agent-sh 0.12.18 → 0.12.20

package/dist/agent/agent-loop.js

@@ -1527,6 +1527,7 @@ export class AgentLoop {
             messages,
             tools: apiTools,
             model: this.currentModel,
+            max_tokens: this.currentMode.maxTokens ?? 65536,
             ...this.reasoningParams(),
         };
         this.bus.emit("llm:request", requestParams);

package/dist/agent/conversation-state.d.ts

@@ -59,6 +59,9 @@ export declare class ConversationState {
     private hasOpenToolCalls;
     private flushPendingNotes;
     getMessages(): ChatCompletionMessageParam[];
+    /** Drop tool messages with no matching preceding tool_call — strict
+     *  providers (DeepSeek) 400, and compaction can leave such orphans. */
+    private dropOrphanToolMessages;
     /**
      * If a stream was interrupted mid-tool-execution, an assistant message
      * with tool_calls can land in history without matching tool results.

package/dist/agent/conversation-state.js

@@ -163,7 +163,24 @@ export class ConversationState {
         this.invalidateMessagesCache();
     }
     getMessages() {
-        return this.normalizeReasoningConsistency(this.stubDanglingToolCalls(this.messages));
+        return this.normalizeReasoningConsistency(this.stubDanglingToolCalls(this.dropOrphanToolMessages(this.messages)));
+    }
+    /** Drop tool messages with no matching preceding tool_call — strict
+     *  providers (DeepSeek) 400, and compaction can leave such orphans. */
+    dropOrphanToolMessages(messages) {
+        const knownIds = new Set();
+        const result = [];
+        for (const msg of messages) {
+            if (msg.role === "assistant" && "tool_calls" in msg && msg.tool_calls) {
+                for (const tc of msg.tool_calls)
+                    knownIds.add(tc.id);
+            }
+            if (msg.role === "tool" && !knownIds.has(msg.tool_call_id)) {
+                continue;
+            }
+            result.push(msg);
+        }
+        return result;
     }
     /**
      * If a stream was interrupted mid-tool-execution, an assistant message
@@ -582,19 +599,23 @@ export class ConversationState {
     slimTurn(messages) {
         const MAX_RESULT_LEN = 1500;
         const result = [];
-        const readOnlyToolIds = new Set();
+        const droppedToolIds = new Set();
         for (const msg of messages) {
             if (msg.role === "assistant" && "tool_calls" in msg && msg.tool_calls) {
                 const kept = msg.tool_calls.filter((tc) => {
                     if (!("function" in tc))
                         return true;
                     if (READ_ONLY_TOOLS.has(tc.function.name)) {
-                        readOnlyToolIds.add(tc.id);
+                        droppedToolIds.add(tc.id);
                         return false;
                     }
                     return true;
                 });
                 if (kept.length === 0) {
+                    // No content + no tool_calls is malformed (DeepSeek 400); drop the husk.
+                    const text = typeof msg.content === "string" ? msg.content.trim() : "";
+                    if (!text)
+                        continue;
                     const { tool_calls: _, ...rest } = msg;
                     result.push(rest);
                 }
@@ -604,7 +625,7 @@ export class ConversationState {
                 continue;
             }
             if (msg.role === "tool") {
-                if (readOnlyToolIds.has(msg.tool_call_id))
+                if (droppedToolIds.has(msg.tool_call_id))
                     continue;
                 const content = typeof msg.content === "string" ? msg.content : "";
                 if (content.length > MAX_RESULT_LEN) {

package/dist/event-bus.d.ts

@@ -69,6 +69,7 @@ export interface ShellEvents {
         messages: unknown[];
         tools?: unknown;
         model?: string;
+        max_tokens?: number;
         reasoning_effort?: string;
     };
     "llm:chunk": {
@@ -318,6 +319,7 @@ export interface ShellEvents {
             id: string;
             reasoning?: boolean;
             contextWindow?: number;
+            maxTokens?: number;
             echoReasoning?: boolean;
         })[];
         /** Provider supports the reasoning_effort parameter. Default: true. */

