fix(web,gateway): close P3 re-review#3 findings — sanitize command errors, harden turn-lock & caps

apps/gateway/src/commands/command-executor-p8012.spec.ts

@@ -1,3 +1,4 @@
+import { Logger } from '@nestjs/common';
 import { describe, it, expect, vi, beforeEach } from 'vitest';
 import { CommandExecutorService } from './command-executor.service.js';
 import type { SlashCommandPayload } from '@mosaicstack/types';
@@ -258,4 +259,39 @@ describe('CommandExecutorService — P8-012 commands', () => {
     expect(result.command).toBe('tools');
     expect(result.message).toContain('tools');
   });
+
+  // Top-level catch sanitization (P3-4 re-review finding #1): a rejected
+  // Redis `set` inside /provider login is the only reachable path into the
+  // top-level catch in `execute()`. The raw exception must be logged
+  // server-side but never handed back to the socket client.
+  it('sanitizes the top-level command catch, logging the raw exception but never returning it to the client', async () => {
+    const distinctiveRawFailure = 'ECONNREFUSED distinctive-raw-redis-failure-token-9f31';
+    const failingRedis = {
+      set: vi.fn().mockRejectedValue(new Error(distinctiveRawFailure)),
+      get: vi.fn(),
+      del: vi.fn(),
+    };
+    const failingService = buildService(failingRedis as unknown as typeof mockRedis);
+    const loggerErrorSpy = vi.spyOn(Logger.prototype, 'error').mockImplementation(() => undefined);
+
+    const payload: SlashCommandPayload = {
+      command: 'provider',
+      args: 'login anthropic',
+      conversationId,
+    };
+    const result = await failingService.execute(payload, userScope);
+
+    expect(result.success).toBe(false);
+    expect(result.command).toBe('provider');
+    expect(result.message).toBe('Command failed due to an internal error.');
+    expect(result.message).not.toContain(distinctiveRawFailure);
+    expect(result.message).not.toContain('ECONNREFUSED');
+
+    // The real exception is still logged server-side.
+    expect(loggerErrorSpy).toHaveBeenCalled();
+    const loggedText = loggerErrorSpy.mock.calls.map((call) => String(call[0])).join(' ');
+    expect(loggedText).toContain(distinctiveRawFailure);
+
+    loggerErrorSpy.mockRestore();
+  });
 });

apps/gateway/src/commands/command-executor.service.ts

@@ -160,7 +160,12 @@ export class CommandExecutorService {
       }
     } catch (err) {
       this.logger.error(`Command /${command} failed: ${err}`);
-      return { command, conversationId, success: false, message: String(err) };
+      return {
+        command,
+        conversationId,
+        success: false,
+        message: 'Command failed due to an internal error.',
+      };
     }
   }
 

apps/web/src/spa/chat/commands-panel.spec.tsx

@@ -234,6 +234,27 @@ describe('CommandsPanel', () => {
     expect(container?.textContent).toContain('Command failed.');
   });
 
+  it('bounds an oversized command result message at the render site as defense-in-depth', async () => {
+    const hostileMessage = 'y'.repeat(50_000);
+    await render(
+      <CommandsPanel
+        manifest={null}
+        results={[
+          { conversationId: 'c1', command: 'model', success: false, message: hostileMessage },
+        ]}
+        approval={null}
+        pendingApproval={null}
+        hasConversation={false}
+        onExecute={vi.fn()}
+        onApprove={vi.fn()}
+        onRunApproved={vi.fn()}
+      />,
+    );
+
+    const text = container?.textContent ?? '';
+    expect(text.length).toBeLessThan(hostileMessage.length);
+  });
+
   it('does not throw when the manifest fields are malformed (non-array commands/skills)', async () => {
     const manifest = {
       commands: 'not-an-array',

apps/web/src/spa/chat/commands-panel.tsx

@@ -1,5 +1,6 @@
 import { useState, type ReactElement } from 'react';
 import type { PendingApproval } from './use-chat-connection';
+import { MAX_COMMAND_MESSAGE_CHARS } from './limits';
 import { asNonEmptyString, asString } from './runtime-guards';
 import type {
   CommandManifest,
@@ -13,6 +14,16 @@ import type {
  * directly, never a raw thrown exception, stack trace, or object value. */
 const COMMAND_FAILURE_COPY = 'Command failed.';
 
