visionclaw 0.1.219-beta.0 → 0.1.219-dev.codex-issue-2805-worker-runtime.1

package/dist/agent/engines/pi-coding/session.js

@@ -4,7 +4,7 @@ import crypto from "crypto";
 import fs from "fs";
 import path from "path";
 import { configuredRuntimeApplicationId } from "../../../config/application-config.js";
-import { getConfigDir, getSkillsDir, getWorkerThreadId, loadSessionId, loadUsageSnapshot, saveSessionId, saveUsageSnapshot } from "../../../config/index.js";
+import { getConfigDir, getSkillsDir, loadSessionId, loadUsageSnapshot, saveSessionId, saveUsageSnapshot } from "../../../config/index.js";
 import { isGeminiEnabled, isTunnelEnabled, isWorkspaceEnabled } from "../../../config/types.js";
 import { isFundLeaguePlayerRuntime } from "../../fund-league/core-contract.js";
 import { logger } from "../../../logger.js";
@@ -14,13 +14,12 @@ import { COMPACTION_ABORT_RESTART_GRACE_MS, DEFAULT_COMPACTION_TIMEOUT_MS, COMPA
 import { logSystemPrompt } from "../system-prompt-log.js";
 import { resolvePiModel } from "../pi/shared/model.js";
 import { resolvePiCompactionSettings, resolvePiMaxBase64ImageBlocks, } from "../pi/shared/context-window.js";
-import { getEffectiveRoute, getLlmEffort } from "../../model-provider.js";
+import { getEffectiveRoute, getLlmEffort, pinnedModelApplies } from "../../model-provider.js";
 import { createLlmGatewayAttemptBinder, isLlmGatewayProvider, } from "../../llm-gateway-request-id.js";
 import { needsGatewayResponsesSanitizer, sanitizeGrokResponsesPayload } from "./grok-responses-payload.js";
 import { convertPiAgentEventToStreamMessages } from "../pi/shared/event-converter.js";
 import { buildPiResultMessage, capturePiMessageUsage, createPiUsageAccumulator } from "../pi/shared/usage.js";
 import { createPiCodingCustomTools } from "./tools.js";
-import { isLabRunThreadId } from "../../../threads/lab-run-thread.js";
 import { pruneSessionImages, sanitizeTranscript } from "../../image-pruner.js";
 import { deriveArchivePath, trimTranscriptAfterCompaction } from "../../session-trimmer.js";
 import { writePiMcpConfig } from "./mcp-config.js";
@@ -28,9 +27,10 @@ import { getMcpScriptRuntimeStatus } from "./mcp-script-runtime.js";
 import { ensureBrowserForTool, isPlaywrightToolName } from "../browser-tool-policy.js";
 import { sanitizeToolSchemasInPayload } from "./tool-schema-sanitizer.js";
 import { getAgentState } from "../../state.js";
-import { resolvePiLabToolNames } from "../../../threads/lab-runtime-policy.js";
+import { resolvePiWorkerToolNames } from "../../../threads/worker-runtime-policy.js";
+import { getWorkerExecutionRuntime } from "../../../threads/worker-execution-runtime.js";
+import { getForegroundTaskTracker } from "../../../threads/foreground-task.js";
 import { CompactionFailureError } from "../../compaction-failure.js";
-import { buildThreadTaskContext } from "../../../threads/thread-task-runtime.js";
 import { ToolOutputGuard } from "../../tool-output-cap.js";
 import { repairTranscriptToolOutputs } from "../../transcript-output-repair.js";
 import { guardPiToolResult } from "../pi/tools.js";
@@ -40,15 +40,20 @@ const MAX_TURNS = 100;
 const PI_ABORT_TIMEOUT_MS = 5_000;
 /** Budget for pi-mcp-adapter to close its stdio children on session_shutdown. */
 const MCP_SHUTDOWN_TIMEOUT_MS = 8_000;
-async function loadMcpAdapter() {
+/**
+ * Pi state lives under `<profile>/pi-agent/general`. The directory name is
+ * kept from the dual-session layout so existing transcripts still resume;
+ * `pi-agent/coding` is no longer read.
