := dir + "/" + name + ".json"
-		if _, err := os.Stat(p); err == nil {
-			return p
-		}
-	}
 	return agentsDir + "/" + name + ".json"
 }
 
diff --git a/cmd/olliesrv/internal/agent/history.go b/cmd/olliesrv/internal/agent/history.go
index 0010c87..5173e51 100644
--- a/cmd/olliesrv/internal/agent/history.go
+++ b/cmd/olliesrv/internal/agent/history.go
@@ -76,7 +76,7 @@ type History struct {
 	turnCostUSD        float64 // >0 when backend reported cost directly (e.g. OpenRouter)
 }
 
-// newHistory creates a new empty Session. The caller is responsible for
+// newHistory creates a new empty History. The caller is responsible for
 // appending the initial user message via appendUserMessage.
 func newHistory(goal string) *History {
 	return &History{goal: goal}
@@ -380,10 +380,16 @@ func cloneMessages(msgs []backend.Message) []backend.Message {
 	out := make([]backend.Message, len(msgs))
 	for i, m := range msgs {
 		out[i] = backend.Message{
+			ID:         m.ID,
 			Role:       m.Role,
 			Content:    m.Content,
+			Reasoning:  m.Reasoning,
 			ToolCallID: m.ToolCallID,
 		}
+		if len(m.ContentBlocks) > 0 {
+			out[i].ContentBlocks = make([]backend.ContentBlock, len(m.ContentBlocks))
+			copy(out[i].ContentBlocks, m.ContentBlocks)
+		}
 		if len(m.ToolCalls) > 0 {
 			out[i].ToolCalls = make([]backend.ToolCall, len(m.ToolCalls))
 			for j, tc := range m.ToolCalls {
@@ -410,9 +416,9 @@ func (s *History) estimateTokens() int {
 	return chars / 4
 }
 
-// stripColdResults summarizes large tool-result messages outside the hot tail
+// stripCold summarizes large tool-result messages outside the hot tail
 // using an LLM call. Messages in the last hotTailSize slots are left verbatim.
-func (s *History) stripColdResults(ctx context.Context, b backend.Backend) {
+func (s *History) stripCold(ctx context.Context, b backend.Backend) {
 	hot := len(s.messages) - hotTailSize
 	for i := range s.messages {
 		if i >= hot {
@@ -425,11 +431,6 @@ func (s *History) stripColdResults(ctx context.Context, b backend.Backend) {
 	}
 }
 
-// stripCold implements the state interface.
-func (s *History) stripCold(ctx context.Context, b backend.Backend) {
-	s.stripColdResults(ctx, b)
-}
-
 // resolveCompactionModel returns the model to use for compaction.
 // Priority: agent config > OLLIE_COMPACTION_MODEL env > per-backend default from models.yaml > session's current model.
 // Returns "" if no override is configured (use the session's current model).
diff --git a/cmd/olliesrv/internal/agent/loop.go b/cmd/olliesrv/internal/agent/loop.go
index 7b635bc..9f8baa6 100644
--- a/cmd/olliesrv/internal/agent/loop.go
+++ b/cmd/olliesrv/internal/agent/loop.go
@@ -294,8 +294,8 @@ func streamResponse(rt *Runtime, ctx TurnCtx, history []backend.Message, respons
 				contentBuf.WriteString(ev.Content)
 				emit(ctx, Event{Role: "assistant", Content: ev.Content, ResponseID: responseID})
 			}
-			toolCalls = append(toolCalls, ev.ToolCalls...)
 			if ev.Done {
+				toolCalls = append(toolCalls, ev.ToolCalls...)
 				if hadReasoning {
 					emit(ctx, Event{Role: "reasoning", Content: "\n</think>\n"})
 					hadReasoning = false
diff --git a/cmd/olliesrv/internal/agent/text_parse.go b/cmd/olliesrv/internal/agent/text_parse.go
index e80ab47..ab39ad6 100644
--- a/cmd/olliesrv/internal/agent/text_parse.go
+++ b/cmd/olliesrv/internal/agent/text_parse.go
@@ -284,5 +284,3 @@ func relaxJSON(s string) string {
 	}
 	return out.String()
 }
-
-// transientWait returns the retry wait for a retryable error and whether it is
diff --git a/cmd/olliesrv/internal/agent/turn.go b/cmd/olliesrv/internal/agent/turn.go
index 435f302..20d7322 100644
--- a/cmd/olliesrv/internal/agent/turn.go
+++ b/cmd/olliesrv/internal/agent/turn.go
@@ -6,7 +6,6 @@ import (
 	"fmt"
 	"os"
 	"runtime/debug"
-	"strconv"
 	"strings"
 