+/** Render-site defense-in-depth: `use-chat-connection.ts` already bounds a
+ * stored command:result message at ingestion, but this component must never
+ * assume every caller went through that path — bounding again here means a
+ * hostile/oversized message can never force an unbounded render. */
+function boundMessage(value: string): string {
+  return value.length > MAX_COMMAND_MESSAGE_CHARS
+    ? value.slice(0, MAX_COMMAND_MESSAGE_CHARS)
+    : value;
+}
+
 interface CommandsPanelProps {
   manifest: CommandManifest | null;
   results: SlashCommandResultPayload[];
@@ -141,9 +152,9 @@ export function CommandsPanel({
               /{asString(result.command)}: {result.success ? 'success' : 'failed'}
               {result.success
                 ? typeof result.message === 'string' && result.message
-                  ? ` — ${result.message}`
+                  ? ` — ${boundMessage(result.message)}`
                   : ''
-                : ` — ${asNonEmptyString(result.message, COMMAND_FAILURE_COPY)}`}
+                : ` — ${boundMessage(asNonEmptyString(result.message, COMMAND_FAILURE_COPY))}`}
             </li>
           ))}
         </ul>

apps/web/src/spa/chat/limits.ts

@@ -16,3 +16,7 @@ export const MAX_COMMAND_RESULTS = 200;
 export const MAX_MANIFEST_ITEMS = 500;
 /** Max executed approval IDs remembered for single-flight dedup. */
 export const MAX_EXECUTED_APPROVAL_IDS = 200;
+/** Max characters retained for a single command:result message — a hostile
+ * or malfunctioning gateway must not be able to push an unbounded curated
+ * success/failure reason into state (or, defensively, onto the page). */
+export const MAX_COMMAND_MESSAGE_CHARS = 1_000;

apps/web/src/spa/chat/use-chat-connection.spec.tsx

@@ -287,6 +287,26 @@ describe('useChatConnection', () => {
     ]);
   });
 
+  it('bounds an oversized command:result message to a fixed cap before it ever reaches state', async () => {
+    await act(async () => {
+      fake.serverEmit('message:ack', { conversationId: 'c1', messageId: 'm1' });
+    });
+
+    const hostileMessage = 'x'.repeat(50_000);
+    await act(async () => {
+      fake.serverEmit('command:result', {
+        conversationId: 'c1',
+        command: 'model',
+        success: false,
+        message: hostileMessage,
+      });
+    });
+
+    const stored = latest?.state.commandResults.at(-1)?.message ?? '';
+    expect(stored.length).toBeLessThan(hostileMessage.length);
+    expect(stored.length).toBeLessThanOrEqual(1000);
+  });
+
   it('pairs a successful command:approval with the pending request so it can be run with its approvalId', async () => {
     await act(async () => {
       fake.serverEmit('message:ack', { conversationId: 'c1', messageId: 'm1' });
@@ -1051,7 +1071,7 @@ describe('useChatConnection', () => {
     expect(latest?.state.manifest?.commands).toHaveLength(MAX_MANIFEST_ITEMS);
   });
 
-  it('caps a system:reload manifest replacement at MAX_MANIFEST_ITEMS for both the raw stored reload and the manifest', async () => {
+  it('caps a system:reload manifest replacement at MAX_MANIFEST_ITEMS for commands, skills, and providers, and drops hostile extra fields instead of spreading the raw payload into state', async () => {
     const commands = Array.from({ length: MAX_MANIFEST_ITEMS + 5 }, (_, i) => ({
       name: `cmd${i}`,
       aliases: [],
@@ -1065,15 +1085,21 @@ describe('useChatConnection', () => {
       description: '',
       available: true,
     }));
+    const providers = Array.from({ length: 550 }, (_, i) => `provider-${i}`);
 
-    await act(async () => {
-      fake.serverEmit('system:reload', {
+    await expect(
+      act(async () => {
+        fake.serverEmitRaw('system:reload', {
           commands,
           skills,
-        providers: ['anthropic'],
+          providers,
           message: 'reloaded',
+          // Hostile field not part of the SystemReloadPayload contract —
+          // must never survive into state.systemReload.
+          maliciousExtra: 'should-not-survive',
         });
-    });
+      }),
+    ).resolves.not.toThrow();
 
     expect(latest?.state.manifest?.commands).toHaveLength(MAX_MANIFEST_ITEMS);
     expect(latest?.state.manifest?.skills).toHaveLength(MAX_MANIFEST_ITEMS);
@@ -1082,6 +1108,8 @@ describe('useChatConnection', () => {
     // must never leave the uncapped raw payload sitting in state.
     expect(latest?.state.systemReload?.commands).toHaveLength(MAX_MANIFEST_ITEMS);
     expect(latest?.state.systemReload?.skills).toHaveLength(MAX_MANIFEST_ITEMS);
+    expect(latest?.state.systemReload?.providers).toHaveLength(500);
+    expect(latest?.state.systemReload).not.toHaveProperty('maliciousExtra');
   });
 
   it('caps availableThinkingLevels before storing a hostile session:info payload', async () => {
@@ -1390,6 +1418,16 @@ describe('useChatConnection', () => {
       latest?.actions.runApprovedCommand();
     });
 