+ */
+const PI_AGENT_DIR_NAME = "general";
+async function loadMcpAdapter(configPath) {
     const module = await import("pi-mcp-adapter");
-    return module.default;
+    return module.createMcpAdapter({ configPath });
 }
 export class PiCodingAgentSession {
     config;
     buildSystemPrompt;
     runtimeSurface;
-    getDualSessionEnabled;
     agentDir;
     sessionDir;
     workspaceCwd;
@@ -63,6 +68,8 @@ export class PiCodingAgentSession {
     _screenshotInjector = null;
     inputClosed = true;
     stopRequested = false;
+    /** Set only by `agent_settled`, the event that means the SDK will not continue. */
+    barrierConfirmed = false;
     lastUsageSnapshot = null;
     sessionId;
     orphanedInjections = 0;
@@ -75,26 +82,23 @@ export class PiCodingAgentSession {
     sessionRelease = Promise.resolve();
     /** In-flight createSession(); its adapter children exist before `this.session` is set. */
     sessionCreation = null;
-    mode;
     set screenshotInjector(injector) {
         this._screenshotInjector = injector;
     }
-    constructor({ config, buildSystemPrompt, runtimeSurface, sessionContext, workspaceCwd, }) {
+    constructor({ config, buildSystemPrompt, runtimeSurface, workspaceCwd, }) {
         this.config = config;
         this.buildSystemPrompt = buildSystemPrompt;
         this.runtimeSurface = runtimeSurface;
-        this.getDualSessionEnabled = sessionContext.getDualSessionEnabled;
-        this.mode = sessionContext.mode;
-        if (runtimeSurface.labRuntimePolicy && !workspaceCwd) {
-            throw new Error("Pi policy-isolated Lab sessions require a workspace cwd");
+        if (runtimeSurface.workerRuntimePolicy && !workspaceCwd) {
+            throw new Error("Pi policy-isolated worker sessions require a workspace cwd");
         }
         this.workspaceCwd = workspaceCwd ?? getConfigDir();
-        this.agentDir = path.join(getConfigDir(), "pi-agent", this.mode);
+        this.agentDir = path.join(getConfigDir(), "pi-agent", PI_AGENT_DIR_NAME);
         this.sessionDir = path.join(this.agentDir, "sessions");
-        const persistedSessionId = loadSessionId(this.mode);
+        const persistedSessionId = loadSessionId();
         this.sessionId = persistedSessionId ?? createSessionId();
-        saveSessionId(this.sessionId, this.mode);
-        const persistedUsage = loadUsageSnapshot(this.mode);
+        saveSessionId(this.sessionId);
+        const persistedUsage = loadUsageSnapshot();
         if (persistedUsage) {
             this.lastUsageSnapshot = { ...persistedUsage, capturedAtMs: Date.now() };
         }
@@ -125,6 +129,7 @@ export class PiCodingAgentSession {
         logSystemPrompt(systemPrompt);
         this.inputClosed = false;
         this.stopRequested = false;
+        this.barrierConfirmed = false;
         this.orphanedInjections = 0;
         return this.streamMessages(content, systemPrompt, options?.effort);
     }
@@ -167,6 +172,9 @@ export class PiCodingAgentSession {
     isStopRequested() {
         return this.stopRequested;
     }
+    executionBarrierConfirmed() {
+        return this.barrierConfirmed;
+    }
     captureSessionId(_id) {
         // Pi Coding Agent owns its session id through SessionManager.
