commandmate 0.33.1 → 0.33.3

package/dist/server/locales/en/worktree.json

@@ -436,12 +436,13 @@
         "endSessionFor": "End session for {name}",
         "filesEmptyHint": "Open a file to enable this",
         "historyAllSplitsHint": "Applies to all splits",
-        "historyChatOnlyHint": "Every split shows Chat, which has no history column",
         "maximizeSplit": "Maximize {split}",
         "maximizeFocusedSplit": "Maximize the focused split",
         "restoreSplits": "Restore the split layout",
         "maximizeShortcutHint": "⌘ / Ctrl + Shift + Enter",
-        "maximizedStatus": "Split {split} maximized"
+        "maximizedStatus": "Split {split} maximized",
+        "loading": "Loading…",
+        "sessionEnded": "The session has ended. Send a message to start a new session."
     },
     "surfaceMode": {
         "groupLabel": "Output surface for {split}",
@@ -504,6 +505,12 @@
             "unclassified": "Dialog not recognized",
             "resolved": "Resolved",
             "unlabeled": "Approval dialog"
+        },
+        "previousSession": {
+            "summary": "Previous session ({count})",
+            "expand": "Show the previous session",
+            "collapse": "Hide the previous session",
+            "hint": "You can view the previous conversation in History."
         }
     },
     "unsentComposer": {

package/dist/server/locales/ja/worktree.json

@@ -436,12 +436,13 @@
         "endSessionFor": "{name}のセッションを終了",
         "filesEmptyHint": "ファイルを開くと有効になります",
         "historyAllSplitsHint": "すべての split に適用",
-        "historyChatOnlyHint": "すべての split がチャット表示のため、履歴列はありません",
         "maximizeSplit": "{split} を最大化",
         "maximizeFocusedSplit": "フォーカス中の split を最大化",
         "restoreSplits": "分割表示に戻す",
         "maximizeShortcutHint": "⌘ / Ctrl + Shift + Enter",
-        "maximizedStatus": "{split} を最大化中"
+        "maximizedStatus": "{split} を最大化中",
+        "loading": "読込中...",
+        "sessionEnded": "セッションは終了しています。送信すると新しいセッションが起動します。"
     },
     "surfaceMode": {
         "groupLabel": "{split} の出力面",
@@ -504,6 +505,12 @@
             "unclassified": "判別できないダイアログ",
             "resolved": "応答済み",
             "unlabeled": "承認ダイアログ"
+        },
+        "previousSession": {
+            "summary": "前回のセッション（{count} 件）",
+            "expand": "前回のセッションを開く",
+            "collapse": "前回のセッションを閉じる",
+            "hint": "前回の会話は履歴で参照できます。"
         }
     },
     "unsentComposer": {

