fix(mcp): surface MCP tool failures as soft errors, not critical aborts

The MCP adapter previously wrapped any error returned by MCPToolManager.ExecuteTool
into a Go error returned from the fantasy.AgentTool.Run interface. The fantasy
agent loop treats those as critical errors and aborts the entire turn —
discarding all prior reasoning, tool calls, and results.

In practice that meant a single misbehaved MCP server returning a JSON-RPC
"-32602 Invalid params" (e.g. a Zod schema mismatch on the server's input
validation) would kill an in-progress turn after the model had already done
dozens of seconds of useful work, with no way for the model to see the
validation message and self-correct.

This mismatched the contract that native Kit tools follow: native tools
return errors via kit.ErrorResult(...), which become soft tool-result errors
that the model reads and can act on (retry with corrected args, try a
different tool, give up gracefully).

Make the MCP path behave the same way:

  - JSON-RPC protocol errors, transport failures, and server-side schema
    rejections are now returned as fantasy.NewTextErrorResponse(...) with
    err == nil, so the agent loop continues and the model sees the failure
    in-band as a tool result it can reason about.
  - Context cancellation (ctx.Err() != nil) remains a critical error so
    callers can abort turns deterministically. This is the only case where
    bubbling up is correct — the caller intentionally tore the turn down
    and the agent must not keep spinning.
  - Server-side soft errors (CallToolResult{ isError: true }) and the
    happy path are unchanged.

The agent loop's MaxSteps cap already bounds the worst case for a
permanently broken MCP server, so there is no risk of unbounded retries.

Side effect: extracted a tiny mcpExecutor interface for the one method the
adapter uses (ExecuteTool), purely so the adapter is unit-testable in
isolation without standing up a full MCPToolManager + connection pool.

Behavior change note for downstream consumers: code that relied on
host.PromptResult / Stream returning a Go error containing
"mcp tool execution failed" will no longer see those errors — the
failure information is now in the assistant's final response (or in the
OnAfterToolResult / OnToolResult hooks, where IsError will be true).
Context cancellation continues to surface as an error from those calls
as before.

cmd/root.go

@@ -784,6 +784,16 @@ func runNormalMode(ctx context.Context) error {
 	}
 	defer func() { _ = kitInstance.Close() }()
 
+	// Build the "System Prompt loaded" notice shown at startup, paralleling the
+	// per-server "MCP server loaded" notifications so users can confirm that a
+	// configured prompt file was found and applied.
+	var systemPromptLoadedMsg string
+	if kitInstance.HasCustomSystemPrompt() {
+		if src := kitInstance.GetSystemPromptSource(); src != "" {
+			systemPromptLoadedMsg = "System Prompt loaded: " + src
+		}
+	}
+
 	// Extract metadata for display and app options.
 	parsedProvider, modelName, serverNames, toolNames, mcpToolCount, extensionToolCount := CollectAgentMetadata(kitInstance, mcpConfig)
 
@@ -801,6 +811,9 @@ func runNormalMode(ctx context.Context) error {
 		}
 
 		DisplayDebugConfig(cli, kitInstance, mcpConfig, parsedProvider)
+		if systemPromptLoadedMsg != "" && cli != nil {
+			cli.DisplayInfo(systemPromptLoadedMsg)
+		}
 	}
 
 	// Load existing messages from resumed/continued sessions.
@@ -840,6 +853,9 @@ func runNormalMode(ctx context.Context) error {
 
 	// Buffer for extension messages during startup (printed after startup banner).
 	var startupExtensionMessages []string
+	if systemPromptLoadedMsg != "" {
+		startupExtensionMessages = append(startupExtensionMessages, systemPromptLoadedMsg)
+	}
 
 	// Set up extension context and emit SessionStart.
 	if kitInstance.Extensions().HasExtensions() {
@@ -919,9 +935,10 @@ func runNormalMode(ctx context.Context) error {
 			source = "project"
 		}
 		skillItems = append(skillItems, ui.SkillItem{
-			Name:   s.Name,
-			Path:   s.Path,
-			Source: source,
+			Name:        s.Name,
+			Path:        s.Path,
+			Source:      source,
+			Description: s.Description,
 		})
 	}
 
@@ -960,9 +977,10 @@ func runNormalMode(ctx context.Context) error {
 				source = "project"
 			}
 			items = append(items, ui.SkillItem{
-				Name:   s.Name,
-				Path:   s.Path,
-				Source: source,
+				Name:        s.Name,
+				Path:        s.Path,
+				Source:      source,
+				Description: s.Description,
 			})
 		}
 		return items

internal/agent/mcp_adapter.go

@@ -9,12 +9,19 @@ import (
 	"github.com/mark3labs/kit/internal/tools"
 )
 