+    // The rejected 201st approval is not silently dropped — it must be
+    // consumed but also surface a stable, visible notice so the user knows
+    // why the command did not run. ChatPage renders state.error as
+    // role="alert".
+    expect(latest?.state.error).toBe(
+      'Approval limit reached for this session. This command was not run.',
+    );
+    expect(latest?.state.approval).toBeNull();
+    expect(latest?.state.pendingApproval).toBeNull();
+
     // A fresh local approval request, replaying the very first approvalId —
     // this is the replay the old eviction policy would have let through a
     // second time because it had forgotten ap-0 ever ran.
@@ -1473,55 +1511,7 @@ describe('useChatConnection', () => {
     expect(fake.emitted.filter((e) => e.event === 'message')).toHaveLength(2);
   });
 
-  it('does not unlock turn B when a stale same-conversation error from turn A arrives', async () => {
-    await act(async () => {
-      fake.serverEmit('message:ack', { conversationId: 'c1', messageId: 'm1' });
-      fake.serverEmit('agent:start', { conversationId: 'c1' });
-      fake.serverEmit('agent:end', { conversationId: 'c1' });
-    });
-    expect(latest?.state.sending).toBe(false);
-
-    await act(async () => {
-      latest?.actions.sendMessage({ content: 'turn B' });
-    });
-    expect(latest?.state.sending).toBe(true);
-
-    // A stale error for turn A, same conversationId, arrives before turn
-    // B's own ack/start — must not unlock turn B.
-    await act(async () => {
-      fake.serverEmit('error', { conversationId: 'c1', error: 'stale turn A failure' });
-    });
-    expect(latest?.state.sending).toBe(true);
-    // A recognized-stale terminal must be a true no-op — its message must
-    // never be displayed/stored, and it must not touch streaming, which
-    // legitimately belongs to the still in-flight turn B.
-    expect(latest?.state.error).toBeNull();
-    expect(latest?.state.streaming).toBe(false);
-
-    await act(async () => {
-      latest?.actions.sendMessage({ content: 'turn C attempt' });
-    });
-    expect(fake.emitted.filter((e) => e.event === 'message')).toHaveLength(1);
-    expect(latest?.state.messages.some((m) => m.text === 'turn C attempt')).toBe(false);
-
-    // Turn B's own current ack/start arms the lock; its own terminal event
-    // (here, its own error) can then legitimately release it.
-    await act(async () => {
-      fake.serverEmit('message:ack', { conversationId: 'c1', messageId: 'm2' });
-      fake.serverEmit('agent:start', { conversationId: 'c1' });
-    });
-    await act(async () => {
-      fake.serverEmit('error', { conversationId: 'c1', error: 'turn B failed' });
-    });
-    expect(latest?.state.sending).toBe(false);
-
-    await act(async () => {
-      latest?.actions.sendMessage({ content: 'turn D' });
-    });
-    expect(fake.emitted.filter((e) => e.event === 'message')).toHaveLength(2);
-  });
-
-  it('a recognized-stale same-conversation error during an unarmed turn B does not clear approvalRequestPending, does not overwrite the frozen pendingApproval, and does not surface its message', async () => {
+  it('settles turn B and releases its lock when a pre-ack error arrives on an already-established conversation, allowing a later send', async () => {
     // Turn A completes normally on c1.
     await act(async () => {
       fake.serverEmit('message:ack', { conversationId: 'c1', messageId: 'm1' });
@@ -1530,7 +1520,41 @@ describe('useChatConnection', () => {
     });
     expect(latest?.state.sending).toBe(false);
 
-    // Turn B is sent but has not yet been armed by its own ack/start.
+    await act(async () => {
+      latest?.actions.sendMessage({ content: 'turn B' });
+    });
+    expect(latest?.state.sending).toBe(true);
+
+    // A same-conversation error arrives before turn B's own ack/start. Once
+    // turn A has already fully settled via its own terminal event, ordered
+    // Socket.IO delivery means this cannot be a leftover of A — the Gateway
+    // has nothing left in flight to emit for a turn it already finished. It
+    // can only be a genuine error for the newly sent turn B (e.g. a
+    // session-creation failure emitted before ack), so it must settle B.
+    await act(async () => {
+      fake.serverEmit('error', { conversationId: 'c1', error: 'turn B session failure' });
+    });
+    expect(latest?.state.sending).toBe(false);
+    expect(latest?.state.streaming).toBe(false);
+    expect(latest?.state.error).toBe('turn B session failure');
+
+    await act(async () => {
+      latest?.actions.sendMessage({ content: 'turn C' });
+    });
+    expect(fake.emitted.filter((e) => e.event === 'message')).toHaveLength(2);
+    expect(latest?.state.messages.some((m) => m.text === 'turn C')).toBe(true);
+  });
+
+  it('invalidates an outstanding approval request when a pre-ack error settles turn B on an already-established conversation', async () => {
+    // Turn A completes normally on c1.
+    await act(async () => {
+      fake.serverEmit('message:ack', { conversationId: 'c1', messageId: 'm1' });
+      fake.serverEmit('agent:start', { conversationId: 'c1' });
+      fake.serverEmit('agent:end', { conversationId: 'c1' });
+    });
+    expect(latest?.state.sending).toBe(false);
+
+    // Turn B is sent but has not yet received its own ack/start.
     await act(async () => {
       latest?.actions.sendMessage({ content: 'turn B' });
     });
@@ -1538,38 +1562,86 @@ describe('useChatConnection', () => {
 
