test(sessions): assert the window key with its value, not the number alone

strings.Contains(out, "40") finds it inside "4096", so the start_line
assertion passed whether or not start_line survived the projection, which
is the opposite of what it claimed. The rendered arguments carry the pairs,
so the assertions name them.

internal/acp/agent.go

@@ -14,6 +14,7 @@ import (
 	"github.com/Gitlawb/zero/internal/providercatalog"
 	"github.com/Gitlawb/zero/internal/providermodelcatalog"
 	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
+	"github.com/Gitlawb/zero/internal/redaction"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/sessions"
 	"github.com/Gitlawb/zero/internal/tools"
@@ -30,10 +31,12 @@ type Deps struct {
 	DiscoverModels func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
 	NewProvider    func(profile config.ProviderProfile) (zeroruntime.Provider, error)
 	RunAgent       func(ctx context.Context, prompt string, provider zeroruntime.Provider, opts agent.Options) (agent.Result, error)
-	// BuildWorkspace builds the SCOPED tool registry and the sandbox engine for a
-	// validated workspace root, so ACP shell tools (bash/exec_command) are confined
-	// exactly like the exec surface — never run unconfined on the host.
-	BuildWorkspace func(workspaceRoot string, resolved config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error)
+	// BuildWorkspace creates the per-turn scoped tool workspace for a validated
+	// workspace root. Its Close method releases any resources the registry owns
+	// (notably MCP server connections) after the turn completes or is cancelled.
+	// Shell tools remain confined exactly like the exec surface — never unconfined
+	// on the host.
+	BuildWorkspace func(ctx context.Context, workspaceRoot string, resolved config.ResolvedConfig, mode agent.PermissionMode) (*Workspace, error)
 	// ResolveWorkspaceRoot validates + normalizes a client-supplied cwd (must be an
 	// existing directory; never the bare root). It is the file-tool confinement root.
 	ResolveWorkspaceRoot func(cwd string) (string, error)
@@ -41,6 +44,27 @@ type Deps struct {
 	AgentInfo            Implementation
 }
 
+// Workspace is the per-turn execution environment passed to the agent. ACP does
+// not retain it across turns because MCP connections belong to the registry that
+// advertised their tools; keeping a stale registry after a cancelled turn would
+// leak its server process and its permission state into the next turn.
+type Workspace struct {
+	Registry       *tools.Registry
+	Sandbox        *sandbox.Engine
+	DeferThreshold int
+	Notices        []string
+	Cleanup        func() error
+}
+
+// Close releases resources created with the workspace. A nil cleanup is valid
+// for core-only workspaces.
+func (w *Workspace) Close() error {
+	if w == nil || w.Cleanup == nil {
+		return nil
+	}
+	return w.Cleanup()
+}
+
 // Agent is the ACP agent server bound to one JSON-RPC connection (one editor).
 type Agent struct {
 	conn *Conn
@@ -241,13 +265,34 @@ func (a *Agent) runTurn(ctx context.Context, sess *acpSession, userText string,
 	if err != nil {
 		return "", RPCError(codeInternalError, "provider: "+err.Error())
 	}
+	mode := sess.currentMode()
 	// Build the SCOPED registry + sandbox engine for this session's workspace so
 	// shell/file tools are confined to the workspace exactly like the exec surface.
-	registry, sandboxEngine, err := a.deps.BuildWorkspace(sess.cwd, resolved)
+	// The workspace can also own MCP connections, which must be closed after every
+	// turn even if the agent run returns an error or the client cancels it.
+	workspace, err := a.deps.BuildWorkspace(ctx, sess.cwd, resolved, mode)
 	if err != nil {
 		return "", RPCError(codeInternalError, "workspace: "+err.Error())
 	}
+	if workspace == nil || workspace.Registry == nil {
+		if workspace != nil {
+			_ = workspace.Close()
+		}
+		return "", RPCError(codeInternalError, "workspace: missing tool registry")
+	}
+	defer func() {
+		if err := workspace.Close(); err != nil {
+			log.Printf("acp: close workspace: %v", err)
+		}
+	}()
+	registry := workspace.Registry
+	sandboxEngine := workspace.Sandbox
 	note := &notifier{conn: a.conn, sessionID: sess.id}
+	for _, notice := range workspace.Notices {
+		if text := strings.TrimSpace(redaction.RedactString(notice, redaction.Options{})); text != "" {
+			note.text("\n\n[zero warning] " + text + "\n")
+		}
+	}
 
 	opts := agent.Options{
 		Cwd:            sess.cwd,
@@ -256,7 +301,8 @@ func (a *Agent) runTurn(ctx context.Context, sess *acpSession, userText string,
 		Model:          resolved.Provider.Model,
 		Registry:       registry,
 		Sandbox:        sandboxEngine,
-		PermissionMode: sess.currentMode(),
+		PermissionMode: mode,
+		DeferThreshold: workspace.DeferThreshold,
 		MaxTurns:       resolved.MaxTurns,
 		Images:         images,
 		OnText:         note.text,

internal/acp/agent_test.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"slices"
 	"strings"
 	"testing"
 	"time"
@@ -35,6 +36,35 @@ func (f fakeProvider) StreamCompletion(_ context.Context, _ zeroruntime.Completi
 	return ch, nil
 }
 
+type acpDeferredPromptTool struct{ name string }
+
+func (t acpDeferredPromptTool) Name() string           { return t.name }
+func (acpDeferredPromptTool) Description() string      { return "deferred MCP test tool" }
+func (acpDeferredPromptTool) Parameters() tools.Schema { return tools.Schema{Type: "object"} }
+func (acpDeferredPromptTool) Deferred() bool           { return true }
+func (acpDeferredPromptTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK, Output: "ok"}
+}
+func (acpDeferredPromptTool) Safety() tools.Safety {
+	return tools.Safety{
+		SideEffect: tools.SideEffectNetwork, Permission: tools.PermissionPrompt,
+		AdvertiseInAuto: true,
+	}
+}
+
+type captureToolsProvider struct {
+	requests chan zeroruntime.CompletionRequest
+}
+
+func (p captureToolsProvider) StreamCompletion(_ context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	p.requests <- request
+	ch := make(chan zeroruntime.StreamEvent, 2)
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: "done"}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+	close(ch)
+	return ch, nil
+}
+
 func testDeps(t *testing.T) Deps {
 	t.Helper()
 	store := sessions.NewStore(sessions.StoreOptions{RootDir: t.TempDir()})
@@ -53,10 +83,10 @@ func testDeps(t *testing.T) Deps {
 			return fakeProvider{text: "Hello from ZERO"}, nil
 		},
 		RunAgent: agent.Run,
-		BuildWorkspace: func(string, config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error) {
+		BuildWorkspace: func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
 			r := tools.NewRegistry()
 			r.Register(tools.NewUpdatePlanTool())
-			return r, nil, nil
+			return &Workspace{Registry: r}, nil
 		},
 		ResolveWorkspaceRoot: func(cwd string) (string, error) { return cwd, nil },
 		Store:                store,
@@ -468,8 +498,8 @@ func TestACPRunTurnWiresSandboxAndScopedRegistry(t *testing.T) {
 	reg := tools.NewRegistry()
 	reg.Register(tools.NewUpdatePlanTool())
 	engine := sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: t.TempDir()})
-	deps.BuildWorkspace = func(string, config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error) {
-		return reg, engine, nil
+	deps.BuildWorkspace = func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
+		return &Workspace{Registry: reg, Sandbox: engine}, nil
 	}
 	var captured agent.Options
 	deps.RunAgent = func(_ context.Context, _ string, _ zeroruntime.Provider, opts agent.Options) (agent.Result, error) {
@@ -497,6 +527,124 @@ func TestACPRunTurnWiresSandboxAndScopedRegistry(t *testing.T) {
 	}
 }
 
+func TestACPRunTurnUsesWorkspaceDeferThresholdAtProviderBoundary(t *testing.T) {
+	for _, tc := range []struct {
+		name      string
+		threshold int
+		want      []string
+	}{
+		{name: "explicit zero disables", threshold: 0, want: []string{"mcp_docs_alpha", "mcp_docs_beta"}},
+		{name: "below threshold stays eager", threshold: 3, want: []string{"mcp_docs_alpha", "mcp_docs_beta"}},
+		{name: "at threshold defers", threshold: 2, want: []string{tools.ToolSearchToolName}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			requests := make(chan zeroruntime.CompletionRequest, 4)
+			deps := testDeps(t)
+			deps.NewProvider = func(config.ProviderProfile) (zeroruntime.Provider, error) {
+				return captureToolsProvider{requests: requests}, nil
+			}
+			deps.BuildWorkspace = func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
+				registry := tools.NewRegistry()
+				registry.Register(acpDeferredPromptTool{name: "mcp_docs_alpha"})
+				registry.Register(acpDeferredPromptTool{name: "mcp_docs_beta"})
+				registry.Register(tools.NewToolSearchTool(registry))
+				return &Workspace{Registry: registry, DeferThreshold: tc.threshold}, nil
+			}
+
+			h := newHarness(t, deps)
+			defer h.stop()
+			ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+			defer cancel()
+			var created NewSessionResult
+			if err := h.client.Call(ctx, MethodSessionNew, NewSessionParams{Cwd: t.TempDir()}, &created); err != nil {
+				t.Fatal(err)
+			}
+			if err := h.client.Call(ctx, MethodSessionPrompt, PromptParams{SessionID: created.SessionID, Prompt: []ContentBlock{TextBlock("hi")}}, &PromptResult{}); err != nil {
+				t.Fatal(err)
+			}
+			request := <-requests
+			got := make([]string, 0, len(request.Tools))
+			for _, definition := range request.Tools {
+				got = append(got, definition.Name)
+			}
+			if !slices.Equal(got, tc.want) {
+				t.Fatalf("provider tools = %#v, want %#v", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestACPRunTurnEmitsWorkspaceSetupNotices(t *testing.T) {
+	deps := testDeps(t)
+	deps.BuildWorkspace = func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
+		registry := tools.NewRegistry()
+		registry.Register(tools.NewUpdatePlanTool())
+		return &Workspace{Registry: registry, Notices: []string{"MCP server docs unavailable, skipped: timeout"}}, nil
+	}
+	h := newHarness(t, deps)
+	defer h.stop()
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	var created NewSessionResult
+	if err := h.client.Call(ctx, MethodSessionNew, NewSessionParams{Cwd: t.TempDir()}, &created); err != nil {
+		t.Fatal(err)
+	}
+	if err := h.client.Call(ctx, MethodSessionPrompt, PromptParams{SessionID: created.SessionID, Prompt: []ContentBlock{TextBlock("hi")}}, &PromptResult{}); err != nil {
+		t.Fatal(err)
+	}
+	got := drainText(t, h.updates)
+	if !strings.Contains(got, "[zero warning] MCP server docs unavailable, skipped: timeout") || !strings.Contains(got, "Hello from ZERO") {
+		t.Fatalf("streamed text = %q", got)
+	}
+}
+
+// TestACPRunTurnClosesWorkspace proves a per-turn workspace never outlives the
+// agent run. This is particularly important for MCP: its registry owns live
+// client connections and, for stdio servers, child processes.
+func TestACPRunTurnClosesWorkspaceAfterSuccessAndFailure(t *testing.T) {
+	for _, tc := range []struct {
+		name   string
+		runErr error
+	}{
+		{name: "success"},
+		{name: "agent failure", runErr: errors.New("provider interrupted")},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			deps := testDeps(t)
+			closed := 0
+			deps.BuildWorkspace = func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
+				registry := tools.NewRegistry()
+				registry.Register(tools.NewUpdatePlanTool())
+				return &Workspace{Registry: registry, Cleanup: func() error {
+					closed++
+					return nil
+				}}, nil
+			}
+			deps.RunAgent = func(context.Context, string, zeroruntime.Provider, agent.Options) (agent.Result, error) {
+				return agent.Result{FinalAnswer: "ok"}, tc.runErr
+			}
+			h := newHarness(t, deps)
+			defer h.stop()
+			ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+			defer cancel()
+			var created NewSessionResult
+			if err := h.client.Call(ctx, MethodSessionNew, NewSessionParams{Cwd: t.TempDir()}, &created); err != nil {
+				t.Fatalf("session/new: %v", err)
+			}
+			err := h.client.Call(ctx, MethodSessionPrompt, PromptParams{SessionID: created.SessionID, Prompt: []ContentBlock{TextBlock("hello")}}, &PromptResult{})
+			if tc.runErr == nil && err != nil {
+				t.Fatalf("session/prompt: %v", err)
+			}
+			if tc.runErr != nil && err == nil {
+				t.Fatal("session/prompt succeeded after agent failure")
+			}
+			if closed != 1 {
+				t.Fatalf("workspace cleanup calls = %d, want 1", closed)
+			}
+		})
+	}
+}
+
 // TestACPRejectsInvalidCwd confirms session/new fails when the workspace root
 // resolver rejects the client cwd (e.g. filesystem root).
 func TestACPRejectsInvalidCwd(t *testing.T) {