+// mcpExecutor is the subset of *tools.MCPToolManager that the adapter
+// actually uses. Extracted as an interface so the adapter is unit-testable
+// without constructing a full manager + connection pool.
+type mcpExecutor interface {
+	ExecuteTool(ctx context.Context, prefixedName, inputJSON string) (*tools.MCPToolResult, error)
+}
+
 // mcpAgentTool adapts an tools.MCPTool to the fantasy.AgentTool interface.
 // This keeps the fantasy dependency confined to the agent layer — the tools
 // package is a pure MCP client library with no LLM framework dependency.
 type mcpAgentTool struct {
 	tool            tools.MCPTool
-	manager         *tools.MCPToolManager
+	exec            mcpExecutor
 	providerOptions fantasy.ProviderOptions
 }
 
@@ -29,10 +36,26 @@ func (t *mcpAgentTool) Info() fantasy.ToolInfo {
 }
 
 // Run executes the MCP tool by delegating to the MCPToolManager.
+//
+// MCP-side failures (JSON-RPC protocol errors, transport failures, schema
+// validation rejections from the server) are surfaced to the model as soft
+// tool errors rather than escalated to a critical agent error. This matches
+// the contract that native Kit tools follow via kit.ErrorResult(...) and
+// lets the model self-correct (e.g. retry with a fixed argument shape) or
+// give up gracefully rather than aborting the turn mid-run.
+//
+// Context cancellation is the one exception: if the caller cancelled the
+// context the turn was aborted intentionally, so we propagate the ctx error
+// to let the agent loop unwind cleanly.
 func (t *mcpAgentTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error) {
-	result, err := t.manager.ExecuteTool(ctx, t.tool.Name, call.Input)
+	result, err := t.exec.ExecuteTool(ctx, t.tool.Name, call.Input)
 	if err != nil {
-		return fantasy.ToolResponse{}, fmt.Errorf("mcp tool execution failed: %w", err)
+		if ctxErr := ctx.Err(); ctxErr != nil {
+			return fantasy.ToolResponse{}, ctxErr
+		}
+		return fantasy.NewTextErrorResponse(
+			fmt.Sprintf("MCP tool %q failed: %s", t.tool.Name, err.Error()),
+		), nil
 	}
 
 	if result.IsError {
@@ -57,8 +80,8 @@ func mcpToolsToAgentTools(mcpTools []tools.MCPTool, manager *tools.MCPToolManage
 	agentTools := make([]fantasy.AgentTool, len(mcpTools))
 	for i, t := range mcpTools {
 		agentTools[i] = &mcpAgentTool{
-			tool:    t,
-			manager: manager,
+			tool: t,
+			exec: manager,
 		}
 	}
 	return agentTools