package/dist/server/src/lib/assistant-response-saver.js

@@ -18,9 +18,43 @@
  * buffer, which is false for alternate-screen tools — those are handled by the
  * response poller instead (Issue #1268 / #1292; see savePendingAssistantResponse).
  */
+var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    var desc = Object.getOwnPropertyDescriptor(m, k);
+    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
+      desc = { enumerable: true, get: function() { return m[k]; } };
+    }
+    Object.defineProperty(o, k2, desc);
+}) : (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    o[k2] = m[k];
+}));
+var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
+    Object.defineProperty(o, "default", { enumerable: true, value: v });
+}) : function(o, v) {
+    o["default"] = v;
+});
+var __importStar = (this && this.__importStar) || (function () {
+    var ownKeys = function(o) {
+        ownKeys = Object.getOwnPropertyNames || function (o) {
+            var ar = [];
+            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
+            return ar;
+        };
+        return ownKeys(o);
+    };
+    return function (mod) {
+        if (mod && mod.__esModule) return mod;
+        var result = {};
+        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
+        __setModuleDefault(result, mod);
+        return result;
+    };
+})();
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.detectBufferReset = detectBufferReset;
 exports.cleanCliResponse = cleanCliResponse;
+exports.advanceCapturedLineForTranscriptTurn = advanceCapturedLineForTranscriptTurn;
 exports.savePendingAssistantResponse = savePendingAssistantResponse;
 const cli_session_1 = require("./session/cli-session");
 const db_1 = require("./db");
@@ -82,6 +116,31 @@ function detectBufferReset(currentLineCount, lastCapturedLine) {
     }
     return { bufferReset: false, reason: null };
 }
+/**
+ * How many rows of a capture count as "read so far".
+ *
+ * Trailing blank rows are discarded, for consistency with the response poller's
+ * `extractResponse`: tmux pads the pane out to its height, and counting that
+ * padding inflates the stored cursor so far past the real content that the
+ * poller's dedup check (`result.lineCount <= lastCapturedLine`) can never fire
+ * again.
+ *
+ * The one definition of the count, shared by the two writers of
+ * `session_states.last_captured_line` in this module (Issue #2437): the
+ * pre-send flush and {@link advanceCapturedLineForTranscriptTurn}. Two spellings
+ * of "how far have I read" would be two cursors that disagree.
+ *
+ * @param output - Raw capture
+ * @returns Row count with trailing blank rows removed
+ */
+function countCapturedLines(output) {
+    const lines = output.split('\n');
+    let trimmedLength = lines.length;
+    while (trimmedLength > 0 && lines[trimmedLength - 1].trim() === '') {
+        trimmedLength--;
+    }
+    return trimmedLength;
+}
 /**
  * Time offset (in milliseconds) for assistant message timestamp
  * Ensures assistant response appears before user message in chronological order
@@ -91,8 +150,23 @@ const ASSISTANT_TIMESTAMP_OFFSET_MS = 1;
 /**
  * Clean CLI tool response based on tool type
  *
+ * Every one of `CLI_TOOL_IDS`' eight tools has a branch (Issue #2437). Until
+ * then four of them — codex, command-code, antigravity and vibe-local — fell
+ * through to `output.trim()`, and codex's branch said so in as many words:
+ * *"Codex doesn't need special cleaning"*. It does. What that branch actually
+ * saved was codex's **idle composer**, ANSI and all, as an assistant reply:
+ *
+ * ```text
+ * › Ask Codex to do anything
+ *
+ *   gpt-6-astra xhigh · ~/share/work/… · Main [default]
+ * ```
+ *
+ * See {@link cleanScrollbackResponse} for how the four share one cleaner
+ * without sharing a boundary rule.
+ *
  * @param output - Raw output from CLI tool
- * @param cliToolId - CLI tool identifier (claude, codex, gemini)
+ * @param cliToolId - CLI tool identifier
  * @returns Cleaned response content
  */
 function cleanCliResponse(output, cliToolId) {
@@ -106,12 +180,101 @@ function cleanCliResponse(output, cliToolId) {
         case 'copilot':
             return (0, response_cleaner_1.cleanCopilotResponse)(output);
         case 'codex':
-            // Codex doesn't need special cleaning
-            return output.trim();
+        case 'command-code':
+        case 'antigravity':
+        case 'vibe-local':
+            return (0, response_cleaner_1.cleanScrollbackResponse)(output, cliToolId);
         default:
             return output.trim();
     }
 }
+/**
+ * Move `last_captured_line` past everything a transcript row now covers
+ * (Issue #2437).
+ *
+ * ## The defect this closes
+ *
+ * The pre-send flush ({@link savePendingAssistantResponse}) reads
+ * `lastCapturedLine`, and saves **everything past it** as the pending reply. The
+ * only writer of that cursor used to be the flush itself and the response
