s-app-scoder-agent 0.6.0 → 0.8.0

package/dist/agent.js

@@ -1,48 +1,43 @@
-import { loadPermissionLists } from "./config.js";
-import { LlmClient, LlmError } from "./llm.js";
+import { loadPermissionLists, loadPolicyRules, loadSavedApprovals, saveApproval } from "./config.js";
+import { LlmClient, LlmError, redactSecrets } from "./llm.js";
 import { AnthropicClient } from "./provider-anthropic.js";
-import { AGENT_TOOLS, getTool } from "./tools.js";
+import { getTool, allToolSpecs, setMaxOutputChars } from "./tools.js";
+import { invalidateFileIndex } from "./tui/at-complete.js";
 import { detectBuildCommand, runBuild } from "./build.js";
-import { trimHistory, summariseOutput, claimsCompletion } from "./context.js";
+import { trimHistory, summariseOutput, expandMentions } from "./context.js";
+import { createGuardrailEngine, CHAT_DONE_MARKER } from "./guardrails.js";
+import { PROMPT_VERSION } from "./prompt.js";
+import { buildSystemPrompt, buildEnvInfo, buildSkillsBlock, buildAgentsBlock, buildMcpBlock } from "./prompt-builder.js";
 import { PermissionChecker } from "./permissions.js";
-import { SessionStore } from "./session.js";
+import { SessionStore, EventSessionStore } from "./session.js";
+import { DEFAULT_COMPACTION, shouldCompact, extractSummaryInput, COMPACTION_INSTRUCTION, applyCompaction } from "./compaction.js";
+import { SnapshotManager } from "./snapshot.js";
+import { contextWindowFor } from "./model-catalog.js";
 import { JsonlLogger } from "./logger.js";
+import { createRuntime, parseCustomProviders, resolveProvider, splitProviderModel } from "./provider-registry.js";
+import { loadRcFile } from "./config.js";
 import { McpManager } from "./mcp.js";
+import { validateSchema } from "./json-schema.js";
+import { watchWorkdir } from "./watch.js";
+import { discoverSkills, findSkill } from "./skills.js";
+import { discoverAgents, findAgent } from "./agents.js";
+import { SteeringInbox } from "./steering.js";
 import * as ui from "./ui.js";
-const SYSTEM_PROMPT = `You are Scoder, an autonomous coding agent running in a terminal.
-You complete coding tasks by using tools: reading/writing/editing files, listing
-files, and running shell commands.
-
-RULES:
-1. Work autonomously. Do NOT ask the user questions mid-task; make reasonable
-   decisions and proceed.
-2. When you believe the task is done, call run_command with the project's build
-   command (or the one the user specified). The loop ends only when the build
-   succeeds.
-3. If the build fails, read the error output carefully, fix the code, and build
-   again. Iterate until it passes.
-4. Prefer small, focused edits. Read a file before editing it.
-5. If an external/environmental problem (missing dependency, wrong environment)
-   makes the task impossible, stop editing and report it clearly in your final
-   message.
-6. Never invent requirements beyond the given task.
-7. IMPORTANT: if a run_command is denied by permission policy, you will receive
-   an error message. Use this as a signal to find a safer alternative or to
-   communicate the limitation to the user.
-8. You have web access ONLY through web_fetch (it fetches exact URLs, it cannot
-   search). Use it when the task depends on external or fast-changing knowledge:
-   library/API documentation, a URL found in an error message or code comment,
-   current package versions, changelogs, release notes, standards. Skip it when
-   the answer is in the working directory or is stable general knowledge. After
-   fetching, ground your edits in what the page actually says.`;
 export class Agent {
     config;
     client;
     messages = [];
-    tools = AGENT_TOOLS.map((t) => t.spec);
+    tools = allToolSpecs();
     ctx;
     permissions;
+    /** T-03: pending permission prompts — id → resolve (TUI/serve answer them). */
+    pendingPermissions = new Map();
+    permSeq = 0;
     session;
+    /** v2 durable event log (session/ + events.jsonl) — crash recovery + steering. */
+    events;
+    /** Git shadow-repo checkpoints before write batches (H-11). */
+    snapshots;
     logger;
     aborted = false;
     streamFallbackWarned = false;
@@ -53,39 +48,161 @@ export class Agent {
     usageTotals = { calls: 0, inputTokens: 0, outputTokens: 0, cachedTokens: 0 };
     mcp = null;
     /** Aborted when abort() is called — cancels in-flight LLM requests (SIGINT). */
+    /** Steering inbox — user input delivered at safe step boundaries (H-08, I5 module). */
+    steerInbox = new SteeringInbox();
     runSignal = new AbortController();
     /** Optional structured-event sink (used by the HTTP/SSE server frontend). */
     onEvent = null;
     /** Set while a tool call is executing, for tool_output event correlation. */
     currentToolCall = null;
+    /** True once any file-writing tool succeeded during the current run. */
+    filesModified = false;
+    /** Discovered skills (project + global). */
+    skills = [];
+    /** M4: hot-reload watcher (rc + skills dir). */
+    watcher = null;
+    /** Content staged via /skill — injected before the next run/turn. */
+    pendingContext = [];
+    /** Session plan (write_todos tool). */
+    plan = [];
+    /** Guardrail Engine — khởi tạo trong constructor (cần this.config). */
+    guardrails;
+    /** Log một quyết định guardrail ra JSONL + event (quan sát được, không ồn ào UI). */
+    logGuardrailDecision(d) {
+        const kind = d.action.kind;
+        this.logger.log({ type: "guardrail", rule: d.rule, action: kind, reason: d.reason });
+        this.emit({ type: "guardrail", rule: d.rule, action: kind, reason: d.reason });
+    }
     constructor(config) {
         this.config = config;
+        this.guardrails = createGuardrailEngine(config.guardrails ?? {}, (d) => this.logGuardrailDecision(d));
         const commonOpts = {
             baseUrl: config.baseUrl,
             apiKey: config.apiKey,
             model: config.model,
             maxTokens: config.maxTokens,
             temperature: config.temperature,
+            requestTimeoutMs: config.limits?.requestTimeoutMs, // G7
         };
-        this.client =
-            config.provider === "anthropic"
-                ? new AnthropicClient(commonOpts)
-                : new LlmClient({ ...commonOpts, stream: config.stream });
+        // I-01: registry-driven runtime — providerDef (custom/builtin) wins; else legacy ternary.
+        // I-05: thinking budget = 10% maxTokens khi reasoning on (Anthropic only).
+        // I-12: eval harness — scripted client override thay vì construct.
+        const thinkingBudget = config.reasoning === "on" ? Math.max(1024, Math.floor(config.maxTokens * 0.25)) : 0;
+        this.client = config.clientFactory?.() ??
+            (config.providerDef?.kind === "anthropic" || (!config.providerDef && config.provider === "anthropic")
+                ? new AnthropicClient({ ...commonOpts, enableCache: config.providerDef?.supportCache ?? true, thinkingBudgetTokens: thinkingBudget })
+                : new LlmClient({ ...commonOpts, stream: config.stream }));
         const permLists = loadPermissionLists(config.workdir);
         this.permissions = new PermissionChecker({
             mode: config.permissionMode,
             allowCommands: permLists.allowCommands,
             denyCommands: permLists.denyCommands,
             nonInteractive: config.yes,
+            rules: config.permissionRules,
+            savedRules: loadSavedApprovals(config.workdir),
+            policyRules: loadPolicyRules(config.workdir),
         });
+        // T-03: route permission prompts through the event stream when an event
+        // sink is attached (TUI/serve) — readline on stdin corrupts alt-screen.