     }
@@ -174,7 +182,7 @@ export class PiCodingAgentSession {
         if (snapshot.contextWindow <= 0)
             return;
         this.lastUsageSnapshot = { ...snapshot, capturedAtMs: Date.now() };
-        saveUsageSnapshot(snapshot, this.mode);
+        saveUsageSnapshot(snapshot);
     }
     capturePostCompactionSnapshot(postCompactionTokens) {
         const contextWindow = this.lastUsageSnapshot?.contextWindow ?? 0;
@@ -217,7 +225,7 @@ export class PiCodingAgentSession {
             void activeSession.abort().catch((err) => {
                 logger.warn(`[pi-coding-session] Compaction abort failed: ${err instanceof Error ? err.message : String(err)}`);
             });
-            restartFallback.disarm ??= armContextActionRestartFallback(`context compaction abort (${this.mode}, pi-coding)`, COMPACTION_ABORT_RESTART_GRACE_MS);
+            restartFallback.disarm ??= armContextActionRestartFallback("context compaction abort (pi-coding)", COMPACTION_ABORT_RESTART_GRACE_MS);
         }, DEFAULT_COMPACTION_TIMEOUT_MS);
         try {
             await activeSession.compact();
@@ -242,7 +250,7 @@ export class PiCodingAgentSession {
         this.currentRuntimeFingerprint = null;
         this.pendingResetSummary = summary ?? null;
         this.sessionId = createSessionId();
-        saveSessionId(this.sessionId, this.mode);
+        saveSessionId(this.sessionId);
     }
     injectSummary(summary) {
         this.disposeCurrentSession();
@@ -347,7 +355,7 @@ export class PiCodingAgentSession {
         if (!transcriptPath || !fs.existsSync(transcriptPath))
             return false;
         if (this.session)
-            await awaitSdkSettled(this.session, this.mode);
+            await awaitSdkSettled(this.session, "pi-coding");
         // Prepare changes on a private copy. Keep an unchanged, idle SDK instance
         // reusable; if maintenance changes anything, release it BEFORE the atomic
         // commit. Only a new SessionManager may load the repaired transcript.
@@ -395,7 +403,7 @@ export class PiCodingAgentSession {
     getMcpServers() {
         return {
             ...this.runtimeSurface.externalMcpServers,
-            ...(this.runtimeSurface.labRuntimePolicy ? {} : this.dynamicMcpServers),
+            ...(this.runtimeSurface.workerRuntimePolicy ? {} : this.dynamicMcpServers),
         };
     }
     getMcpFingerprint() {
@@ -411,13 +419,11 @@ export class PiCodingAgentSession {
             baseUrl: model.baseUrl,
             reasoning: model.reasoning,
             contextWindow: model.contextWindow,
-            dualSession: this.getDualSessionEnabled(),
         });
     }
     prepareMcpConfig() {
-        // pi-mcp-adapter reads this process-global env var to locate mcp-adapter.json.
-        // Running multiple Pi Coding profiles in one process is intentionally unsupported.
-        process.env.PI_CODING_AGENT_DIR = this.agentDir;
+        // Each SDK session receives its own explicit configuration path. Changing
+        // a process environment variable would race other foreground conversations.
         writePiMcpConfig({
             agentDir: this.agentDir,
             servers: this.getMcpServers(),
@@ -473,7 +479,7 @@ export class PiCodingAgentSession {
         const mcpFingerprint = this.prepareMcpConfig();
         const { model, apiKey } = resolvePiModel(this.config, { useCase: "main_coding" });
         const runtimeFingerprint = this.getRuntimeFingerprint(model);
-        const tierEffort = getLlmEffort(this.config, this.mode === "coding" ? "main_coding" : "main_general");
+        const tierEffort = getLlmEffort(this.config, "main_general");
         const thinkingLevel = resolveThinkingLevel(model, effort ?? tierEffort);
         if (this.session &&
             this.currentSystemPrompt === renderedSystemPrompt &&
@@ -525,20 +531,20 @@ export class PiCodingAgentSession {
             }),
         });
         const sessionManager = this.getSessionManager();
-        const mcpAdapter = await loadMcpAdapter();
+        const mcpAdapter = await loadMcpAdapter(path.join(this.agentDir, "mcp-adapter.json"));
         const profileDir = getConfigDir();
-        const labPolicy = this.runtimeSurface.labRuntimePolicy?.policy;
+        const workerPolicy = this.runtimeSurface.workerRuntimePolicy?.policy;
         // Share the Claude `.claude/skills` directory so PiCodingAgentSession and
         // ClaudeAgentSession see the same set of skills (installed by
         // `installBuiltinSkills()` and the `skill` tool). Also use the profile
         // directory as the Pi resource-loader `agentDir` and enable context-file
         // loading so GPT/Pi sessions pick up profile-level CLAUDE.md/AGENTS.md in
         // the same profile cwd that Claude Code sees.