     // An approval request is outstanding on c1 — approveCommand has no
     // dependency on `sending`/`streaming`, so this is legitimate even while
-    // turn B is unarmed.
+    // turn B has not yet started.
     await act(async () => {
       latest?.actions.approveCommand({ command: 'deploy', args: 'prod' });
     });
     expect(latest?.state.approvalRequestPending).toBe(true);
-    expect(latest?.state.pendingApproval).toEqual({ command: 'deploy', args: 'prod' });
-    expect(fake.emitted.filter((e) => e.event === 'command:approve')).toHaveLength(1);
 
-    // A stale error from turn A, same conversationId, arrives before turn
-    // B's own ack/start. This must be a true no-op: it must not release
-    // `sending`, must not display/store its message, must not touch
-    // `streaming`, and — critically — must not clear
-    // `approvalRequestPending`/`pendingApproval`, which would re-arm the
-    // approve UI for a request that is still outstanding.
+    // A genuine pre-ack error for turn B is terminal — same as an
+    // active-turn error, it invalidates any approval request still awaiting
+    // a response, since the Gateway that just errored is unlikely to still
+    // answer it.
     await act(async () => {
-      fake.serverEmit('error', { conversationId: 'c1', error: 'stale turn A failure' });
+      fake.serverEmit('error', { conversationId: 'c1', error: 'turn B session failure' });
     });
 
-    expect(latest?.state.sending).toBe(true);
-    expect(latest?.state.approvalRequestPending).toBe(true);
-    expect(latest?.state.pendingApproval).toEqual({ command: 'deploy', args: 'prod' });
-    expect(latest?.state.error).toBeNull();
-    expect(latest?.state.streaming).toBe(false);
+    expect(latest?.state.sending).toBe(false);
+    expect(latest?.state.approvalRequestPending).toBe(false);
+    expect(latest?.state.error).toBe('turn B session failure');
 
-    // A second approval attempt while the first is still outstanding must
-    // still be rejected — exactly one command:approve total, and the
-    // original frozen command+args must be unchanged.
+    // A fresh approval request can be issued again after the invalidation.
     await act(async () => {
       latest?.actions.approveCommand({ command: 'deploy', args: 'staging' });
     });
-    expect(fake.emitted.filter((e) => e.event === 'command:approve')).toHaveLength(1);
-    expect(latest?.state.pendingApproval).toEqual({ command: 'deploy', args: 'prod' });
+    expect(fake.emitted.filter((e) => e.event === 'command:approve')).toHaveLength(2);
+  });
+
+  it("does not settle turn B when a duplicate agent:end from already-settled turn A is redelivered after B's own ack but before B's own start", async () => {
+    // Turn A completes normally on c1.
+    await act(async () => {
+      fake.serverEmit('message:ack', { conversationId: 'c1', messageId: 'm1' });
+      fake.serverEmit('agent:start', { conversationId: 'c1' });
+      fake.serverEmit('agent:end', { conversationId: 'c1' });
+    });
+    expect(latest?.state.sending).toBe(false);
+
+    // Turn B is sent and its own ack arrives. Under the old
+    // ack-arms-the-lock design this alone made the lock releasable by any
+    // same-conversation terminal — the precise bug: it could not yet
+    // distinguish B's own eventual agent:end from a late duplicate delivery
+    // of A's already-consumed one.
+    await act(async () => {
+      latest?.actions.sendMessage({ content: 'turn B' });
+    });
+    await act(async () => {
+      fake.serverEmit('message:ack', { conversationId: 'c1', messageId: 'm2' });
+    });
+    expect(latest?.state.sending).toBe(true);
+
+    // A duplicate/straggler agent:end for turn A (already fully settled
+    // above) is redelivered for the same conversation before B's own
+    // agent:start ever arrived. Only B's own accepted agent:start may move
+    // it into the active phase that a real agent:end may settle — this
+    // duplicate must be a true no-op.
+    await act(async () => {
+      fake.serverEmit('agent:end', { conversationId: 'c1' });
+    });
+    expect(latest?.state.sending).toBe(true);
+    expect(latest?.state.streaming).toBe(false);
+
+    // Turn C must still be blocked — the lock is still genuinely held by B.
+    await act(async () => {
+      latest?.actions.sendMessage({ content: 'turn C attempt' });
+    });
+    expect(fake.emitted.filter((e) => e.event === 'message')).toHaveLength(1);
+    expect(latest?.state.messages.some((m) => m.text === 'turn C attempt')).toBe(false);
+
+    // B's own start, then its own end, legitimately unlocks it.
+    await act(async () => {
+      fake.serverEmit('agent:start', { conversationId: 'c1' });
+    });
+    expect(latest?.state.streaming).toBe(true);
+
+    await act(async () => {
+      fake.serverEmit('agent:end', { conversationId: 'c1' });
+    });
+    expect(latest?.state.sending).toBe(false);
+
+    await act(async () => {
+      latest?.actions.sendMessage({ content: 'turn D' });
+    });
+    expect(fake.emitted.filter((e) => e.event === 'message')).toHaveLength(2);
   });
 
