agent-from-zero 0.35.4 → 0.35.5

package/README.md

@@ -115,6 +115,12 @@ export MINI_AGENT_MODEL=gpt-4.1-mini
 
 输入 `/mcp` 选一个服务器，会看到和 Claude Code 一样的详情（状态、登录、协议版本、URL、配置文件位置、能力、工具数）和操作：View tools、Re-authenticate、Clear authentication、Reconnect、Disable。也可以直接敲 `/mcp tools|auth|clear-auth|reconnect|disable|enable <名字>`。
 
+### 下一句预测
+
+和 Claude Code 一样，每轮对话结束后，mini-agent 会猜你下一句大概要说什么（比如 “跑一下测试”、“提交并推送”），灰色显示在空的输入框里。按 Tab（或 →）变成真正的输入，可以改，也可以直接 Enter 发送。开始打字它就消失。
+
+生成预测要多调一次模型，但请求内容和刚结束的主调用相同，只在最后加了一条指令，所以基本都能命中 DeepSeek 的前缀缓存，费用很低，也会算进 `/cost`。定时任务和其他实例发来的消息触发的回合不做预测。不想要可以在 `settings.json` 里设 `"promptSuggestions": false`，或者单次运行时设环境变量 `MINI_AGENT_PROMPT_SUGGESTIONS=0`。
+
 ### 多开实例互相通信
 
 在几个终端里各开一个 mini-agent（比如一个在后端仓库、一个在前端仓库），它们会自动互相发现，不用配置。

package/dist/config.js

@@ -136,6 +136,10 @@ export const CONFIG = {
     memory: {
         autoExtract: projectSettings.memory?.autoExtract ?? globalSettings.memory?.autoExtract ?? false,
     },
+    // Next-prompt suggestions (Claude Code's placeholder you accept with Tab). On
+    // by default: the call reuses the main conversation's cached prefix, so it is
+    // cheap. MINI_AGENT_PROMPT_SUGGESTIONS=0 turns it off for one run.
+    promptSuggestions: process.env.MINI_AGENT_PROMPT_SUGGESTIONS === "0" ? false : (projectSettings.promptSuggestions ?? globalSettings.promptSuggestions ?? true),
 };
 // Persist a single key to the global settings file, preserving all other keys.
 // Used by /model to remember the user's last choice across sessions.