internal/agent/mcp_adapter_test.go

@@ -0,0 +1,158 @@
+package agent
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+	"time"
+
+	"charm.land/fantasy"
+
+	"github.com/mark3labs/kit/internal/tools"
+)
+
+// stubExecutor lets each test script the (result, err) pair returned by
+// ExecuteTool. The adapter holds an mcpExecutor interface, so this is the
+// only seam the tests need.
+type stubExecutor struct {
+	result *tools.MCPToolResult
+	err    error
+	// called records the last invocation for assertion.
+	called bool
+	name   string
+	input  string
+}
+
+func (s *stubExecutor) ExecuteTool(_ context.Context, prefixedName, inputJSON string) (*tools.MCPToolResult, error) {
+	s.called = true
+	s.name = prefixedName
+	s.input = inputJSON
+	return s.result, s.err
+}
+
+func newMCPAgentTool(exec mcpExecutor, name string) *mcpAgentTool {
+	return &mcpAgentTool{
+		tool: tools.MCPTool{Name: name},
+		exec: exec,
+	}
+}
+
+// Manager-side Go errors (JSON-RPC protocol errors, transport failures,
+// schema validation rejections from the MCP server) must be surfaced to
+// the model as soft tool errors so the agent loop can keep going. Aborting
+// the turn would discard all prior tool results — see issue #N.
+func TestMCPAgentTool_RPCErrorBecomesSoftError(t *testing.T) {
+	exec := &stubExecutor{
+		err: errors.New("MCP error -32602: Invalid params: missing field \"task\""),
+	}
+	tool := newMCPAgentTool(exec, "pubmed__search")
+
+	resp, err := tool.Run(context.Background(), fantasy.ToolCall{
+		ID:    "call-1",
+		Name:  "pubmed__search",
+		Input: `{"query":"foo"}`,
+	})
+
+	if err != nil {
+		t.Fatalf("expected nil error (soft), got %v", err)
+	}
+	if !resp.IsError {
+		t.Fatalf("expected IsError=true, got false")
+	}
+	if !strings.Contains(resp.Content, "pubmed__search") {
+		t.Errorf("expected tool name in error content, got %q", resp.Content)
+	}
+	if !strings.Contains(resp.Content, "-32602") {
+		t.Errorf("expected underlying error text in content, got %q", resp.Content)
+	}
+}
+
+// Context cancellation is the one error that must remain critical: it
+// means the caller intentionally aborted, and the agent loop needs to
+// unwind cleanly rather than burning more steps.
+func TestMCPAgentTool_CtxCancelStaysCritical(t *testing.T) {
+	exec := &stubExecutor{
+		// Real managers typically return ctx.Err() (or a wrapper) when the
+		// context is cancelled mid-call.
+		err: context.Canceled,
+	}
+	tool := newMCPAgentTool(exec, "slow__tool")
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	resp, err := tool.Run(ctx, fantasy.ToolCall{Name: "slow__tool"})
+
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("expected context.Canceled, got %v", err)
+	}
+	if resp.IsError || resp.Content != "" {
+		t.Errorf("expected empty response on critical error, got IsError=%v Content=%q", resp.IsError, resp.Content)
+	}
+}
+
+// Deadline-exceeded behaves the same as cancellation: ctx.Err() is
+// non-nil, so the adapter must propagate the critical error rather than
+// converting the executor's error into a soft response.
+func TestMCPAgentTool_CtxDeadlineStaysCritical(t *testing.T) {
+	exec := &stubExecutor{err: context.DeadlineExceeded}
+	tool := newMCPAgentTool(exec, "slow__tool")
+
+	ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(-time.Second))
+	defer cancel()
+
+	resp, err := tool.Run(ctx, fantasy.ToolCall{Name: "slow__tool"})
+	if !errors.Is(err, context.DeadlineExceeded) {
+		t.Fatalf("expected context.DeadlineExceeded, got %v", err)
+	}
+	if resp.IsError || resp.Content != "" {
+		t.Errorf("expected empty response on critical error, got IsError=%v Content=%q", resp.IsError, resp.Content)
+	}
+}
+
+// Server-side soft errors (CallToolResult{ isError: true }) must continue
+// to flow through as soft errors — this was the existing behavior and
+// must not regress.
+func TestMCPAgentTool_ServerIsErrorRemainsSoftError(t *testing.T) {
+	exec := &stubExecutor{
+		result: &tools.MCPToolResult{
+			IsError: true,
+			Content: "search service is rate limited; try again in 30s",
+		},
+	}
+	tool := newMCPAgentTool(exec, "pubmed__search")
+
+	resp, err := tool.Run(context.Background(), fantasy.ToolCall{Name: "pubmed__search"})
+	if err != nil {
+		t.Fatalf("expected nil error, got %v", err)
+	}
+	if !resp.IsError {
+		t.Fatalf("expected IsError=true, got false")
+	}
+	if resp.Content != "search service is rate limited; try again in 30s" {
+		t.Errorf("expected pass-through content, got %q", resp.Content)
+	}
+}
+
+// Happy path: ordinary successful tool result is passed through unchanged.
+func TestMCPAgentTool_SuccessIsPassthrough(t *testing.T) {
+	exec := &stubExecutor{
+		result: &tools.MCPToolResult{
+			IsError: false,
+			Content: `{"hits":3}`,
+		},
+	}
+	tool := newMCPAgentTool(exec, "pubmed__search")
+
+	resp, err := tool.Run(context.Background(), fantasy.ToolCall{Name: "pubmed__search"})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if resp.IsError {
+		t.Fatalf("expected IsError=false")
+	}
+	if resp.Content != `{"hits":3}` {
+		t.Errorf("expected pass-through content, got %q", resp.Content)
+	}
+}

internal/app/app.go

@@ -78,6 +78,13 @@ type App struct {
 	// (~1 frame) so new updates are always let through once the TUI has had a
 	// chance to process the pending event.
 	widgetUpdatePending atomic.Bool
+
+	// steerDrainFn is the test seam used by releaseBusyAfterCompact to pull
+	// any steer messages that arrived during compaction. In production it is
+	// nil and the helper falls back to a.opts.Kit.DrainSteer(); tests that
+	// need to exercise the steer-drain path without standing up a full
+	// *kit.Kit can set this field directly to inject fake items.
+	steerDrainFn func() []queueItem
 }
 
 // New creates a new App with the provided options and pre-loaded messages.
@@ -356,6 +363,10 @@ func (a *App) AddContextMessage(text string) {
 // tea.Program. customInstructions is optional text appended to the summary
 // prompt (e.g. "Focus on the API design decisions").
 //
+// Any prompts queued via Run/RunWithFiles or steering messages injected via
+// Steer/SteerWithFiles while compaction is running are flushed automatically
+// once compaction completes (see releaseBusyAfterCompact).
+//
 // Satisfies ui.AppController.
 func (a *App) CompactConversation(customInstructions string) error {
 	a.mu.Lock()
@@ -377,11 +388,7 @@ func (a *App) CompactConversation(customInstructions string) error {
 
 	go func() {
 		defer a.wg.Done()
-		defer func() {
-			a.mu.Lock()
-			a.busy = false
-			a.mu.Unlock()
-		}()
+		defer a.releaseBusyAfterCompact()
 