internal/agent/prompt_budget_test.go

@@ -1,9 +1,11 @@
 package agent
 
 import (
+	"strings"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
 // These ceilings are a deliberate ratchet on Zero's fixed per-turn overhead — the
@@ -33,7 +35,24 @@ const (
 	// The ratchet did its job: it caught the creep and forced a decision instead
 	// of a drift. That is the point of it, so keep raising it deliberately rather
 	// than reflexively.
+	//
+	// THIS COUNTS THE SCHEMAS ONLY, NOT THE HOST'S SHELL GUIDANCE. Until
+	// 2026-09-07 it counted both, and that made it a different measurement on
+	// every machine: exec_command appends shell guidance whose size depends on
+	// which shell was detected, so the same tree measured 3498 on Linux, ~3600 on
+	// a Windows host with pwsh 7, and 3653 on Windows PowerShell 5.1, which is
+	// stock Windows. The 5.1 case has been over this ceiling since view_image
+	// landed, so `go test ./...` failed for anyone on a stock Windows box while
+	// CI stayed green, because the runners have pwsh 7 and take the cheaper
+	// branch. A budget that passes or fails on which PowerShell happens to be
+	// installed is not ratcheting anything.
 	maxEagerToolSchemaTokens = 3650
+	// The host-specific half, ratcheted on its own so it cannot creep either.
+	// Measured 2026-09-07: Windows PowerShell 5.1 is the most expensive host at
+	// 155 tokens (722 chars), pwsh 7 is 110, cmd.exe is 87, and every POSIX host
+	// is 0 because no guidance is appended there. Set above the worst case, not
+	// above whatever this machine happens to be.
+	maxExecCommandShellGuidanceTokens = 200
 )
 
 func TestSystemPromptTokenBudget(t *testing.T) {
@@ -56,8 +75,35 @@ func TestEagerToolSchemaTokenBudget(t *testing.T) {
 	// Options{} keeps DeferThreshold at 0, so deferral is inactive and every core
 	// tool is exposed eagerly — exactly what a plugin-free session sends each turn.
 	exposed, _ := partitionTools(registry, PermissionModeAuto, Options{}, map[string]bool{})
-	got := estimateToolDefTokens(exposed)
-	t.Logf("eager core tool schemas: %d tokens across %d tools", got, len(exposed))
+	total := estimateToolDefTokens(exposed)
+	// exec_command's description carries this host's shell guidance, and that
+	// guidance is the one part of the eager set that is not the same everywhere.
+	// Charged separately below rather than folded in here, or this ceiling would
+	// mean a different thing on every machine. See the constants above.
+	//
+	// Removed from the description BEFORE estimating rather than subtracted from
+	// the total afterwards. ApproxTextTokens is an integer division, so
+	// floor((schema+guidance)/4) - floor(guidance/4) comes out one token high
+	// whenever the two remainders add past four, and which side of that line a
+	// host lands on depends on the length of its guidance: exactly the
+	// host-dependence this number is supposed to be rid of.
+	guidance := tools.HostExecCommandShellGuidance()
+	schemas := make([]zeroruntime.ToolDefinition, len(exposed))
+	copy(schemas, exposed)
+	stripped := 0
+	for index := range schemas {
+		without := strings.TrimSuffix(schemas[index].Description, guidance)
+		if without != schemas[index].Description {
+			stripped++
+		}
+		schemas[index].Description = without
+	}
+	if guidance != "" && stripped != 1 {
+		t.Fatalf("SETUP INVALID: this host's shell guidance came off %d descriptions, want exactly exec_command's; internal/tools pins it as that description's tail", stripped)
+	}
+	got := estimateToolDefTokens(schemas)
+	guidanceTokens := ApproxTextTokens(guidance)
+	t.Logf("eager core tool schemas: %d tokens across %d tools (%d total, %d host shell guidance)", got, len(exposed), total, guidanceTokens)
 	if got > maxEagerToolSchemaTokens {
 		// NOT "defer a tool": this test pins DeferThreshold at 0, so marking a
 		// tool deferred leaves it exposed here and changes nothing. Deferral is
@@ -66,6 +112,9 @@ func TestEagerToolSchemaTokenBudget(t *testing.T) {
 		// deliberate raise.
 		t.Fatalf("eager tool schemas are %d tokens, over the %d ceiling — trim a schema, drop a core tool, or raise the ceiling deliberately (deferring will NOT help: this test disables deferral)", got, maxEagerToolSchemaTokens)
 	}
+	if guidanceTokens > maxExecCommandShellGuidanceTokens {
+		t.Fatalf("this host's exec_command shell guidance is %d tokens, over the %d ceiling — trim the guidance in internal/tools/shell_runtime.go or raise the ceiling deliberately", guidanceTokens, maxExecCommandShellGuidanceTokens)
+	}
 }
 
 func TestAgentAdvertisesBothEditTools(t *testing.T) {

internal/cli/acp.go

@@ -10,9 +10,11 @@ import (
 	"github.com/Gitlawb/zero/internal/acp"
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/execution"
+	"github.com/Gitlawb/zero/internal/mcp"
 	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
+	"github.com/Gitlawb/zero/internal/redaction"
 	"github.com/Gitlawb/zero/internal/sandbox"
-	"github.com/Gitlawb/zero/internal/tools"
 )
 
 const acpUsage = `zero acp — serve the Agent Client Protocol (ACP) over stdio
@@ -57,20 +59,8 @@ func runACP(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int
 		// surface — no ACP-specific credential handling needed.
 		NewProvider: deps.newProvider,
 		RunAgent:    agent.Run,
-		// Build the SCOPED registry + sandbox engine per workspace, exactly like the
-		// exec surface, so ACP shell/file tools are confined — never run unconfined.
-		BuildWorkspace: func(workspaceRoot string, resolved config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error) {
-			scope, err := sandbox.NewScope(workspaceRoot, resolved.Sandbox.AdditionalWriteRoots)
-			if err != nil {
-				return nil, nil, err
-			}
-			engine, err := buildExecSandboxEngine(workspaceRoot, resolved, deps, scope)
-			if err != nil {
-				return nil, nil, err
-			}
-			registry := newCoreRegistryScoped(workspaceRoot, scope)
-			registerLocalControlTools(registry, workspaceRoot, resolved.LocalControl)
-			return registry, engine, nil
+		BuildWorkspace: func(ctx context.Context, workspaceRoot string, resolved config.ResolvedConfig, mode agent.PermissionMode) (*acp.Workspace, error) {
+			return buildACPWorkspace(ctx, workspaceRoot, resolved, mode, deps)
 		},
 		ResolveWorkspaceRoot: acpWorkspaceRootResolver(deps),
 		Store:                deps.newSessionStore(),
@@ -85,6 +75,76 @@ func runACP(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int
 	return exitSuccess
 }
 
+// buildACPWorkspace matches exec's registry construction for one ACP turn. MCP
+// servers use the same sandbox-prepared execution runner, project configuration is
+// gated by the validated workspace's trust state, and their runtime is released
+// when acp.Agent completes the turn. ACP deliberately stays at low autonomy: an
+// editor connection never upgrades MCP permissions beyond an interactive prompt.
+func buildACPWorkspace(ctx context.Context, workspaceRoot string, resolved config.ResolvedConfig, mode agent.PermissionMode, deps appDeps) (*acp.Workspace, error) {
+	scope, err := sandbox.NewScope(workspaceRoot, resolved.Sandbox.AdditionalWriteRoots)
+	if err != nil {
+		return nil, err
+	}
+	engine, err := buildExecSandboxEngine(workspaceRoot, resolved, deps, scope)
+	if err != nil {
+		return nil, err
+	}
+	registry := newCoreRegistryScoped(workspaceRoot, scope)
+
+	workspace := &acp.Workspace{
+		Registry: registry, Sandbox: engine,
+		DeferThreshold: resolved.Tools.DeferThreshold,
+	}
+	if mode != agent.PermissionModePlan {
+		// MCP stdio servers are subprocesses. Passing the engine to their runner
+		// keeps them inside the exact sandbox / lifecycle path used by zero exec.
+		runtime, trustSkip, err := registerMCPToolsForWorkspaceWithOptions(
+			ctx, workspaceRoot, registry, deps, mcp.AutonomyLow, workspaceRoot,
+			mcp.RegisterOptions{Execution: execution.NewRunner(engine), AdvertiseInAuto: true},
+		)
+		if err != nil {
+			// RegisterTools may have connected an earlier server before reporting a
+			// later failure, so do not orphan a partial runtime on this error path.
+			if runtime != nil {
+				_ = runtime.Close()
+			}
+			return nil, err
+		}
+		workspace.Cleanup = runtime.Close
+		workspace.Notices = acpMCPSetupNotices(trustSkip, runtime)
+	}
+	registerLocalControlTools(registry, workspaceRoot, resolved.LocalControl)
+	// MCP tools are deferred-eligible. Register their loader only after every
+	// ACP-visible tool is present, using the same mode the agent receives.
+	registerToolSearchIfEligible(registry, resolved.Tools.DeferThreshold, mode, nil, nil)
+	return workspace, nil
+}
+
+func acpMCPSetupNotices(skip trustSkip, runtime mcpToolRuntime) []string {
+	var notices []string
+	if skip.excludedProjectConfig {
+		if skip.trustCheckErrored {
+			notices = append(notices, "The workspace-trust store could not be read; project MCP servers were ignored (fail-closed). Run 'zero trust' to enable them.")
+		} else {
+			notices = append(notices, "Project MCP servers were ignored in this untrusted workspace. Run 'zero trust' to enable them.")
+		}
+	}
+	if runtime == nil {
+		return notices
+	}
+	for _, skipped := range runtime.Skipped() {
+		if skipped.UnconfiguredDefault {
+			continue
+		}
+		message := fmt.Sprintf("MCP server %s unavailable, skipped", skipped.Name)
+		if skipped.Err != nil {
+			message += ": " + redaction.ErrorMessage(skipped.Err, redaction.Options{})
+		}
+		notices = append(notices, redaction.RedactString(message, redaction.Options{}))
+	}
+	return notices
+}
+
 // acpWorkspaceRootResolver validates a client-supplied cwd into a confinement
 // root. It reuses exec's resolveWorkspaceRoot (abs+clean, must be an existing
 // dir) and additionally rejects the filesystem root and the home directory — an

internal/cli/acp_test.go

@@ -5,10 +5,19 @@ import (
 	"context"
 	"errors"
 	"io"
+	"os"
+	"path/filepath"
 	"strings"
 	"sync"
 	"testing"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/agent"
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
 )
 
 type acpTestReader func([]byte) (int, error)
@@ -108,6 +117,198 @@ func TestRunACPIdleCancellationExitsCleanly(t *testing.T) {
 	}
 }
 
