agent-from-zero 0.34.2 → 0.35.0

package/dist/status.js

@@ -0,0 +1,80 @@
+import { CONFIG } from "./config.js";
+import chalk from "chalk";
+import wrapAnsi from "wrap-ansi";
+import { displayWidth } from "./editor.js";
+import { cacheHitRate } from "./cost.js";
+import { getAccountStatus } from "./account-status.js";
+function show(metric, format) {
+    return metric.state === "ok" ? format(metric.value) : chalk.dim(`unavailable — ${metric.reason}`);
+}
+const money = (values) => values.map((m) => `${m.currency} ${m.amount.toFixed(4)}`).join(" · ");
+function accountLines(status) {
+    return [
+        { text: chalk.bold("Vendor account") },
+        { label: "Provider", text: chalk.cyan.bold(status.provider) },
+        { label: "Balance", text: show(status.balance, (values) => chalk.green.bold(money(values))) },
+        { label: "Reported cost", text: show(status.cost, (values) => chalk.yellow.bold(money(values))) },
+        { label: "Account usage", text: show(status.usage, (u) => chalk.cyan(`input ${u.input} (${u.cached} cached) · output ${u.output}`)) },
+        { label: "Period (UTC)", text: status.period },
+        { label: "Checked at", text: chalk.dim(status.checkedAt) },
+        ...(status.note ? [{ text: chalk.dim(status.note) }] : []),
+    ];
+}
+// Both frontends use the same ANSI/CJK-aware layout. Narrow terminals stack
+// labels above values; wider ones align values in a second column.
+function renderLines(lines, width) {
+    return lines.flatMap(({ label, text }) => {
+        const indent = label && width >= 44 ? 18 : 0;
+        const value = wrapAnsi(text, Math.max(1, width - indent), { hard: true, trim: false }).split("\n");
+        if (!label)
+            return value;
+        if (!indent)
+            return [chalk.dim(`${label}:`), ...value];
+        return value.map((line, i) => (i === 0 ? chalk.dim(`${label}:`.padEnd(indent)) : " ".repeat(indent)) + line);
+    });
+}
+export function formatAccountStatus(status) {
+    return renderLines(accountLines(status), 92).join("\n");
+}
+function statusCard(lines, columns) {
+    const width = Math.max(1, Math.min(96, Math.floor(columns)));
+    if (width < 6)
+        return renderLines(lines, width).join("\n");
+    const inner = width - 4; // two borders and one space of padding per side
+    const body = renderLines(lines, inner).map((line) => chalk.dim("│ ") + line + " ".repeat(Math.max(0, inner - displayWidth(line))) + chalk.dim(" │"));
+    return [chalk.dim(`╭${"─".repeat(width - 2)}╮`), ...body, chalk.dim(`╰${"─".repeat(width - 2)}╯`)].join("\n");
+}
+// Shared by Ink and readline. No model call, conversation insertion or logging.
+export async function statusCommand(line, meter, options = {}) {
+    const arg = line.trim().split(/\s+/).slice(1).join(" ");
+    if (arg && !["local", "deepseek", "openai", "anthropic"].includes(arg))
+        return "Usage: /status [local|deepseek|openai|anthropic]";
+    // A URL may itself contain credentials/query tokens; show only its origin.
+    let endpoint = "custom endpoint";
+    try {
+        endpoint = new URL(CONFIG.baseURL).origin;
+    }
+    catch { /* leave generic */ }
+    const usage = meter.snapshot();
+    const local = [
+        { text: chalk.bold("mini-agent · Status") },
+        { text: "" },
+        { label: "Model", text: chalk.cyan.bold(CONFIG.model) },
+        { label: "Endpoint", text: endpoint },
+        { text: "" },
+        { text: chalk.bold("Current session") },
+        { label: "Model calls", text: chalk.cyan(String(usage.calls)) },
+        { label: "Input tokens", text: chalk.cyan(String(usage.inputUncached + usage.inputCached)) },
+        { label: "Cached input", text: `${usage.inputCached} (${(cacheHitRate(usage) * 100).toFixed(0)}% hit rate)` },
+        { label: "Output tokens", text: chalk.cyan(String(usage.output)) },
+        { label: "Estimated cost", text: chalk.yellow.bold(`$${usage.cost.toFixed(4)}`) },
+        { text: "" },
+        { text: chalk.dim("Main/subagent streams only; excludes permission review/Jev and other auxiliary calls.") },
+        { text: chalk.dim("Estimates use settings.pricing, not account billing.") },
+    ];
+    const columns = options.columns ?? (process.stdout.columns || 80);
+    if (arg === "local")
+        return statusCard(local, columns);
+    const account = await getAccountStatus({ baseURL: CONFIG.baseURL, apiKey: CONFIG.apiKey, env: options.env ?? process.env, provider: arg ? arg : undefined, request: options.request, signal: options.signal });
+    return statusCard([...local, { text: "" }, ...accountLines(account)], columns);
+}