 		// Subscribe to SDK events for streaming compaction summary to the TUI.
 		sendFn := func(msg tea.Msg) {
@@ -420,6 +427,9 @@ func (a *App) CompactConversation(customInstructions string) error {
 // CompactAsync is like CompactConversation but calls onComplete/onError
 // callbacks instead of sending TUI events. Used by the extension API's
 // ctx.Compact() which needs callback-based notification.
+//
+// Like CompactConversation, any prompts/steer messages received during
+// compaction are flushed automatically once compaction finishes.
 func (a *App) CompactAsync(customInstructions string, onComplete func(), onError func(string)) error {
 	a.mu.Lock()
 	if a.closed {
@@ -440,11 +450,7 @@ func (a *App) CompactAsync(customInstructions string, onComplete func(), onError
 
 	go func() {
 		defer a.wg.Done()
-		defer func() {
-			a.mu.Lock()
-			a.busy = false
-			a.mu.Unlock()
-		}()
+		defer a.releaseBusyAfterCompact()
 
 		// Subscribe to SDK events for streaming compaction summary to the TUI.
 		sendFn := func(msg tea.Msg) {
@@ -489,6 +495,81 @@ func (a *App) CompactAsync(customInstructions string, onComplete func(), onError
 	return nil
 }
 
+// releaseBusyAfterCompact is the deferred tail that runs at the end of every
+// compaction goroutine (success, error, or panic-after-recover paths). It
+// flips a.busy back to false, but before doing so it checks whether any
+// prompts piled up while compaction was running:
+//
+//   - Run/RunWithFiles append to a.queue when a.busy is set.
+//   - Steer/SteerWithFiles deposit messages into the SDK steer channel via
+//     Kit.InjectSteerWithFiles when a.busy is set.
+//
+// Without this hand-off the queue would sit idle until the user submits
+// another prompt — see issue #27. If we find anything pending we keep busy
+// set, splice the steer messages to the front of the queue, and start a
+// fresh drainQueue goroutine to deliver them as a single batched turn.
+func (a *App) releaseBusyAfterCompact() {
+	// Pull steer messages outside the app mutex; DrainSteer takes its own
+	// internal lock and we don't want to nest the two. The test seam
+	// (a.steerDrainFn) takes precedence so unit tests can inject fake
+	// steer items without a real *kit.Kit.
+	var steerItems []queueItem
+	switch {
+	case a.steerDrainFn != nil:
+		steerItems = a.steerDrainFn()
+	case a.opts.Kit != nil:
+		if leftover := a.opts.Kit.DrainSteer(); len(leftover) > 0 {
+			steerItems = make([]queueItem, len(leftover))
+			for i, sm := range leftover {
+				steerItems[i] = queueItem{Prompt: sm.Text, Files: sm.Files}
+			}
+		}
+	}
+
+	a.mu.Lock()
+	// If the app was closed while compaction was running, drop everything
+	// and just clear busy. Run/Steer would have rejected new items already
+	// after Close(), but this guards against in-flight items that slipped
+	// in just before closed was set.
+	if a.closed {
+		a.queue = a.queue[:0]
+		a.busy = false
+		a.mu.Unlock()
+		return
+	}
+
+	// Combine steer-channel items (front) with the in-memory queue (back).
+	// Steer messages are placed first so they retain their "act now"
+	// semantics relative to ordinary queued prompts that arrived later.
+	pending := append(steerItems, a.queue...)
+	a.queue = a.queue[:0]
+
+	if len(pending) == 0 {
+		a.busy = false
+		a.mu.Unlock()
+		return
+	}
+
+	// Hand off to drainQueue: it will pick up the first item directly and
+	// scoop the rest from a.queue on its first iteration.
+	first := pending[0]
+	if len(pending) > 1 {
+		a.queue = append(a.queue, pending[1:]...)
+	}
+	// Stay busy across the goroutine swap.
+	a.wg.Add(1)
+	a.mu.Unlock()
+
+	// Notify the UI that steer-channel messages were consumed so the
+	// steering badge can clear; ordinary queued prompts will be reflected
+	// by the QueueUpdatedEvent that drainQueue emits as it picks them up.
+	if len(steerItems) > 0 {
+		a.sendEvent(SteerConsumedEvent{})
+	}
+
+	go a.drainQueue(first)
+}
+
 // --------------------------------------------------------------------------
 // Non-interactive execution
 // --------------------------------------------------------------------------