-        const skillsDir = labPolicy
+        const skillsDir = workerPolicy
             ? path.join(getConfigDir(), ".claude", "skills")
             : getSkillsDir();
-        const allowedSkills = new Set(labPolicy?.skills.allowed ?? []);
-        const additionalSkillPaths = labPolicy
+        const allowedSkills = new Set(workerPolicy?.skills.allowed ?? []);
+        const additionalSkillPaths = workerPolicy
             ? (fs.existsSync(skillsDir)
                 ? fs.readdirSync(skillsDir, { withFileTypes: true })
                     .filter((entry) => entry.isDirectory() && allowedSkills.has(entry.name))
@@ -551,26 +557,26 @@ export class PiCodingAgentSession {
             settingsManager,
             extensionFactories: [mcpAdapter, this.createVisionClawRuntimeExtension()],
             additionalSkillPaths,
-            noSkills: labPolicy !== undefined && additionalSkillPaths.length === 0,
+            // Disable automatic resource discovery while retaining the host's explicit
+            // skills and inline extensions. A nonempty skill allowlist must not reopen
+            // the SDK's global/profile/project default skill directories.
+            noSkills: workerPolicy !== undefined,
+            noExtensions: workerPolicy !== undefined,
             noPromptTemplates: true,
             noThemes: true,
-            noContextFiles: labPolicy !== undefined,
+            noContextFiles: workerPolicy !== undefined,
+            ...(workerPolicy ? { appendSystemPrompt: [] } : {}),
             systemPrompt: renderedSystemPrompt,
         });
         await loader.reload();
         const customTools = createPiCodingCustomTools({
-            dualSession: labPolicy ? false : this.getDualSessionEnabled(),
-            mode: this.mode,
             workspaceEnabled: isWorkspaceEnabled(this.config),
             geminiEnabled: isGeminiEnabled(this.config),
             tunnelEnabled: isTunnelEnabled(this.config),
             desktopEnabled: this.runtimeSurface.desktop.available,
             eventTriggerEnabled: this.runtimeSurface.credentialBacked.eventTrigger,
-            finAppEnabled: this.runtimeSurface.credentialBacked.finApp,
             referenceDataEnabled: this.runtimeSurface.credentialBacked.financialTool,
-            labRunEnabled: labPolicy
-                ? labPolicy.tools.allowed.some((name) => name === "query_market_data" || name === "lab_run_control")
-                : isLabRunThreadId(getWorkerThreadId()),
+            adapterTools: this.runtimeSurface.adapterTools,
             // A league player is identified by the credentials it actually holds, not
             // by an env var the provisioner may not have set: claimed pool hosts run
             // from a systemd unit that never gets VISIONCLAW_APPLICATION_ID, which
@@ -579,7 +585,7 @@ export class PiCodingAgentSession {
             // wins where the runtime is explicitly pinned.
             fundLeagueEnabled: isFundLeaguePlayerRuntime()
                 || configuredRuntimeApplicationId() === "ai-fund",
-            allowedToolNames: labPolicy?.tools.allowed,
+            allowedToolNames: workerPolicy?.tools.allowed,
             workspaceCwd: this.workspaceCwd,
         });
         const { session } = await createAgentSession({
@@ -599,11 +605,11 @@ export class PiCodingAgentSession {
         });
         try {
             await session.bindExtensions({});
-            if (labPolicy) {
-                const allowed = new Set(resolvePiLabToolNames(labPolicy.tools.allowed));
+            if (workerPolicy) {
+                const allowed = new Set(resolvePiWorkerToolNames(workerPolicy.tools.allowed));
                 const mcpTools = session.getAllTools()
                     .map((tool) => tool.name)
-                    .filter((name) => Object.keys(labPolicy.mcp.staticServers)
+                    .filter((name) => Object.keys(workerPolicy.mcp.staticServers)
                     .some((server) => name.startsWith(`${server}__`)));
                 session.setActiveToolsByName([...allowed, ...mcpTools]);
             }
@@ -621,7 +627,7 @@ export class PiCodingAgentSession {
         this.currentMcpFingerprint = mcpFingerprint;
         this.currentRuntimeFingerprint = runtimeFingerprint;
         this.session = session;
-        saveSessionId(this.sessionId, this.mode);
+        saveSessionId(this.sessionId);
         this.logLoadedResources(session, "create", mcpFingerprint, this.sessionOpenMode);
         return session;
     }
@@ -714,7 +720,8 @@ export class PiCodingAgentSession {
             // gpt/claude paths are known-good and stay byte-identical.
             pi.on("before_provider_request", (event) => {
                 let payload = event.payload;
-                if (this.config.pinnedModel) {
+                // Same predicate as the model id: a frozen execution keeps its admitted pin.