 	"ollie/cmd/olliesrv/internal/backend"
@@ -71,9 +70,8 @@ func (ag *Agent) Submit(ctx context.Context, input string) {
 func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 	ag.emit(Event{Role: "user", Content: input})
 
-	// Snapshot session state before this turn modifies it. Restored on failure
-	// so the session is clean for the next attempt.
-	snapSession := ag.history
+	// Snapshot messages before this turn modifies them. Restored on context
+	// overflow so the retry starts from a clean state.
 	var snapMessages []backend.Message
 	if ag.history != nil {
 		snapMessages = cloneMessages(ag.history.messages)
@@ -120,11 +118,8 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 	}
 
 	turnCtx := TurnCtx{
-		Output:      ag.output,
-		Save:        ag.save,
-		ResultCache: &ag.resultCache,
-		BgTracker:   &ag.bgProcs,
-		ToolServer:  ag.runtime.ToolServer,
+		BgTracker:  &ag.bgProcs,
+		ToolServer: ag.runtime.ToolServer,
 	}
 
 	var replyBuf strings.Builder
@@ -156,12 +151,6 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 				CostUSD:             costUSD,
 			}, est != 0)
 			ag.notifyChange()
-			if maxCostStr := os.Getenv("OLLIE_MAX_SESSION_COST"); maxCostStr != "" {
-				if limit, ferr := strconv.ParseFloat(maxCostStr, 64); ferr == nil && limit > 0 && ag.history.SessionCostUSD >= limit {
-					ag.emit(infoEvent(fmt.Sprintf("spending cap $%.2f reached — stopping", limit)))
-					ag.Interrupt(ErrInterrupted)
-				}
-			}
 		}
 		ag.emit(ev)
 	}
@@ -213,24 +202,6 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 		}
 	}
 
-	// Spending cap: reject before spending more tokens.
-	if maxCostStr := os.Getenv("OLLIE_MAX_SESSION_COST"); maxCostStr != "" && ag.history != nil {
-		if limit, err := strconv.ParseFloat(maxCostStr, 64); err == nil && limit > 0 {
-			if ag.history.SessionCostUSD >= limit {
-				ag.emit(Event{Role: "error", Content: fmt.Sprintf("spending cap $%.2f reached (session total $%.4f)", limit, ag.history.SessionCostUSD)})
-				ag.SetState("idle")
-				actCancel(nil)
-				ag.currentAction.CompareAndSwap(handle, nil)
-				if snapSession == nil {
-					ag.history = nil
-				} else {
-					ag.history.messages = snapMessages
-				}
-				return ""
-			}
-		}
-	}
-
 	if ag.history != nil {
 		ag.history.resetTurnAccumulators()
 	}
diff --git a/cmd/olliesrv/internal/backend/kiro.go b/cmd/olliesrv/internal/backend/kiro.go
index f2380bf..8c20479 100644
--- a/cmd/olliesrv/internal/backend/kiro.go
+++ b/cmd/olliesrv/internal/backend/kiro.go
@@ -114,7 +114,7 @@ func (b *KiroBackend) fetchModels(ctx context.Context) *kiroListModelsResponse {
 }
 