+        this.permissions.promptHandler = (req) => this.programmaticPermissionPrompt(req);
         this.ctx = {
             workdir: config.workdir,
             permissions: this.permissions,
             permissionMode: config.permissionMode,
             onToolOutput: this.handleToolOutput,
+            plan: {
+                get: () => this.plan,
+                set: (items) => {
+                    this.plan = items;
+                    const done = items.filter((t) => t.status === "completed").length;
+                    ui.info(`plan: ${done}/${items.length} completed`);
+                    this.emit({ type: "todo", items: [...items] });
+                    this.logger.log({ type: "todo", items });
+                },
+            },
+            formatter: config.formatter,
+            sandbox: config.sandbox,
+            allowSubagents: config.allowSubagents,
         };
+        // G7: thresholds từ rc limits.
+        if (config.limits?.maxOutputChars)
+            setMaxOutputChars(config.limits.maxOutputChars);
         this.session = new SessionStore(config.workdir, config.noSave !== true);
+        this.events = new EventSessionStore(config.workdir, config.noSave !== true, this.session.id);
+        this.snapshots = new SnapshotManager(config.workdir, config.snapshots !== false);
+        // I-06: task tool → child Agent run (same process, own session, event
+        // relay, abort cascade via parent's runSignal). Depth-1 only: the runner
+        // refuses when the calling agent was constructed with allowSubagents:false
+        // (children set this; the runner checks ctx, not a global flag — global
+        // registration alone allowed grandchildren through). Runner lives on the
+        // per-instance ctx — concurrent Agents (server multi-run) stay isolated.
+        if (config.allowSubagents !== false) {
+            const runner = async (ctx, args) => {
+                if (ctx.allowSubagents === false) {
+                    return { ok: false, output: "ERROR: task tool not available — subagents cannot spawn further subagents (depth limit)." };
+                }
+                const def = args.agent ? findAgent(config.workdir, args.agent) : findAgent(config.workdir, "explore");
+                if (args.agent && !def) {
+                    return { ok: false, output: `ERROR: agent "${args.agent}" not found — available: ${discoverAgents(config.workdir).filter((a) => a.mode !== "subagent" || a.name === "explore").map((a) => a.name).join(", ")}` };
+                }
+                const child = new Agent({
+                    ...config,
+                    allowSubagents: false,
+                    model: def?.model ?? config.model,
+                    maxSteps: def?.steps ?? Math.min(config.maxSteps, 20),
+                    noSave: true, // subagent runs are ephemeral — no session pollution
+                    // plan/explore agents stay read-only regardless of parent mode.
+                    permissionMode: def?.name === "plan" || def?.name === "explore" ? "readonly" : config.permissionMode,
+                });
+                // Abort cascade: parent abort kills child's in-flight LLM calls.
+                this.runSignal.signal.addEventListener("abort", () => child.abort(), { once: true });
+                // Event relay: child activity appears as subagent-tagged tool output.
+                const tag = `subagent:${def?.name ?? "task"}`;
+                child.onEvent = (ev) => {
+                    if (ev.type === "tool_call")
+                        this.emit({ type: "tool_output", name: tag, stream: "stdout", delta: `→ ${ev.name}\n` });
+                    else if (ev.type === "llm_delta")
+                        this.emit({ type: "tool_output", name: tag, stream: "stdout", delta: ev.delta });
+                };
+                try {
+                    const res = await child.run(args.prompt);
+                    return {
+                        ok: res.success,
+                        output: res.success
+                            ? `[subagent ${def?.name ?? "task"} done — ${res.steps} steps]\n${res.summary}`
+                            : `[subagent ${def?.name ?? "task"} FAILED — ${res.steps} steps]\n${res.summary}`,
+                    };
+                }
+                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;
+        }
         this.logger = JsonlLogger.forSession(config.workdir, this.session.id, config.logFile);
+        this.skills = discoverSkills(config.workdir);
+        // M4: hot-reload — .scoderrc (MCP servers) + .scoder/skills/ đổi giữa
+        // session → reload ở turn sau (debounced, không watcher trong test).
+        if (config.noSave !== true && process.env.SCODER_NO_WATCH !== "1") {
+            this.watcher = watchWorkdir(config.workdir, (kind) => {
+                if (kind === "skills") {
+                    this.skills = discoverSkills(config.workdir);
+                    this.logger.log({ type: "hot_reload", kind, skills: this.skills.length });
+                }
+                else if (kind === "config") {
+                    // MCP servers mới: merge vào manager hiện tại (turn sau advertise).
+                    const rc = loadRcFile(config.workdir);
+                    const servers = rc.mcpServers ?? {};
+                    if (this.mcp && Object.keys(servers).length > 0) {
+                        void this.mcp.applyConfig(servers);
+                    }
+                }
+            });
+        }
     }
     /**
      * Handle SIGINT: abort gracefully after current tool finishes.
@@ -95,6 +212,32 @@ export class Agent {
         this.aborted = true;
         this.runSignal.abort(new DOMException("interrupted by user", "AbortError"));
     }
+    /**
+     * Steer: user input delivered at the NEXT safe step boundary (pattern from
+     * opencode "steer by default"). Injected as a user message before the next
+     * LLM call; resets the doom-loop text-allowance once for the whole batch.
+     */
+    steer(text) {
+        this.steerInbox.steer(text);
+        const t = text.trim();
+        if (!t)
+            return;
+        this.events.append({ type: "steer.enqueued", mode: "steer", text: t });
+        this.emit({ type: "steer_queued", mode: "steer", text: t });
+    }
+    /** Queue: same inbox, but only relevant when the agent is idle (queued run). */
+    queue(text) {
+        this.steerInbox.queue(text);
+        const t = text.trim();
+        if (!t)
+            return;
+        this.events.append({ type: "steer.enqueued", mode: "queue", text: t });
+        this.emit({ type: "steer_queued", mode: "queue", text: t });
+    }
+    /** Drain the steering inbox; delivered at the top of the next step. */
+    drainSteerInbox() {
+        return this.steerInbox.drain().map((x) => x.text);
+    }
     /** Reset the run-signal so a new run() can use a fresh controller. */
     resetRunSignal() {
         this.runSignal = new AbortController();
@@ -107,6 +250,49 @@ export class Agent {
     emit(ev) {
         this.onEvent?.(ev);
     }
+    /**
+     * T-03: permission ask via the event stream. Emits permission_request and
+     * waits for resolvePermission(); when no sink is attached, falls back to
+     * deny (the TUI always attaches one; programmatic callers answer directly).
+     * Timeout ⇒ deny + auto-resolve (no leaked promise).