package/dist/team.js

@@ -151,14 +151,20 @@ export function teammateCount() {
     return teammates.size;
 }
 // Record a newly spawned teammate. The caller supplies the promise for its loop.
-export function registerTeammate(name, role, done) {
-    teammates.set(name, { name, role, status: "running", state: "active", startedAt: Date.now(), done });
+export function registerTeammate(name, role, done, progress) {
+    teammates.set(name, { name, role, status: "running", state: "active", startedAt: Date.now(), done, progress });
 }
 // Mark a teammate finished. `ok=false` means its loop ended on a non-Done reason.
 export function finishTeammate(name, ok) {
     const t = teammates.get(name);
-    if (t)
+    if (t) {
         t.status = ok ? "done" : "failed";
+        t.progress?.finish();
+    }
+}
+// Unlike /team, this snapshot does not read mailboxes on each UI repaint.
+export function listTeammateViews() {
+    return [...teammates.values()].flatMap((t) => t.progress ? [t.progress.view(`team:${t.name}`, t.name, t.role, t.status === "running" ? t.state === "idle" ? "idle" : "running" : t.status)] : []);
 }
 // How many teammates are still running — at least one running teammate means the
 // team isn't disbanded yet.

package/dist/tools.js

@@ -8,6 +8,7 @@ import { recordMutation } from "./undo.js"; // capture the before-state so /undo
 import { parseTodos, setTodos, summarizeTodos } from "./todos.js"; // the agent's own checklist (Day 36)
 import { startBackground, readBackground } from "./background.js"; // long-running commands that outlive the turn (Day 37)
 import { scheduleJob, cancelJob, listJobs } from "./cron.js"; // cron scheduler (Day s14): scheduled, recurring work
+import { childEnv } from "./auto-debug.js"; // shells never inherit review tracing
 // ============ Session state ============
 // The foundation of "read before edit": which files has this session actually read?
 // Files that were never read must not be edited — otherwise the model edits from
@@ -200,7 +201,7 @@ On error: failures return the error output — analyze it and try a different ap
     // returned. The AbortSignal lets the loop kill the child on interruption
     // instead of waiting out the 30s timeout.
     run: (args, signal) => new Promise((resolve) => {
-        const child = spawn(args.command, { shell: true, stdio: ["ignore", "pipe", "pipe"] }); // shell:true → bash semantics
+        const child = spawn(args.command, { shell: true, stdio: ["ignore", "pipe", "pipe"], env: childEnv() }); // shell:true → bash semantics
         let stdout = ""; // captured stdout
         let stderr = ""; // captured stderr (for the model, not the user's terminal)
         let settled = false; // resolve exactly once
@@ -428,6 +429,17 @@ export function toolDefinitions() {
 // order. (See loop.ts for the batching.)
 const READ_ONLY_TOOLS = new Set(["read_file", "search"]);
 export const isReadOnlyTool = (name) => READ_ONLY_TOOLS.has(name);
+// A server's own claim that a tool loses nothing: it only reads (readOnlyHint),
+// or only adds — a message, an issue — without changing or deleting what exists
+// (destructiveHint: false), so a mistake can be deleted afterwards. Any server,
+// no per-service list. Trusted for auto-mode review only, NOT for the concurrent
+// batching above: that needs our own guarantee. The MCP default is destructive,
+// so an unannotated tool still goes to review.
+export function declaresRecoverable(name) {
+    const hints = tools[name]?.annotations;
+    return hints?.readOnlyHint === true ? "read-only tool (server annotation)"
+        : hints?.destructiveHint === false ? "additive tool (server annotation: not destructive)" : null;
+}
 // Single entry point: every failure becomes text fed back to the model.
 // It never throws — the main loop must never die because of a tool. Async since
 // Day 13 so run_bash can spawn without blocking; the signal lets Ctrl+C kill a