+func TestBuildACPWorkspaceRegistersSandboxedMCPToolsAndClosesThem(t *testing.T) {
+	setTrustConfigRoot(t)
+	workspaceRoot := t.TempDir()
+	if err := workspacetrust.Trust(workspaceRoot); err != nil {
+		t.Fatalf("trust workspace: %v", err)
+	}
+	grantStore, err := sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: t.TempDir() + "/grants.json"})
+	if err != nil {
+		t.Fatalf("new grant store: %v", err)
+	}
+
+	var gotExclude, registered, closed bool
+	deps := fillAppDeps(appDeps{
+		resolveMCPConfig: func(root string, excludeProject bool) (config.MCPConfig, error) {
+			if root != workspaceRoot {
+				t.Fatalf("MCP workspace root = %q, want %q", root, workspaceRoot)
+			}
+			gotExclude = excludeProject
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"docs": {Type: "stdio", Command: "fake-docs"},
+			}}, nil
+		},
+		newMCPStore:     func() (*mcp.PermissionStore, error) { return nil, nil },
+		newSandboxStore: func() (*sandbox.GrantStore, error) { return grantStore, nil },
+		registerMCPTools: func(ctx context.Context, registry *tools.Registry, cfg config.MCPConfig, options mcp.RegisterOptions) (mcpToolRuntime, error) {
+			if ctx == nil {
+				t.Fatal("MCP registration received nil context")
+			}
+			if len(cfg.Servers) != 1 || cfg.Servers["docs"].Command != "fake-docs" {
+				t.Fatalf("MCP config = %+v", cfg)
+			}
+			if options.Autonomy != mcp.AutonomyLow {
+				t.Fatalf("MCP autonomy = %q, want low", options.Autonomy)
+			}
+			if options.Execution == nil {
+				t.Fatal("MCP stdio server was not given the sandbox execution runner")
+			}
+			if options.WorkspaceRoot != workspaceRoot {
+				t.Fatalf("MCP execution workspace = %q, want %q", options.WorkspaceRoot, workspaceRoot)
+			}
+			if !options.AdvertiseInAuto {
+				t.Fatal("ACP MCP tools must be advertised in auto mode without being pre-approved")
+			}
+			registered = true
+			registry.Register(cliFakeDeferredTool{
+				name: "mcp_docs_search", permission: tools.PermissionPrompt,
+				advertiseInAuto: options.AdvertiseInAuto,
+			})
+			return closeFunc(func() error {
+				closed = true
+				return nil
+			}), nil
+		},
+	})
+
+	workspace, err := buildACPWorkspace(context.Background(), workspaceRoot, config.ResolvedConfig{
+		Tools: config.ToolsConfig{DeferThreshold: 1},
+	}, agent.PermissionModeAuto, deps)
+	if err != nil {
+		t.Fatalf("buildACPWorkspace: %v", err)
+	}
+	if gotExclude {
+		t.Fatal("trusted ACP workspace excluded project MCP configuration")
+	}
+	if !registered {
+		t.Fatal("ACP workspace did not register configured MCP tools")
+	}
+	if _, ok := workspace.Registry.Get("mcp_docs_search"); !ok {
+		t.Fatal("MCP tool is absent from ACP agent registry")
+	}
+	if _, ok := workspace.Registry.Get(tools.ToolSearchToolName); !ok {
+		t.Fatal("ACP registry is missing tool_search for deferred MCP tools")
+	}
+	if workspace.DeferThreshold != 1 {
+		t.Fatalf("workspace defer threshold = %d, want 1", workspace.DeferThreshold)
+	}
+	if err := workspace.Close(); err != nil {
+		t.Fatalf("close ACP workspace: %v", err)
+	}
+	if !closed {
+		t.Fatal("ACP workspace did not close its MCP runtime")
+	}
+}
+
+func TestBuildACPWorkspacePreservesRedactedMCPSetupNotices(t *testing.T) {
+	setTrustConfigRoot(t)
+	workspaceRoot := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(workspaceRoot, ".zero"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	projectConfig := `{"mcp":{"servers":{"project-docs":{"type":"stdio","command":"project-docs"}}}}`
+	if err := os.WriteFile(filepath.Join(workspaceRoot, ".zero", "config.json"), []byte(projectConfig), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	secret := "ghp_ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789"
+	grantStore, err := sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: filepath.Join(t.TempDir(), "grants.json")})
+	if err != nil {
+		t.Fatal(err)
+	}
+	deps := fillAppDeps(appDeps{
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"docs": {Type: "stdio", Command: "fake-docs"},
+			}}, nil
+		},
+		newMCPStore:     func() (*mcp.PermissionStore, error) { return nil, nil },
+		newSandboxStore: func() (*sandbox.GrantStore, error) { return grantStore, nil },
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return fakeMCPRuntimeWithSkips{skipped: []mcp.SkippedServer{{
+				Name: "docs", Err: errors.New("token=" + secret),
+			}}}, nil
+		},
+	})
+
+	workspace, err := buildACPWorkspace(context.Background(), workspaceRoot, config.ResolvedConfig{}, agent.PermissionModeAuto, deps)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer workspace.Close()
+	joined := strings.Join(workspace.Notices, "\n")
+	if !strings.Contains(joined, "untrusted workspace") || !strings.Contains(joined, "MCP server docs unavailable") {
+		t.Fatalf("workspace notices = %#v", workspace.Notices)
+	}
+	if strings.Contains(joined, secret) || !strings.Contains(joined, "[REDACTED]") {
+		t.Fatalf("workspace notices leaked secret: %q", joined)
+	}
+}
+
+func TestBuildACPWorkspaceSkipsMCPInPlanMode(t *testing.T) {
+	grantStore, err := sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: t.TempDir() + "/grants.json"})
+	if err != nil {
+		t.Fatalf("new grant store: %v", err)
+	}
+	called := false
+	deps := fillAppDeps(appDeps{
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
+			called = true
+			return config.MCPConfig{}, nil
+		},
+		newSandboxStore: func() (*sandbox.GrantStore, error) { return grantStore, nil },
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			called = true
+			return nil, nil
+		},
+	})
+
+	workspace, err := buildACPWorkspace(context.Background(), t.TempDir(), config.ResolvedConfig{}, agent.PermissionModePlan, deps)
+	if err != nil {
+		t.Fatalf("buildACPWorkspace: %v", err)
+	}
+	if called {
+		t.Fatal("plan-mode ACP workspace must not resolve or start MCP servers")
+	}
+	if err := workspace.Close(); err != nil {
+		t.Fatalf("close plan workspace: %v", err)
+	}
+}
+
+func TestBuildACPWorkspaceClosesPartialMCPRuntimeOnRegistrationError(t *testing.T) {
+	grantStore, err := sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: t.TempDir() + "/grants.json"})
+	if err != nil {
+		t.Fatalf("new grant store: %v", err)
+	}
+	closed := false
+	deps := fillAppDeps(appDeps{
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"partial": {Type: "stdio", Command: "fake-partial"},
+			}}, nil
+		},
+		newMCPStore:     func() (*mcp.PermissionStore, error) { return nil, nil },
+		newSandboxStore: func() (*sandbox.GrantStore, error) { return grantStore, nil },
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return closeFunc(func() error {
+				closed = true
+				return nil
+			}), errors.New("second MCP server failed")
+		},
+	})
+
+	workspace, err := buildACPWorkspace(context.Background(), t.TempDir(), config.ResolvedConfig{}, agent.PermissionModeAuto, deps)
+	if err == nil {
+		t.Fatal("buildACPWorkspace succeeded after MCP registration error")
+	}
+	if workspace != nil {
+		t.Fatal("buildACPWorkspace returned a workspace after MCP registration error")
+	}
+	if !closed {
+		t.Fatal("partial MCP runtime was not closed after registration error")
+	}
+}
+
 type acpNotifyingReadCloser struct {
 	io.ReadCloser
 	readStarted chan<- struct{}

internal/cli/deferred_wiring_test.go

@@ -18,14 +18,20 @@ import (
 // cliFakeDeferredTool is deferred-eligible (implements Deferred() bool), mirroring
 // an MCP registry tool, so it counts toward the deferral threshold.
 type cliFakeDeferredTool struct {
-	name string
+	name            string
+	permission      tools.Permission
+	advertiseInAuto bool
 }
 
 func (t cliFakeDeferredTool) Name() string             { return t.name }
 func (t cliFakeDeferredTool) Description() string      { return "fake deferred tool" }
 func (t cliFakeDeferredTool) Parameters() tools.Schema { return tools.Schema{Type: "object"} }
 func (t cliFakeDeferredTool) Safety() tools.Safety {
-	return tools.Safety{SideEffect: tools.SideEffectNetwork, Permission: tools.PermissionAllow}
+	permission := t.permission
+	if permission == "" {
+		permission = tools.PermissionAllow
+	}
+	return tools.Safety{SideEffect: tools.SideEffectNetwork, Permission: permission, AdvertiseInAuto: t.advertiseInAuto}
 }
 func (t cliFakeDeferredTool) Run(context.Context, map[string]any) tools.Result {
 	return tools.Result{Status: tools.StatusOK, Output: "ok"}

internal/cli/mcp_tools.go

@@ -31,6 +31,14 @@ type mcpToolListItem struct {
 // into the one-line trust notice (mirroring the hooks and plugins chokepoints);
 // otherwise a workspace whose only project config is MCP would be gated silently.
 func registerMCPToolsForWorkspace(ctx context.Context, workspaceRoot string, registry *tools.Registry, deps appDeps, autonomy mcp.PermissionAutonomy, trustRoot string, runners ...*execution.Runner) (mcpToolRuntime, trustSkip, error) {
+	options := mcp.RegisterOptions{}
+	if len(runners) > 0 {
+		options.Execution = runners[0]
+	}
+	return registerMCPToolsForWorkspaceWithOptions(ctx, workspaceRoot, registry, deps, autonomy, trustRoot, options)
+}
+
+func registerMCPToolsForWorkspaceWithOptions(ctx context.Context, workspaceRoot string, registry *tools.Registry, deps appDeps, autonomy mcp.PermissionAutonomy, trustRoot string, options mcp.RegisterOptions) (mcpToolRuntime, trustSkip, error) {
 	excludeProject, trustCheckErrored := resolveTrust(trustRoot)
 	skip := trustSkip{
 		excludedProjectConfig: excludeProject && projectMCPConfigExists(workspaceRoot),
@@ -47,16 +55,10 @@ func registerMCPToolsForWorkspace(ctx context.Context, workspaceRoot string, reg
 	if err != nil {
 		return nil, skip, err
 	}
-	var runner *execution.Runner
-	if len(runners) > 0 {
-		runner = runners[0]
-	}
-	runtime, err := deps.registerMCPTools(ctx, registry, cfg, mcp.RegisterOptions{
-		PermissionStore: store,
-		Autonomy:        autonomy,
-		Execution:       runner,
-		WorkspaceRoot:   workspaceRoot,
-	})
+	options.PermissionStore = store
+	options.Autonomy = autonomy
+	options.WorkspaceRoot = workspaceRoot
+	runtime, err := deps.registerMCPTools(ctx, registry, cfg, options)
 	return runtime, skip, err
 }
 

internal/mcp/registry.go

@@ -23,6 +23,9 @@ const defaultConnectTimeout = 8 * time.Second
 type RegisterOptions struct {
 	PermissionStore *PermissionStore
 	Autonomy        PermissionAutonomy
+	// AdvertiseInAuto exposes prompt-gated MCP tools to model surfaces whose
+	// default mode is auto while preserving the execution-time prompt.
+	AdvertiseInAuto bool
 	ClientFactory   func(context.Context, Server) (ToolClient, error)
 	// ConnectTimeout bounds the per-server connect+list at startup. Zero uses
 	// defaultConnectTimeout.
@@ -271,9 +274,10 @@ func newRegistryTool(server Server, remote RemoteTool, client ToolClient, option
 		client:     client,
 		parameters: SchemaFromMCP(remote.InputSchema),
 		safety: tools.Safety{
-			SideEffect: tools.SideEffectNetwork,
-			Permission: permission,
-			Reason:     fmt.Sprintf("MCP tool %s/%s runs through the configured %s server.", server.Name, remote.Name, server.Type),
+			SideEffect:      tools.SideEffectNetwork,
+			Permission:      permission,
+			Reason:          fmt.Sprintf("MCP tool %s/%s runs through the configured %s server.", server.Name, remote.Name, server.Type),
+			AdvertiseInAuto: options.AdvertiseInAuto,
 		},
 	}
 }

internal/mcp/registry_test.go

@@ -342,6 +342,19 @@ func TestRegistryToolIsDeferredEligible(t *testing.T) {
 	}
 }
 
