s-app-scoder-agent 0.7.0 → 0.8.1

package/README.md

@@ -272,7 +272,7 @@ The autonomous loop is wrapped by a configurable rule engine (`src/guardrails.ts
 | `chat-done` | Ends conversational turns (marker-based) without a build |
 | `completion-claim` | Never trusts "the task is complete" — forces an independent build verification |
 | `text-only-streak` | Ends the run after 2 consecutive tool-less steps with no file changes (the "chào" failure mode) |
-| `continue-nudge` | Nudges plain-text answers back to work |
+| `continue-nudge` | Nudges plain-text answers back to work (no build demand when no files changed — suggests the `<<CHAT_DONE>>` exit instead) |
 | `doom-loop` | Intervenes after 3× identical consecutive tool-call batches |
 | `build-stop` | Ends the run when the exact build command succeeds; re-verifies on failure |
 | `history-trim` | Char-budget history trimming with whole-turn drops + old-tool-result pruning |
@@ -371,9 +371,10 @@ scoder tui "fix the bug"       # auto-submit an initial task
 Features:
 
 - **Alternate screen buffer** (vim/htop style) — the app owns the full screen: the header stays pinned at the top, input + status bar at the bottom, the chat pane scrolls internally, and your terminal content is restored on exit. Terminals without alt-screen support (`TERM=dumb`, non-TTY) automatically render inline instead.
-- **Build Rail** (pinned top bar) — brand, build lamp (`● build ok` / `● build fail` / `no build yet`), permission mode, and a step fuel gauge (`▰▰▰▱▱▱▱▱▱▱`) that fills as the agent burns through its step budget- **Phase breadcrumb + live metrics** (second header row, while running) — `plan › explore › edit › build › verify` (current phase highlighted) on the left; tokens in→out, **tok/s**, elapsed time, and guardrail count on the right
-- **Chat pane** — your messages, streaming assistant text with a live block cursor (`▋`), and every tool call with live stdout/stderr preview (`✔/✘` status, spinner while running, wall-clock duration per tool), prefixed with a flight-recorder step number (`03▸`)
-- **Header row 2** — phase breadcrumb (`plan › explore › edit › build › verify`), live metrics (tokens in→out, **tok/s**, elapsed, guardrail count), and a human-readable one-line notice of the latest guardrail intervention (e.g. `⚠ agent reminder: agent trả lời text mà không dùng tool — nhắc tiếp tục làm việc`)
+- **Build Rail** (pinned top bar, single row) — brand, build lamp (`✓ ok` / `✗ fail` / `no build yet`), permission mode, and a center region that shows either the phase breadcrumb (`plan › explore › edit › build › verify`) or the latest guardrail notice spanning the full available width (e.g. `⚠ trả lời text chưa dùng tool — nhắc tiếp tục`)
+- **Sidebar (≥92 cols)** — SESSION (model + workdir, no permission line), METRICS (step fuel gauge, tokens, tok/s, ctx input/limit, cost, elapsed, guardrail count), PHASES, BUILD, MCP, PLAN (live checklist at the bottom; resets on each new user message)
+- **Chat pane** — your messages with a full colored border + `you ▸` label, streaming assistant text with a live block cursor (`▋`), and every tool call with live stdout/stderr preview (`✔/✘` status, spinner while running, wall-clock duration per tool), prefixed with a flight-recorder step number (`03▸`)
+- **Header row 2** — (no longer present; phase breadcrumb + metrics moved into the sidebar / status bar to avoid duplication)
 - **Diff cards** (`✎`) — every file write/edit renders as a unified diff (`+` green / `-` red / `@@` cyan hunk headers); the same diff also flows out of the HTTP SSE stream
 - **Status bar** — model, token usage (input/output/cached), scroll position, keybind hints
 - **Slash commands** — `/help`, `/clear`, `/cost`, `/model`, `/skills`, `/skill`, `/mcp`, `/review`, `/permission`, `/exit` (identical to the REPL)