package/dist/tui.js

@@ -78,37 +78,52 @@ export function revealReasoning() {
     process.stdout.write(chalk.dim(reasoningTrace.join("\n\n— — —\n\n").replace(/^/gm, "  ")) + "\n");
     return true;
 }
-// ---- collapsed tool calls (the activity trace) -------------------------------
-// A long task fires dozens of read_file/search calls; printing one line each
-// floods the screen. Instead the loop prints a compact per-round tally and we
-// keep the full call lines here for Ctrl+T to reveal. Cleared per turn, so
-// Ctrl+T shows the calls behind the LAST answer. (Diffs, prompts, and results
-// still print live — only the one-line announcements are folded.)
 const toolCallLog = [];
-const MAX_TOOL_LOG = 500; // bound the memory
+const MAX_TOOL_LOG = 100;
+const MAX_DETAIL_CHARS = 32_000;
+let toolCallCount = 0;
+function boundedDetail(text) {
+    return text.length > MAX_DETAIL_CHARS ? text.slice(0, MAX_DETAIL_CHARS) + "\n… (display truncated)" : text;
+}
 // Record a tool call's full line for Ctrl+T, and stash its args for Tab.
-export function recordToolCall(line, full) {
-    toolCallLog.push(line);
+export function recordToolCall(line, full, action) {
+    const record = { line, full: boundedDetail(full), ...action, args: action ? boundedDetail(action.args) : undefined, startedAt: Date.now() };
+    toolCallCount++;
+    toolCallLog.push(record);
     if (toolCallLog.length > MAX_TOOL_LOG)
         toolCallLog.shift();
-    collapsed.push({ full, expanded: false }); // Tab still reveals the most recent call's full args
+    collapsed.push({ full: record.full, expanded: false });
     if (collapsed.length > MAX_COLLAPSED)
         collapsed.shift();
+    return record;
+}
+export function recordToolResult(record, result) {
+    record.endedAt = Date.now();
+    record.result = boundedDetail(result);
 }
+export function recordToolDetail(id, text) {
+    const record = [...toolCallLog].reverse().find((r) => r.id === id);
+    if (record)
+        record.full = boundedDetail(record.full + "\n" + text);
+}
+export function getToolActivity() { return toolCallLog; }
+export function getToolCallCount() { return toolCallCount; }
 export function clearToolCalls() {
     toolCallLog.length = 0;
+    collapsed.length = 0;
+    toolCallCount = 0;
 }
 // Ctrl+T handler: print every tool call behind the last answer. Returns false
 // (nothing to show) so the editor can ignore the key.
 export function revealToolCalls() {
-    if (!toolCallLog.length)
+    const text = getToolCalls();
+    if (!text)
         return false;
-    process.stdout.write(chalk.dim(`tool calls this turn (${toolCallLog.length}):\n`));
-    process.stdout.write(toolCallLog.join("\n") + "\n");
+    process.stdout.write(text + "\n");
     return true;
 }
 // Ink variants of the two reveals: RETURN the formatted block (or null) instead
-// of writing to stdout, so the Ink REPL can commit it as a note in its tree.
+// of writing to stdout, so Ink can show it in a transient, scrollable panel.
 export function getReasoning() {
     if (!reasoningTrace.length)
         return null;
@@ -117,7 +132,8 @@ export function getReasoning() {
 export function getToolCalls() {
     if (!toolCallLog.length)
         return null;
-    return chalk.dim(`tool calls this turn (${toolCallLog.length}):\n`) + toolCallLog.join("\n");
+    const omitted = toolCallCount > toolCallLog.length ? ` — showing latest ${toolCallLog.length}` : "";
+    return chalk.dim(`tool calls this turn (${toolCallCount})${omitted}:\n`) + toolCallLog.map((r) => `${r.line}\n${r.full}\nresult: ${r.result ?? "(in progress)"}`).join("\n\n");
 }
 // ---- cleanup -----------------------------------------------------------------
 // Called on exit. Nothing fancy to restore now (we no longer hijack the bottom