+func TestRegistryToolCanAdvertiseInAutoWithoutPreapproval(t *testing.T) {
+	tool := newRegistryTool(
+		Server{Name: "docs", Type: "stdio"},
+		RemoteTool{Name: "search"},
+		&fakeToolClient{},
+		RegisterOptions{Autonomy: AutonomyLow, AdvertiseInAuto: true},
+	)
+	safety := tool.Safety()
+	if safety.Permission != tools.PermissionPrompt || !safety.AdvertiseInAuto {
+		t.Fatalf("MCP safety = %#v", safety)
+	}
+}
+
 // TestRegistryToolReportsMCPServerName verifies the registryTool reports its true
 // configured server name (not the sanitized tool-name token) so the deferred-tools
 // discovery label names a multi-token server correctly via tools.DeferredSource.

internal/sessions/exec_prompt_tool_call_identity_test.go

@@ -0,0 +1,463 @@
+package sessions
+
+import (
+	"fmt"
+	"strings"
+	"testing"
+)
+
+// A CALL'S PAYLOAD MUST NOT RIDE INTO A LATER PROMPT ANY MORE THAN A RESULT'S.
+//
+// toolResultOutcome drops result bodies, and calls were left verbatim, so an
+// interrupted write_file put its content into the next turn's prompt while the
+// matching result body did not. Same secret, one door left open. The identity
+// still has to survive, since a resumed turn knowing WHICH file was written is
+// the whole point of admitting calls at all.
+//
+// Pinned as whole lines, compared as fields, for the same reason the result test
+// is: a substring search for one fixture secret passes when a truncated or
+// reworded copy leaks, and the renderer walks a map so field order varies.
+func TestResumePromptCarriesToolCallIdentityWithoutPayload(t *testing.T) {
+	const secret = "AWS_SECRET_ACCESS_KEY=wJalrXUtnFEMI/K7MDENG"
+	const token = "Bearer sk-live-4eC39HqLyjWDarjtT1zdp7dc"
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "rotate the deploy key"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c1", "name": "write_file", "arguments": `{"path":"deploy/prod.env","content":"` + secret + `"}`})},
+		{Sequence: 3, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "write_file", "status": "ok", "output": "written"})},
+		{Sequence: 4, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c2", "name": "edit_file", "arguments": `{"path":"deploy/prod.env","old_string":"` + secret + `","new_string":"rotated"}`})},
+		{Sequence: 5, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "edit_file", "status": "ok"})},
+		{Sequence: 6, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c3", "name": "exec_command", "arguments": `{"cmd":"curl -H '` + token + `' https://api.example/rotate","workdir":"deploy"}`})},
+		{Sequence: 7, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "exec_command", "status": "ok"})},
+		{Sequence: 8, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c4", "name": "apply_patch", "arguments": `{"patch":"*** Begin Patch\n*** Update File: deploy/prod.env\n-` + secret + `\n+rotated\n*** End Patch"}`})},
+		{Sequence: 9, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "apply_patch", "status": "ok"})},
+		{Sequence: 10, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c5", "name": "read_file", "arguments": `{"path":"deploy/prod.env","offset":1,"limit":40}`})},
+		{Sequence: 11, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "read_file", "status": "ok", "output": secret})},
+		{Sequence: 12, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error: upstream timeout"})},
+	}
+	out := resumePrompt(t, events)
+
+	// The payloads: none of them, in any form.
+	for _, leaked := range []string{secret, secret[:16], token, token[:14], "rotated", "Begin Patch", "curl -H"} {
+		if strings.Contains(out, leaked) {
+			t.Errorf("tool call payload %q reached the resume prompt:\n%s", leaked, out)
+		}
+	}
+
+	// The identities: every one of them, from the call side.
+	want := map[int]struct {
+		kind   string
+		fields []string
+	}{
+		2:  {"tool_call", []string{"c1", "write_file", `{"path":"deploy/prod.env"}`}},
+		4:  {"tool_call", []string{"c2", "edit_file", `{"path":"deploy/prod.env"}`}},
+		6:  {"tool_call", []string{"c3", "exec_command", `{"workdir":"deploy"}`}},
+		8:  {"tool_call", []string{"c4", "apply_patch"}},
+		10: {"tool_call", []string{"c5", "read_file"}},
+	}
+	seen := map[int]bool{}
+	for _, line := range strings.Split(out, "\n") {
+		if !strings.HasPrefix(line, "- #") || !strings.Contains(line, "tool_call:") {
+			continue
+		}
+		var sequence int
+		var kind string
+		rest, _ := strings.CutPrefix(line, "- #")
+		if _, err := fmt.Sscanf(rest, "%d %s", &sequence, &kind); err != nil {
+			t.Fatalf("unparsable context line %q: %v", line, err)
+		}
+		expected, wanted := want[sequence]
+		if !wanted {
+			t.Errorf("unexpected tool_call line %q", line)
+			continue
+		}
+		seen[sequence] = true
+		_, payload, _ := strings.Cut(rest, ": ")
+		fields := strings.Fields(payload)
+		for _, field := range expected.fields {
+			if !containsField(fields, field) {
+				t.Errorf("line #%d lost identity field %q: %q", sequence, field, line)
+			}
+		}
+	}
+	for sequence := range want {
+		if !seen[sequence] {
+			t.Errorf("tool_call #%d did not reach the resume prompt at all, so the interrupted work is invisible again", sequence)
+		}
+	}
+	// read_file keeps its window too, or a resumed turn re-reads a part it had.
+	if !strings.Contains(out, `"offset":1`) || !strings.Contains(out, `"limit":40`) {
+		t.Errorf("read_file lost its offset/limit window:\n%s", out)
+	}
+}
+
+func containsField(fields []string, want string) bool {
+	for _, field := range fields {
+		if field == want {
+			return true
+		}
+	}
+	// The reduced arguments object may carry more than one identity key, in map
+	// order, so a single-key expectation is also satisfied by an object that
+	// contains that key.
+	if strings.HasPrefix(want, "{") {
+		key := strings.TrimSuffix(strings.TrimPrefix(want, "{"), "}")
+		for _, field := range fields {
+			if strings.HasPrefix(field, "{") && strings.Contains(field, key) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// ARGUMENTS THAT CANNOT BE READ ARE NOT PASSED THROUGH AS TEXT.
+//
+// An allow-list only protects what it can see. Arguments that do not decode as
+// an object are dropped rather than rendered, since text this file could not
+// inspect is text it cannot vouch for.
+func TestResumePromptDropsUnreadableToolCallArguments(t *testing.T) {
+	const secret = "sk-live-unparseable-9f8e7d6c"
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "go"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c1", "name": "bash", "arguments": `not json at all ` + secret})},
+		{Sequence: 3, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error"})},
+	}
+	out := resumePrompt(t, events)
+	if strings.Contains(out, secret) {
+		t.Errorf("unparseable arguments were rendered as text:\n%s", out)
+	}
+	if !strings.Contains(out, "bash") {
+		t.Errorf("the call's identity was lost along with its unreadable arguments:\n%s", out)
+	}
+}
+
+// AND A TOOL THIS FILE HAS NEVER HEARD OF STILL KEEPS ITS PATH.
+//
+// The allow-list is by key, not by tool name, so an MCP tool or a future core
+// tool whose argument is a path or a url carries it into the resume prompt
+// without being enumerated here, while any body field it has is dropped.
+func TestResumePromptKeepsIdentityForUnknownTools(t *testing.T) {
+	const body = "large opaque payload that must not be replayed"
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "go"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c1", "name": "mcp_uploader", "arguments": `{"url":"https://files.example/x","blob":"` + body + `"}`})},
+		{Sequence: 3, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error"})},
+	}
+	out := resumePrompt(t, events)
+	if strings.Contains(out, body) {
+		t.Errorf("an unlisted body field was replayed:\n%s", out)
+	}
+	if !strings.Contains(out, "https://files.example/x") {
+		t.Errorf("an unlisted tool lost its url identity:\n%s", out)
+	}
+}
+
+// AND AN ALLOW-LISTED KEY IS NOT A SAFE VALUE.
+//
+// web_fetch accepts a credential in the query string and redacts the URL it
+// reports back, so an interrupted fetch had its token dropped from the result
+// and kept verbatim in the call. This projection is what admits the call into a
+// later turn's prompt, so the token rode into the next resume or fork. The
+// identity the resumed turn actually needs is the host and path, and those
+// survive.
+func TestResumePromptRedactsCredentialsInsideRetainedValues(t *testing.T) {
+	const token = "secret-value-9f8e7d6c5b4a"
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "check the feed"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c1", "name": "web_fetch", "arguments": `{"url":"https://api.example/data?access_token=` + token + `&page=2"}`})},
+		{Sequence: 3, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c2", "name": "web_fetch", "arguments": `{"url":"https://reader:[email protected]/feed"}`})},
+		{Sequence: 4, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error: upstream timeout"})},
+	}
+	out := resumePrompt(t, events)
+
+	for _, leaked := range []string{token, token[:12], "hunter2"} {
+		if strings.Contains(out, leaked) {
+			t.Errorf("credential %q reached the resume prompt:\n%s", leaked, out)
+		}
+	}
+	// The non-secret identity is the whole reason calls are admitted at all.
+	for _, kept := range []string{"api.example", "/data", "page=2", "/feed"} {
+		if !strings.Contains(out, kept) {
+			t.Errorf("URL identity %q was lost to the scrub:\n%s", kept, out)
+		}
+	}
+}
+
+// The scrub must not eat ordinary identities. A path, a glob and a query are
+// what the resumed turn navigates by, and none of them is a credential.
+func TestResumePromptKeepsOrdinaryIdentitiesThroughTheScrub(t *testing.T) {
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "go"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c1", "name": "read_file", "arguments": `{"path":"internal/sessions/exec_session.go","offset":1,"limit":40}`})},
+		{Sequence: 3, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c2", "name": "grep", "arguments": `{"pattern":"func toolCallIdentity","glob":"src/**/*.tsx"}`})},
+		{Sequence: 4, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error"})},
+	}
+	out := resumePrompt(t, events)
+	for _, kept := range []string{"internal/sessions/exec_session.go", "func toolCallIdentity", "src/**/*.tsx"} {
+		if !strings.Contains(out, kept) {
+			t.Errorf("identity %q was lost to the scrub:\n%s", kept, out)
+		}
+	}
+	// Numbers are not strings and are nothing to scrub; they must survive whole.
+	if !strings.Contains(out, "40") {
+		t.Errorf("a numeric read window did not survive the scrub:\n%s", out)
+	}
+}
+
+// AND A FIELD THIS FILE HAS NEVER HEARD OF DOES NOT RIDE ALONG.
+//
+// toolResultOutcome builds its output from the fields it keeps, so an unknown
+// one cannot reach a prompt through it. The call side edited arguments in place
+// and returned every sibling key, which is the opposite contract: a producer
+// recording one more top-level field, or the same payload with no arguments at