package/dist/agent.js

@@ -176,6 +176,10 @@ export class Agent {
                 catch (err) {
                     return { ok: false, output: `ERROR: subagent failed: ${err.message}` };
                 }
+                finally {
+                    // Child agents are ephemeral: close their MCP connections (if any).
+                    child.dispose();
+                }
             };
             this.ctx.taskRunner = runner;
         }
@@ -626,7 +630,12 @@ export class Agent {
             this.tools = [...allToolSpecs(), ...mcpSpecs];
         }
     }
-    /** Disconnect MCP servers at the end of a run. */
+    /**
+     * Disconnect MCP servers when the agent is disposed — connections are kept
+     * alive for the whole agent/session lifetime (one "connected" line, no
+     * re-spawn per message). Ephemeral owners (server runs, one-shot CLI) call
+     * dispose() when their agent is done.
+     */
     shutdownMcp() {
         this.mcp?.killAll();
         this.mcp = null;
@@ -689,13 +698,8 @@ export class Agent {
             buildCmd,
             permission: this.config.permissionMode,
         });
-        try {
-            await this.connectMcp();
-            return await this.runLoop(0, buildCmd, null);
-        }
-        finally {
-            this.shutdownMcp();
-        }
+        await this.connectMcp();
+        return this.runLoop(0, buildCmd, null);
     }
     /**
      * One conversational turn in an ongoing session (used by the REPL):
@@ -742,13 +746,8 @@ export class Agent {
             this.messages.push({ role: "user", content });
         this.pendingContext = [];
         this.messages.push({ role: "user", content: task });
-        try {
-            await this.connectMcp();
-            return await this.runLoop(0, buildCmd, null);
-        }
-        finally {
-            this.shutdownMcp();
-        }
+        await this.connectMcp();
+        return this.runLoop(0, buildCmd, null);
     }
     /**
      * Resume a previous session: restore messages/step count and continue the loop.
@@ -782,13 +781,8 @@ export class Agent {
             : null;
         // v2: opening a resumed run stamps a fresh run.started (event log continuity).
         this.events.append({ type: "run.started", task: this.currentTask });
-        try {
-            await this.connectMcp();
-            return await this.runLoop(session.steps ?? 0, buildCmd, lastBuild);
-        }
-        finally {
-            this.shutdownMcp();
-        }
+        await this.connectMcp();
+        return this.runLoop(session.steps ?? 0, buildCmd, lastBuild);
     }
     /**
      * The shared autonomous loop:

package/dist/guardrails.js

@@ -167,6 +167,20 @@ const continueNudgeRule = { name: "continue-nudge",
             return null;
         if (claimsCompletion(ctx.text))
             return null; // completion-claim rule phụ trách
+        // No file changed yet → never demand a build for what may be a pure Q&A or
+        // an explanation-only answer. Give the model the explicit conversational
+        // exit instead; text-only-streak still ends the run if it keeps stalling.
+        if (!ctx.filesModified) {
+            return {
+                action: {
+                    kind: "inject",
+                    reminder: `Continue working on the task. Use tools only if code changes are needed. ` +
+                        `If the user only asked a question and no code change is needed, do NOT run the build — ` +
+                        `end your reply with ${CHAT_DONE_MARKER} on its own last line instead.`,
+                },
+                reason: "plain text answer without tool calls (no files modified)",
+            };
+        }
         return {
             action: {
                 kind: "inject",

package/dist/index.js

@@ -220,6 +220,9 @@ async function runOnce(cfg, task) {
         ui.error(err.message);
         return 2;
     }
+    finally {
+        agent.dispose();
+    }
 }
 /**
  * `scoder init` — first-run wizard: asks for the essentials, writes
@@ -323,11 +326,13 @@ async function repl(cfg) {
             config: cfg,
             replaceAgent: (next) => {
                 agent.abort();
+                agent.dispose();
                 agent = next ?? new Agent(cfg);
                 first = true;
             },
             resumeAndRun: async (s) => {
                 agent.abort();
+                agent.dispose();
                 agent = new Agent(cfg);
                 first = false;
                 await agent.resume(s);
@@ -357,6 +362,7 @@ async function repl(cfg) {
             ui.error(err.message);
         }
     }
+    agent.dispose();
     ui.info("bye");
     process.exit(0);
 }
@@ -419,9 +425,11 @@ async function main() {
         });
         try {
             const result = await agent.resume(s, args.task ?? undefined);
+            agent.dispose();
             process.exit(result.success ? 0 : 1);
         }
         catch (err) {
+            agent.dispose();
             ui.error(err.message);
             process.exit(2);
         }