internal/app/app_test.go

@@ -763,3 +763,209 @@ func TestFormatMaxTokensTruncatedMessage_NoKit(t *testing.T) {
 		}
 	}
 }
+
+// --------------------------------------------------------------------------
+// releaseBusyAfterCompact (issue #27)
+// --------------------------------------------------------------------------
+
+// TestReleaseBusyAfterCompact_flushesQueuedMessages is a regression test for
+// issue #27: messages queued via Run() while /compact is running used to sit
+// in a.queue indefinitely until the user typed another prompt. After the fix
+// the deferred releaseBusyAfterCompact tail picks up any pending items and
+// dispatches drainQueue automatically.
+//
+// We simulate the compaction completion path directly (bypassing the SDK)
+// by toggling busy=true, populating the queue exactly as Run() would have
+// during compaction, and then invoking releaseBusyAfterCompact.
+func TestReleaseBusyAfterCompact_flushesQueuedMessages(t *testing.T) {
+	stub := newStubWithFuncs(
+		func(ctx context.Context) (*kit.TurnResult, error) {
+			return turnResult("compacted then drained"), nil
+		},
+	)
+	app := newTestApp(stub)
+	defer app.Close()
+
+	// Simulate the state at the start of the compaction tail: busy is set
+	// and a couple of prompts have piled up in the queue while we were
+	// summarising. (Run() would have appended them and returned a queue
+	// length > 0 to the caller.)
+	app.mu.Lock()
+	app.busy = true
+	app.queue = append(app.queue,
+		queueItem{Prompt: "queued during compact #1"},
+		queueItem{Prompt: "queued during compact #2"},
+	)
+	app.mu.Unlock()
+
+	// Invoke the deferred tail directly. It should kick off drainQueue.
+	app.releaseBusyAfterCompact()
+
+	// drainQueue runs in a goroutine. Wait for the app to come back to idle.
+	ok := waitForCondition(2*time.Second, func() bool {
+		app.mu.Lock()
+		defer app.mu.Unlock()
+		return !app.busy
+	})
+	if !ok {
+		t.Fatal("app did not become idle after releaseBusyAfterCompact: queue not drained")
+	}
+
+	// Wait for any in-flight goroutine to finish before reading state.
+	app.wg.Wait()
+
+	if got := app.QueueLength(); got != 0 {
+		t.Fatalf("expected empty queue after drain, got %d", got)
+	}
+	if n := stub.callCount(); n == 0 {
+		t.Fatalf("expected stub PromptFunc to fire at least once after compact, got %d calls", n)
+	}
+}
+
+// TestReleaseBusyAfterCompact_idleWhenQueueEmpty verifies that with no
+// pending messages the helper just clears busy and does NOT spawn a
+// drainQueue goroutine (no spurious agent turn).
+func TestReleaseBusyAfterCompact_idleWhenQueueEmpty(t *testing.T) {
+	stub := newStub()
+	app := newTestApp(stub)
+	defer app.Close()
+
+	app.mu.Lock()
+	app.busy = true
+	app.mu.Unlock()
+
+	app.releaseBusyAfterCompact()
+
+	app.mu.Lock()
+	busy := app.busy
+	app.mu.Unlock()
+	if busy {
+		t.Fatal("expected busy=false after releaseBusyAfterCompact with empty queue")
+	}
+
+	// Give any rogue goroutine a moment to (incorrectly) call PromptFunc.
+	time.Sleep(50 * time.Millisecond)
+	if n := stub.callCount(); n != 0 {
+		t.Fatalf("expected 0 PromptFunc calls when queue empty, got %d", n)
+	}
+}
+
+// TestReleaseBusyAfterCompact_splicesSteerAheadOfQueue exercises the SDK
+// steer-drain branch of releaseBusyAfterCompact (issue #27 follow-up).
+//
+// Production wires a.opts.Kit.DrainSteer() to pull messages that arrived via
+// Steer/SteerWithFiles during compaction, but Options.Kit is *kit.Kit (a
+// concrete struct) so unit tests cannot stand up a real instance without a
+// full LLM backend. The test uses the unexported steerDrainFn seam to inject
+// fake steer items, then asserts that:
+//
+//   - Steer items are dispatched ahead of any prompts that piled up in
+//     a.queue (steer retains "act now" priority over ordinary queued
+//     prompts), and
+//   - the helper still hands off to drainQueue so the steer item actually
+//     fires (the previous behaviour left them stranded — see #27).
+func TestReleaseBusyAfterCompact_splicesSteerAheadOfQueue(t *testing.T) {
+	var pmu sync.Mutex
+	var firstPrompt string
+	stub := newStubWithFuncs(
+		func(ctx context.Context) (*kit.TurnResult, error) {
+			return turnResult("steer dispatched"), nil
+		},
+	)
+	// Wrap PromptFunc so we can capture the prompt text the stub receives
+	// (newStubWithFuncs's fns ignore prompt; we need it to verify ordering).
+	capturingPrompt := func(ctx context.Context, prompt string) (*kit.TurnResult, error) {
+		pmu.Lock()
+		if firstPrompt == "" {
+			firstPrompt = prompt
+		}
+		pmu.Unlock()
+		return stub.fn(ctx, prompt)
+	}
+	app := New(Options{PromptFunc: capturingPrompt}, nil)
+	defer app.Close()
+
+	// Inject fake steer items via the test seam. In production the same
+	// items would have been delivered through Kit.InjectSteerWithFiles
+	// during /compact and pulled by DrainSteer here.
+	app.steerDrainFn = func() []queueItem {
+		return []queueItem{
+			{Prompt: "steer-1"},
+			{Prompt: "steer-2"},
+		}
+	}
+
+	// Simulate the state at the end of compaction: busy is set and a couple
+	// of regular Run() prompts have piled up after the steer messages.
+	app.mu.Lock()
+	app.busy = true
+	app.queue = append(app.queue,
+		queueItem{Prompt: "queued-1"},
+		queueItem{Prompt: "queued-2"},
+	)
+	app.mu.Unlock()
+
+	app.releaseBusyAfterCompact()
+
+	// Wait for the dispatched batch to complete.
+	ok := waitForCondition(2*time.Second, func() bool {
+		app.mu.Lock()
+		defer app.mu.Unlock()
+		return !app.busy
+	})
+	if !ok {
+		t.Fatal("app did not become idle after steer-spliced releaseBusyAfterCompact")
+	}
+	app.wg.Wait()
+
+	// drainQueue picks up `first` directly and batches the rest. With
+	// PromptFunc set, executeBatch invokes us with items[0] only — that
+	// item must be the first steer message, proving steer items were
+	// spliced ahead of the previously queued prompts.
+	pmu.Lock()
+	got := firstPrompt
+	pmu.Unlock()
+	if got != "steer-1" {
+		t.Fatalf("expected first dispatched prompt to be steer item %q (steer items must come before queued prompts), got %q",
+			"steer-1", got)
+	}
+
+	// Queue should be fully drained and PromptFunc must have actually fired.
+	if n := app.QueueLength(); n != 0 {
+		t.Fatalf("expected empty queue after drain, got %d entries", n)
+	}
+	if n := stub.callCount(); n == 0 {
+		t.Fatal("expected stub PromptFunc to fire at least once after splice")
+	}
+}
+
+// TestReleaseBusyAfterCompact_dropsQueueWhenClosed verifies that if the app
+// was closed during compaction the helper discards any pending items rather
+// than spawning drainQueue against a torn-down App.
+func TestReleaseBusyAfterCompact_dropsQueueWhenClosed(t *testing.T) {
+	stub := newStub()
+	app := newTestApp(stub)
+
+	app.mu.Lock()
+	app.busy = true
+	app.queue = append(app.queue, queueItem{Prompt: "would have run"})
+	app.closed = true
+	app.mu.Unlock()
+
+	app.releaseBusyAfterCompact()
+
+	app.mu.Lock()
+	busy := app.busy
+	qLen := len(app.queue)
+	app.mu.Unlock()
+	if busy {
+		t.Fatal("expected busy=false even when closed")
+	}
+	if qLen != 0 {
+		t.Fatalf("expected queue cleared on closed app, got %d entries", qLen)
+	}
+	time.Sleep(20 * time.Millisecond)
+	if n := stub.callCount(); n != 0 {
+		t.Fatalf("expected 0 PromptFunc calls on closed app, got %d", n)
+	}
+}