+// all, put that field straight into the next turn. The projection names what
+// survives, so both halves now drop what they do not name.
+func TestResumePromptProjectsTopLevelToolCallFields(t *testing.T) {
+	const body = "opaque top-level payload that must not be replayed"
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "go"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c1", "name": "write_file", "arguments": `{"path":"deploy/prod.env"}`, "rawInput": body,
+		})},
+		// The same shape with no arguments at all, which used to return the whole
+		// payload untouched.
+		{Sequence: 3, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c2", "name": "read_file", "stashedContent": body,
+		})},
+		{Sequence: 4, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error"})},
+	}
+	out := resumePrompt(t, events)
+
+	for _, leaked := range []string{body, "rawInput", "stashedContent"} {
+		if strings.Contains(out, leaked) {
+			t.Errorf("an unnamed top-level field %q reached the resume prompt:\n%s", leaked, out)
+		}
+	}
+	// The identity itself still survives, from both calls.
+	for _, kept := range []string{"c1", "write_file", "deploy/prod.env", "c2", "read_file"} {
+		if !strings.Contains(out, kept) {
+			t.Errorf("call identity %q was lost to the projection:\n%s", kept, out)
+		}
+	}
+}
+
+// AND A PERMITTED KEY IS NOT PERMISSION FOR WHATEVER HANGS UNDER IT.
+//
+// The projection scrubbed strings and passed everything else through, which
+// reads as "a number is nothing to scrub" and is true of a number and not of an
+// object. MCP schemas allow object and array properties, so a whole payload can
+// arrive under a permitted key. Redacting a serialized object would not help:
+// ordinary private text has no credential shape to match.
+func TestResumePromptDropsNestedArgumentContainers(t *testing.T) {
+	const body = "private body text with no credential shape at all"
+	const token = "secret-value-9f8e7d6c5b4a"
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "go"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c1", "name": "mcp_search",
+			"arguments": `{"query":{"api_key":"` + token + `","content":"` + body + `"}}`,
+		})},
+		{Sequence: 3, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c2", "name": "web_fetch",
+			"arguments": `{"url":["https://api.example/data?access_token=` + token + `"]}`,
+		})},
+		{Sequence: 4, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error"})},
+	}
+	out := resumePrompt(t, events)
+
+	for _, leaked := range []string{body, body[:20], token, token[:12], "api_key"} {
+		if strings.Contains(out, leaked) {
+			t.Errorf("nested argument content %q reached the resume prompt:\n%s", leaked, out)
+		}
+	}
+	// The call still says which tool ran, so a resumed turn can ask again.
+	for _, kept := range []string{"c1", "mcp_search", "c2", "web_fetch"} {
+		if !strings.Contains(out, kept) {
+			t.Errorf("call identity %q was lost along with the container:\n%s", kept, out)
+		}
+	}
+}
+
+// THE SEARCH ALIASES ARE TOOL-SPECIFIC, AND ONE OF THEM IS A TRAP.
+//
+// grep accepts `search` for its pattern and edit_file accepts the same word for
+// the text being replaced. Admitting it globally for the sake of the first would
+// replay file contents through the second.
+func TestResumePromptKeepsGrepSearchAndDropsEditSearch(t *testing.T) {
+	const oldText = "OLD FILE CONTENTS THAT MUST NOT BE REPLAYED"
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "go"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c1", "name": "grep", "arguments": `{"search":"func toolCallIdentity","path":"internal/sessions"}`,
+		})},
+		{Sequence: 3, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c2", "name": "edit_file", "arguments": `{"path":"a.go","search":"` + oldText + `","replace":"new"}`,
+		})},
+		{Sequence: 4, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error"})},
+	}
+	out := resumePrompt(t, events)
+
+	if !strings.Contains(out, "func toolCallIdentity") {
+		t.Errorf("grep lost the pattern it was searching for:\n%s", out)
+	}
+	if strings.Contains(out, oldText) || strings.Contains(out, oldText[:20]) {
+		t.Errorf("edit_file replayed the text it was replacing:\n%s", out)
+	}
+	// Both calls keep the path, which is identity for either tool.
+	for _, kept := range []string{"internal/sessions", "a.go"} {
+		if !strings.Contains(out, kept) {
+			t.Errorf("path identity %q was lost:\n%s", kept, out)
+		}
+	}
+}
+
+// AND THE SUPPORTED SEARCH AND WINDOW FORMS SURVIVE, not only the canonical
+// ones. A resumed turn that knows the path but not which slice was asked for
+// has half the identity and will read again.
+func TestResumePromptKeepsSupportedAliasesAndWindows(t *testing.T) {
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "go"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c1", "name": "glob", "arguments": `{"match":"src/**/*.go"}`,
+		})},
+		{Sequence: 3, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c2", "name": "read_file", "arguments": `{"path":"large.go","byte_offset":4096,"byte_limit":512}`,
+		})},
+		{Sequence: 4, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{
+			"id": "c3", "name": "read_file", "arguments": `{"path":"big.go","start_line":40,"end_line":80,"max_lines":41}`,
+		})},
+		{Sequence: 5, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error"})},
+	}
+	out := resumePrompt(t, events)
+
+	// KEY AND VALUE TOGETHER, NOT THE NUMBER ALONE. Searching for "40" finds
+	// it inside "4096", so the start_line assertion passed whether or not
+	// start_line survived at all, which is the opposite of what it claims.
+	for _, kept := range []string{
+		"src/**/*.go",
+		`"byte_offset":4096`, `"byte_limit":512`,
+		`"start_line":40`, `"end_line":80`, `"max_lines":41`,
+	} {
+		if !strings.Contains(out, kept) {
+			t.Errorf("supported identity or window %s was lost:\n%s", kept, out)
+		}
+	}
+}
+
+// THE SHARED TABLE MUST NEVER NAME A MUTATING TOOL'S BODY ARGUMENT.
+//
+// This is the rule the grep/edit_file collision taught, kept as a test so the
+// next alias added for a search tool cannot quietly open the same door. The
+// aliases below are the ones the mutating tools accept for the content they
+// write; they belong in the per-tool table or nowhere.
+func TestIdentityKeysNeverNameAMutatingToolBodyArgument(t *testing.T) {
+	bodyAliases := map[string][]string{
+		"edit_file":   {"old_string", "old", "search", "find", "old_str", "new_string", "new", "replace", "replacement", "new_str"},
+		"write_file":  {"content", "contents", "text", "body", "data", "file_content"},
+		"apply_patch": {"patch"},
+	}
+	for tool, aliases := range bodyAliases {
+		for _, alias := range aliases {
+			if toolCallIdentityKeys[alias] {
+				t.Errorf("%s accepts %q for content it writes, and the shared identity table admits it: a resumed prompt would replay the body", tool, alias)
+			}
+		}
+	}
+	// The per-tool table is where an ambiguous name belongs, and grep's use of
+	// `search` is the case that proves the mechanism is wired.
+	if !toolCallIdentityKeysByTool["grep"]["search"] {
+		t.Error("grep lost its search alias, so an interrupted grep resumes without the pattern")
+	}
+}
+
+// THROUGH THE STORE, NOT JUST THE RENDERER.
+//
+// Every test above hands FormatExecPrompt an events slice it built in memory,
+// so none of them shows the projection still applies to events that were
+// written, reloaded and prepared by the real resume and fork paths. That is
+// where it has to hold: the leak this guards against is a later turn reading a
+// recorded call, and the recording is what makes it later.
+//
+// Includes a call the tool would have REJECTED. The agent records OnToolCall
+// before execution, so a read_file whose path is an object is persisted even
+// though the tool refuses it, and a projection that only considered
+// well-formed calls would carry it through.
+func TestPersistedToolCallsAreProjectedOnResumeAndFork(t *testing.T) {
+	const body = "private body text with no credential shape at all"
+	const token = "secret-value-9f8e7d6c5b4a"
+	const oldText = "OLD FILE CONTENTS THAT MUST NOT BE REPLAYED"
+
+	store := NewStore(StoreOptions{RootDir: t.TempDir()})
+	session, err := store.Create(CreateInput{SessionID: "projection_persisted"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	appends := []AppendEventInput{
+		{Type: EventMessage, Payload: map[string]any{"role": "user", "content": "go"}},
+		{Type: EventToolCall, Payload: map[string]any{
+			"id": "c1", "name": "mcp_search",
+			"arguments": `{"query":{"api_key":"` + token + `","content":"` + body + `"}}`,
+		}},
+		{Type: EventToolCall, Payload: map[string]any{
+			"id": "c2", "name": "edit_file",
+			"arguments": `{"path":"a.go","search":"` + oldText + `","replace":"new"}`,
+		}},
+		// Rejected by read_file, recorded anyway.
+		{Type: EventToolCall, Payload: map[string]any{
+			"id": "c3", "name": "read_file",
+			"arguments": `{"path":{"nested":"` + body + `"},"start_line":40}`,
+		}},
+		{Type: EventToolCall, Payload: map[string]any{
+			"id": "c4", "name": "grep",
+			"arguments": `{"search":"func toolCallIdentity","path":"internal/sessions"}`,
+		}},
+		{Type: EventError, Payload: map[string]any{"message": "provider error: upstream timeout"}},
+	}
+	if _, err := store.AppendEvents(session.SessionID, appends); err != nil {
+		t.Fatal(err)
+	}
+
+	for _, mode := range []struct {
+		name    string
+		options PrepareExecOptions
+	}{
+		{"resume", PrepareExecOptions{Store: store, Resume: session.SessionID}},
+		{"fork", PrepareExecOptions{Store: store, Fork: session.SessionID, SessionID: "projection_persisted_fork"}},
+	} {
+		t.Run(mode.name, func(t *testing.T) {
+			prepared, err := PrepareExec(mode.options)
+			if err != nil {
+				t.Fatalf("PrepareExec: %v", err)
+			}
+			prompt := FormatExecPrompt("continue", prepared)
+
+			for _, leaked := range []string{body, body[:20], token, token[:12], oldText, oldText[:20], "api_key"} {
+				if strings.Contains(prompt, leaked) {
+					t.Errorf("%q survived persistence into the %s prompt:\n%s", leaked, mode.name, prompt)
+				}
+			}
+			// The identities that make the tail worth carrying at all.
+			for _, kept := range []string{"mcp_search", "edit_file", "read_file", "grep", "func toolCallIdentity", "internal/sessions", "a.go"} {
+				if !strings.Contains(prompt, kept) {
+					t.Errorf("identity %q was lost from the %s prompt:\n%s", kept, mode.name, prompt)
+				}
+			}
+		})
+	}
+
+	// AND THE STORE STILL HOLDS THE ORIGINALS. The projection is a view for the
+	// prompt, not an edit to the record: /rewind and an audit read the events.
+	events, err := store.ReadEvents(session.SessionID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	stored := ""
+	for _, event := range events {
+		stored += string(event.Payload)
+	}
+	for _, want := range []string{body, token, oldText} {
+		if !strings.Contains(stored, want) {
+			t.Errorf("the projection modified the persisted events: %q is no longer in the store", want)
+		}
+	}
+}