package/dist/prompt.js

@@ -3,18 +3,14 @@
  *  - versioning rõ ràng (PROMPT_VERSION được ghi vào session JSONL khi chạy),
  *  - test được độc lập (không cần instantiate Agent).
  *
- * V3 — stricter scope rules, clarification-before-action, safety boundaries,
- * MCP/skills as passive listings only:
- *  - ORIENT: read relevant files before ANY edit
- *  - PLAN: specific and actionable, not generic
- *  - EDIT: match existing style, minimal diffs
- *  - VERIFY: read the FULL error, fix the cause not the symptom
- *  - Safety & Boundaries: explicit prohibitions on git hooks, CI configs,
- *    dependencies, infrastructure configs, and destructive commands
- *  - Tool selection: concrete descriptions with explicit usage rules
- *  - Skills/MCP: passive listing, never auto-load
+ * V4 — conversational exits + no wasted builds:
+ *  - restore the <<CHAT_DONE>> marker rule (V3 dropped it, so the chat-done
+ *    guardrail could never fire and every Q&A turn was nudged into a build)
+ *  - build only after file edits; never run the build when nothing changed
+ *  - V3 baseline: stricter scope rules, clarification-before-action, safety
+ *    boundaries, MCP/skills as passive listings only
  */
-export const PROMPT_VERSION = 3;
+export const PROMPT_VERSION = 4;
 export const SYSTEM_PROMPT = `You are Scoder, an autonomous coding agent running in a terminal.
 You complete coding tasks end-to-end: exploring the code, editing files, and
 verifying your work with the build command. You act; you don't just advise.
@@ -30,9 +26,9 @@ verifying your work with the build command. You act; you don't just advise.
 3. EDIT: Make small, focused changes. Match the file's existing style
    (indentation, quote style, naming conventions) — do NOT reformat unrelated
    parts. Never refactor unrelated code, never add speculative features.
-4. VERIFY: Read the FULL error output before editing — fix the CAUSE, not the
-   symptom. Errors can be misleading (e.g. 'X not found' might be caused by a
-   missing import upstream). Then run the build command.
+4. VERIFY: If you changed files, run the build command. Read the FULL error
+   output before editing — fix the CAUSE, not the symptom. Errors can be
+   misleading (e.g. 'X not found' might be caused by a missing import upstream).
 5. REPORT: Your final message states what changed, why, and anything left
    unresolved. Be honest: if something didn't work, say exactly what failed.
 
@@ -46,7 +42,8 @@ verifying your work with the build command. You act; you don't just advise.
 - **write_file**: Create or overwrite. Parent dirs created automatically. Use for
   new files or when write is cleaner than targeted edits.
 - **run_command**: Run shell command. Captures stdout + stderr. THIS IS THE BUILD
-  COMMAND — every task ends with run_command(<buildCmd>).
+  COMMAND — run it only when you changed files; a coding task ends with
+  run_command(<buildCmd>).
 - **write_todos**: Plan + live progress for multi-step work.
 - **web_fetch**: Exact URLs only (it cannot search). Use for library/API docs,
   versions, changelogs when the answer isn't in the working directory.
@@ -74,11 +71,20 @@ verifying your work with the build command. You act; you don't just advise.
    - Improve code quality outside the task scope
 2. If the request is ambiguous or the scope is unclear: ask ONE clarifying question
    before taking action. Do NOT guess and proceed — clarify first.
-3. When you believe the task is done, call run_command with the build command.
-   The loop ends only when the build succeeds.
-4. If run_command is denied by permission policy, find a safer alternative or
+3. When you believe the task is done:
+   - If you changed files, call run_command with the build command. The coding
+     loop ends only when the build succeeds.
+   - If you changed nothing (you answered a question or explained code), do NOT
+     run the build command — a build with no changes proves nothing and wastes
+     time.
+4. CONVERSATIONAL INPUT: if the user's message is purely conversational — a
+   greeting, thanks, or a question needing no code changes and no shell
+   commands — do NOT use tools, do NOT edit files, and do NOT run a build.
+   Reply briefly in the user's language and end your reply with the exact marker
+   <<CHAT_DONE>> on its own last line.
+5. If run_command is denied by permission policy, find a safer alternative or
    report the limitation clearly. Do not retry the same denied command.
-5. If an environmental problem makes the task impossible, stop and report it.
+6. If an environmental problem makes the task impossible, stop and report it.
 