+     */
+    programmaticPermissionPrompt(req) {
+        if (!this.onEvent)
+            return Promise.resolve("no");
+        const id = `perm-${++this.permSeq}`;
+        return new Promise((resolve) => {
+            const timer = setTimeout(() => {
+                if (this.pendingPermissions.delete(id))
+                    resolve("no");
+            }, req.timeoutSecs * 1000);
+            this.pendingPermissions.set(id, (a) => {
+                clearTimeout(timer);
+                resolve(a);
+            });
+            this.emit({
+                type: "permission_request",
+                id,
+                kind: req.kind,
+                label: req.label,
+                ...(req.command !== undefined ? { command: req.command } : {}),
+                ...(req.diff !== undefined ? { diff: req.diff } : {}),
+                ...(req.alwaysAllowPrefix !== undefined ? { alwaysAllowPrefix: req.alwaysAllowPrefix } : {}),
+                timeoutSecs: req.timeoutSecs,
+            });
+        });
+    }
+    /**
+     * T-03: answer a pending permission_request (called from the TUI modal /
+     * HTTP API). Unknown id → silently ignored (already timed out).
+     */
+    resolvePermission(id, answer) {
+        const res = this.pendingPermissions.get(id);
+        if (res) {
+            this.pendingPermissions.delete(id);
+            res(answer);
+        }
+    }
+    /** Aborted when abort() is called — cancels in-flight LLM requests (SIGINT). */
     /**
      * Live tool-output handler: renders chunks in the terminal (dimmed block) and
      * forwards them as `tool_output` events. Opens the stream block lazily on the
@@ -126,12 +312,14 @@ export class Agent {
     }
     /** Push one tool-result message back into the conversation. */
     pushToolResult(tc, content) {
-        this.messages.push({
+        const msg = {
             role: "tool",
             tool_call_id: tc.id,
             name: tc.name,
             content,
-        });
+        };
+        this.messages.push(msg);
+        this.events.append({ type: "message.added", message: msg });
     }
     /**
      * Handle a rejected tool call (unknown tool, policy denial, user denial):
@@ -140,10 +328,151 @@ export class Agent {
      */
     rejectToolCall(step, tc, output, logExtra) {
         ui.toolResult(tc.name, false, output);
-        this.emit({ type: "tool_result", step, name: tc.name, ok: false, output });
+        this.emit({ type: "tool_result", step, id: tc.id, name: tc.name, ok: false, output });
         this.logger.log({ type: "tool", step, tool: tc.name, ok: false, ...logExtra });
         this.pushToolResult(tc, output);
     }
+    /** I-09: tools safe to run concurrently (pure reads, no permission ask). */
+    static READONLY_TOOLS = new Set(["read_file", "list_files", "grep", "glob", "load_skill"]);
+    /**
+     * I-09: bounded-concurrency map preserving input order.
+     */
+    async parallelMap(items, fn, limit) {
+        const results = new Array(items.length);
+        let next = 0;
+        const workers = Array.from({ length: Math.min(limit, items.length) }, async () => {
+            while (next < items.length) {
+                const i = next++;
+                results[i] = await fn(items[i], i);
+            }
+        });
+        await Promise.all(workers);
+        return results;
+    }
+    /**
+     * Execute one tool call end-to-end: validate args, render, permission gate,
+     * run (builtin or MCP), log, and emit result. Returns `post` side-effects to
+     * run by the caller AFTER the whole batch completes — in parallel mode these
+     * must not interleave with other calls' work (messages order, abort checks).
+     */
+    async executeToolCall(tc, steps, stepToolResults) {
+        const tool = getTool(tc.name);
+        let args;
+        try {
+            args = tc.arguments ? JSON.parse(tc.arguments) : {};
+        }
+        catch {
+            this.pushToolResult(tc, `ERROR: invalid JSON arguments: ${tc.arguments.slice(0, 200)}`);
+            return {};
+        }
+        ui.toolCall(tc.name, args);
+        this.currentToolCall = { step: steps, id: tc.id, name: tc.name };
+        this.emit({ type: "tool_call", step: steps, id: tc.id, name: tc.name, args });
+        if (!tool && !this.mcp?.hasTool(tc.name)) {
+            this.rejectToolCall(steps, tc, `ERROR: unknown tool "${tc.name}". Available: ${allToolSpecs().map((t) => t.function.name).join(", ")}${this.mcp ? " + MCP tools (mcp__<server>__<tool>)" : ""}`, { error: "unknown tool" });
+            return {};
+        }
+        let result;
+        try {
+            // Permission gate for shell commands
+            if (tc.name === "run_command" && typeof args.command === "string") {
+                const decision = this.permissions.classifyShell(args.command);
+                if (decision.decision === "deny") {
+                    const output = `ERROR: command denied by permission policy (${decision.reason}): ${String(args.command).slice(0, 200)}`;
+                    this.rejectToolCall(steps, tc, output, { args, denied: decision.reason });
+                    return {};
+                }
+                if (decision.decision === "ask") {
+                    const allowed = await this.permissions.confirm(args.command);
+                    if (!allowed) {
+                        const output = `ERROR: user denied permission to run: ${String(args.command).slice(0, 200)}`;
+                        this.rejectToolCall(steps, tc, output, { args, denied: "user" });
+                        return {};
+                    }
+                }
+            }
+            const t0 = Date.now();
+            // Route MCP tools first (they carry the mcp__ prefix, not in AGENT_TOOLS)
+            if (tool) {
+                result = await tool.run(this.ctx, args);
+            }
+            else {
+                // MCP tools — G5: đi qua permission engine như tool thường
+                // (policy floor → rules → mode fallback). Không có nhánh bypass.
+                const mcpDecision = this.permissions.classifyMcpTool(tc.name);
+                if (mcpDecision?.decision === "deny") {
+                    const output = `ERROR: MCP tool "${tc.name}" denied by permission policy (${mcpDecision.reason}).`;
+                    this.rejectToolCall(steps, tc, output, { args, denied: mcpDecision.reason });
+                    return {};
+                }
+                if (this.config.permissionMode === "readonly" && mcpDecision?.decision !== "allow") {
+                    result = {
+                        ok: false,
+                        output: `ERROR: readonly mode — MCP tool "${tc.name}" is disabled. Allow it explicitly with a permission rule (mcp__${tc.name.replace(/^mcp__/, "").replace(/__/, "__")} or a wildcard).`,
+                    };
+                }
+                else if (mcpDecision?.decision === "ask" && !this.config.yes) {
+                    const allowed = await this.confirmMcpTool(tc.name);
+                    if (!allowed) {
+                        const output = `ERROR: user denied permission to call MCP tool: ${tc.name}`;
+                        this.rejectToolCall(steps, tc, output, { args, denied: "user" });
+                        return {};
+                    }
+                }
+                result = await this.invokeMcpTool(tc.name, args);
+            }
+            this.logger.log({ type: "tool", step: steps, tool: tc.name, args, ok: result.ok, elapsedMs: Date.now() - t0 });
+        }
+        catch (err) {
+            result = { ok: false, output: `ERROR: ${err.message}` };
+            this.logger.log({ type: "tool", step: steps, tool: tc.name, args, ok: false, error: err.message });
+        }
+        // User pressed [q]uit in an edit-confirmation prompt — stop the whole run.
+        if (this.permissions.abortRequested) {
+            this.aborted = true;
+            ui.warn("edit confirmation aborted by user ([q]uit) — stopping the run");
+        }
+        const filesModified = result.ok && (tc.name === "write_file" || tc.name === "edit_file");
+        return {
+            post: () => {
+                // Track file modifications — a completion claim with no edits needs no
+                // build verification, and conversational turns never touch files.
+                if (filesModified)
+                    this.filesModified = true;
+                // T-01: @-autocomplete index invalidate on tool writes.
