linkshell-cli 0.5.19 → 0.5.21

package/dist/cli/src/runtime/acp/agent-workspace.js

@@ -708,6 +708,8 @@ const CLAUDE_REMOTE_HIDDEN_COMMANDS = new Set([
     "resume",
     "rewind",
     "settings",
+    "setup-bedrock",
+    "setup-vertex",
     "statusline",
     "stickers",
     "tasks",
@@ -757,6 +759,7 @@ const CLAUDE_BUILT_IN_COMMANDS = [
     { name: "bug", description: "Alias for /feedback" },
     { name: "fewer-permission-prompts", description: "Reduce common permission prompts", argsMode: "none" },
     { name: "focus", description: "Toggle the focus view", argsMode: "none" },
+    { name: "goal", description: "Set or clear a session goal the agent works toward", argsMode: "raw" },
     { name: "heapdump", description: "Write a JavaScript heap snapshot for diagnostics", argsMode: "none" },
     { name: "help", description: "Show help and available commands", argsMode: "none" },
     { name: "hooks", description: "View hook configurations", argsMode: "none" },
@@ -977,7 +980,20 @@ function mergeCommands(...groups) {
             map.set(key, {
                 ...existing,
                 ...command,
+                // Preserve curated metadata the discovered set doesn't carry: the SDK's
+                // SlashCommand has no destructive/category/argsMode/executionKind
+                // fields, and runtimeCommands hardcodes argsMode:"raw" + executionKind:
+                // "prompt". The merge order is (curated, discovered), so without these
+                // a discovered /clear would clobber the curated destructive flag
+                // (red-dot warning), the grouping, AND the curated argsMode:"none" that
+                // makes the web palette execute it on pick (vs insert-and-wait). For
+                // Codex it would also flip native commands to "prompt". Curated wins on
+                // collision; genuinely new discovered commands fall through to "raw".
+                destructive: command.destructive ?? existing?.destructive,
+                category: command.category ?? existing?.category,
                 disabledReason: command.disabledReason ?? existing?.disabledReason,
+                argsMode: existing?.argsMode ?? command.argsMode,
+                executionKind: existing?.executionKind ?? command.executionKind,
             });
         }
     }
@@ -1023,7 +1039,7 @@ function runtimeCommands(provider, value) {
     })
         .filter((entry) => Boolean(entry));
 }
-function parseModelListCapabilities(value) {
+function parseModelListCapabilities(provider, value) {
     const raw = asRecord(value);
     const modelsValue = Array.isArray(value) ? value :
         Array.isArray(raw?.data) ? raw.data :
@@ -1076,12 +1092,14 @@ function parseModelListCapabilities(value) {
     const reasoningEfforts = efforts.size
         ? ALL_REASONING_EFFORTS.filter((effort) => efforts.has(effort))
         : undefined;
-    if (models.length === 0 && !defaultModel && !reasoningEfforts?.length)
+    const commands = raw && Array.isArray(raw.commands) ? runtimeCommands(provider, raw.commands) : undefined;
+    if (models.length === 0 && !defaultModel && !reasoningEfforts?.length && !commands?.length)
         return undefined;
     return {
         ...(models.length > 0 ? { models } : {}),
         ...(defaultModel ? { defaultModel } : {}),
         ...(reasoningEfforts?.length ? { reasoningEfforts } : {}),
+        ...(commands?.length ? { commands } : {}),
     };
 }
 function parseTimestamp(value) {
@@ -1115,6 +1133,11 @@ function parseRemoteSessions(value) {
         const id = firstString(source, ["id", "threadId", "sessionId", "agentSessionId"]);
         if (!id)
             continue;
+        // Only accept a transcript-derived status the provider actually reported
+        // (Claude's listSessions sets "running"/"idle" from the on-disk transcript).
+        // Anything else is left undefined so the sync site keeps its own status.