package/dist/ink/app.js

@@ -10,6 +10,8 @@ import { initFormState, reduceForm, renderForm, collectAnswers } from "../form.j
 import { CONFIG, saveGlobalSetting } from "../config.js"; // session allowlist + /model save
 import { effortMenu, setEffort } from "../effort.js";
 import { useSlashCompletion } from "./completion.js";
+import { generatePromptSuggestion } from "../suggestion.js"; // the predicted next message (Claude Code's prompt suggestion)
+import { lastRequestTools } from "../loop.js"; // resend the main call's tools so the guess hits the prompt cache
 import { commandNames, commandTokenRanges, inlineCommandGhost } from "../completion.js"; // which /command tokens to paint + the mid-text ghost suggestion
 import { TerminateReason, listAgentViews } from "../loop.js"; // how a turn can end
 import { compactHistory, compactThreshold, contextWindowFor } from "../context.js"; // /compact + /model info
@@ -178,6 +180,8 @@ export function App({ session, runTurn, clipboard }) {
     const [detailOffset, setDetailOffset] = useState(0);
     const [live, setLive] = useState(null); // the streaming answer
     const [input, setInput] = useState(""); // the current input buffer
+    const [suggestion, setSuggestion] = useState(null); // the predicted next message: a dim placeholder, Tab accepts it
+    const suggestionSeq = useRef(0); // bumped by every new turn, so a late guess for an old turn is dropped
     const [talkTo, setTalkTo] = useState(null); // /peers → talk: the input box sends to another session instead of this agent
     const replyTo = useRef([]); // who gets this turn's final answer (the senders of "user" peer messages)
     const [images, setImages] = useState([]);
@@ -225,6 +229,8 @@ export function App({ session, runTurn, clipboard }) {
     }, rows - 10, columns);
     // A "/" later in the text: no list, just the rest of the best match as dim ghost text (Tab accepts).
     const ghost = completing && cursor === input.length ? inlineCommandGhost(input, { skills: typeaheadSkills }) : null;
+    // The predicted next message, shown only in an empty box while nothing else wants attention.
+    const placeholder = suggestion && !input && !busy && !pending && !talkTo && !details && subAgentFocus === null ? suggestion : null;
     useEffect(() => {
         const id = setInterval(() => setTick((t) => t + 1), 1000);
         return () => clearInterval(id);
@@ -253,6 +259,11 @@ export function App({ session, runTurn, clipboard }) {
         clearToolCalls(); // ...and Ctrl+T THIS turn's tool calls
         const senders = replyTo.current; // taken now, so a failed turn can't leak them into the next one
         replyTo.current = [];
+        // Guess the next message only after turns YOU started — not cron, not
+        // another agent's message, not your message relayed from another window.
+        const suggestAfter = CONFIG.promptSuggestions && !senders.length && (content === null || display?.kind === "user");
+        const seq = ++suggestionSeq.current;
+        setSuggestion(null);
         setBusy(true);
         setSessionState("busy", CONFIG.model); // other sessions see "busy" in list_peers
         const controller = new AbortController();
@@ -294,6 +305,12 @@ export function App({ session, runTurn, clipboard }) {
             if (result.reason !== TerminateReason.Done && !(result.reason === TerminateReason.UserInterrupt && followUps.size))
                 note(chalk.yellow(`⚠️ ${EXIT_NOTES[result.reason] ?? result.reason}`));
             saveSession(sessionId, model, messages, pendingTitle.current); // snapshot after every turn — crash-safe by construction
+            if (suggestAfter && result.reason === TerminateReason.Done) {
+                void generatePromptSuggestion(client, CONFIG.model, messages, lastRequestTools()).then((guess) => {
+                    if (guess && seq === suggestionSeq.current)
+                        setSuggestion(guess); // no newer turn started meanwhile (the box shows it only while idle)
+                });
+            }
             if (CONFIG.memory.autoExtract && result.reason === TerminateReason.Done) {
                 try {
                     const got = await extractMemories(client, CONFIG.subAgentModel || model, messages);
@@ -952,6 +969,13 @@ export function App({ session, runTurn, clipboard }) {
                 return;
             }
         }
+        if ((key.tab || (key.rightArrow && !input)) && placeholder) { // Tab (or →) turns the predicted next message into real text; Enter then sends it
+            setInput(placeholder);
+            setCursor(placeholder.length);
+            setHistIdx(null);
+            setSuggestion(null);
+            return;
+        }
         if (key.tab && ghost) { // accept the inline suggestion as "/name "
             const next = input.slice(0, ghost.start) + `/${ghost.full} `;
             setInput(next);
@@ -1193,7 +1217,7 @@ export function App({ session, runTurn, clipboard }) {
                                     const at = line.slice(cuOff, cuOff + 1) || " ";
                                     const after = line.slice(cuOff + 1);
                                     const base = lineStarts[i] ?? 0;
-                                    return (_jsxs(Box, { children: [i === 0 && _jsx(Text, { color: promptColor, children: prompt }), _jsx(Text, { children: paint(before, base, "b") }), _jsx(Text, { inverse: true, color: isLit(base + cuOff) ? SKILL_COLOR : undefined, children: ghost ? ghost.suffix[0] : at }), _jsx(Text, { children: paint(after, base + cuOff + at.length, "a") }), ghost && _jsx(Text, { dimColor: true, children: ghost.suffix.slice(1) })] }, i));
+                                    return (_jsxs(Box, { children: [i === 0 && _jsx(Text, { color: promptColor, children: prompt }), _jsx(Text, { children: paint(before, base, "b") }), _jsx(Text, { inverse: true, color: isLit(base + cuOff) ? SKILL_COLOR : undefined, children: ghost ? ghost.suffix[0] : placeholder && i === 0 ? [...placeholder][0] : at }), _jsx(Text, { children: paint(after, base + cuOff + at.length, "a") }), ghost && _jsx(Text, { dimColor: true, children: ghost.suffix.slice(1) }), placeholder && i === 0 && _jsxs(Text, { dimColor: true, wrap: "truncate-end", children: [[...placeholder].slice(1).join(""), _jsx(Text, { dimColor: true, children: "  ⇥ tab" })] })] }, i));
                                 }
                                 return (_jsxs(Box, { children: [i === 0 && _jsx(Text, { color: promptColor, children: prompt }), _jsx(Text, { children: paint(line, lineStarts[i] ?? 0, "l") })] }, i));
                             });

package/dist/ink/commands.js

@@ -46,6 +46,7 @@ export const SESSION_HELP = `commands:
   exit       leave (Ctrl+C does the same)
 
 keys (during or after a turn):
+  Tab / →    accept the predicted next message (the dim placeholder after a turn); Enter sends it
   /          show command suggestions below the input (while idle)
   ↑↓ / Tab   choose / complete a suggestion; Enter selects; Esc closes
   Enter      while working, queue a follow-up for the next tool boundary