+                if (filesModified)
+                    invalidateFileIndex();
+                this.endToolStream();
+                this.currentToolCall = null;
+                ui.toolResult(tc.name, result.ok, result.output);
+                this.emit({
+                    type: "tool_result",
+                    step: steps,
+                    id: tc.id,
+                    name: tc.name,
+                    ok: result.ok,
+                    output: result.output,
+                    ...(result.diff ? { diff: result.diff } : {}),
+                    ...(result.image ? { image: result.image } : {}),
+                });
+                stepToolResults.set(tc.id, { ok: result.ok, output: result.output });
+                this.pushToolResult(tc, summariseOutput(result.output));
+            },
+        };
+    }
+    /** MCP tool call with output clamp (extracted for I-09 readability). */
+    async invokeMcpTool(name, args) {
+        const reg = this.mcp?.getTool(name);
+        if (!reg)
+            return { ok: false, output: `ERROR: unknown tool "${name}"` };
+        // M2: validate args theo inputSchema — lỗi local, không gọi server.
+        const schemaErrors = validateSchema(args, reg.def.inputSchema);
+        if (schemaErrors.length > 0) {
+            return { ok: false, output: `ERROR: invalid arguments for ${name}:\n${schemaErrors.join("\n")}` };
+        }
+        const r = await reg.conn.callTool(reg.toolName, args);
+        return { ok: r.ok, output: r.output.length > 20_000 ? r.output.slice(0, 20_000) + "\n... [truncated]" : r.output };
+    }
     /**
      * Record usage of one LLM call: render token counts, log an `llm_usage`
      * event, and accumulate totals into session state.
@@ -185,17 +514,27 @@ export class Agent {
     /**
      * One LLM turn: streaming if enabled (renders deltas live), with automatic
      * fallback to non-streaming when the provider/stream fails.
+     *
+     * H-09 incomplete-stream continuation: when a stream dies mid-flight but a
+     * COMPLETE tool-call delta was already accumulated, surface the partial
+     * result (flagged `incomplete`) instead of discarding it — the run loop
+     * executes the tool calls and continues with a synthetic instruction.
      */
     async chatWithStream() {
         const signal = this.runSignal.signal;
         if (!this.config.stream) {
             return this.client.chat(this.messages, this.tools, { signal });
         }
+        let streamedText = "";
+        let final;
+        let started = false;
+        let sawToolCallDelta = false;
         try {
-            let streamedText = "";
-            let final;
-            let started = false;
             for await (const chunk of this.client.streamChat(this.messages, this.tools, { signal })) {
+                if (chunk.reasoningDelta) {
+                    // I-05: reasoning hiển thị dim qua UI stream; không vào llm_delta.
+                    this.emit({ type: "llm_reasoning", delta: chunk.reasoningDelta });
+                }
                 if (chunk.contentDelta) {
                     if (!started) {
                         ui.llmStreamStart("assistant");
@@ -205,6 +544,8 @@ export class Agent {
                     this.emit({ type: "llm_delta", delta: chunk.contentDelta });
                     streamedText += chunk.contentDelta;
                 }
+                if (chunk.toolCallDelta)
+                    sawToolCallDelta = true;
                 if (chunk.final)
                     final = chunk.final;
             }
@@ -219,7 +560,35 @@ export class Agent {
             // No finish_reason seen — fall through to non-streaming
             return this.client.chat(this.messages, this.tools);
         }
-        catch (err) {
+        catch (caught) {
+            if (started)
+                ui.llmStreamEnd();
+            // A timeout is a real failure, not a streaming quirk — falling back to
+            // non-streaming would just wait the full timeout a second time.
+            const e = caught;
+            const isTimeout = e?.name === "TimeoutError" || /timed out|TimeoutError/i.test(e?.message ?? "");
+            const err = isTimeout
+                ? (caught instanceof LlmError ? caught : new LlmError(`LLM request timed out: ${e?.message ?? caught}`))
+                : caught;
+            // Incomplete-stream continuation: the stream died (timeout/disconnect)
+            // after a complete tool-call delta → execute what we have rather than
+            // failing the whole step (pattern from opencode runner).
+            if (sawToolCallDelta &&
+                (isTimeout || /ECONNRESET|socket hang up|network|terminated/i.test(e?.message ?? ""))) {
+                const partial = this.client.lastPartialToolCalls ?? [];
+                if (partial.length > 0) {
+                    ui.warn("stream interrupted after complete tool-call(s) — continuing with what was received");
+                    return {
+                        message: {
+                            role: "assistant",
+                            content: streamedText || null,
+                            tool_calls: partial,
+                        },
+                        usage: undefined,
+                        incomplete: true,
+                    };
+                }
+            }
             if (err instanceof LlmError && !this.streamFallbackWarned) {
                 this.streamFallbackWarned = true;
                 ui.warn(`streaming unavailable (${err.message.slice(0, 120)}) — falling back to non-streaming`);
@@ -235,21 +604,48 @@ export class Agent {
         const servers = this.config.mcpServers ?? {};
         if (Object.keys(servers).length === 0)
             return;
+        // Idempotent: /mcp may connect while idle — never spawn a second set of
+        // server processes (the first manager's children would leak).
+        if (this.mcp)
+            return;
         if (this.config.permissionMode === "readonly") {
             ui.warn("readonly mode: MCP tools are disabled (they bypass the permission system)");
         }
         this.mcp = new McpManager(servers);
+        this.mcp.onStatus = (line) => {
+            this.logger.log({ type: "mcp_status", line });
+            this.emit({ type: "mcp_status", line });
+            ui.info(line); // setQuiet suppresses this in the TUI — alt-screen stays clean
+        };
+        // Sidebar TUI: trạng thái máy trạng thái per-server (M1) — event structured
+        // (server/state) để reducer giữ map, không append thêm notice vào chat.
+        this.mcp.onServerState = (name, st, detail) => {
+            const line = `MCP "${name}" ${st}${detail ? ` — ${detail}` : ""}`;
+            this.logger.log({ type: "mcp_status", line, server: name, state: st });
+            this.emit({ type: "mcp_status", line, server: name, state: st });
+        };
         await this.mcp.connectAll();
         const mcpSpecs = this.mcp.toToolSpecs();
         if (mcpSpecs.length > 0) {
-            this.tools = [...AGENT_TOOLS.map((t) => t.spec), ...mcpSpecs];
+            this.tools = [...allToolSpecs(), ...mcpSpecs];
         }
     }
-    /** Disconnect MCP servers at the end of a run. */
+    /**
+     * Disconnect MCP servers when the agent is disposed — connections are kept
+     * alive for the whole agent/session lifetime (one "connected" line, no
+     * re-spawn per message). Ephemeral owners (server runs, one-shot CLI) call
+     * dispose() when their agent is done.
+     */
     shutdownMcp() {
         this.mcp?.killAll();
         this.mcp = null;
     }
+    /** M4: đóng hot-reload watcher (khi agent bị hủy vĩnh viễn). */
+    dispose() {
+        this.watcher?.close();
+        this.watcher = null;
+        this.shutdownMcp();
+    }
     /**
      * Autonomous closed loop:
      * task -> LLM plans -> tool calls -> build -> on failure: read error, patch, retry.
@@ -259,6 +655,8 @@ export class Agent {
             throw new Error("No API key configured. Set SCODER_API_KEY or OPENAI_API_KEY (or put apiKey in .scoderrc.json).");
         }
         const buildCmd = this.buildCmd;
+        // @file mention expansion — task trước khi build messages.