 ## Quality bar
 

package/dist/server.js

@@ -83,7 +83,10 @@ export function startServer(baseConfig, opts = {}) {
             record.finishedAt = new Date().toISOString();
             for (const sub of record.subscribers)
                 sub();
-        });
+        })
+            // MCP connections live for the agent lifetime — each server run owns an
+            // ephemeral agent, so dispose it here (never leak child processes).
+            .finally(() => agent.dispose());
     };
     const handleEventsSSE = (req, res, record) => {
         res.writeHead(200, {

package/dist/tui/app.js

@@ -61,6 +61,12 @@ export function TuiApp({ config, initialTask }) {
     const historyRef = useRef([]);
     const runningRef = useRef(false);
     const agentRef = useRef(null);
+    // MCP connections live for the session (one connect, no per-message re-spawn).
+    // Dispose on unmount so child server processes never outlive the TUI.
+    useEffect(() => () => {
+        agentRef.current?.dispose();
+        agentRef.current = null;
+    }, []);
     // ── Prompt-gate review flow (/review) ──────────────────────────────────────
     // State machine relaying the universal-prompt-gate-mcp text-fallback flow:
     // call → pending_level_selection → user picks level → pending_user_selection
@@ -214,6 +220,7 @@ export function TuiApp({ config, initialTask }) {
         isRunning: () => runningRef.current,
         replaceAgent: (next) => {
             agentRef.current?.abort();
+            agentRef.current?.dispose();
             agentRef.current = next;
             if (!next) {
                 runningRef.current = false;

package/dist/tui/chat-views.js

@@ -114,8 +114,9 @@ function TodoItemView() {
 export function ChatItemView({ item }) {
     switch (item.kind) {
         case "user":
-            // Clean user bubble: accent arrow + bold text
-            return (_jsxs(Box, { marginTop: 1, paddingLeft: 1, children: [_jsx(Text, { color: theme.user, bold: true, children: "\u276F " }), _jsx(Text, { bold: true, children: item.text })] }));
+            // User messages stand out with a full colored border + bold label so they're
+            // immediately visible at a glance, distinct from the agent's ◆ markdown.
+            return (_jsx(Box, { marginTop: 1, marginLeft: 1, children: _jsxs(Box, { borderStyle: "single", borderColor: theme.user, paddingX: 1, paddingTop: 0, flexDirection: "column", children: [_jsx(Text, { color: theme.user, bold: true, children: "you \u25B8" }), _jsx(Text, { bold: true, children: item.text })] }) }));
         case "assistant": {
             // Streaming: dim with pulsing cursor; finalized: markdown with agent mark.
             if (item.streaming || !item.text) {