+ * poller — the Stop path that writes the agent's OWN Markdown into History
+ * (`hooks/sources/<tool>/history.ts`) never touched it. So on this ordering, which
+ * costs nothing to hit at a 2-second poll interval:
+ *
+ * ```text
+ * 1. turn ends  → the Stop path writes the turn as a Markdown row   (cursor unmoved)
+ * 2. `/send` arrives BEFORE the next poll tick
+ * 3. the flush saves "everything since the cursor" = the whole finished turn
+ * ```
+ *
+ * The second copy is the pane's scrape of the very same turn, so the chat
+ * surface shows one reply twice. A cleaner cannot help: what is duplicated is
+ * the real body, not chrome.
+ *
+ * ## Why a capture rather than an arithmetic bump
+ *
+ * The cursor is a row index into the pane, and the transcript reader never looks
+ * at the pane at all — it reads a JSONL file. The only honest way to say "the
+ * pane up to here is already in History" is to ask the pane how tall it is now.
+ * The capture is the cached one ({@link captureSessionOutput}, 5s TTL shared
+ * with the poller), so a Stop that lands between two ticks usually costs no
+ * `tmux capture-pane` at all.
+ *
+ * ## What it deliberately does not do
+ *
+ * - **Alternate-screen tools are skipped.** For claude, opencode and copilot the
+ *   line count is a screen-row constant rather than a cursor (Issue #1268), and
+ *   {@link savePendingAssistantResponse} refuses to run for them at all — there
+ *   is no cursor here to advance and writing one would be a lie.
+ * - **The cursor only ever moves forward.** A capture shorter than the stored
+ *   value is a buffer reset, and `detectBufferReset` owns that reading; moving
+ *   the cursor backwards from here would hand the flush a range it has already
+ *   saved.
+ * - **It never throws, and never changes the caller's answer.** The transcript
+ *   row is written whether or not this succeeds; a failure here costs one
+ *   duplicated reply, and a failure that propagated would cost the row.
+ *
+ * @param target - The instance whose turn was just written to History
+ * @returns The cursor's new value, or null when it was left where it was
+ */
+async function advanceCapturedLineForTranscriptTurn(target) {
+    const { worktreeId, cliToolId } = target;
+    const resolvedInstanceId = target.instanceId ?? cliToolId;
+    try {
+        if ((0, types_1.usesAlternateScreen)(cliToolId)) {
+            return null;
+        }
+        const output = await (0, cli_session_1.captureSessionOutput)(worktreeId, cliToolId, SESSION_OUTPUT_BUFFER_SIZE, target.instanceId);
+        if (!output) {
+            return null;
+        }
+        const currentLineCount = countCapturedLines(output);
+        const { getDbInstance } = await Promise.resolve().then(() => __importStar(require('./db/db-instance')));
+        const db = getDbInstance();
+        const lastCapturedLine = (0, db_1.getSessionState)(db, worktreeId, resolvedInstanceId)?.lastCapturedLine || 0;
+        if (currentLineCount <= lastCapturedLine) {
+            return null;
+        }
+        (0, db_1.updateSessionState)(db, worktreeId, cliToolId, currentLineCount, target.instanceId);
+        logger.info('transcript:cursor-advanced', {
+            worktreeId,
+            instanceId: resolvedInstanceId,
+            from: lastCapturedLine,
+            to: currentLineCount,
+        });
+        return currentLineCount;
+    }
+    catch (error) {
+        // Never the caller's problem: the transcript row is already written, and the
+        // worst this costs is the duplicate it was meant to prevent.
+        logger.debug('transcript:cursor-advance-failed', {
+            worktreeId,
+            instanceId: resolvedInstanceId,
+            error: error instanceof Error ? error.message : String(error),
+        });
+        return null;
+    }
+}
 /**
  * Save pending assistant response before a new user message
  *
@@ -170,16 +333,9 @@ async function savePendingAssistantResponse(db, worktreeId, cliToolId, userMessa
         if (!output) {
             return null;
         }
-        // 3. Calculate current line count
-        // Trim trailing empty lines for consistency with response-poller's extractResponse.
-        // Without this, tmux buffer padding inflates the line count, causing the poller's
-        // dedup check (result.lineCount <= lastCapturedLine) to always trigger.
+        // 3. Calculate current line count (see countCapturedLines for the trim).
         const lines = output.split('\n');
-        let trimmedLength = lines.length;
-        while (trimmedLength > 0 && lines[trimmedLength - 1].trim() === '') {
-            trimmedLength--;
-        }
-        const currentLineCount = trimmedLength;
+        const currentLineCount = countCapturedLines(output);
         // 4. Detect buffer reset (Issue #59 fix)
         const { bufferReset, reason } = detectBufferReset(currentLineCount, lastCapturedLine);
         if (bufferReset) {