package/dist/ui.js

@@ -100,17 +100,15 @@ export function formatElapsed(ms) {
     const h = Math.floor(m / 60);
     return `${h}h${String(m % 60).padStart(2, "0")}m`;
 }
-// The live spinner text: the model (so you can SEE which one is answering — proof
-// a /model switch took effect) + a thinking word + elapsed + tokens streamed so
-// far + how to bail. Mirrors Claude Code's "Mulling… (1m32s · ↓ 4.0k tokens)".
-export function spinnerText(word, elapsedSec, subAgent, model, tokens) {
+// Live activity, elapsed time and streamed tokens. The persistent status line
+// owns the model name, so each animation need not repeat it.
+export function spinnerText(word, elapsedSec, subAgent, tokens) {
     const head = subAgent ? "sub-agent" : word;
-    const tag = model && !subAgent ? chalk.dim(`${model} · `) : ""; // show the model on top-level calls
     const bits = [`${elapsedSec}s`];
     if (tokens && tokens > 0)
         bits.push(`↓ ${formatTokens(tokens)} tokens`);
     bits.push("Ctrl+C to interrupt");
-    return `${tag}${head}… ${chalk.dim(`(${bits.join(" · ")})`)}`;
+    return `${head}… ${chalk.dim(`(${bits.join(" · ")})`)}`;
 }
 // The persistent status line — model, project, branch, context use, spend, time.
 // Pure (dir + branch passed in) so it's testable without a filesystem or git.

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-from-zero",
-  "version": "0.34.2",
-  "description": "Build a Claude Code-style CLI agent from scratch, with AI as your pair programmer — companion code for a 10-day tutorial (DeepSeek by default, any OpenAI-compatible model)",
+  "version": "0.35.0",
+  "description": "Build a Claude Code-style CLI agent from scratch, with AI as your pair programmer \u2014 companion code for a 10-day tutorial (DeepSeek by default, any OpenAI-compatible model)",
   "type": "module",
   "bin": {
     "mini-agent": "dist/agent.js"
@@ -15,9 +15,10 @@
     "ink": "tsx --env-file=.env src/ink/main.tsx",
     "typecheck": "tsc --noEmit",
     "eval": "tsx --env-file=.env eval/run.ts",
+    "auto:report": "tsx eval/auto-report.ts",
     "build": "tsc -p tsconfig.build.json",
     "prepare": "npm run build",
-    "test": "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/memory.test.ts && tsx tests/cost.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/approval.test.ts && tsx tests/form.test.ts && tsx tests/model.test.ts && tsx tests/reasoning.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/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/board.test.ts && tsx tests/drop.test.ts && tsx tests/editline.test.ts && tsx tests/cron.test.ts"
   },
   "engines": {
     "node": ">=20.6"