internal/ui/model.go

@@ -129,9 +129,10 @@ type AppController interface {
 // SkillItem holds display metadata about a loaded skill for the startup
 // [Skills] section. Built by the CLI layer from the SDK's []*kit.Skill.
 type SkillItem struct {
-	Name   string // Skill name (e.g. "btca-cli").
-	Path   string // Absolute path to the skill file.
-	Source string // "project" or "user" (global).
+	Name        string // Skill name (e.g. "btca-cli").
+	Path        string // Absolute path to the skill file.
+	Source      string // "project" or "user" (global).
+	Description string // Short summary used in autocomplete and help.
 }
 
 // MCPPromptInfo describes an MCP prompt for display in the TUI (autocomplete,
@@ -912,6 +913,20 @@ func NewAppModel(appCtrl AppController, opts AppModelOptions) *AppModel {
 		}
 	}
 
+	// Merge skills into autocomplete as /skill:<name> commands. Skills accept
+	// optional trailing args, so HasArgs is true — Enter populates the input
+	// with "/skill:name " rather than auto-submitting.
+	if ic, ok := m.input.(*InputComponent); ok && len(opts.SkillItems) > 0 {
+		for _, s := range opts.SkillItems {
+			ic.commands = append(ic.commands, commands.SlashCommand{
+				Name:        "/skill:" + s.Name,
+				Description: formatSkillDescription(s),
+				Category:    "Skills",
+				HasArgs:     true,
+			})
+		}
+	}
+
 	// Merge MCP prompts into autocomplete as /<server>:<prompt> commands.
 	if ic, ok := m.input.(*InputComponent); ok && len(opts.MCPPrompts) > 0 {
 		for _, p := range opts.MCPPrompts {
@@ -3395,13 +3410,46 @@ func (m *AppModel) refreshPromptTemplates() {
 	}
 }
 