internal/sessions/exec_prompt_tool_context_test.go

@@ -0,0 +1,227 @@
+package sessions
+
+import (
+	"encoding/json"
+	"fmt"
+	"slices"
+	"strings"
+	"testing"
+)
+
+func toolContextPayload(t *testing.T, value map[string]any) json.RawMessage {
+	t.Helper()
+	raw, err := json.Marshal(value)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return json.RawMessage(raw)
+}
+
+// interruptedTurnEvents is a turn that read two files and then died on a
+// provider error, which is the shape #913 reports: no assistant answer was ever
+// produced, so nothing in the conversation describes the work.
+func interruptedTurnEvents(t *testing.T) []Event {
+	t.Helper()
+	return []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "add retries to the http client"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c1", "name": "read_file", "arguments": `{"path":"internal/http/client.go"}`})},
+		{Sequence: 3, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "read_file", "content": "package http"})},
+		{Sequence: 4, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c2", "name": "read_file", "arguments": `{"path":"internal/http/retry.go"}`})},
+		{Sequence: 5, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "read_file", "content": "var defaultBackoff = 250ms"})},
+		{Sequence: 6, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error: upstream timeout"})},
+	}
+}
+
+func resumePrompt(t *testing.T, events []Event) string {
+	t.Helper()
+	return FormatExecPrompt("continue", PreparedExec{
+		Mode:          ModeResume,
+		Session:       Metadata{SessionID: "s1"},
+		ContextEvents: events,
+	})
+}
+
+// AN INTERRUPTED TURN HAS TO SAY WHAT IT DID.
+//
+// A turn that ends normally describes its own work in the assistant's answer,
+// so the next turn inherits a prose record. A turn killed by a provider error
+// produces no answer, and the tool events that were the only record of the work
+// were filtered out of the resume context. The next turn was told a request had
+// been made and an error had happened, and nothing else, so it re-read from
+// scratch (#913).
+func TestResumePromptCarriesInterruptedToolWork(t *testing.T) {
+	out := resumePrompt(t, interruptedTurnEvents(t))
+
+	for _, want := range []string{"internal/http/client.go", "internal/http/retry.go"} {
+		if !strings.Contains(out, want) {
+			t.Errorf("the resume prompt does not name %s, so the next turn cannot know it was already read:\n%s", want, out)
+		}
+	}
+	// The conversation spine is still there.
+	for _, want := range []string{"add retries to the http client", "upstream timeout"} {
+		if !strings.Contains(out, want) {
+			t.Errorf("the resume prompt lost conversation context %q:\n%s", want, out)
+		}
+	}
+}
+
+// AND THE ORDER STAYS THE ORDER IT HAPPENED IN.
+//
+// The sequence number is rendered into each line, so a list that jumps
+// backwards reads as a corrupted history rather than a merge artifact.
+func TestResumePromptKeepsEventsInSequenceOrder(t *testing.T) {
+	out := resumePrompt(t, interruptedTurnEvents(t))
+
+	last := -1
+	for _, line := range strings.Split(out, "\n") {
+		if !strings.HasPrefix(line, "- #") {
+			continue
+		}
+		var seq int
+		if _, err := fmt.Sscanf(line, "- #%d", &seq); err != nil {
+			t.Fatalf("unparsable context line %q", line)
+		}
+		if seq <= last {
+			t.Fatalf("context is out of order at %q (previous #%d):\n%s", line, last, out)
+		}
+		last = seq
+	}
+	if last < 0 {
+		t.Fatal("SETUP INVALID: the prompt rendered no context lines")
+	}
+}
+
+// TOOL EVENTS MUST NOT EVICT CONVERSATION.
+//
+// This is what the original filter was added for (#460): tool events vastly
+// outnumber messages, so admitting them into the same trailing budget would let
+// one tool-heavy turn push every earlier message out of the context. A separate
+// allowance is what keeps both properties at once.
+func TestToolEventsNeverEvictConversation(t *testing.T) {
+	var events []Event
+	seq := 0
+	next := func() int { seq++; return seq }
+	// An early message that must survive, then a flood of tool work.
+	events = append(events, Event{Sequence: next(), Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "EARLIEST-REQUEST-MARKER"})})
+	for i := 0; i < 500; i++ {
+		arguments := fmt.Sprintf(`{"path":"noise/%03d.go"}`, i)
+		events = append(events, Event{Sequence: next(), Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"name": "read_file", "arguments": arguments})})
+		events = append(events, Event{Sequence: next(), Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "read_file", "content": "noise"})})
+	}
+	events = append(events, Event{Sequence: next(), Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "LATEST-REQUEST-MARKER"})})
+
+	out := resumePrompt(t, events)
+	for _, want := range []string{"EARLIEST-REQUEST-MARKER", "LATEST-REQUEST-MARKER"} {
+		if !strings.Contains(out, want) {
+			t.Errorf("a tool-heavy turn evicted conversation event %q, which is the regression the filter exists to prevent", want)
+		}
+	}
+
+	selected := promptContextEvents(events)
+	tools := 0
+	for _, event := range selected {
+		if event.Type == EventToolCall || event.Type == EventToolResult {
+			tools++
+		}
+	}
+	if tools == 0 {
+		t.Error("no tool work survived at all, so an interrupted tool-heavy turn still says nothing about what it did")
+	}
+	if len(selected) > 80 {
+		t.Errorf("selected %d events, over the 80 the prompt budget allows", len(selected))
+	}
+}
+
+// A session with no tool work renders exactly as it did before.
+func TestResumePromptUnchangedWithoutToolEvents(t *testing.T) {
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "hello"})},
+		{Sequence: 2, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "assistant", "content": "hi"})},
+	}
+	out := resumePrompt(t, events)
+	if strings.Count(out, "- #") != 2 {
+		t.Fatalf("a tool-free session rendered %d context lines, want 2:\n%s", strings.Count(out, "- #"), out)
+	}
+}
+
+// TOOL OUTPUT MUST NOT RIDE ALONG INTO A LATER PROMPT.
+//
+// The tail is there to say what the interrupted turn did, and the call already
+// says that. Carrying the result body would put up to 500 bytes of raw tool
+// output into a prompt on a later turn, and nothing redacts on the way in.
+func TestResumePromptCarriesToolOutcomeWithoutOutput(t *testing.T) {
+	secret := "AKIAIOSFODNN7EXAMPLE-and-more-file-contents"
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: toolContextPayload(t, map[string]any{"role": "user", "content": "look at the config"})},
+		{Sequence: 2, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c1", "name": "read_file", "arguments": `{"path":"deploy/prod.env"}`})},
+		{Sequence: 3, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "read_file", "status": "ok", "output": secret})},
+		{Sequence: 4, Type: EventToolCall, Payload: toolContextPayload(t, map[string]any{"id": "c2", "name": "read_file", "arguments": `{"path":"deploy/missing.env"}`})},
+		{Sequence: 5, Type: EventToolResult, Payload: toolContextPayload(t, map[string]any{"name": "read_file", "status": "error", "output": "no such file"})},
+		{Sequence: 6, Type: EventError, Payload: toolContextPayload(t, map[string]any{"message": "provider error: upstream timeout"})},
+	}
+	out := resumePrompt(t, events)
+
+	// THE TOOL LINES ARE PINNED WHOLE, NOT SEARCHED.
+	//
+	// Two reasons the obvious assertions are too weak to say what the trim
+	// promises. Searching the prompt for the fixture's full secret only rejects
+	// that exact string, so a change that carried a truncated or partly
+	// redacted prefix of the output would pass it. And asking whether "ok" and
+	// "error" each appear somewhere among the result lines passes just as well
+	// when the two statuses are swapped, or both attached to the wrong call.
+	//
+	// Pinning the whole of every tool line covers both: nothing but the tool
+	// name and the outcome survives a result, each outcome sits on its own
+	// line, and the line order pairs each one with the call above it.
+	//
+	// Compared as a set of fields per line rather than as a string, because the
+	// renderer walks the payload map and Go randomizes that order, so the same
+	// events render their fields in a different order on every run.
+	want := []struct {
+		sequence int
+		kind     string
+		fields   []string
+	}{
+		{2, "tool_call", []string{"c1", "read_file", `{"path":"deploy/prod.env"}`}},
+		{3, "tool_result", []string{"read_file", "ok"}},
+		{4, "tool_call", []string{"c2", "read_file", `{"path":"deploy/missing.env"}`}},
+		{5, "tool_result", []string{"read_file", "error"}},
+	}
+	var toolLines []string
+	for _, line := range strings.Split(out, "\n") {
+		if strings.HasPrefix(line, "- #") && (strings.Contains(line, "tool_call:") || strings.Contains(line, "tool_result:")) {
+			toolLines = append(toolLines, line)
+		}
+	}
+	if len(toolLines) != len(want) {
+		t.Fatalf("the resume prompt carries %d tool lines, want %d:\n%s", len(toolLines), len(want), out)
+	}
+	for index, line := range toolLines {
+		var sequence int
+		var kind string
+		rest, found := strings.CutPrefix(line, "- #")
+		if !found {
+			t.Fatalf("unparsable context line %q", line)
+		}
+		if _, err := fmt.Sscanf(rest, "%d %s", &sequence, &kind); err != nil {
+			t.Fatalf("unparsable context line %q: %v", line, err)
+		}
+		kind = strings.TrimSuffix(kind, ":")
+		_, payload, _ := strings.Cut(rest, ": ")
+		fields := strings.Fields(payload)
+		slices.Sort(fields)
+		expected := slices.Clone(want[index].fields)
+		slices.Sort(expected)
+		if sequence != want[index].sequence || kind != want[index].kind || !slices.Equal(fields, expected) {
+			t.Errorf("tool line %d is not what the trim promises.\ngot:  #%d %s %v\nwant: #%d %s %v\nfull prompt:\n%s",
+				index, sequence, kind, fields, want[index].sequence, want[index].kind, expected, out)
+		}
+	}
+	// Named on its own so a leak reports as a leak rather than as the
+	// formatting drift the comparison above would also catch. A PREFIX, because
+	// the danger is output reaching a later prompt at all, not this exact
+	// fixture string reaching it.
+	if strings.Contains(out, secret[:16]) {
+		t.Errorf("tool output reached the resume prompt:\n%s", out)
+	}
+}

internal/sessions/exec_session.go

@@ -4,8 +4,11 @@ import (
 	"encoding/json"
 	"fmt"
 	"log"
+	"sort"
 	"strings"
 	"unicode/utf8"
+
+	"github.com/Gitlawb/zero/internal/redaction"
 )
 
 type ExecMode string
@@ -188,23 +191,307 @@ func FormatExecPrompt(prompt string, prepared PreparedExec) string {
 	}, "\n")
 }
 