   it('removes every listener and tears down the socket on cleanup, using no network', async () => {

apps/web/src/spa/chat/use-chat-connection.ts

@@ -1,6 +1,7 @@
 import { useEffect, useReducer, useRef } from 'react';
 import { destroySocket, getSocket } from '@/lib/socket';
 import {
+  MAX_COMMAND_MESSAGE_CHARS,
   MAX_COMMAND_RESULTS,
   MAX_EXECUTED_APPROVAL_IDS,
   MAX_MANIFEST_ITEMS,
@@ -121,6 +122,14 @@ function isValidToolCallId(value: unknown): value is string {
  * startup failure instead of leaving `sending`/`pendingSend` stuck forever. */
 const CONVERSATION_START_FAILURE = 'Unable to start this conversation. Please try again.';
 
+/** Shown when the non-evicting executed-approval dedup cache is already at
+ * MAX_EXECUTED_APPROVAL_IDS and a distinct, never-before-seen approval is
+ * denied as a result (see `runApprovedCommand`'s saturation branch below).
+ * The approval is still consumed (the UI lock is released) but the user
+ * must see why the command did not run rather than have it silently
+ * dropped. */
+const APPROVAL_LIMIT_MESSAGE = 'Approval limit reached for this session. This command was not run.';
+
 /** True only for the narrow case a malformed-conversationId `error`/`agent:end`
  * must be treated as a terminal startup failure: no conversation has ever been
  * established yet, and a send is still pending one. Once a conversation is
@@ -131,17 +140,6 @@ function isUnrecoverableStartupFailure(state: ChatConnectionState): boolean {
   return state.conversationId === null && state.pendingSend;
 }
 
-/** True when a scoped `agent:end`/`error` is a recognized-stale terminal: a
- * send is genuinely in flight (`sending`) but the current turn's own
- * ack/start has not armed `armedTurnToken` yet — so this event must be a
- * leftover from an earlier turn on the same conversation, not this one's own
- * terminal. Recognizing this must make the event a true no-op: it may not
- * finalize/clear transient state (text, thinking, streaming, error,
- * approvalRequestPending) that belongs to the still in-flight turn. */
-function isRecognizedStaleTerminal(state: ChatConnectionState): boolean {
-  return state.sending && state.armedTurnToken !== state.turnToken;
-}
-
 /** Deterministic, unique-per-event fallback id for a malformed `toolCallId`.
  * Sourced from a monotonically increasing counter carried in state (`toolSeq`)
  * rather than `tools.length`, so it stays collision-free even once `tools` is
@@ -159,6 +157,15 @@ function capList<T>(arr: unknown, max: number): T[] {
   return (arr.length > max ? arr.slice(0, max) : arr) as T[];
 }
 
+/** Like `capList`, but for a server-provided string collection (e.g.
+ * `system:reload.providers`): a non-string entry is dropped rather than
+ * invalidating the whole list, and the result is capped at `max`. */
+function capStringList(arr: unknown, max: number): string[] {
+  if (!Array.isArray(arr)) return [];
+  const strings = arr.filter((item): item is string => typeof item === 'string');
+  return strings.length > max ? strings.slice(0, max) : strings;
+}
+
 /** Normalizes a raw message:ack payload before it is stored — `messageId` is
  * given a stable, visible "unknown" fallback rather than storing a raw
  * non-string value (which would either render as "[object Object]" or,
@@ -198,7 +205,10 @@ function sanitizeApproval(
 