+        task = expandMentions(task, this.config.workdir);
         ui.banner("scoder", [
             ["task", task],
             ["model", `${this.client.model} (${this.config.provider}${this.config.stream ? ", stream" : ""})`],
@@ -267,16 +665,30 @@ export class Agent {
             ["steps", `max ${this.config.maxSteps}  ·  permission ${this.config.permissionMode}`],
         ]);
         this.currentTask = task;
+        this.events.append({ type: "run.started", task });
         this.resetRunSignal();
         this.resetUsageTotals();
+        this.filesModified = false;
+        this.plan = [];
+        const skillsBlock = buildSkillsBlock(this.skills); // I3
         this.messages = [
             {
                 role: "system",
-                content: SYSTEM_PROMPT +
-                    `\n\nWorking directory: ${this.config.workdir}\nBuild command (stop condition): ${buildCmd}`,
+                content: buildSystemPrompt({
+                    workdir: this.config.workdir,
+                    buildCmd,
+                    envInfo: buildEnvInfo(this.config.workdir),
+                    skillsBlock,
+                    mcpBlock: buildMcpBlock(this.config),
+                }) + buildAgentsBlock(discoverAgents(this.config.workdir), this.config.allowSubagents !== false),
             },
+            ...this.pendingContext.map((content) => ({ role: "user", content })),
             { role: "user", content: task },
         ];
+        this.pendingContext = [];
+        // Phiên bản system prompt được ghi vào log — giúp so sánh chất lượng
+        // giữa các phiên bản prompt khi phân tích JSONL.
+        this.logger.log({ type: "prompt_version", version: PROMPT_VERSION });
         this.emit({
             type: "run_start",
             task,
@@ -284,36 +696,58 @@ export class Agent {
             provider: this.config.provider,
             workdir: this.config.workdir,
             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):
      * appends the task to the existing message history instead of resetting it,
      * then runs the autonomous loop. First call behaves like run().
      */
+    /**
+     * T-07: trim the conversation back one exchange (last user message +
+     * everything after) — fork/edit-resend. The caller re-sends an edited
+     * task via runTurn().
+     */
+    trimLastTurn() {
+        // Cut = last REAL user message (skip guardrail reminder tails injected at
+        // finish); drop it and everything after. Reminder tails BEFORE the cut
+        // point belong to earlier turns and stay (history intact).
+        for (let i = this.messages.length - 1; i >= 0; i--) {
+            const m = this.messages[i];
+            if (m.role !== "user")
+                continue;
+            if (this.isReminderTail(String(m.content ?? "")))
+                continue;
+            this.messages = this.messages.slice(0, i);
+            return;
+        }
+        this.messages = [];
+    }
+    /** True for guardrail-injected user messages (not real user tasks). */
+    isReminderTail(content) {
+        return (content.includes("<system-reminder>") ||
+            /^Continue working on (the task|the original task)\. Build must pass before you finish\.$/.test(content.trim()));
+    }
     async runTurn(task) {
         if (this.messages.length === 0) {
             return this.run(task);
         }
         const buildCmd = this.buildCmd;
+        task = expandMentions(task, this.config.workdir);
         this.currentTask = `${this.currentTask}\n${task}`.trim();
         this.resetRunSignal();
         this.resetUsageTotals();
+        this.filesModified = false;
+        this.plan = [];
+        for (const content of this.pendingContext)
+            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.
@@ -325,11 +759,16 @@ export class Agent {
         this.currentTask = session.task;
         this.resetRunSignal();
         this.resetUsageTotals();
+        this.filesModified = false;
+        this.skills = discoverSkills(this.config.workdir);
         this.messages = session.messages ?? [];
         ui.header(`RESUME ${session.id}`);
         ui.info(`previous task: ${session.task} (${session.steps} step(s) done)`);
         if (followUp) {
-            this.messages.push({ role: "user", content: followUp });
+            for (const content of this.pendingContext)
+                this.messages.push({ role: "user", content });
+            this.pendingContext = [];
+            this.messages.push({ role: "user", content: expandMentions(followUp, this.config.workdir) });
         }
         else {
             this.messages.push({
@@ -340,13 +779,10 @@ export class Agent {
         const lastBuild = session.lastBuild
             ? { success: session.lastBuild.success, command: session.lastBuild.command, output: session.lastBuild.output }
             : null;
-        try {
-            await this.connectMcp();
-            return await this.runLoop(session.steps ?? 0, buildCmd, lastBuild);
-        }
-        finally {
-            this.shutdownMcp();
-        }
+        // v2: opening a resumed run stamps a fresh run.started (event log continuity).
+        this.events.append({ type: "run.started", task: this.currentTask });
+        await this.connectMcp();
+        return this.runLoop(session.steps ?? 0, buildCmd, lastBuild);
     }
     /**
      * The shared autonomous loop:
@@ -355,16 +791,48 @@ export class Agent {
     async runLoop(startSteps, buildCmd, initialBuild = null) {
         let steps = startSteps;
         let lastBuild = initialBuild;
+        // G1: guardrail tracker state (textOnlyStreak, sigRepeat, toolError*,
+        // lowTokenStreak, wrapUpInjected) sống trong GuardrailEngine — reset mỗi run.
+        this.guardrails.reset();
+        // H-10 compaction config (false = disabled → mechanical trim only).
+        // I-02: contextWindow từ model catalog (lookup sync, cache/fallback) thay
+        // vì default 128k cứng — compaction trigger chính xác theo model.
+        const catalogWindow = contextWindowFor(this.config.provider, this.config.model);
+        const compactionCfg = this.config.compaction === false
+            ? null
+            : {
+                ...DEFAULT_COMPACTION,
+                contextWindow: this.config.contextWindow ?? catalogWindow,
+                ...(this.config.compaction ?? {}),
+            };
         while (steps < this.config.maxSteps && !this.aborted) {
+            // ── Safe step boundary: deliver steered user input (H-08) ────────────
+            const steered = this.drainSteerInbox();
+            if (steered.length > 0) {
+                for (const s of steered) {
+                    this.messages.push({ role: "user", content: s });
+                    this.events.append({ type: "message.added", message: { role: "user", content: s } });
+                }
+                // One allowance reset for the whole steering batch (opencode behavior)
+                // — engine exposes a hook so the runLoop doesn't touch tracker state.
+                this.guardrails.notifySteered();
+                this.emit({ type: "steer_delivered", count: steered.length });
+                ui.info(`steer: +${steered.length} user message(s) delivered`);
+            }
             steps++;
             ui.step(steps, this.config.maxSteps, "thinking...");
+            this.events.claimExecution(steps);
             this.emit({ type: "step", step: steps, maxSteps: this.config.maxSteps });
-            const { message, usage } = await this.chatWithStream().catch((err) => {
+            // Metrics: thời gian phần LLM của bước này (điều chỉnh chất lượng +
+            // so sánh provider; hiển thị trong `scoder cost`).