package/dist/server/src/lib/cli-tools/vibe-local.js

@@ -20,6 +20,7 @@ const db_1 = require("../db");
 const logger_1 = require("../../lib/logger");
 const cli_tool_timing_config_1 = require("../../config/cli-tool-timing-config");
 const install_hints_1 = require("./install-hints");
+const agent_session_lifecycle_1 = require("../../lib/session/agent-session-lifecycle");
 const logger = (0, logger_1.createLogger)('cli-tools/vibe-local');
 /**
  * Extract error message from unknown error type (DRY)
@@ -75,6 +76,15 @@ class VibeLocalTool extends base_1.BaseCLITool {
             }
             logger.warn('vibe-local-session-relaunch', { sessionName });
         }
+        // Issue #1759 / #2444: the one line every tool's creation path owes the
+        // rest of the system. vibe-local emits no structured events, so the
+        // generation fence itself is inert here — but `beginAgentSession` is also
+        // where a superseded session's chat rows are retired, and vibe-local was
+        // the only tool that never called it. Placed after the healthy-reuse
+        // `return` above and before the pane is touched, exactly as codex places
+        // its call: a reuse is the same process and must archive nothing, a
+        // relaunch into this same pane is a new one and must.
+        (0, agent_session_lifecycle_1.beginAgentSession)({ worktreeId, cliToolId: this.id, instanceId });
         try {
             // Issue #2070: creation only. On the relaunch path the pane already
             // exists and holds the transcript of the process that died in it; the