 /** Normalizes a raw command:result payload before it is stored — `success`
  * is coerced to a literal boolean so a truthy non-boolean value (e.g. an
- * object) can never be displayed as a successful result. */
+ * object) can never be displayed as a successful result, and a curated
+ * `message` is bounded to MAX_COMMAND_MESSAGE_CHARS so a hostile/malformed
+ * gateway cannot push an unbounded string into state (the render site in
+ * commands-panel.tsx applies the same bound again as defense-in-depth). */
 function sanitizeCommandResult(
   payload: SlashCommandResultPayload,
   conversationId: string,
@@ -207,7 +217,10 @@ function sanitizeCommandResult(
     conversationId,
     command: asString(payload.command, 'unknown'),
     success: payload.success === true,
-    message: typeof payload.message === 'string' ? payload.message : undefined,
+    message:
+      typeof payload.message === 'string'
+        ? payload.message.slice(0, MAX_COMMAND_MESSAGE_CHARS)
+        : undefined,
   };
 }
 
@@ -266,20 +279,32 @@ export interface ChatConnectionState {
    * tools remain, so fallback ids stay unique across the MAX_TOOLS cap
    * boundary. */
   toolSeq: number;
-  /** Monotonically increasing id for the current in-flight turn, minted at
-   * local send time. Wire payloads (agent:end/error) carry only a
-   * conversationId, not a turn-scoped correlation id — two different turns
-   * on the SAME conversation are indistinguishable on the wire, so this
-   * local counter is what actually tells them apart. */
-  turnToken: number;
-  /** The `turnToken` value (if any) that has been armed to accept a normal
-   * terminal release, set only by an accepted (conversation-scoped)
-   * message:ack or agent:start for the CURRENT token. A same-conversation
-   * agent:end/error may only set `sending: false` when this equals
-   * `turnToken` — otherwise it is a stale terminal event from an earlier
-   * turn on the same conversation and must not unlock a still-in-flight
-   * later turn. Fails closed: an unarmed token can never be unlocked. */
-  armedTurnToken: number | null;
+  /** Explicit per-turn phase for the CURRENT (most recently sent) turn. Wire
+   * payloads (agent:end/error) carry only a conversationId, not a
+   * turn-scoped correlation id — two different turns on the SAME
+   * conversation are indistinguishable on the wire by conversationId alone,
+   * so this local phase is what actually tells them apart:
+   *
+   * - 'pending': minted by `local/send`; the turn has been sent but has not
+   *   yet received its own `agent:start`. A same-conversation `agent:end`
+   *   arriving in this phase cannot be trusted as this turn's own — Socket.IO
+   *   delivers in order and the Gateway always emits `agent:start` before a
+   *   legitimate `agent:end`, so this can only be a stale/duplicate end left
+   *   over from an earlier turn (a `message:ack` alone does NOT advance this
+   *   phase, precisely to close that hole). A same-conversation `error`,
+   *   however, IS trusted in this phase: the Gateway can legitimately emit a
+   *   session-creation error before ack/start, and once an earlier turn on
+   *   this conversation has already fully settled (via its own terminal
+   *   event) ordered delivery guarantees nothing further remains in flight
+   *   for it — so any later scoped error can only belong to this turn.
+   * - 'active': set only by this turn's own accepted `agent:start`. Only
+   *   from this phase may a scoped `agent:end` settle the turn.
+   * - 'settled': set once this turn's own terminal (`agent:end` or `error`)
+   *   has been processed. A further scoped `agent:end`/`error` in this phase
+   *   is a true no-op — it must not re-finalize or re-release anything.
+   *
+   * Fails closed: `agent:end` can only ever settle from 'active'. */
+  turnPhase: 'pending' | 'active' | 'settled';
 }
 
 export interface ChatConnectionActions {
@@ -319,8 +344,7 @@ const initialState: ChatConnectionState = {
   messages: [],
   messageSeq: 0,
   toolSeq: 0,
-  turnToken: 0,
-  armedTurnToken: null,
+  turnPhase: 'settled',
 };
 
 type Action =
@@ -340,6 +364,7 @@ type Action =
   | { type: 'local/send'; content: string }
   | { type: 'local/approve-request'; command: string; args?: string }
   | { type: 'local/consume-approval' }
+  | { type: 'local/approval-saturated' }
   | { type: 'local/disconnect' };
 