+            const stepStart = Date.now();
+            const { message, usage, incomplete } = await this.chatWithStream().catch((err) => {
                 // SIGINT during the LLM call — stop gracefully instead of crashing
                 if (this.aborted) {
                     return {
                         message: { role: "assistant", content: null, tool_calls: undefined },
                         usage: undefined,
+                        incomplete: undefined,
                     };
                 }
                 throw err;
@@ -372,163 +840,214 @@ export class Agent {
             if (this.aborted)
                 break;
             this.recordLlmUsage(usage);
-            this.logger.log({ type: "step", step: steps });
-            // No tool calls => assistant produced a text answer
-            if (!message.tool_calls || message.tool_calls.length === 0) {
-                const text = message.content ?? "";
+            this.logger.log({ type: "step", step: steps, durationMs: Date.now() - stepStart, toolCalls: message.tool_calls?.length ?? 0 });
+            // H-09: after an interrupted stream, nudge the model to continue rather
+            // than restart (synthetic continuation instruction, opencode pattern).
+            if (incomplete) {
+                this.messages.push({
+                    role: "user",
+                    content: "[system] The previous response was cut off mid-stream. The tool call(s) above were received in full — continue exactly where you left off; do not repeat completed work.",
+                });
+                this.events.append({ type: "retry.scheduled", reason: "incomplete-stream continuation" });
+            }
+            // ── Chuẩn bị context cho Guardrail Engine ─────────────────────────────
+            // Chat marker được làm sạch ở đây để rule nhận text đã chuẩn hoá.
+            const hasToolCalls = !!message.tool_calls && message.tool_calls.length > 0;
+            let text = message.content ?? "";
+            let chatDone = false;
+            if (!hasToolCalls && text.includes(CHAT_DONE_MARKER)) {
+                chatDone = true;
+                text = text.split(CHAT_DONE_MARKER)[0].trim();
+            }
+            if (!hasToolCalls) {
                 this.messages.push({ role: "assistant", content: text });
-                // If the assistant says it's done without building, force a build check.
-                if (claimsCompletion(text) && steps < this.config.maxSteps) {
-                    ui.info("agent claims completion -> verifying with build");
-                    lastBuild = await this.checkBuild(buildCmd);
-                    if (lastBuild.success) {
-                        return this.finish(true, steps, buildCmd, text);
-                    }
-                    this.messages.push({
-                        role: "user",
-                        content: `Build still FAILED after your claimed completion. Output:\n${summariseOutput(lastBuild.output)}\nFix and continue.`,
-                    });
-                    continue;
-                }
+                this.events.append({ type: "message.added", message: { role: "assistant", content: text } });
                 ui.llm("agent", text);
-                // Treat a plain text answer with no build as interim commentary; nudge loop.
+            }
+            else {
                 this.messages.push({
-                    role: "user",
-                    content: "Continue working on the task. Use tools. Build must pass before you finish.",
+                    role: "assistant",
+                    content: message.content ?? null,
+                    tool_calls: message.tool_calls,
                 });
-                continue;
+                this.events.append({ type: "message.added", message: this.messages[this.messages.length - 1] });
             }
-            this.messages.push({
-                role: "assistant",
-                content: message.content ?? null,
-                tool_calls: message.tool_calls,
-            });
             // Execute each tool call (validate args before running)
             const stepToolResults = new Map();
-            for (const tc of message.tool_calls) {
-                if (this.aborted)
-                    break;
-                const tool = getTool(tc.name);
-                let args;
+            // H-11: checkpoint before any batch that may write files.
+            const batchMayWrite = (message.tool_calls ?? []).some((tc) => tc.name === "write_file" || tc.name === "edit_file");
+            if (batchMayWrite)
+                this.snapshots.capture(`step ${steps}: ${(message.tool_calls ?? []).map((t) => t.name).join(",")}`);
+            // I-09: parallelize read-only tool calls (concurrency cap 4); mutating
+            // calls (and any batch containing an ask-gated call) run sequentially.
+            const calls = message.tool_calls ?? [];
+            const needsAsk = calls.some((tc) => {
+                if (tc.name !== "run_command")
+                    return false;
                 try {
-                    args = tc.arguments ? JSON.parse(tc.arguments) : {};
+                    const a = JSON.parse(tc.arguments || "{}");
+                    if (typeof a.command !== "string")
+                        return false;
+                    const d = this.permissions.classifyShell(a.command);
+                    return d.decision === "ask" || d.decision === "deny";
                 }
                 catch {
-                    this.pushToolResult(tc, `ERROR: invalid JSON arguments: ${tc.arguments.slice(0, 200)}`);
-                    continue;
-                }
-                ui.toolCall(tc.name, args);
-                this.currentToolCall = { step: steps, id: tc.id, name: tc.name };
-                this.emit({ type: "tool_call", step: steps, id: tc.id, name: tc.name, args });
-                if (!tool && !(this.mcp?.hasTool(tc.name))) {
-                    this.rejectToolCall(steps, tc, `ERROR: unknown tool "${tc.name}". Available: ${AGENT_TOOLS.map((t) => t.spec.function.name).join(", ")}${this.mcp ? " + MCP tools (mcp__<server>__<tool>)" : ""}`, { error: "unknown tool" });
-                    continue;
+                    return true;
                 }
-                let result;
-                try {
-                    // Permission gate for shell commands
-                    if (tc.name === "run_command" && typeof args.command === "string") {
-                        const decision = this.permissions.classify(args.command);
-                        if (decision.decision === "deny") {
-                            result = {
-                                ok: false,
-                                output: `ERROR: command denied by permission policy (${decision.reason}): ${String(args.command).slice(0, 200)}`,
-                            };
-                            this.rejectToolCall(steps, tc, result.output, { args, denied: decision.reason });
-                            continue;
-                        }
-                        if (decision.decision === "ask") {
-                            const allowed = await this.permissions.confirm(args.command);
-                            if (!allowed) {
-                                result = {
-                                    ok: false,
-                                    output: `ERROR: user denied permission to run: ${String(args.command).slice(0, 200)}`,
-                                };
-                                this.rejectToolCall(steps, tc, result.output, { args, denied: "user" });
-                                continue;
-                            }
-                        }
-                    }
-                    const t0 = Date.now();
-                    // Route MCP tools first (they carry the mcp__ prefix, not in AGENT_TOOLS)
-                    if (tool) {
-                        result = await tool.run(this.ctx, args);
+            });
+            if (calls.length > 1 && !needsAsk) {
+                const readOnly = calls.every((tc) => Agent.READONLY_TOOLS.has(tc.name));
+                if (readOnly) {
+                    const results = await this.parallelMap(calls, (tc) => this.executeToolCall(tc, steps, stepToolResults), this.config.limits?.toolConcurrency ?? 4);
+                    for (const r of results) {
+                        if (r.post)
+                            r.post();
                     }
-                    else if (this.config.permissionMode === "readonly") {
-                        // MCP tools are not covered by the permission system (no command
-                        // classification, no diff preview) — deny them entirely in readonly.
-                        result = {
-                            ok: false,
-                            output: `ERROR: readonly mode — MCP tool "${tc.name}" is disabled. MCP tools bypass the built-in permission policy, so they cannot run in readonly mode.`,
-                        };
+                }
+                else {
+                    for (const tc of calls) {
+                        if (this.aborted)
+                            break;
+                        const r = await this.executeToolCall(tc, steps, stepToolResults);
+                        if (r.post)
+                            r.post();
                     }
-                    else {
-                        const reg = this.mcp?.getTool(tc.name);
-                        if (!reg) {
-                            result = { ok: false, output: `ERROR: unknown tool "${tc.name}"` };
+                }
+            }
+            else {
+                for (const tc of calls) {
+                    if (this.aborted)
+                        break;
+                    const r = await this.executeToolCall(tc, steps, stepToolResults);
+                    if (r.post)
+                        r.post();
+                }
+            }
+            // ── Guardrail Engine — đánh giá toàn bộ rule sau khi tools chạy xong ──
+            // Thông tin build stop condition: có tool call nào chạy ĐÚNG buildCmd?