package/dist/server/src/lib/hooks/hook-settings-generator.js

@@ -43,7 +43,7 @@
  * @module lib/hooks/hook-settings-generator
  */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.PERMISSION_DENY_RULES = exports.TOOL_USE_MATCHER = exports.NOTIFICATION_MATCHER = exports.AUTH_TOKEN_ENV_VAR = exports.PERMISSION_REQUEST_TIMEOUT_SECONDS = exports.HOOK_TIMEOUT_SECONDS = exports.PERMISSION_REQUEST_PATH = exports.AGENT_EVENT_PATH = void 0;
+exports.PERMISSION_DENY_RULES = exports.TMUX_UNPINNED_DENY_RULES = exports.TOOL_USE_MATCHER = exports.NOTIFICATION_MATCHER = exports.AUTH_TOKEN_ENV_VAR = exports.PERMISSION_REQUEST_TIMEOUT_SECONDS = exports.HOOK_TIMEOUT_SECONDS = exports.PERMISSION_REQUEST_PATH = exports.AGENT_EVENT_PATH = void 0;
 exports.isHookInjectionEnabled = isHookInjectionEnabled;
 exports.isAuthTokenExpected = isAuthTokenExpected;
 exports.getHookSettingsDirectory = getHookSettingsDirectory;
@@ -127,6 +127,62 @@ exports.NOTIFICATION_MATCHER = 'permission_prompt|idle_prompt';
  * receiver takes.
  */
 exports.TOOL_USE_MATCHER = permission_request_payload_1.ASK_USER_QUESTION_TOOL;
+/**
+ * tmux spellings that reach whichever server the ambient environment points at
+ * (Issue #2442).
+ *
+ * On 2026-09-08 22:53 JST all 42 `mcbd-*` sessions on the default tmux server
+ * were destroyed by a socket-less `tmux kill-server` run from an ad-hoc Bash
+ * block — the second such wipe after 2026-08-02 (#1624). The block had set
+ * `TMUX_TMPDIR` first, which reads as isolation and is not: tmux consults it
+ * only while `$TMUX` is unset, and every CommandMate agent runs inside a pane.
+ *
+ * A tmux invocation carries its target as a *global option before the
+ * subcommand* — `-L <name>` (socket name) or `-S <path>` (socket path) — and
+ * both outrank `$TMUX`. So "which server does this hit" is decidable from the
+ * argv shape alone, and that is what these rules read: a command whose first
+ * word is `tmux` and whose second is a destroying or server-global subcommand
+ * names no socket and therefore hits the shared one.
+ *
+ * The pinned spelling of every rule below is deliberately *not* covered, which
+ * is what keeps the isolated-probe recipes in `docs/design/` working. "Not
+ * covered" means "no new deny rule matches it", not "pre-approved": such a
+ * command still goes through the session's normal permission handling.
+ *
+ * Aliases get their own entry because Claude's prefix matching is **token-exact
+ * and adjacency-sensitive**, measured on 2.1.266 in
+ * `docs/design/agent-hooks-permission-deny-verification.md` §6: a rule naming
+ * `tmux list-p` left `tmux list-panes` untouched, and one naming
+ * `tmux show-options -g` left both `tmux show-options -gv …` and
+ * `tmux show-options -t 0 -g` untouched. A single truncated rule covering a
+ * family is therefore not available — it would load without error and match
+ * nothing. `bind` / `unbind` / `set` are tmux's own documented aliases for
+ * `bind-key` / `unbind-key` / `set-option`; `kill-server` and `kill-session`
+ * have none.
+ *
+ * Shapes that escape these rules (absolute path, `bash -c`, another global
+ * option before the subcommand, tmux's unambiguous-prefix abbreviations,
+ * combined global flags) are measured and listed in §6.4 of the same document
+ * rather than guessed at here. This layer narrows the accident, it does not
+ * close the class — `env-clean` (#1740) is the after-the-fact detector.
+ */
+exports.TMUX_UNPINNED_DENY_RULES = [
+    // Destroys every session on the server it reaches. The 2026-08-02 and
+    // 2026-09-08 wipes were both this word.
+    'Bash(tmux kill-server:*)',
+    // `-t` included on purpose: an exact-match `kill-session -t '=name:'` is safe
+    // only about *what* it names, never about *where* — and the static guard's
+    // exemption for that spelling is why agent recipes must pin the socket.
+    'Bash(tmux kill-session:*)',
+    // Server-global mutations outlive the caller and are shared with every session
+    // on that server.
+    'Bash(tmux set-option -g:*)',
+    'Bash(tmux set -g:*)',
+    'Bash(tmux bind-key:*)',
+    'Bash(tmux bind:*)',
+    'Bash(tmux unbind-key:*)',
+    'Bash(tmux unbind:*)',
+];
 /**
  * Bash spellings this server refuses to let an agent run (Issue #1739).
  *
@@ -162,11 +218,15 @@ exports.TOOL_USE_MATCHER = permission_request_payload_1.ASK_USER_QUESTION_TOOL;
  * signals every process the user owns and `kill -9 -<pgid>` a whole group; the
  * default `kill` (SIGTERM) is untouched, so the graceful spelling of the same
  * intent is the one that survives.
+ *
+ * Issue #2442 adds the tmux half of the same story: see
+ * {@link TMUX_UNPINNED_DENY_RULES}.
  */
 exports.PERMISSION_DENY_RULES = [
     'Bash(pkill:*)',
     'Bash(killall:*)',
     'Bash(kill -9:*)',
+    ...exports.TMUX_UNPINNED_DENY_RULES,
 ];
 /** Relay script shipped with the package (`files: ["scripts/hooks/"]`). */
 const RELAY_SCRIPT_RELATIVE_PATH = (0, path_1.join)('scripts', 'hooks', 'cmate-agent-event.sh');