 /**
@@ -371,11 +396,13 @@ function resolveScopedConversation(
   if (state.conversationId === null && state.pendingSend) {
     // The very first scoped event for a brand-new conversation can only belong
     // to the currently in-flight turn — no prior same-conversation turn exists to
-    // be confused with. Arm the current turnToken so a terminal event (error,
-    // agent:end) arriving before ack/start can still properly release the send lock.
+    // be confused with. This only establishes conversation identity; it does not
+    // advance `turnPhase` — only this turn's own `agent:start` may do that (see
+    // the 'server/agent:start' case), and only this turn's own `error` may settle
+    // it directly from 'pending' (see the 'server/error' case).
     return {
       active: true,
-      state: { ...state, conversationId, pendingSend: false, armedTurnToken: state.turnToken },
+      state: { ...state, conversationId, pendingSend: false },
     };
   }
   return { active: false, state: null };
@@ -413,17 +440,15 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
           conversationId,
           pendingSend: false,
           ack: sanitizeAck(payload, conversationId),
-          // The very first scoped event ever received for a brand-new
-          // conversation can only belong to the currently in-flight turn —
-          // no prior same-conversation turn can exist to be confused with.
-          armedTurnToken: state.turnToken,
+          // Deliberately does NOT advance `turnPhase` — an ack alone must
+          // never be enough to make a later `agent:end` settle this turn.
+          // Only this turn's own `agent:start` may do that.
         };
       }
       if (payload.conversationId !== state.conversationId) return state;
       return {
         ...state,
         ack: sanitizeAck(payload, state.conversationId),
-        armedTurnToken: state.turnToken,
       };
     }
 
@@ -439,7 +464,9 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
         thinkingDroppedChars: 0,
         tools: [],
         error: null,
-        armedTurnToken: resolved.state.turnToken,
+        // Only this turn's own accepted agent:start may move it into the
+        // 'active' phase that a real agent:end may later settle.
+        turnPhase: 'active',
       };
     }
 
@@ -539,18 +566,24 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
             sending: false,
             streaming: false,
             approvalRequestPending: false,
-            armedTurnToken: null,
+            turnPhase: 'settled',
             error: CONVERSATION_START_FAILURE,
           };
         }
         return state;
       }
       const next = resolved.state;
-      if (isRecognizedStaleTerminal(next)) {
-        // A same-conversation agent:end recognized as stale must be a true
-        // no-op — it must not finalize the stale turn's leftover
-        // text/thinking into a message or reset streaming, both of which
-        // legitimately belong to the still in-flight later turn.
+      if (next.turnPhase !== 'active') {
+        // Only an 'active' turn (one that has received its own agent:start)
+        // may be settled by an agent:end. A 'pending' phase means this
+        // turn has not started yet — Socket.IO delivers in order and the
+        // Gateway always emits agent:start before a legitimate agent:end, so
+        // this can only be a stale/duplicate end left over from an earlier
+        // turn. A 'settled' phase means this turn's own end already fired —
+        // this is a redelivered duplicate. Either way this must be a true
+        // no-op: it must not finalize leftover text/thinking into a message
+        // or reset streaming, both of which legitimately belong to the
+        // still in-flight (or already-finalized) current turn.
         return next;
       }
       const hasContent = next.text.length > 0 || next.thinking.length > 0;
@@ -569,19 +602,11 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
             MAX_MESSAGES,
           )
         : next.messages;
-      // The wire carries only conversationId, not a turn id — a stale
-      // agent:end left over from an EARLIER turn on this SAME conversation
-      // is indistinguishable from this turn's own by conversationId alone.
-      // Only release `sending` when this turn's own accepted ack/start armed
-      // the current token; otherwise this is treated as a stale echo and the
-      // lock stays held (fail closed) so it cannot unlock a later, still
-      // in-flight turn.
-      const canRelease = next.armedTurnToken === next.turnToken;
       return {
         ...next,
         streaming: false,
-        sending: canRelease ? false : next.sending,
-        armedTurnToken: canRelease ? null : next.armedTurnToken,
+        sending: false,
+        turnPhase: 'settled',
         text: '',
         thinking: '',
         textDroppedChars: 0,
@@ -671,19 +696,19 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
 