package/dist/tui/chrome.js

@@ -23,7 +23,7 @@ export function gaugeColor(step, maxSteps) {
 /**
  * BuildRail — polished top bar (SINGLE row).
  *
- * Layout:  [⚑ scoder] ── [phase breadcrumb / guardrail warning] ── [perm] [build] [steps]
+ * Layout:  [⚑ scoder] ── [phase breadcrumb / guardrail warning] ── [perm] [build]
  *
  * Terminal ≥ SIDEBAR_MIN_COLS: sidebar handles METRICS + PHASES.
  * Terminal < SIDEBAR_MIN_COLS: phase breadcrumb + metrics move into this row.
@@ -37,15 +37,15 @@ export function BuildRail({ state, config }) {
     const lampColor = state.buildLamp === "ok" ? theme.success : state.buildLamp === "fail" ? theme.error : theme.dim;
     const lampLabel = state.buildLamp === "ok" ? "✓" : state.buildLamp === "fail" ? "✗" : "○";
     const lampText = state.buildLamp === "ok" ? "ok" : state.buildLamp === "fail" ? "fail" : "—";
-    // Step badge
-    const stepLabel = state.step > 0
-        ? `${state.step}${state.maxSteps > 0 ? `/${state.maxSteps}` : ""}`
-        : "idle";
-    // Guardrail: shown inline if it fired (small warning before the phase breadcrumb)
+    // Guardrail: shown inline if it fired (small warning before the phase breadcrumb).
     const hasGuardrail = Boolean(state.lastGuardrail);
+    // Guardrail text gets the whole center region — no fixed 40-char cap that
+    // truncates while the header still has empty space. Reserve room for the
+    // brand on the left and permission/build lamps on the right.
+    const centerBudget = Math.max(24, (columns ?? 100) - 34);
     // Phase breadcrumb or narrow-terminal metrics (right of the center)
     const showNarrowMetrics = !wide && (state.status === "running" || state.phases.length > 0);
-    return (_jsxs(Box, { borderStyle: "bold", borderColor: theme.border, paddingX: 1, justifyContent: "space-between", alignItems: "center", flexShrink: 0, children: [_jsxs(Text, { bold: true, color: theme.brand, children: ["\u2691 ", _jsx(Text, { bold: true, children: "scoder" })] }), _jsx(Box, { flexGrow: 1, justifyContent: "center", children: hasGuardrail ? (_jsxs(Text, { color: theme.warn, bold: true, children: ["\u26A0 ", fit(state.lastGuardrail, 40)] })) : state.status === "running" ? (showNarrowMetrics ? (_jsx(PhaseBreadcrumb, { phases: state.phases })) : (_jsx(Text, { dimColor: true, children: "plan \u203A explore \u203A edit \u203A build \u203A verify" }))) : (_jsx(Text, { dimColor: true, children: "ready" })) }), _jsxs(Text, { children: [_jsxs(Text, { color: permColor, children: ["[", perm, "]"] }), _jsx(Text, { children: " " }), _jsx(Text, { color: lampColor, children: lampLabel }), _jsxs(Text, { dimColor: true, children: [" ", lampText] }), _jsx(Text, { children: " " }), _jsx(Text, { dimColor: true, children: stepLabel })] })] }));
+    return (_jsxs(Box, { borderStyle: "bold", borderColor: theme.border, paddingX: 1, justifyContent: "space-between", alignItems: "center", flexShrink: 0, children: [_jsxs(Text, { bold: true, color: theme.brand, children: ["\u2691 ", _jsx(Text, { bold: true, children: "scoder" })] }), _jsx(Box, { flexGrow: 1, justifyContent: "center", children: hasGuardrail ? (_jsxs(Text, { color: theme.warn, bold: true, children: ["\u26A0 ", fit(state.lastGuardrail, centerBudget)] })) : state.status === "running" ? (showNarrowMetrics ? (_jsx(PhaseBreadcrumb, { phases: state.phases })) : (_jsx(Text, { dimColor: true, children: "plan \u203A explore \u203A edit \u203A build \u203A verify" }))) : (_jsx(Text, { dimColor: true, children: "ready" })) }), _jsxs(Text, { children: [_jsxs(Text, { color: permColor, children: ["[", perm, "]"] }), _jsx(Text, { children: " " }), _jsx(Text, { color: lampColor, children: lampLabel }), _jsxs(Text, { dimColor: true, children: [" ", lampText] })] })] }));
 }
 /** Breadcrumb tiến trình: plan › explore › edit › build › verify. */
 const PHASE_SEQUENCE = ["plan", "explore", "edit", "build", "verify"];