+                if (pinnedModelApplies(this.config)) {
                     payload = sanitizeToolSchemasInPayload(payload);
                 }
                 // Grok and DeepSeek: hoist tool-result screenshots onto a user message
@@ -726,14 +733,32 @@ export class PiCodingAgentSession {
             });
             pi.on("tool_result", (event) => guardPiToolResult(event, this.toolOutputGuard));
             pi.on("tool_call", async (event) => {
+                // Every call that reaches execution is counted, each call of a batch
+                // separately; once exhausted no further tool starts.
+                const refusal = this.runtimeSurface.executionBudget?.beforeToolCall();
+                if (refusal)
+                    return { block: true, reason: refusal };
+                // Foreman: the foreground limit of a conversational task, checked
+                // before the call starts. A refused call never ran and is not counted.
+                const foreground = this.runtimeSurface.workerRuntimePolicy
+                    ? undefined
+                    : getForegroundTaskTracker()?.admit(event.toolName);
+                if (foreground && !foreground.allowed)
+                    return { block: true, reason: foreground.reason };
+                getWorkerExecutionRuntime()?.toolBoundary(event.toolCallId, event.toolName, "started");
                 const browserToolName = getBrowserToolName(event);
                 if (!browserToolName || !this.runtimeSurface.browser.available)
                     return {};
-                await ensureBrowserForTool({
+                const gate = await ensureBrowserForTool({
                     toolName: browserToolName,
                     config: this.config,
                 });
-                return {};
+                if (gate.proceed)
+                    return {};
+                return {
+                    block: true,
+                    reason: gate.reason ?? "The selected browser did not open its debugging port.",
+                };
             });
         };
     }
@@ -768,7 +793,6 @@ export class PiCodingAgentSession {
             subtype: "init",
             session_id: session.sessionId,
             model: piModel.id,
-            mode: this.mode,
             open_mode: this.sessionOpenMode,
             mcp_servers: Object.keys(this.getMcpServers()).sort().map((name) => ({ name, status: "connected" })),
         };
@@ -787,7 +811,7 @@ export class PiCodingAgentSession {
         // started was still active — the very collision this barrier prevents.
         // Waiting for idle first guarantees the listener below only ever observes
         // events belonging to the run this call starts.
-        await awaitSdkSettled(session, this.mode);
+        await awaitSdkSettled(session, "pi-coding");
         const eventQueue = [];
         let resolveWaiting = null;
         let done = false;
@@ -795,6 +819,8 @@ export class PiCodingAgentSession {
         const isDone = () => done;
         const usage = createPiUsageAccumulator();
         let turnCount = 0;
+        const budget = this.runtimeSurface.executionBudget;
+        budget?.assertRunning();
         const compactionTrim = { needed: false };
         let compactionFailure = null;
         let unsubscribed = false;
@@ -808,8 +834,38 @@ export class PiCodingAgentSession {
                 const msg = event.message;
                 capturePiMessageUsage(usage, msg);
             }
+            if (event.type === "turn_start" && budget) {
+                // One Pi turn is one model attempt; retries start a new turn.
+                try {
+                    budget.beginModelAttempt();
+                }
+                catch {
+                    // The meter latched and asked the host to stop; abort this run now.
+                }
+                if (budget.stopReason) {
+                    session.abort().catch((err) => {
+                        logger.warn(`[pi-coding-session] Abort failed after budget exhaustion: ${err instanceof Error ? err.message : String(err)}`);
+                    });
+                }
+            }
             if (event.type === "turn_end") {
                 turnCount++;
+                if (budget) {
+                    const ended = event.message;
+                    if (ended?.role === "assistant" && (ended.stopReason === "error" || ended.stopReason === "aborted")) {
+                        budget.recordFailedModelAttempt();
+                    }
+                    if (ended?.usage)
+                        budget.recordTokens(ended.usage.input, ended.usage.output);
+                    // A tool-use turn needs one more attempt to read its results; with
+                    // none left, stop now instead of starting a turn that is refused.