     case 'server/system:reload': {
       const { payload } = action;
-      // Computed once and reused for both `systemReload` and `manifest` —
-      // spreading the raw `payload` into `systemReload` first and only
-      // capping the copy handed to `manifest` left the raw, uncapped
-      // commands/skills sitting in `state.systemReload`. The sanitized
-      // fields are placed after the spread below so they always win.
+      // `systemReload` is built from exactly the SystemReloadPayload
+      // contract fields (commands, skills, providers, message) after
+      // runtime normalization — the raw payload is never spread into state,
+      // so a hostile/malfunctioning gateway cannot smuggle extra fields in.
       const commands = capList<CommandDef>(payload.commands, MAX_MANIFEST_ITEMS);
       const skills = capList<SkillCommandDef>(payload.skills, MAX_MANIFEST_ITEMS);
+      const providers = capStringList(payload.providers, MAX_MANIFEST_ITEMS);
       return {
         ...state,
         systemReload: {
-          ...payload,
           commands,
           skills,
+          providers,
           message: asString(payload.message, 'Commands reloaded.'),
         },
         manifest: {
@@ -705,31 +730,33 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
             sending: false,
             streaming: false,
             approvalRequestPending: false,
-            armedTurnToken: null,
+            turnPhase: 'settled',
             error: CONVERSATION_START_FAILURE,
           };
         }
         return state;
       }
       const next = resolved.state;
-      if (isRecognizedStaleTerminal(next)) {
-        // Same true-no-op guard as `server/agent:end` above: a
-        // same-conversation error recognized as stale must not display/store
-        // its message, touch streaming, or clear approvalRequestPending —
-        // doing so would re-arm the approve UI for a request that belongs to
-        // the still in-flight later turn while the first remains outstanding.
-        return next;
+      if (next.turnPhase === 'settled') {
+        // No in-flight current turn: the error is still visibly surfaced,
+        // but there is no turn/approval lock left to mutate.
+        return { ...next, error: asString(payload.error, 'An error occurred.') };
       }
-      // Same fail-closed turn-token guard as `server/agent:end` above: a
-      // same-conversation error left over from an earlier, already-finished
-      // turn must not release a later turn's still-in-flight send lock.
-      const canRelease = next.armedTurnToken === next.turnToken;
+      // A current turn-scoped error is terminal from either 'pending'
+      // (pre-ack — the Gateway can emit a session-creation error before
+      // ack) or 'active' (a prompt error after ack/start) state. Once an
+      // earlier turn on this conversation has already fully settled (its
+      // own terminal event processed, moving turnPhase to 'settled' before
+      // this turn was even sent), ordered Socket.IO delivery guarantees
+      // nothing further remains in flight for it — so a scoped error seen
+      // while turnPhase is 'pending' or 'active' can only belong to THIS
+      // turn, and must settle and release the lock.
       return {
         ...next,
         error: asString(payload.error, 'An error occurred.'),
         streaming: false,
-        sending: canRelease ? false : next.sending,
-        armedTurnToken: canRelease ? null : next.armedTurnToken,
+        sending: false,
+        turnPhase: 'settled',
         // A turn-scoped error also invalidates any approval request awaiting
         // a response — the gateway that just errored is unlikely to still
         // answer it, and the synchronous approveLockRef mirrors this field.
@@ -752,12 +779,11 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
         messageSeq: state.messageSeq + 1,
         error: null,
         sending: true,
-        // Sending a new turn mints a fresh token and immediately clears
-        // terminal eligibility — only THIS turn's own accepted ack/start may
-        // arm it, so a same-conversation terminal left over from the turn
-        // that just finished can never be mistaken for this one's.
-        turnToken: state.turnToken + 1,
-        armedTurnToken: null,
+        // Sending a new turn begins a fresh 'pending' phase — only THIS
+        // turn's own accepted agent:start may move it to 'active', so a
+        // same-conversation agent:end left over from the turn that just
+        // finished can never be mistaken for this one's.
+        turnPhase: 'pending',
         pendingSend: state.conversationId === null ? true : state.pendingSend,
       };
     }
@@ -771,7 +797,7 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
         sending: false,
         pendingSend: false,
         approvalRequestPending: false,
-        armedTurnToken: null,
+        turnPhase: 'settled',
       };
     }
 
@@ -792,6 +818,14 @@ function reduce(state: ChatConnectionState, action: Action): ChatConnectionState
       return { ...state, approval: null, pendingApproval: null };
     }
 
+    case 'local/approval-saturated': {
+      // The non-evicting executed-approval cache is full and this is a
+      // distinct, never-before-seen approval — it is denied (fail closed),
+      // but unlike a plain replay of an already-executed ID, this must be
+      // visibly surfaced rather than silently dropped.
+      return { ...state, approval: null, pendingApproval: null, error: APPROVAL_LIMIT_MESSAGE };
+    }
+
     default:
       return state;
   }
@@ -973,9 +1007,10 @@ export function useChatConnection(): ChatConnectionValue {
         // unmounted (lifecycle reset), not recoverable by re-approving — a
         // false positive/availability cost but a genuine replay of any ID
         // ever seen by this hook can never execute a second time. Consuming
-        // (rather than silently no-op'ing) releases the UI lock so the
-        // denial is visible/recoverable by remounting.
-        dispatch({ type: 'local/consume-approval' });
+        // (rather than silently no-op'ing) releases the UI lock, and unlike
+        // the plain-replay branch above, sets a visible error notice so the
+        // denial is surfaced rather than silently dropped.
+        dispatch({ type: 'local/approval-saturated' });
         return;
       }
       executedApprovalIds.current.add(approval.approvalId);