package/dist/tui/sidebar.js

@@ -15,11 +15,9 @@ function SectionLine({ label, value, color }) {
 }
 // ── Session ──────────────────────────────────────────────────────────────────
 function SessionSection({ state, config }) {
-    const perm = state.permission || config.permissionMode;
-    const permColor = perm === "readonly" ? theme.warn : perm === "ask" ? theme.accent : theme.dim;
     const model = state.model || config.model || "—";
     const dir = config.workdir.split("/").filter(Boolean).pop() ?? config.workdir;
-    return (_jsxs(Section, { title: "SESSION", children: [_jsx(SectionLine, { label: "model", value: fit(model, INNER - 9) }), _jsx(SectionLine, { label: "perm", value: perm, color: permColor }), _jsx(SectionLine, { label: "dir", value: fit(dir, INNER - 9) })] }));
+    return (_jsxs(Section, { title: "SESSION", children: [_jsx(SectionLine, { label: "model", value: fit(model, INNER - 9) }), _jsx(SectionLine, { label: "dir", value: fit(dir, INNER - 9) })] }));
 }
 function PlanRow({ t, width }) {
     if (t.status === "completed") {
@@ -51,7 +49,7 @@ function MetricsSection({ state }) {
         ? Math.round(state.usage.outputTokens / (elapsedMs / 1000))
         : null;
     // Live estimate based on accumulated streaming chars (chars / ~4 ≈ tokens).
-    const liveTokPerSec = state.streamingOutputChars > 0 && elapsedMs != null && elapsedMs > 500
+    const liveTokPerSec = state.streamingOutputChars > 0 && elapsedMs != null && elapsedMs > 250
         ? Math.round((state.streamingOutputChars / 4) / (elapsedMs / 1000))
         : null;
     const win = contextWindowFor(state.provider, state.model);
@@ -61,7 +59,7 @@ function MetricsSection({ state }) {
     const ctxDisplay = ctxPct != null
         ? `${ctxFmt(state.usage.inputTokens)}/${ctxFmt(win)} (${ctxPct}%)`
         : null;
-    return (_jsx(Section, { title: "METRICS", children: !hasUsage && elapsedSec == null ? (_jsx(Text, { dimColor: true, children: "  starting\u2026" })) : (_jsxs(Box, { flexDirection: "column", flexShrink: 0, children: [gauge && (_jsxs(Text, { wrap: "truncate", children: [_jsx(Text, { dimColor: true, children: "  steps " }), _jsx(Text, { color: gaugeColor(state.step, state.maxSteps), children: gauge }), _jsxs(Text, { dimColor: true, children: [" ", state.step, "/", state.maxSteps] })] })), hasUsage && (_jsx(SectionLine, { label: "toks", value: `${fmt(state.usage.inputTokens)} → ${fmt(state.usage.outputTokens)}` })), tokPerSec != null && (_jsx(SectionLine, { label: "speed", value: `${fmt(tokPerSec)} tok/s`, color: theme.accent })), liveTokPerSec != null && tokPerSec == null && (_jsx(SectionLine, { label: "speed", value: `~${fmt(liveTokPerSec)} tok/s`, color: theme.accent })), ctxDisplay != null && _jsx(SectionLine, { label: "ctx", value: ctxDisplay }), cost > 0 && _jsx(SectionLine, { label: "cost", value: `$${cost < 0.01 ? cost.toFixed(4) : cost.toFixed(2)}` }), elapsedSec != null && (_jsx(SectionLine, { label: "elapsed", value: `${Math.floor(elapsedSec / 60)}:${String(elapsedSec % 60).padStart(2, "0")}` })), state.usage.cachedTokens > 0 && (_jsx(SectionLine, { label: "cached", value: fmt(state.usage.cachedTokens) })), state.guardrailCount > 0 && (_jsx(SectionLine, { label: "guard", value: String(state.guardrailCount), color: theme.warn }))] })) }));
+    return (_jsx(Section, { title: "METRICS", children: !hasUsage && elapsedSec == null ? (_jsx(Text, { dimColor: true, children: "  starting\u2026" })) : (_jsxs(Box, { flexDirection: "column", flexShrink: 0, children: [gauge && (_jsxs(Text, { wrap: "truncate", children: [_jsx(Text, { dimColor: true, children: "  steps " }), _jsx(Text, { color: gaugeColor(state.step, state.maxSteps), children: gauge }), _jsxs(Text, { dimColor: true, children: [" ", state.step, "/", state.maxSteps] })] })), hasUsage && (_jsx(SectionLine, { label: "toks", value: `${fmt(state.usage.inputTokens)} → ${fmt(state.usage.outputTokens)}` })), !hasUsage && state.streamingOutputChars > 0 && (_jsx(SectionLine, { label: "toks", value: `~${fmt(Math.round(state.streamingOutputChars / 4))} (streaming)`, color: theme.accent })), tokPerSec != null && (_jsx(SectionLine, { label: "speed", value: `${fmt(tokPerSec)} tok/s`, color: theme.accent })), liveTokPerSec != null && tokPerSec == null && (_jsx(SectionLine, { label: "speed", value: `~${fmt(liveTokPerSec)} tok/s`, color: theme.accent })), ctxDisplay != null && _jsx(SectionLine, { label: "ctx", value: ctxDisplay }), cost > 0 && _jsx(SectionLine, { label: "cost", value: `$${cost < 0.01 ? cost.toFixed(4) : cost.toFixed(2)}` }), elapsedSec != null && (_jsx(SectionLine, { label: "elapsed", value: `${Math.floor(elapsedSec / 60)}:${String(elapsedSec % 60).padStart(2, "0")}` })), state.usage.cachedTokens > 0 && (_jsx(SectionLine, { label: "cached", value: fmt(state.usage.cachedTokens) })), state.guardrailCount > 0 && (_jsx(SectionLine, { label: "guard", value: String(state.guardrailCount), color: theme.warn }))] })) }));
 }
 // ── Phases (plan › explore › edit › build › verify — vertical) ───────────────
 const PHASES = ["plan", "explore", "edit", "build", "verify"];
@@ -97,5 +95,5 @@ function McpSection({ state, config }) {
 // ── Root ─────────────────────────────────────────────────────────────────────
 /** Sidebar phải — mọi trạng thái sống của run, cập nhật theo AgentEvent. */
 export function Sidebar({ state, config }) {
-    return (_jsxs(Box, { flexDirection: "column", borderStyle: "single", borderColor: theme.border, width: SIDEBAR_WIDTH, paddingX: 1, flexShrink: 0, overflow: "hidden", children: [_jsx(SessionSection, { state: state, config: config }), _jsx(PlanSection, { state: state }), _jsx(MetricsSection, { state: state }), _jsx(PhaseSection, { state: state }), _jsx(BuildSection, { state: state, config: config }), _jsx(McpSection, { state: state, config: config })] }));
+    return (_jsxs(Box, { flexDirection: "column", borderStyle: "single", borderColor: theme.border, width: SIDEBAR_WIDTH, paddingX: 1, flexShrink: 0, overflow: "hidden", children: [_jsx(SessionSection, { state: state, config: config }), _jsx(McpSection, { state: state, config: config }), _jsx(MetricsSection, { state: state }), _jsx(PlanSection, { state: state }), _jsx(PhaseSection, { state: state }), _jsx(BuildSection, { state: state, config: config })] }));
 }