package/dist/extensions/agent-backend.js

@@ -36,6 +36,7 @@ export default function agentBackend(ctx) {
                     provider: id,
                     providerConfig: { apiKey: p.apiKey, baseURL: p.baseURL },
                     contextWindow: mc?.contextWindow ?? p.contextWindow,
+                    maxTokens: mc?.maxTokens ?? (mc?.contextWindow ? Math.min(Math.floor(mc.contextWindow * 0.4), 65536) : undefined),
                     reasoning: mc?.reasoning,
                     supportsReasoningEffort: p.supportsReasoningEffort,
                     echoReasoning: mc?.echoReasoning,
@@ -149,7 +150,7 @@ export default function agentBackend(ctx) {
             }
             else {
                 modelIds.push(m.id);
-                caps.set(m.id, { reasoning: m.reasoning, contextWindow: m.contextWindow, echoReasoning: m.echoReasoning });
+                caps.set(m.id, { reasoning: m.reasoning, contextWindow: m.contextWindow, maxTokens: m.maxTokens, echoReasoning: m.echoReasoning });
             }
         }
         providerRegistry.set(p.id, {
@@ -169,6 +170,7 @@ export default function agentBackend(ctx) {
                 provider: p.id,
                 providerConfig: { apiKey: p.apiKey ?? "", baseURL: p.baseURL },
                 contextWindow: mc?.contextWindow,
+                maxTokens: mc?.maxTokens,
                 reasoning: mc?.reasoning,
                 supportsReasoningEffort: p.supportsReasoningEffort,
                 echoReasoning: mc?.echoReasoning,
@@ -212,6 +214,7 @@ export default function agentBackend(ctx) {
                 provider: name,
                 providerConfig: { apiKey: p.apiKey, baseURL: p.baseURL },
                 contextWindow: mc?.contextWindow ?? p.contextWindow,
+                maxTokens: mc?.maxTokens ?? (mc?.contextWindow ? Math.min(Math.floor(mc.contextWindow * 0.4), 65536) : undefined),
                 reasoning: mc?.reasoning,
                 supportsReasoningEffort: p.supportsReasoningEffort,
                 echoReasoning: mc?.echoReasoning,

package/dist/extensions/tui-renderer.js

@@ -654,26 +654,16 @@ export default function activate(ctx) {
             return [];
         const boxW = Math.min(120, width - 2);
         const contentW = boxW - 4;
-        let body;
-        if (diff.isNewFile) {
-            const lines = diff.hunks.flatMap(h => h.lines.map(l => l.text));
-            const preview = getSettings().newFilePreviewLines;
-            const head = lines.slice(0, preview);
-            const truncated = head.map(l => l.length > contentW ? l.slice(0, contentW - 1) + "…" : l);
-            const more = lines.length > preview
-                ? [`${p.dim}… ${lines.length - preview} more lines${p.reset}`]
-                : [];
-            body = ["", ...truncated, ...more, ""];
-        }
-        else {
-            const diffLines = renderDiff(diff, {
-                width: contentW,
-                filePath,
-                maxLines: getSettings().diffMaxLines,
-                trueColor: true,
-            });
-            body = diffLines.length > 1 ? ["", ...diffLines.slice(1), ""] : diffLines;
-        }
+        const maxLines = diff.isNewFile
+            ? getSettings().newFilePreviewLines
+            : getSettings().diffMaxLines;
+        const diffLines = renderDiff(diff, {
+            width: contentW,
+            filePath,
+            maxLines,
+            trueColor: true,
+        });
+        const body = diffLines.length > 1 ? ["", ...diffLines.slice(1), ""] : diffLines;
         return renderBoxFrame(body, {
             width: boxW,
             style: "rounded",

package/dist/settings.d.ts

@@ -9,6 +9,8 @@ export interface ModelCapabilityConfig {
     reasoning?: boolean;
     /** Context window size in tokens for this specific model. */
     contextWindow?: number;
+    /** Max output tokens for this model. */
+    maxTokens?: number;
     /** Echo reasoning_content back on assistant turns. Required by DeepSeek. */
     echoReasoning?: boolean;
 }
@@ -148,6 +150,7 @@ export interface ResolvedProvider {
     modelCapabilities?: Map<string, {
         reasoning?: boolean;
         contextWindow?: number;
+        maxTokens?: number;
         echoReasoning?: boolean;
     }>;
     /** Borrow another registered provider's reasoning request shape by id. */

package/dist/settings.js

@@ -148,8 +148,8 @@ export function resolveProvider(name) {
         }
         else {
             modelIds.push(m.id);
-            if (m.reasoning !== undefined || m.contextWindow !== undefined || m.echoReasoning !== undefined) {
-                caps.set(m.id, { reasoning: m.reasoning, contextWindow: m.contextWindow, echoReasoning: m.echoReasoning });
+            if (m.reasoning !== undefined || m.contextWindow !== undefined || m.maxTokens !== undefined || m.echoReasoning !== undefined) {
+                caps.set(m.id, { reasoning: m.reasoning, contextWindow: m.contextWindow, maxTokens: m.maxTokens, echoReasoning: m.echoReasoning });
             }
         }
     }

package/dist/types.d.ts

@@ -41,6 +41,8 @@ export interface AgentMode {
     };
     /** Context window size in tokens (for usage display). */
     contextWindow?: number;
+    /** Max output tokens for this mode. */
+    maxTokens?: number;
     /** Model supports reasoning/thinking tokens. */
     reasoning?: boolean;
     /** Provider supports the reasoning_effort parameter. */

package/dist/utils/llm-client.js

@@ -44,7 +44,7 @@ export class LlmClient {
             model: opts.model ?? this.model,
             messages: opts.messages,
             tools: opts.tools?.length ? opts.tools : undefined,
-            max_tokens: opts.max_tokens ?? 8192,
+            max_tokens: opts.max_tokens ?? 65536,
             stream: true,
             stream_options: { include_usage: true },
             ...(opts.reasoning_effort

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-sh",
-  "version": "0.12.18",
+  "version": "0.12.20",
   "description": "A shell-first terminal where AI is one keystroke away",
   "type": "module",
   "main": "dist/core.js",