-// refreshSkillItems reloads skill items from the provider callback.
-// Called on ContentReloadEvent.
+// refreshSkillItems reloads skill items from the provider callback and
+// updates the autocomplete entries. Called on ContentReloadEvent.
 func (m *AppModel) refreshSkillItems() {
 	if m.getSkillItems == nil {
 		return
 	}
-	m.skillItems = m.getSkillItems()
+	newItems := m.getSkillItems()
+	m.skillItems = newItems
+
+	if ic, ok := m.input.(*InputComponent); ok {
+		// Remove old Skills commands and add fresh ones.
+		var kept []commands.SlashCommand
+		for _, sc := range ic.commands {
+			if sc.Category != "Skills" {
+				kept = append(kept, sc)
+			}
+		}
+		for _, s := range newItems {
+			kept = append(kept, commands.SlashCommand{
+				Name:        "/skill:" + s.Name,
+				Description: formatSkillDescription(s),
+				Category:    "Skills",
+				HasArgs:     true,
+			})
+		}
+		ic.commands = kept
+	}
+}
+
+// formatSkillDescription returns the autocomplete description for a skill,
+// prefixed with [project] or [user] so users can tell colliding names apart.
+func formatSkillDescription(s SkillItem) string {
+	prefix := "[user]"
+	if s.Source == "project" {
+		prefix = "[project]"
+	}
+	if s.Description == "" {
+		return prefix
+	}
+	return prefix + " " + s.Description
 }
 
 // refreshMCPPrompts reloads MCP prompts from the provider callback and

pkg/kit/kit.go

@@ -58,6 +58,9 @@ type Kit struct {
 	// When false, per-model system prompts from modelSettings/customModels
 	// can replace the default prompt on model switch.
 	hasCustomSystemPrompt bool
+	// systemPromptSource holds the raw configured value (file path or text)
+	// when hasCustomSystemPrompt is true; empty when the built-in default is in use.
+	systemPromptSource string
 
 	// Hook registries — interception layer (see hooks.go).
 	beforeToolCall  *hookRegistry[BeforeToolCallHook, BeforeToolCallResult]
@@ -632,6 +635,21 @@ func (m *Kit) SetModel(ctx context.Context, modelString string) error {
 	return nil
 }
 
+// HasCustomSystemPrompt reports whether the user explicitly configured a system
+// prompt via --system-prompt, a config file entry, or SDK Options.SystemPrompt.
+// When false, the built-in default (or a per-model override) is in use and can
+// be replaced transparently on model switch.
+func (m *Kit) HasCustomSystemPrompt() bool {
+	return m.hasCustomSystemPrompt
+}
+
+// GetSystemPromptSource returns the raw configured value — a file path or
+// inline text — when HasCustomSystemPrompt is true; returns an empty string
+// when the built-in default prompt is active.
+func (m *Kit) GetSystemPromptSource() string {
+	return m.systemPromptSource
+}
+
 // composeSystemPrompt takes a base system prompt and composes it with the
 // current runtime context: AGENTS.md content, skills metadata, and date/cwd.
 // This mirrors the composition done during Kit.New() initialization.
@@ -1179,6 +1197,7 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		maxSteps              int
 		streaming             bool
 		hasCustomSystemPrompt bool
+		systemPromptSource    string
 	)
 
 	if err := func() error {
@@ -1285,13 +1304,27 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		// explicitly set system-prompt, use the per-model prompt as the
 		// base instead of the global default.
 		{
-			basePrompt := viper.GetString("system-prompt")
+			rawPromptInput := viper.GetString("system-prompt")
+
+			// Resolve a file path to its content so PromptBuilder receives the
+			// actual prompt text rather than a literal path string. Without this,
+			// when system-prompt is set to a file path in the config file or via
+			// --system-prompt, the path itself becomes the effective system prompt
+			// sent to the model (LoadSystemPrompt only ran later, after viper had
+			// been overwritten with the augmented base text).
+			basePrompt, _ := config.LoadSystemPrompt(rawPromptInput)
+			if basePrompt == "" {
+				basePrompt = rawPromptInput
+			}
 
 			// Track whether the user explicitly configured a custom system
 			// prompt. When they haven't (basePrompt is the built-in default
 			// or empty), per-model system prompts can replace it on switch.
 			userSetSystemPrompt := basePrompt != "" && basePrompt != defaultSystemPrompt
 			hasCustomSystemPrompt = userSetSystemPrompt
+			if hasCustomSystemPrompt {
+				systemPromptSource = rawPromptInput
+			}
 