package/dist/tui/state.js

@@ -69,12 +69,24 @@ export function visibleWindow(items, viewportRows, margin = 50) {
 /**
  * Plan mới nhất cho sidebar — todo items được append mỗi lần update nhưng
  * sidebar chỉ cần snapshot hiện tại (last-wins), null khi chưa có plan.
+ *
+ * Reset realtime: mỗi lần user gửi message mới, plan reset — chỉ hiện plan
+ * được viết SAU user message cuối cùng. Tránh stale plan từ task trước hiện
+ * trong task mới.
  */
 export function latestPlan(items) {
+    // Tìm user message cuối cùng — plan chỉ reset tại đây.
+    let lastUserIdx = -1;
     for (let i = items.length - 1; i >= 0; i--) {
-        const it = items[i];
-        if (it.kind === "todo")
-            return it;
+        if (items[i].kind === "user") {
+            lastUserIdx = i;
+            break;
+        }
+    }
+    // Chỉ tìm todo SAU user cuối.
+    for (let i = items.length - 1; i > lastUserIdx; i--) {
+        if (items[i].kind === "todo")
+            return items[i];
     }
     return null;
 }
@@ -120,13 +132,11 @@ function applyGuardrail(state, ev) {
         "chat-done": "lượt chat thuần — kết thúc không cần build",
         "completion-claim": "agent báo hoàn thành — buộc chạy build xác minh",
         "text-only-streak": "agent chỉ nói không làm — dừng run",
-        "continue-nudge": "agent trả lời text mà không dùng tool — nhắc tiếp tục làm việc",
+        "continue-nudge": "trả lời text chưa dùng tool — nhắc tiếp tục",
         "doom-loop": "lặp lẫn lộn cùng tool call — yêu cầu đổi hướng",
         "build-stop": "build pass — kết thúc run",
     };
     state.lastGuardrail = HUMAN[ev.rule] ?? `${ev.rule}: ${ev.reason}`;
-    if (state.lastGuardrail.length > 72)
-        state.lastGuardrail = state.lastGuardrail.slice(0, 69) + "…";
 }
 /**
  * Fold one AgentEvent into the state (mutates `state`, returns it for

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "s-app-scoder-agent",
-  "version": "0.7.0",
+  "version": "0.8.1",
   "description": "Autonomous coding agent CLI — writes, edits, and builds code autonomously until the project compiles without errors.",
   "type": "module",
   "main": "dist/index.js",