+            // Match theo arguments đã parse (đúng hơn substring: "npm run build && rm -rf /"
+            // vẫn tính là build call, nhưng lệnh echo chứa chuỗi buildCmd thì không).
+            const buildCall = hasToolCalls
+                ? (message.tool_calls ?? []).find((tc) => tc.name === "run_command" &&
+                    (() => {
+                        try {
+                            const a = JSON.parse(tc.arguments || "{}");
+                            return typeof a.command === "string" && a.command.includes(buildCmd);
                         }
-                        else {
-                            const r = await reg.conn.callTool(reg.toolName, args);
-                            result = { ok: r.ok, output: r.output.length > 20_000 ? r.output.slice(0, 20_000) + "\n... [truncated]" : r.output };
+                        catch {
+                            return false;
                         }
-                    }
-                    this.logger.log({ type: "tool", step: steps, tool: tc.name, args, ok: result.ok, elapsedMs: Date.now() - t0 });
+                    })())
+                : undefined;
+            let buildPassedThisStep = false;
+            if (buildCall) {
+                const r = stepToolResults.get(buildCall.id);
+                if (r && r.ok)
+                    buildPassedThisStep = true;
+            }
+            const decisions = await this.guardrails.evaluate({
+                step: steps,
+                maxSteps: this.config.maxSteps,
+                text,
+                message,
+                hasToolCalls,
+                toolResults: stepToolResults,
+                buildPassedThisStep,
+                buildCallPresent: !!buildCall,
+                filesModified: this.filesModified,
+                chatDone,
+                outputTokensThisStep: usage?.completion_tokens,
+                totalTokensIn: this.usageTotals.inputTokens, // G6 budget rule
+                totalTokensOut: this.usageTotals.outputTokens,
+            });
+            // Áp dụng decision: finish đầu tiên thắng; inject gom lại đẩy một lần.
+            const reminders = [];
+            let doTrim = false;
+            for (const d of decisions) {
+                const a = d.action;
+                if (a.kind === "inject") {
+                    if (d.rule === "wrap-up")
+                        this.guardrails.markWrapUpInjected(); // chỉ inject một lần / run
+                    reminders.push(a.reminder);
+                    continue;
                 }
-                catch (err) {
-                    result = { ok: false, output: `ERROR: ${err.message}` };
-                    this.logger.log({ type: "tool", step: steps, tool: tc.name, args, ok: false, error: err.message });
+                if (a.kind === "trim") {
+                    doTrim = true;
+                    continue;
                 }
-                // User pressed [q]uit in an edit-confirmation prompt — stop the whole run.
-                if (this.permissions.abortRequested) {
-                    this.aborted = true;
-                    ui.warn("edit confirmation aborted by user ([q]uit) — stopping the run");
+                if (a.kind === "verify-build") {
+                    // Xác minh độc lập với lời của model (async — loop tự chạy build).
+                    lastBuild = await this.checkBuild(buildCmd);
+                    if (lastBuild.success) {
+                        if (doTrim)
+                            this.messages = trimHistory(this.messages);
+                        return this.finish(true, steps, buildCmd, a.successSummary);
+                    }
+                    reminders.push(`${a.failReminder}\n${summariseOutput(lastBuild.output)}\nFix and continue.`);
+                    continue;
+                }
+                if (a.kind === "finish") {
+                    // Giữ nguyên thứ tự tin nhắn như bản inline cũ: reminder (wrap-up)
+                    // đã đẩy trước khi kết thúc.
+                    for (const r of reminders)
+                        this.messages.push({ role: "user", content: r });
+                    if (doTrim)
+                        this.messages = trimHistory(this.messages);
+                    return this.finish(a.success, steps, buildCmd, a.summary, a.chatOnly ?? false);
                 }
-                this.endToolStream();
-                this.currentToolCall = null;
-                ui.toolResult(tc.name, result.ok, result.output);
-                this.emit({
-                    type: "tool_result",
-                    step: steps,
-                    name: tc.name,
-                    ok: result.ok,
-                    output: result.output,
-                    ...(result.diff ? { diff: result.diff } : {}),
-                });
-                stepToolResults.set(tc.id, { ok: result.ok, output: result.output });
-                this.pushToolResult(tc, summariseOutput(result.output));
             }
-            // After a build-command run, treat success as the terminal condition.
-            // Match the FULL build command (not just its first word): matching "npm"
-            // would falsely trigger on "npm install" mid-task and end the agent.
-            const buildCall = message.tool_calls.find((tc) => tc.name === "run_command" && tc.arguments.includes(buildCmd));
-            if (buildCall) {
-                // Reuse the tool result when the agent ran the exact build command —
-                // avoids executing the build a second time just to verify.
-                const result = stepToolResults.get(buildCall.id);
-                if (result && result.ok) {
-                    return this.finish(true, steps, buildCmd, message.content ?? "Build passed.");
+            for (const r of reminders)
+                this.messages.push({ role: "user", content: r });
+            if (doTrim)
+                this.messages = trimHistory(this.messages);
+            // H-10: LLM compaction when history approaches the context window.
+            if (compactionCfg && shouldCompact(this.messages, compactionCfg)) {
+                try {
+                    ui.info("compaction: summarizing older history…");
+                    const summaryInput = extractSummaryInput(this.currentTask, this.messages);
+                    const summaryRes = await this.client.chat([
+                        ...this.messages,
+                        { role: "user", content: `${summaryInput}\n\n${COMPACTION_INSTRUCTION}` },
+                    ], [], // no tools for the summarizer call
+                    { signal: this.runSignal.signal });
+                    this.messages = applyCompaction(this.messages, redactSecrets(summaryRes.message.content ?? "History summarized."), compactionCfg);
+                    this.events.append({ type: "compaction.ran", tokensBefore: 0, summary: (summaryRes.message.content ?? "").slice(0, 2000) });
+                    ui.info(`compaction: done (history now ${this.messages.length} messages)`);
                 }
-                lastBuild = await this.checkBuild(buildCmd);
-                if (lastBuild.success) {
-                    return this.finish(true, steps, buildCmd, message.content ?? "Build passed.");
+                catch (err) {
+                    // Compaction is best-effort — mechanical trim stays as fallback.
+                    ui.warn(`compaction failed (${err.message.slice(0, 100)}) — using mechanical trim`);
+                    this.messages = trimHistory(this.messages);
                 }
-                this.messages.push({
-                    role: "user",
-                    content: `Build FAILED (stop condition not met). Error output:\n${summariseOutput(lastBuild.output)}\nAnalyze the errors, fix the code, and rebuild.`,
-                });
             }
-            this.messages = trimHistory(this.messages);
             // Persist session state after every step (enables resume after crash/SIGINT)
             this.saveSession(steps, buildCmd, lastBuild);
+            this.events.releaseExecution(steps);
         }
         // Max steps reached or aborted
         const reason = this.aborted ? "interrupted by user (SIGINT)" : `reached max steps (${this.config.maxSteps})`;
         this.saveSession(steps, buildCmd, lastBuild);
         return this.finish(false, steps, buildCmd, `Agent stopped: ${reason}`);
     }
-    /** Persist current agent state to .scoder/sessions/<id>.json. */
+    /** Persist current agent state (v2 event log + manifest; crash-resumable). */
     saveSession(steps, buildCmd, lastBuild) {
         this.lastBuildResult = lastBuild;
+        // v2: durable event log is authoritative; JSON manifest is a cheap summary.