+                    if (ended?.stopReason === "toolUse" && budget.stopBeforeNextAttempt()) {
+                        session.abort().catch((err) => {
+                            logger.warn(`[pi-coding-session] Abort failed after budget exhaustion: ${err instanceof Error ? err.message : String(err)}`);
+                        });
+                    }
+                }
+                // The SDK per-send cap stays as an extra guard; the task budget is authoritative.
                 if (turnCount >= MAX_TURNS) {
                     session.abort().catch((err) => {
                         logger.warn(`[pi-coding-session] Abort failed after max turns: ${err instanceof Error ? err.message : String(err)}`);
@@ -821,19 +877,6 @@ export class PiCodingAgentSession {
                     this.capturePostCompactionSnapshot(estimatePostCompactionTokens(event.result.summary));
                     compactionTrim.needed = true;
                     compactionFailure = null;
-                    // The summary replaced the transcript; re-inject the thread's tasks
-                    // and checkpoints when the run continues on it (#1996), the role of
-                    // the Claude engine's SessionStart(compact) hook. A threshold
-                    // compaction ends the run; steering then would force an extra turn,
-                    // and the next wake or session switch carries the block anyway.
-                    // TODO: an in-wake continuation send after a threshold compaction,
-                    // an overflow compaction that stops instead of retrying, or a manual
-                    // compaction goes without the block until the next wake or switch.
-                    // This block is also not reconciled first (this subscriber is
-                    // synchronous); it shows the latest commit plus unfolded results.
-                    const threadTaskBlock = event.willRetry ? buildThreadTaskContext() : null;
-                    if (threadTaskBlock)
-                        this.injectMessage([{ type: "text", text: threadTaskBlock }]);
                 }
                 else if (!event.aborted && !event.willRetry && event.errorMessage) {
                     const errorCode = extractCompactionErrorCode(event.errorMessage);
@@ -852,6 +895,11 @@ export class PiCodingAgentSession {
                 }
             }
             if (event.type === "tool_execution_end") {
+                if (budget && event.isError)
+                    budget.recordToolFailure();
+                const ended = event;
+                if (ended.toolCallId)
+                    getWorkerExecutionRuntime()?.toolBoundary(ended.toolCallId, ended.toolName ?? "unknown", "finished");
                 // TODO: detect MCP-proxied Playwright browser actions when Pi reports them through the generic mcp tool name.
                 const toolName = event.toolName;
                 if (toolName && this._screenshotInjector) {
@@ -878,6 +926,7 @@ export class PiCodingAgentSession {
                 // `_isAgentRunActive` BEFORE emitting it, so by the time we observe
                 // this the next send can no longer collide.
                 this.inputClosed = true;
+                this.barrierConfirmed = true;
                 done = true;
                 resolveWaiting?.();
             }
@@ -887,7 +936,7 @@ export class PiCodingAgentSession {
                 resolveWaiting?.();
             }
         });
-        session.sendUserMessage(contentToPiContent(effectiveContent)).catch((err) => {
+        const sendPromise = session.sendUserMessage(contentToPiContent(effectiveContent)).catch((err) => {
             runError = err;
             done = true;
             resolveWaiting?.();
@@ -904,6 +953,9 @@ export class PiCodingAgentSession {
                 await new Promise((r) => { resolveWaiting = r; });
                 resolveWaiting = null;
             }
+            // agent_settled is emitted from the SDK's finally block, before the
+            // send promise may reject. Include finalization failures in the result.
+            await sendPromise;
             unsubscribe();
             unsubscribed = true;
             if (compactionTrim.needed) {
@@ -915,7 +967,7 @@ export class PiCodingAgentSession {
                     logger.warn(`[pi-coding-session] Post-compaction JSONL trim failed: ${err instanceof Error ? err.message : String(err)}`);
                 }
             }
-            const assistantError = usage.lastAssistantError;
+            const assistantError = usage.terminalAssistantError;
             yield buildPiResultMessage({
                 error: runError ?? compactionFailure ?? (assistantError ? new Error(assistantError) : null),
                 inputTokens: usage.inputTokens,