package/dist/loop.js

@@ -44,6 +44,16 @@ export var TerminateReason;
     TerminateReason["FatalApiError"] = "fatal_api_error";
     TerminateReason["UserInterrupt"] = "user_interrupt";
 })(TerminateReason || (TerminateReason = {}));
+// The tool list of the last TOP-LEVEL model call. The next-prompt suggestion
+// (suggestion.ts) resends the same tools so its request shares the main
+// conversation's prefix — and the provider's prompt cache with it.
+let lastTopLevelTools = [];
+export const lastRequestTools = () => lastTopLevelTools;
+function rememberTools(opts, tools) {
+    if (!opts.subAgent)
+        lastTopLevelTools = tools;
+    return tools;
+}
 // Resolve the output sink for a loop: the caller's sink, or stdout (the old
 // behaviour) when none was passed. Centralised so every site reads the same way.
 function sink(opts) {
@@ -456,7 +466,7 @@ opts) {
         // The tool manual depends on the agent kind (Lead / teammate / sub-agent)
         // — see toolsFor(). This is also where nested spawning is prevented: a
         // teammate's manual simply omits spawn_teammate and task.
-        { model: opts.model, messages: effortMessages(opts.model, messages), ...effortParameters(opts.model), tools: toolsFor(opts), stream: true, stream_options: { include_usage: true } }, { signal });
+        { model: opts.model, messages: effortMessages(opts.model, messages), ...effortParameters(opts.model), tools: rememberTools(opts, toolsFor(opts)), stream: true, stream_options: { include_usage: true } }, { signal });
         let content = ""; // accumulated answer text
         const calls = []; // tool calls under assembly, indexed by delta.index
         // Stash thinking without appending one indicator per model round. The live