+        this.events.writeManifest({
+            task: this.currentTask,
+            model: this.client.model,
+            provider: this.config.provider,
+            buildCmd,
+            steps,
+            maxSteps: this.config.maxSteps,
+            usage: {
+                llmCalls: this.usageTotals.calls,
+                inputTokens: this.usageTotals.inputTokens,
+                outputTokens: this.usageTotals.outputTokens,
+                cachedTokens: this.usageTotals.cachedTokens,
+            },
+        });
+        // v1 snapshot kept in sync (back-compat for /sessions readers, cost, resume).
         this.session.save({
             task: this.currentTask,
             model: this.client.model,
@@ -557,9 +1076,134 @@ export class Agent {
         this.logger.log({ type: "build", success: res.success, command: buildCmd, output: res.output.slice(0, 10_000) });
         return res;
     }
-    finish(success, steps, buildCmd, summary) {
+    // ── Frontend-facing APIs (REPL/TUI slash commands) ─────────────────────────
+    /** Skills visible to this project (for /skills). */
+    getSkills() {
+        return this.skills;
+    }
+    /**
+     * Stage a skill's content for injection before the next run/turn (/skill).
+     * Non-destructive: nothing is sent to the LLM until the user submits a task.
+     */
+    stageSkill(name) {
+        const skill = findSkill(this.config.workdir, name);
+        if (!skill) {
+            const names = this.skills.map((s) => s.name).join(", ");
+            return { ok: false, message: `skill "${name}" not found${names ? ` — available: ${names}` : " — no skills discovered (try /skills)"}` };
+        }
+        this.pendingContext.push(`[User loaded skill "${skill.name}" — follow these instructions for the upcoming task]\n\n${skill.body}`);
+        return { ok: true, message: `skill "${skill.name}" staged — it will apply to your next task` };
+    }
+    /** MCP server overview (for /mcp): config always; tools once connected. */
+    async mcpStatus() {
+        const configs = (this.config.mcpServers ?? {});
+        // M3: http server hiển thị url thay vì command.
+        const target = (cfg) => cfg.url ?? cfg.command ?? "(unknown)";
+        if (!this.mcp) {
+            return Object.entries(configs).map(([name, cfg]) => ({ name, command: target(cfg), connected: false, tools: [] }));
+        }
+        await this.connectMcp();
+        const byServer = new Map(this.mcp.listToolsByServer().map((s) => [s.server, s.tools]));
+        return Object.entries(configs).map(([name, cfg]) => ({
+            name,
+            command: target(cfg),
+            connected: byServer.has(name),
+            tools: byServer.get(name) ?? [],
+        }));
+    }
+    /** Call one MCP tool directly by its full name (mcp__server__tool) (/mcp exec). */
+    async callMcpTool(fullName, args) {
+        if (this.config.permissionMode === "readonly") {
+            return { ok: false, output: "ERROR: readonly mode — MCP tools are disabled (they bypass the permission policy)." };
+        }
+        await this.connectMcp();
+        const reg = this.mcp?.getTool(fullName);
+        if (!reg) {
+            return { ok: false, output: `ERROR: unknown MCP tool "${fullName}" — run /mcp to list servers and tools.` };
+        }
+        // Text shorthand (`/mcp exec <tool> free text`): bind to the tool's first
+        // required string property from its inputSchema, so users don't have to
+        // hand-write JSON for the common one-argument case.
+        if ("__text" in args && typeof args.__text === "string") {
+            const props = (reg.def.inputSchema?.properties ?? {});
+            const required = (reg.def.inputSchema?.required ?? []);
+            const target = required.find((k) => props[k]?.type === "string") ??
+                Object.keys(props).find((k) => props[k]?.type === "string");
+            if (!target) {
+                return { ok: false, output: `ERROR: "${fullName}" has no string parameter — pass JSON args instead.` };
+            }
+            args = { [target]: args.__text };
+        }
+        try {
+            return await reg.conn.callTool(reg.toolName, args);
+        }
+        catch (err) {
+            return { ok: false, output: `ERROR: ${err.message}` };
+        }
+    }
+    /** Ask user to approve an MCP tool call; "always" persists a project-scoped rule. */
+    async confirmMcpTool(fullName) {
+        return this.permissions.confirmMcpTool(fullName, () => {
+            try {
+                saveApproval(this.config.workdir, { action: fullName, resource: "*", effect: "allow" });
+            }
+            catch { /* persistence failure must not block the run */ }
+        });
+    }
+    /** Switch permission mode at runtime (/permission) — rebuilds the checker. */ setPermissionMode(mode) {
+        if (mode === this.config.permissionMode)
+            return;
+        this.config = { ...this.config, permissionMode: mode };
+        const permLists = loadPermissionLists(this.config.workdir);
+        this.permissions = new PermissionChecker({
+            mode,
+            allowCommands: permLists.allowCommands,
+            denyCommands: permLists.denyCommands,
+            nonInteractive: this.config.yes,
+            rules: this.config.permissionRules,
+            savedRules: loadSavedApprovals(this.config.workdir),
+            policyRules: loadPolicyRules(this.config.workdir),
+        });
+        this.ctx.permissionMode = mode;
+        ui.info(`permission mode → ${mode}`);
+    }
+    /**
+     * T-04: runtime model hot-swap. Same provider → setModel on the live
+     * client (zero rebuild, mid-session). Different provider (provider/model
+     * syntax) → rebuild the runtime from the registry. Returns a notice line.
+     */
+    setModel(model) {
+        const [prov, id] = splitProviderModel(model);
+        if (!prov || prov === this.config.provider) {
+            const m = id ?? model;
+            this.client.setModel(m);
+            this.config = { ...this.config, model: m };
+            return `model → ${m} (${this.config.provider})`;
+        }
+        // Cross-provider: rebuild the client runtime.
+        const resolved = resolveProvider(prov, parseCustomProviders(loadRcFile(this.config.workdir).providers));
+        if ("error" in resolved)
+            return `ERROR: ${resolved.error}`;
+        this.client = createRuntime(resolved.def, {
+            baseUrl: this.config.baseUrl,
+            apiKey: this.config.apiKey,
+            model: id,
+            maxTokens: this.config.maxTokens,
+            temperature: this.config.temperature,
+            enableCache: resolved.def.supportCache,
+        });
+        this.config = { ...this.config, provider: prov, model: id, providerDef: resolved.def };
+        return `model → ${id} (provider ${prov})`;
+    }
+    finish(success, steps, buildCmd, summary, chatOnly = false) {
+        this.events.append({ type: "finish", success, summary: summary.slice(0, 2000) });
         if (success) {
-            ui.success(`Task complete in ${steps} step(s). Build passed: ${buildCmd}`);
+            if (chatOnly) {
+                ui.success(`Done in ${steps} step(s).`);
+            }
+            else {
+                ui.success(`Task complete in ${steps} step(s). Build passed: ${buildCmd}`);
+            }
         }
         else {
             ui.error(`Task NOT complete. ${summary}`);