 			// Check for per-model system prompt override when no explicit
 			// global system-prompt was configured by the user.
@@ -1500,6 +1533,7 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		opts:                  opts,
 		mcpConfig:             mcpConfig,
 		hasCustomSystemPrompt: hasCustomSystemPrompt,
+		systemPromptSource:    systemPromptSource,
 		beforeToolCall:        beforeToolCall,
 		afterToolResult:       afterToolResult,
 		beforeTurn:            beforeTurn,

pkg/kit/kit_test.go

@@ -3,6 +3,7 @@ package kit_test
 import (
 	"context"
 	"os"
+	"strings"
 	"testing"
 
 	"github.com/spf13/viper"
@@ -306,3 +307,92 @@ func TestSessionManagement(t *testing.T) {
 // resetViper wipes viper's global state so a test case doesn't leak
 // viper.Set() calls into the next one. Used via defer in subtests.
 func resetViper() { viper.Reset() }
+
+// TestNewSystemPromptFilePath is a regression test for issue #25.
+//
+// When Options.SystemPrompt (or the --system-prompt flag / config entry) is a
+// file path, Kit must resolve the path to its file contents *before* the
+// PromptBuilder composes the runtime context. Previously the path string
+// itself was used verbatim as the base prompt, so the LLM received the path —
+// not the prompt — as its system message.
+func TestNewSystemPromptFilePath(t *testing.T) {
+	if os.Getenv("ANTHROPIC_API_KEY") == "" {
+		t.Skip("Skipping test: ANTHROPIC_API_KEY not set")
+	}
+	defer resetViper()
+
+	const promptContent = "You are a strict regression-test persona. Marker: KIT-25-OK"
+
+	tmpFile, err := os.CreateTemp(t.TempDir(), "kit-system-prompt-*.md")
+	if err != nil {
+		t.Fatalf("failed to create temp prompt file: %v", err)
+	}
+	if _, err := tmpFile.WriteString(promptContent); err != nil {
+		t.Fatalf("failed to write temp prompt file: %v", err)
+	}
+	if err := tmpFile.Close(); err != nil {
+		t.Fatalf("failed to close temp prompt file: %v", err)
+	}
+
+	ctx := context.Background()
+	host, err := kit.New(ctx, &kit.Options{
+		Model:        "anthropic/claude-sonnet-4-5-20250929",
+		SystemPrompt: tmpFile.Name(),
+		Quiet:        true,
+		NoSession:    true,
+	})
+	if err != nil {
+		t.Fatalf("Failed to create Kit with system-prompt file: %v", err)
+	}
+	defer func() { _ = host.Close() }()
+
+	if !host.HasCustomSystemPrompt() {
+		t.Error("HasCustomSystemPrompt() = false; want true when --system-prompt is set")
+	}
+	if got, want := host.GetSystemPromptSource(), tmpFile.Name(); got != want {
+		t.Errorf("GetSystemPromptSource() = %q; want %q", got, want)
+	}
+
+	// The composed system prompt is written back to viper after PromptBuilder
+	// runs. It must contain the file's contents, not the file path.
+	composed := viper.GetString("system-prompt")
+	if !strings.Contains(composed, promptContent) {
+		t.Errorf("composed system-prompt does not contain file contents\n  composed = %q\n  want substring = %q", composed, promptContent)
+	}
+	if strings.TrimSpace(composed) == tmpFile.Name() {
+		t.Errorf("composed system-prompt is the file path verbatim (%q); LoadSystemPrompt was not applied before PromptBuilder", composed)
+	}
+}
+
+// TestNewSystemPromptInline confirms that inline system-prompt strings still
+// flow through unchanged after the file-path resolution change.
+func TestNewSystemPromptInline(t *testing.T) {
+	if os.Getenv("ANTHROPIC_API_KEY") == "" {
+		t.Skip("Skipping test: ANTHROPIC_API_KEY not set")
+	}
+	defer resetViper()
+
+	const inline = "You are a concise inline-prompt persona."
+
+	ctx := context.Background()
+	host, err := kit.New(ctx, &kit.Options{
+		Model:        "anthropic/claude-sonnet-4-5-20250929",
+		SystemPrompt: inline,
+		Quiet:        true,
+		NoSession:    true,
+	})
+	if err != nil {
+		t.Fatalf("Failed to create Kit with inline system-prompt: %v", err)
+	}
+	defer func() { _ = host.Close() }()
+
+	if !host.HasCustomSystemPrompt() {
+		t.Error("HasCustomSystemPrompt() = false; want true for inline prompt")
+	}
+	if got := host.GetSystemPromptSource(); got != inline {
+		t.Errorf("GetSystemPromptSource() = %q; want %q", got, inline)
+	}
+	if composed := viper.GetString("system-prompt"); !strings.Contains(composed, inline) {
+		t.Errorf("composed system-prompt missing inline content; got %q", composed)
+	}
+}

specs/README.md

@@ -1,5 +0,0 @@
-# Specs
-
-| Spec | Status | Description |
-|------|--------|-------------|
-| [unified-bubbletea-architecture](unified-bubbletea-architecture.md) | Draft | Replace micro-program pattern with single Bubble Tea program + thick app layer |