package/dist/suggestion.js

@@ -0,0 +1,55 @@
+import { effortMessages, effortProfile } from "./effort.js"; // same message shape as the main loop sends
+import { recordUsage } from "./cost.js"; // the suggestion call shows up in /cost like any other
+// ---- Next-prompt suggestion (Claude Code's prompt suggestion) -------------------
+// When a turn ends, guess what you'll type next and show it as a dim placeholder
+// in the empty input box: Tab turns it into real text, Enter sends it. It's
+// most useful at the obvious moments — "run the tests", "commit it", "yes, do
+// the same for the other file" — so the guess must be short and concrete, or
+// nothing at all.
+//
+// Cost: the request is the SAME history plus the SAME tools as the main call
+// that just finished, with one instruction appended. That shared prefix is
+// what the provider's prompt cache keys on (DeepSeek caches automatically),
+// so almost all of the input is billed at the cached rate. tool_choice "none"
+// keeps the tools in the prompt (cache) without letting the model call one.
+const SUGGESTION_TIMEOUT_MS = 15_000;
+const MAX_CHARS = 120;
+const SUGGESTION_PROMPT = `[Next-prompt suggestion — this message is from the app, not the user]
+Predict the user's NEXT message in this conversation: what they would most likely type now, given what just happened.
+- Write it AS THE USER, addressed to the assistant (e.g. "run the tests", "commit and push", "也改一下另一个文件").
+- Short and concrete: at most 12 words, or 30 characters for Chinese/Japanese. Same language the user writes in.
+- Only a natural next step the user would actually want. No questions about preferences, no thanks, no pleasantries.
+- If there is no clear next step, reply exactly: NONE
+Reply with ONLY the message text — no quotes, no explanation.`;
+// Keep only something that reads like a user's short next message. Anything
+// else — an answer, a plan, the assistant's own voice — is worse than nothing.
+export function cleanSuggestion(raw) {
+    let s = raw.trim().replace(/^["'“”「」`]+|["'“”「」`]+$/g, "").trim();
+    if (!s || /^none\.?$/i.test(s) || s.includes("\n"))
+        return null;
+    if (s.length > MAX_CHARS)
+        return null;
+    if (/^(i'll|i will|i've|i have|let me|sure|okay,? i|here's|the user)\b/i.test(s))
+        return null; // the assistant talking, not the user
+    if (/^(好的|我来|我会|我已经|用户)/.test(s))
+        return null;
+    s = s.replace(/\s+/g, " ");
+    return s;
+}
+export async function generatePromptSuggestion(client, model, messages, tools = [], signal) {
+    try {
+        const res = await client.chat.completions.create({
+            model, // the conversation's model: its cache holds this prefix
+            messages: [...effortMessages(model, messages), { role: "user", content: SUGGESTION_PROMPT }],
+            ...(tools.length ? { tools, tool_choice: "none" } : {}),
+            ...(effortProfile(model)?.parameter === "deepseek" ? { thinking: { type: "disabled" } } : {}), // a one-line guess needs no reasoning
+            max_tokens: 60,
+            stream: false,
+        }, { timeout: SUGGESTION_TIMEOUT_MS, signal });
+        recordUsage(res.usage);
+        return cleanSuggestion(res.choices?.[0]?.message?.content ?? "");
+    }
+    catch {
+        return null; // a failed guess is simply no guess
+    }
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-from-zero",
-  "version": "0.35.4",
+  "version": "0.35.5",
   "type": "module",
   "bin": {
     "mini-agent": "dist/agent.js"
@@ -17,7 +17,7 @@
     "auto:report": "tsx eval/auto-report.ts",
     "build": "tsc -p tsconfig.build.json",
     "prepare": "npm run build",
-    "test": "export MINI_AGENT_REVIEW_DEBUG=0 && tsx tests/tools.test.ts && tsx tests/permissions.test.ts && tsx tests/context.test.ts && tsx tests/hooks.test.ts && tsx tests/mcp.test.ts && tsx tests/judge.test.ts && tsx tests/auto.test.ts && tsx tests/auto-context.test.ts && tsx tests/auto-providers.test.ts && tsx tests/auto-cascade.test.ts && tsx tests/auto-rules.test.ts && tsx tests/auto-rule-cascade.test.ts && tsx tests/auto-debug.test.ts && tsx tests/auto-report.test.ts && tsx tests/auto-clarification.test.ts && tsx tests/memory.test.ts && tsx tests/cost.test.ts && tsx tests/cli.test.ts && tsx tests/print.test.ts && tsx tests/status.test.ts && tsx tests/validate.test.ts && tsx tests/plan.test.ts && tsx tests/diff.test.ts && tsx tests/undo.test.ts && tsx tests/tiering.test.ts && tsx tests/mcp-http.test.ts && tsx tests/mcp-oauth.test.ts && tsx tests/changes.test.ts && tsx tests/mentions.test.ts && tsx tests/session.test.ts && tsx tests/title.test.ts && tsx tests/ui.test.ts && tsx tests/activity.test.ts && tsx tests/follow-up.test.ts && tsx tests/follow-up-ui.test.tsx && tsx tests/agents.test.ts && tsx tests/images.test.ts && tsx tests/approval.test.ts && tsx tests/form.test.ts && tsx tests/model.test.ts && tsx tests/reasoning.test.ts && tsx tests/effort.test.ts && tsx tests/completion.test.ts && tsx tests/pasted-text.test.ts && tsx tests/skills.test.ts && tsx tests/markdown.test.ts && tsx tests/editor.test.ts && tsx tests/todos.test.ts && tsx tests/background.test.ts && tsx tests/team.test.ts && tsx tests/peers.test.ts && tsx tests/board.test.ts && tsx tests/drop.test.ts && tsx tests/editline.test.ts && tsx tests/cron.test.ts"
+    "test": "export MINI_AGENT_REVIEW_DEBUG=0 && tsx tests/tools.test.ts && tsx tests/permissions.test.ts && tsx tests/context.test.ts && tsx tests/hooks.test.ts && tsx tests/mcp.test.ts && tsx tests/judge.test.ts && tsx tests/auto.test.ts && tsx tests/auto-context.test.ts && tsx tests/auto-providers.test.ts && tsx tests/auto-cascade.test.ts && tsx tests/auto-rules.test.ts && tsx tests/auto-rule-cascade.test.ts && tsx tests/auto-debug.test.ts && tsx tests/auto-report.test.ts && tsx tests/auto-clarification.test.ts && tsx tests/memory.test.ts && tsx tests/cost.test.ts && tsx tests/cli.test.ts && tsx tests/print.test.ts && tsx tests/status.test.ts && tsx tests/validate.test.ts && tsx tests/plan.test.ts && tsx tests/diff.test.ts && tsx tests/undo.test.ts && tsx tests/tiering.test.ts && tsx tests/mcp-http.test.ts && tsx tests/mcp-oauth.test.ts && tsx tests/changes.test.ts && tsx tests/mentions.test.ts && tsx tests/session.test.ts && tsx tests/title.test.ts && tsx tests/ui.test.ts && tsx tests/activity.test.ts && tsx tests/follow-up.test.ts && tsx tests/follow-up-ui.test.tsx && tsx tests/agents.test.ts && tsx tests/images.test.ts && tsx tests/approval.test.ts && tsx tests/form.test.ts && tsx tests/model.test.ts && tsx tests/reasoning.test.ts && tsx tests/effort.test.ts && tsx tests/completion.test.ts && tsx tests/pasted-text.test.ts && tsx tests/skills.test.ts && tsx tests/markdown.test.ts && tsx tests/editor.test.ts && tsx tests/todos.test.ts && tsx tests/background.test.ts && tsx tests/team.test.ts && tsx tests/peers.test.ts && tsx tests/suggestion.test.ts && tsx tests/board.test.ts && tsx tests/drop.test.ts && tsx tests/editline.test.ts && tsx tests/cron.test.ts"
   },
   "engines": {
     "node": ">=20.6"