 func (b *KiroBackend) ContextLength(ctx context.Context) int {
-	if b.ctxLen > 0 {
+	if b.ctxLen > 0 && b.ctxModel == b.model {
 		return b.ctxLen
 	}
 	resp := b.fetchModels(ctx)
@@ -124,11 +124,13 @@ func (b *KiroBackend) ContextLength(ctx context.Context) int {
 	for _, m := range resp.Models {
 		if m.TokenLimits != nil && (b.model == "auto" || m.ModelID == b.model) {
 			b.ctxLen = m.TokenLimits.MaxInputTokens
+			b.ctxModel = b.model
 			return b.ctxLen
 		}
 	}
 	if resp.DefaultModel != nil && resp.DefaultModel.TokenLimits != nil {
 		b.ctxLen = resp.DefaultModel.TokenLimits.MaxInputTokens
+		b.ctxModel = b.model
 		return b.ctxLen
 	}
 	return 0
diff --git a/cmd/olliesrv/internal/backend/openai.go b/cmd/olliesrv/internal/backend/openai.go
index a902dd6..92608db 100644
--- a/cmd/olliesrv/internal/backend/openai.go
+++ b/cmd/olliesrv/internal/backend/openai.go
@@ -105,10 +105,10 @@ func (b *OpenAIBackend) DefaultModel() string {
 	switch b.name {
 	case "openrouter":
 		return "deepseek/deepseek-v3.2"
-	case "anthropic":
-		return "claude-sonnet-4-5"
+	case "gemini":
+		return geminiDefaultModel
 	default:
-		return "qwen3.5:9b"
+		return "gpt-4o"
 	}
 }
 
diff --git a/cmd/olliesrv/internal/backend/openai_test.go b/cmd/olliesrv/internal/backend/openai_test.go
index fc0cf42..d2701ac 100644
--- a/cmd/olliesrv/internal/backend/openai_test.go
+++ b/cmd/olliesrv/internal/backend/openai_test.go
@@ -423,9 +423,9 @@ func TestDefaultModel(t *testing.T) {
 		want string
 	}{
 		{"openrouter", "deepseek/deepseek-v3.2"},
-		{"anthropic", "claude-sonnet-4-5"},
-		{"openai", "qwen3.5:9b"},
-		{"unknown", "qwen3.5:9b"},
+		{"gemini", geminiDefaultModel},
+		{"openai", "gpt-4o"},
+		{"unknown", "gpt-4o"},
 	}
 	for _, tt := range tests {
 		if got := (&OpenAIBackend{baseBackend: baseBackend{name: tt.name}}).DefaultModel(); got != tt.want {
diff --git a/cmd/olliesrv/internal/fs/handlers_agent.go b/cmd/olliesrv/internal/fs/handlers_agent.go
index a503fcf..8d947e6 100644
--- a/cmd/olliesrv/internal/fs/handlers_agent.go
+++ b/cmd/olliesrv/internal/fs/handlers_agent.go
@@ -329,22 +329,7 @@ var agentCtlHandlers = map[string]rdwrHandler{
 		return []byte(ctx.Session.CachedListModels() + "\n"), nil
 	},
 	"tools": func(ctx HandlerCtx, _ []string) ([]byte, error) {
-		if ctx.Session.ToolsConn() == nil {
-			return []byte("(no tool server)\n"), nil
-		}
-		loaded, err := ctx.Session.ToolsConn().ListTools()
-		if err != nil {
-			return nil, err
-		}
-		var sb strings.Builder
-		for _, ti := range loaded {
-			if ti.Description != "" {
-				fmt.Fprintf(&sb, "%-20s %s\n", ti.Name, ti.Description)
-			} else {
-				sb.WriteString(ti.Name + "\n")
-			}
-		}
-		return []byte(sb.String()), nil
+		return readAgentTools(ctx)
 	},
 	"tool_load": func(ctx HandlerCtx, args []string) ([]byte, error) {
 		if len(args) == 0 {