+        const reportedStatus = firstString(source, ["status"]);
+        const status = reportedStatus === "running" || reportedStatus === "idle" ? reportedStatus : undefined;
         result.push({
             id,
             cwd: firstString(source, ["cwd", "workingDirectory", "workspacePath"]),
@@ -1123,6 +1146,7 @@ function parseRemoteSessions(value) {
             createdAt: parseTimestamp(source.createdAt ?? source.created_at),
             lastActivityAt: parseTimestamp(source.lastActivityAt ?? source.updatedAt ?? source.modifiedAt ?? source.lastModified ?? source.updated_at),
             usage: source.usage,
+            ...(status ? { status } : {}),
         });
     }
     return result;
@@ -1687,7 +1711,7 @@ export class AgentWorkspaceProxy {
         if (typeof listModels === "function") {
             try {
                 const result = await listModels.call(client);
-                runtimeCapabilities = parseModelListCapabilities(result);
+                runtimeCapabilities = parseModelListCapabilities(provider, result);
             }
             catch (error) {
                 if (this.input.verbose) {
@@ -1750,6 +1774,26 @@ export class AgentWorkspaceProxy {
                     const usage = usageFields
                         ? { ...usageFields, updatedAt: remote.lastActivityAt ?? now }
                         : existing?.usage;
+                    // Status precedence. While LinkShell is driving a turn for this
+                    // conversation, the in-memory status is authoritative — the transcript
+                    // lags (the SDK hasn't flushed the in-flight turn to disk), so trusting
+                    // it would flicker a live "running" turn back to idle. Otherwise the
+                    // transcript-derived remote.status is authoritative for Claude (it
+                    // reports running/idle straight from disk): surface "running" while an
+                    // external process drives it, and let its "idle" clear a previously
+                    // observed "running" once the external turn finishes. We only fall
+                    // back to the existing status to keep error/waiting_permission sticky
+                    // across a re-sync — deriveTranscriptStatus never emits those, so they
+                    // live only in `existing`; defaulting everything else to idle avoids a
+                    // purely-external session getting stuck at "running" forever.
+                    const linkshellDriving = this.currentTurnIds.has(conversationId);
+                    const status = linkshellDriving
+                        ? existing?.status ?? "running"
+                        : remote.status === "running"
+                            ? "running"
+                            : existing?.status === "error" || existing?.status === "waiting_permission"
+                                ? existing.status
+                                : "idle";
                     const conversation = {
                         id: conversationId,
                         agentSessionId,
@@ -1760,7 +1804,7 @@ export class AgentWorkspaceProxy {
                         reasoningEffort: existing?.reasoningEffort,
                         permissionMode: existing?.permissionMode,
                         collaborationMode: existing?.collaborationMode,
-                        status: existing?.status ?? "idle",
+                        status,
                         archived: existing?.archived ?? false,
                         lastMessagePreview: existing?.lastMessagePreview,
                         lastActivityAt: remote.lastActivityAt ?? existing?.lastActivityAt ?? now,
@@ -1854,6 +1898,22 @@ export class AgentWorkspaceProxy {
         const model = firstString(thread, ["model", "modelId", "currentModel"]);
         if (model && !conversation.model)
             conversation.model = model;
+        // Seed a scroll-back cursor for Claude, which hydrates via readSession (full
+        // transcript) yet pages via listTurns. The timeline only keeps the last
+        // MAX_TIMELINE_ITEMS, so anything older is dropped here but reachable via
+        // agent.v2.history.request. Cursor = total turn count (end-exclusive): the
+        // first page overlaps the displayed tail (harmlessly deduped by the client's
+        // mergeTimeline) and each subsequent page walks strictly older, terminating
+        // when the cursor reaches 0. This integer-index cursor is Claude-specific —
+        // Codex uses an opaque string cursor captured by the listTurns branch above,
+        // so gate strictly on the provider rather than on listTurns presence (Codex
+        // also exposes listTurns) to avoid poisoning Codex's cursor.
+        if (conversation.provider === "claude" &&
+            !this.historyCursors.has(conversation.id)) {
+            const turnCount = Array.isArray(thread?.turns) ? thread.turns.length : 0;
+            if (turnCount > 0)
+                this.historyCursors.set(conversation.id, String(turnCount));
+        }
         const hydratedItems = timelineItemsFromProviderThread(source, conversation.id);
         if (hydratedItems.length === 0)
             return;