+// promptContextEvents chooses what a resumed turn is told about the session so
+// far.
+//
+// A RESUMED TURN NEEDS TO KNOW WHAT THE PREVIOUS ONE DID, NOT ONLY WHAT IT SAID.
+//
+// Conversation events alone were selected here, so a turn that read six files
+// and then died before answering left this behind:
+//
+//   - #1 message: add retries to the http client
+//   - #6 error: provider error: upstream timeout
+//
+// Nothing names the files, so the next turn re-reads them from scratch (#913).
+// On a turn that ENDS NORMALLY the assistant's answer describes the work, which
+// is why this was survivable; an interrupted turn produces no such answer, and
+// the record of the work goes with it.
+//
+// ONLY THE UNSUMMARIZED TAIL, NOT EVERY TOOL EVENT. Filtering tool events out
+// was deliberate (#460) and is right for work the assistant has already
+// described: forty read_file results add length and no information next to the
+// answer that explains them. It is wrong only for work with no answer after it,
+// which is precisely what an interrupted turn leaves behind. So the events after
+// the last assistant message come along, and everything the assistant already
+// spoke for stays filtered.
+//
+// The tail also gets its own allowance rather than sharing the conversation
+// budget, so a tool-heavy interrupted turn can never push an earlier message out
+// of the context. That is the property #460 added this filter for and it still
+// holds.
 func promptContextEvents(events []Event) []Event {
 	const maxPromptContextEvents = 80
+	// Enough to describe an interrupted turn's work, small enough that the
+	// conversation still dominates the prompt.
+	const maxPromptContextTailEvents = 24
 
-	filtered := make([]Event, 0, len(events))
-	for _, event := range events {
-		switch event.Type {
-		case EventMessage, EventCompaction, EventSessionFork, EventSessionChild, EventSpecialistStart, EventSpecialistStop, EventError:
-			filtered = append(filtered, event)
+	lastSpoken := -1
+	for index, event := range events {
+		if event.Type == EventMessage && payloadRole(event.Payload) == "assistant" {
+			lastSpoken = index
 		}
 	}
-	if len(filtered) == 0 {
-		filtered = append(filtered, events...)
+
+	conversation := make([]Event, 0, len(events))
+	tail := make([]Event, 0, maxPromptContextTailEvents)
+	for index, event := range events {
+		switch event.Type {
+		case EventMessage, EventCompaction, EventSessionFork, EventSessionChild, EventSpecialistStart, EventSpecialistStop, EventError:
+			conversation = append(conversation, event)
+		case EventToolCall:
+			if index > lastSpoken {
+				tail = append(tail, toolCallIdentity(event))
+			}
+		case EventToolResult:
+			if index > lastSpoken {
+				tail = append(tail, toolResultOutcome(event))
+			}
+		}
 	}
-	if len(filtered) > maxPromptContextEvents {
-		filtered = filtered[len(filtered)-maxPromptContextEvents:]
+	if len(conversation) == 0 && len(tail) == 0 {
+		// An unrecognized event stream still says more than nothing.
+		conversation = append(conversation, events...)
 	}
-	return filtered
+	if len(conversation) > maxPromptContextEvents {
+		conversation = conversation[len(conversation)-maxPromptContextEvents:]
+	}
+	// Tail slots exist only when the conversation uses fewer than the 80-event
+	// cap. A session whose conversation already fills it carries no interrupted
+	// tool work, which is what every session did before tool events were admitted
+	// at all, and what the #460 cap is there to hold. Reserving a minimum tail by
+	// trimming conversation further is a product decision, not this change.
+	tailBudget := min(maxPromptContextTailEvents, maxPromptContextEvents-len(conversation))
+	if tailBudget < 0 {
+		tailBudget = 0
+	}
+	if len(tail) > tailBudget {
+		tail = tail[len(tail)-tailBudget:]
+	}
+	if len(tail) == 0 {
+		return conversation
+	}
+	// Merged back into execution order: the sequence numbers are rendered, so a
+	// list that jumps backwards would read as a corrupted history.
+	merged := make([]Event, 0, len(conversation)+len(tail))
+	merged = append(merged, conversation...)
+	merged = append(merged, tail...)
+	sort.SliceStable(merged, func(i, j int) bool { return merged[i].Sequence < merged[j].Sequence })
+	return merged
+}
+
+// toolCallIdentityKeys are the argument fields that say WHAT a call was about
+// without carrying what it was about to write, run, or search for.
+//
+// Path, directory, URL, name and pattern fields identify the work: the file that
+// was read, the tree that was listed, the expression that was searched. Body
+// fields carry payload: write_file's content, edit_file's old and new strings,
+// apply_patch's hunks, a shell command with whatever credential was on its
+// line. The result side already drops payload through toolResultOutcome, and
+// admitting calls into resume context without the same care put an interrupted
+// write_file's content into the next turn's prompt while its result body did
+// not. Same fact, one door left open.
+//
+// AN ALLOW-LIST, NOT A DENY-LIST, so an argument this file has never heard of is
+// dropped rather than replayed. A new tool with a new body field is then a
+// missing path in a resume prompt, which is visible, instead of a new leak, which
+// is not. The keys cover every alias the tools accept for the identity fields.
+var toolCallIdentityKeys = map[string]bool{
+	// files and directories
+	"path": true, "file": true, "file_path": true, "filepath": true, "filename": true,
+	"dir": true, "directory": true, "cwd": true, "workdir": true,
+	// what a search was for; these are what the interrupted turn was looking at
+	"pattern": true, "glob": true, "query": true, "regex": true, "expression": true,
+	"match": true,
+	// fetches and named resources
+	"url": true, "name": true,
+	// read windows, so a resumed turn knows which part it already had. Every
+	// form read_file accepts, because a resumed turn that knows the path but not
+	// which slice was asked for has half the identity and will read again.
+	"offset": true, "limit": true,
+	"start_line": true, "end_line": true, "max_lines": true,
+	"byte_offset": true, "byte_limit": true,
+}
+
+// toolCallIdentityKeysByTool holds the argument names whose meaning depends on
+// which tool was called, so they cannot live in the table above.
+//
+// grep accepts `search` for its pattern, and edit_file accepts the same word
+// for the text being replaced. Admitting it globally for the sake of the first
+// would replay file contents through the second, which is the whole thing this
+// projection exists to stop. The rule this encodes: an argument name that any
+// MUTATING tool accepts for body content is scoped to the tools that mean
+// something else by it, never added to the shared table.
+var toolCallIdentityKeysByTool = map[string]map[string]bool{
+	"grep": {"search": true},
+}
+
+// retainedIdentityValue decides whether a value under a permitted key may be
+// kept, and what it looks like if so.
+//
+// AN ALLOW-LISTED KEY IS NOT A SAFE VALUE. A url is a valid place for a
+// credential to appear: web_fetch accepts a query token and redacts the URL it
+// reports back, so an interrupted fetch of
+// https://api.example/data?access_token=... had its token dropped from the
+// result and kept verbatim in the call. This projection is what admits call
+// arguments into a later turn's prompt, so the same scrub belongs here or the
+// credential is replayed on resume. Host and path survive it, which is the
+// identity the resumed turn needs.
+//
+// AND A KEY IS NOT PERMISSION FOR WHATEVER HANGS UNDER IT. Scrubbing strings
+// and passing everything else through read as "numbers are nothing to scrub",
+// which is true of a number and not of an object: MCP schemas allow object and
+// array properties, so {"query":{"api_key":"...","content":"private body"}}
+// arrives under a permitted key and carries a whole payload with it. Redacting
+// a serialized object would not help either, since ordinary private text has no
+// credential shape to match.
+//
+// So the retained shapes are named: a string, scrubbed, and the scalars a read
+// window is made of. A container is dropped. The call keeps its name and id, so
+// a resumed turn still knows which tool ran and can ask again; what it does not
+// get is arbitrary nested content it was never meant to see.
+func retainedIdentityValue(value any) (any, bool) {
+	switch typed := value.(type) {
+	case string:
+		return redaction.RedactString(typed, redaction.Options{}), true
+	case float64, int, int64, bool, json.Number:
+		return typed, true
+	default:
+		// Objects, arrays, null, and anything a future decoder invents.
+		return nil, false
+	}
+}
+
+// toolCallIdentityKeyAllowed reports whether key is an identity field for the
+// tool that was called.
+func toolCallIdentityKeyAllowed(toolName, key string) bool {
+	lowered := strings.ToLower(key)
+	if toolCallIdentityKeys[lowered] {
+		return true
+	}
+	return toolCallIdentityKeysByTool[strings.ToLower(strings.TrimSpace(toolName))][lowered]
+}
+
+// toolCallIdentityFields are the top-level payload fields a projected call
+// keeps: which call it was, and which tool. Everything else on the payload is
+// dropped, including a field added by a future producer that this file has
+// never seen.
+var toolCallIdentityFields = []string{"id", "name"}
+
+// toolCallIdentity keeps a tool call's identity and drops its payload, the
+// symmetric half of toolResultOutcome.
+//
+// The arguments travel as a JSON string inside the payload. They are decoded,
+// reduced to the identity keys, and re-encoded; anything that does not decode as
+// an object is removed outright rather than passed through as text, since text
+// that could not be read is text that cannot be checked.
+//
+// REBUILT, NOT PATCHED, WHICH IS WHAT MAKES IT THE SYMMETRIC HALF.
+// toolResultOutcome constructs its output from the two fields it keeps, so a
+// field it has never heard of cannot reach a prompt through it. Editing
+// arguments in place and returning the rest of the payload is the opposite
+// contract: it drops what it recognises as a body and forwards every sibling
+// key, so a producer recording one more top-level field would put that field
+// into the next turn. The projection names what survives.
+func toolCallIdentity(event Event) Event {
+	var decoded map[string]json.RawMessage
+	if err := json.Unmarshal(event.Payload, &decoded); err != nil {
+		event.Payload = json.RawMessage(`{}`)
+		return event
+	}
+	projected := map[string]json.RawMessage{}
+	for _, field := range toolCallIdentityFields {
+		if value, present := decoded[field]; present {
+			projected[field] = value
+		}
+	}
+	// The tool decides what some argument names mean, so the projection has to
+	// know which tool this was before it can say which keys are identity.
+	toolName := ""
+	if raw, present := decoded["name"]; present {
+		_ = json.Unmarshal(raw, &toolName)
+	}
+	kept := map[string]any{}
+	if raw, present := decoded["arguments"]; present {
+		var argumentsText string
+		if err := json.Unmarshal(raw, &argumentsText); err == nil {
+			var arguments map[string]any
+			if err := json.Unmarshal([]byte(argumentsText), &arguments); err == nil {
+				for key, value := range arguments {
+					if !toolCallIdentityKeyAllowed(toolName, key) {
+						continue
+					}
+					if retained, ok := retainedIdentityValue(value); ok {
+						kept[key] = retained
+					}
+				}
+			}
+		}
+	}
+	if len(kept) > 0 {
+		if reduced, err := json.Marshal(kept); err == nil {
+			if quoted, err := json.Marshal(string(reduced)); err == nil {
+				projected["arguments"] = quoted
+			}
+		}
+	}
+	rebuilt, err := json.Marshal(projected)
+	if err != nil {
+		event.Payload = json.RawMessage(`{}`)
+		return event
+	}
+	event.Payload = json.RawMessage(rebuilt)
+	return event
+}
+
+// toolResultOutcome strips a tool result down to WHICH tool ran and HOW IT
+// ENDED, dropping the output body.
+//
+// The tail exists so a resumed turn knows what the interrupted one did, and the
+// CALL already carries that: the tool name and its arguments, which is the path
+// for a read. The result adds only a 500-byte prefix of the output, which is
+// worth little beside the call and is the one part of an event that can carry
+// file contents. Nothing redacts on the way into a prompt, so keeping it would
+// re-emit raw tool output into a later turn on the strength of a truncation
+// limit alone.
+//
+// The status stays, because dropping it would be worse than dropping the whole
+// result: a bare call reads as work that succeeded, so a failed read would come
+// back as a file the next turn believes it already has.
+func toolResultOutcome(event Event) Event {
+	var decoded struct {
+		Name   string `json:"name"`
+		Status string `json:"status"`
+	}
+	if err := json.Unmarshal(event.Payload, &decoded); err != nil {
+		// Undecodable: drop the payload rather than pass an unknown shape through.
+		event.Payload = json.RawMessage(`{}`)
+		return event
+	}
+	trimmed, err := json.Marshal(map[string]string{
+		"name":   decoded.Name,
+		"status": decoded.Status,
+	})
+	if err != nil {
+		event.Payload = json.RawMessage(`{}`)
+		return event
+	}
+	event.Payload = json.RawMessage(trimmed)
+	return event
+}
+
+// payloadRole reads the "role" field of a message payload, returning "" when the
+// payload is not a decodable object or carries no role.
+func payloadRole(payload json.RawMessage) string {
+	if len(payload) == 0 {
+		return ""
+	}
+	var decoded struct {
+		Role string `json:"role"`
+	}
+	if err := json.Unmarshal(payload, &decoded); err != nil {
+		return ""
+	}
+	return strings.TrimSpace(decoded.Role)
 }
 
 func forkTitle(title string) string {

internal/sessions/store_test.go

@@ -610,8 +610,60 @@ func TestFormatExecPromptKeepsConversationMessagesWhenNoisyEventsFollow(t *testi
 			t.Fatalf("expected prompt to contain %q, got %q", want, prompt)
 		}
 	}
+	// The tool results here follow the last assistant answer, so they are work
+	// nothing has spoken for yet. They are carried on purpose: that is the whole
+	// of what an interrupted turn leaves behind (#913). The property this test
+	// was written for (#460) is that CONVERSATION survives a noisy turn, and the
+	// assertions above still hold it.
+	// The tool results here follow the last assistant answer, so they are work
+	// nothing has spoken for yet, and an interrupted turn leaves exactly this
+	// behind (#913). Their IDENTITY is carried; their OUTPUT is not, because the
+	// call beside them already names what was touched and the body is the part
+	// that can carry file contents into a later prompt.
+	if !strings.Contains(prompt, "tool_result: read_file") {
+		t.Fatalf("expected unanswered tool work to be named, got %q", prompt)
+	}
 	if strings.Contains(prompt, "noisy tool result") {
-		t.Fatalf("expected prompt to omit noisy non-conversation events, got %q", prompt)
+		t.Fatalf("tool output body reached the prompt, got %q", prompt)
+	}
+	if count := strings.Count(prompt, "tool_result:"); count > 24 {
+		t.Fatalf("carried %d tool events, over the tail allowance of 24: %q", count, prompt)
+	}
+}
+
+// AND WORK THE ASSISTANT ALREADY SPOKE FOR STAYS FILTERED.
+//
+// The counterpart to the case above, and what keeps that one from being
+// satisfied by carrying every tool event ever recorded. Once an answer describes
+// the work, repeating the raw results adds length and no information.
+func TestFormatExecPromptOmitsToolWorkAnAnswerAlreadyCovers(t *testing.T) {
+	events := []Event{
+		{Sequence: 1, Type: EventMessage, Payload: json.RawMessage(`{"role":"user","content":"first user request"}`)},
+	}
+	for sequence := 2; sequence <= 20; sequence++ {
+		events = append(events, Event{Sequence: sequence, Type: EventToolResult, Payload: json.RawMessage(`{"name":"read_file","output":"already summarized tool result"}`)})
+	}
+	events = append(events,
+		Event{Sequence: 21, Type: EventMessage, Payload: json.RawMessage(`{"role":"assistant","content":"I read the files and here is the summary"}`)},
+		Event{Sequence: 22, Type: EventMessage, Payload: json.RawMessage(`{"role":"user","content":"latest user request"}`)},
+	)
+
+	prompt := FormatExecPrompt("continue", PreparedExec{
+		Mode:          ModeResume,
+		Session:       Metadata{SessionID: "session-already-answered"},
+		ContextEvents: events,
+	})
+
+	if strings.Contains(prompt, "already summarized tool result") {
+		t.Fatalf("tool work the assistant already described was repeated verbatim, got %q", prompt)
+	}
+	if strings.Contains(prompt, "tool_result:") {
+		t.Fatalf("an answered turn still listed its tool events, got %q", prompt)
+	}
+	for _, want := range []string{"first user request", "here is the summary", "latest user request"} {
+		if !strings.Contains(prompt, want) {
+			t.Fatalf("expected prompt to contain %q, got %q", want, prompt)
+		}
 	}
 }
 

internal/tools/exec_command.go

@@ -75,14 +75,10 @@ func NewScopedExecCommandTool(workspaceRoot string, scope PathScope, manager *ex
 	if manager == nil {
 		manager = defaultExecSessionManager
 	}
-	description := "Runs a command in a PTY, returning output or a session ID for ongoing interaction."
-	if runtimeGOOS() == "windows" {
-		description += "\n\n" + shellGuidanceForGOOS(runtimeGOOS())
-	}
 	return execCommandTool{
 		baseTool: baseTool{
 			name:        ExecCommandToolName,
-			description: description,
+			description: execCommandDescription(runtimeGOOS()),
 			parameters: Schema{
 				Type: "object",
 				Properties: map[string]PropertySchema{
@@ -674,3 +670,19 @@ func execDisplaySummary(commandText string, sessionID int, exited bool, exitCode
 func runtimeGOOS() string {
 	return runtime.GOOS
 }
+
+// execCommandDescription is exec_command's description as a host running goos
+// would see it. The shell guidance is appended on Windows only.
+//
+// Taking goos as an argument rather than reading it is what lets the guidance
+// contract be checked for every platform from any one of them. internal/agent's
+// token ratchet removes exactly this appended text before charging the schemas,
+// so the condition here and hostExecCommandShellGuidance have to agree on all
+// three platforms, not only on whichever one the test happens to run on.
+func execCommandDescription(goos string) string {
+	description := "Runs a command in a PTY, returning output or a session ID for ongoing interaction."
+	if goos == "windows" {
+		description += "\n\n" + shellGuidanceForGOOS(goos)
+	}
+	return description
+}

internal/tools/exec_command_guidance_test.go

@@ -0,0 +1,83 @@
+package tools
+
+import (
+	"runtime"
+	"strings"
+	"testing"
+)
+
+// THE BUDGET TEST REMOVES THIS TEXT, SO IT HAS TO BE EXACTLY WHAT IS ADDED.
+//
+// internal/agent's eager-schema ratchet charges the tool schemas and this
+// host's shell guidance separately, and it separates them by stripping
+// HostExecCommandShellGuidance off exec_command's description before
+// estimating. That is only valid while the guidance is appended verbatim as the
+// tail of that description, and only while a platform that reports no guidance
+// really appends none. If the two drift apart the ratchet keeps reporting a
+// number, just not the one it claims.
+//
+// Checked for all three platforms rather than for the running one: the
+// appending is itself conditional on GOOS, so a host-only check is the shape of
+// bug this pins. Off Windows it would pass while reporting nothing at all.
+func TestExecCommandDescriptionMatchesReportedGuidance(t *testing.T) {
+	for _, goos := range []string{"windows", "linux", "darwin"} {
+		t.Run(goos, func(t *testing.T) {
+			description := execCommandDescription(goos)
+			guidance := hostExecCommandShellGuidance(goos)
+
+			if goos == "windows" {
+				if guidance == "" {
+					t.Fatal("exec_command appends shell guidance on Windows, but the accessor reports none; the budget would charge it to the schemas")
+				}
+				if !strings.HasSuffix(description, guidance) {
+					t.Fatalf("the guidance is not the tail of the description, so removing it from the schema total is wrong.\nguidance: %q\ndescription tail: %q",
+						guidance, description[max(0, len(description)-len(guidance)-40):])
+				}
+			} else if guidance != "" {
+				t.Fatalf("no shell guidance is appended off Windows, but the accessor reported %q", guidance)
+			}
+
+			// Whatever guidance THIS platform would produce, none of it may be
+			// left in the description once the reported guidance is removed.
+			// Off Windows the accessor reports nothing, so this is the check
+			// that the description really carries nothing: drop the GOOS
+			// condition in execCommandDescription and a POSIX host starts
+			// appending its own guidance (about 150 characters on darwin) with
+			// the ratchet still charging every byte of it to the schemas.
+			platformGuidance := shellGuidanceForGOOS(goos)
+			if platformGuidance == "" {
+				t.Fatalf("SETUP INVALID: %s produces no shell guidance at all, so the check below cannot fail", goos)
+			}
+			if remainder := strings.TrimSuffix(description, guidance); strings.Contains(remainder, platformGuidance) {
+				t.Fatalf("exec_command's description still carries %s shell guidance after removing the %d characters the accessor reports, so the ratchet charges the rest to the tool schemas.\nremainder: %q",
+					goos, len(guidance), remainder)
+			}
+		})
+	}
+}
+
+// And the running host agrees with the table above, so the exported accessor
+// and the real constructor cannot drift away from the platform contract.
+func TestHostExecCommandShellGuidanceMatchesTheDescription(t *testing.T) {
+	tool := NewExecCommandTool(t.TempDir(), nil)
+	if got, want := tool.Description(), execCommandDescription(runtime.GOOS); got != want {
+		t.Fatalf("the constructed tool's description is not what the platform contract says.\ngot:  %q\nwant: %q", got, want)
+	}
+	if got, want := HostExecCommandShellGuidance(), hostExecCommandShellGuidance(runtime.GOOS); got != want {
+		t.Fatalf("the exported accessor disagrees with the platform contract.\ngot:  %q\nwant: %q", got, want)
+	}
+}
+
+// And the guidance a Windows host gets is never empty, whichever shell was
+// detected, since an empty one would mean the model gets no syntax rule at all.
+func TestWindowsShellGuidanceIsNeverEmpty(t *testing.T) {
+	for _, shell := range []shellRuntime{
+		{GOOS: "windows", Executable: `C:\Windows\System32\WindowsPowerShell\v1.0\powershell.exe`, Kind: shellKindPowerShell, Syntax: "PowerShell"},
+		{GOOS: "windows", Executable: `C:\Program Files\PowerShell\7\pwsh.exe`, Kind: shellKindPowerShell, Syntax: "PowerShell"},
+		{GOOS: "windows", Executable: `C:\Windows\System32\cmd.exe`, Kind: shellKindCmd, Syntax: "cmd.exe"},
+	} {
+		if guidance := shellGuidanceForRuntime(shell); strings.TrimSpace(guidance) == "" {
+			t.Errorf("%s produced no shell guidance", shell.Executable)
+		}
+	}
+}

internal/tools/shell_runtime.go

@@ -181,6 +181,28 @@ func shellGuidanceForRuntime(shell shellRuntime) string {
 	return guidance
 }
 
+// HostExecCommandShellGuidance returns the shell guidance appended to
+// exec_command's description on this host, or "" where none is appended.
+//
+// Exported for the per-turn token ratchet in internal/agent. That budget has to
+// separate what every host pays for the tool schemas from what THIS host pays
+// for its shell guidance, because the guidance varies by shell and by
+// PowerShell version, and a single number covering both silently means a
+// different thing on each machine.
+func HostExecCommandShellGuidance() string {
+	return hostExecCommandShellGuidance(runtimeGOOS())
+}
+
+// hostExecCommandShellGuidance is the guidance a host running goos appends. It
+// is the pair of execCommandDescription, so the two can be checked against each
+// other for every platform rather than only for the running one.
+func hostExecCommandShellGuidance(goos string) string {
+	if goos != "windows" {
+		return ""
+	}
+	return shellGuidanceForGOOS(goos)
+}
+
 // HostShellEnvironmentGuidance returns the concise, model-facing shell rule
 // for the current host.
 func HostShellEnvironmentGuidance() string {