lastlight 0.1.15 → 0.2.0

package/dist/engine/agent-executor.js

@@ -0,0 +1,1140 @@
+import { resolve, basename, join, relative } from "path";
+import { appendFileSync, cpSync, existsSync, mkdirSync, readFileSync, rmSync, symlinkSync, writeFileSync } from "fs";
+import { randomUUID } from "crypto";
+import { createTaskSandbox, setupTaskWorktree, SANDBOX_IMAGE_QA, } from "../sandbox/index.js";
+import { refreshGitAuth } from "./git-auth.js";
+import { AGENTIC_PROFILE_FOR, GITHUB_PERMISSION_PROFILES, loadAgentContext, } from "./profiles.js";
+import { AgenticShim, truncateForLog, safeStringify } from "./event-shim.js";
+import { projectSlugForCwd } from "../session-log.js";
+import { BuildAssetStore } from "../state/build-assets.js";
+import { ALLOW_ALL_SENTINEL, DEFAULT_ALLOWLIST, mergeAllowlist } from "../sandbox/egress-allowlist.js";
+import { getDockerSandboxOtelEnv, getOtelEnvForSandbox, recordError, recordExecutionMetrics, safeSpanAttributes, withSpan } from "../telemetry/index.js";
+import { recordPiEvent } from "../telemetry/pi-events.js";
+const DEFAULT_MODEL = "anthropic/claude-sonnet-4-6";
+const DOCKER_WORKSPACE_DIR = "/home/agent/workspace";
+// Directory holding one skill bundle per phase. Deliberately NOT named
+// `.agents/skills` (pi's auto-discovery path): we map each phase's bundle
+// explicitly via --skill / skillPaths so two phases sharing a workspace —
+// sequential today, parallel via worktrees tomorrow — can never clobber each
+// other's catalogue. The agent keeps cwd = the repo (no `cd` preamble on
+// every command); the bundle is staged at the workspace ROOT — a sibling of
+// the repo, never in its git tree — and reached by an absolute path. On
+// docker/none that root is genuinely outside the repo. gondolin mounts only
+// cwd, so there the bundle is staged under the repo and added to the
+// checkout's local `.git/info/exclude` (never committed; see `excludeFromGit`).
+const SKILL_BUNDLE_ROOT = ".lastlight-skills";
+const THINKING_LEVELS = new Set([
+    "off", "minimal", "low", "medium", "high", "xhigh",
+]);
+/**
+ * Stage this phase's declared skills into a per-phase bundle directory at
+ * `<workspaceRoot>/.lastlight-skills/<phaseKey>/<basename>/` and return the
+ * staged skill dirs, so the caller can point pi at them explicitly (via
+ * `--skill` for docker or `skillPaths` for the in-process backends). Each
+ * skill is a directory containing SKILL.md plus any `scripts/` / `references/`
+ * / `assets/` — the whole tree comes along.
+ *
+ * Keyed by phase so concurrent phases in one workspace never touch each
+ * other's bundle: only the phase's own `<phaseKey>` subtree is cleared, so a
+ * clean slate per phase doesn't disturb a sibling phase mid-run.
+ *
+ * `mode` controls how each skill lands:
+ *   - "symlink": one symlink per skill → host source. gondolin/none, where pi
+ *     reads skill files host-side / through the cwd mount. Zero-copy.
+ *   - "copy": recursive copy. docker, where the agent's tools run inside the
+ *     container and host symlink targets wouldn't resolve; the copy lands
+ *     under the bind-mounted workspace.
+ *
+ * Returns `undefined` when the phase declares no skills (after clearing its
+ * bundle), so a phase with no `skills:` gets no catalogue at all.
+ */
+export function stageSkillBundle(workspaceRoot, phaseKey, skillPaths, mode) {
+    const bundleDir = join(workspaceRoot, SKILL_BUNDLE_ROOT, phaseKey);
+    if (existsSync(bundleDir)) {
+        rmSync(bundleDir, { recursive: true, force: true });
+    }
+    if (!skillPaths?.length)
+        return undefined;
+    mkdirSync(bundleDir, { recursive: true });
+    const staged = [];
+    for (const hostPath of skillPaths) {
+        const dest = join(bundleDir, basename(hostPath));
+        if (mode === "symlink") {
+            symlinkSync(hostPath, dest, "dir");
+        }
+        else {
+            cpSync(hostPath, dest, { recursive: true, dereference: true });
+        }
+        staged.push(dest);
+    }
+    return staged;
+}
+/**
+ * Sanitized per-phase key for the skill bundle directory. Phase name first
+ * (unique even for loop iterations like `reviewer_fix_1`), then workflow name,
+ * then a constant fallback — so the bundle is always isolated per phase.
+ */
+function skillBundleKey(config) {
+    const raw = config.telemetry?.phaseName || config.telemetry?.workflowName || "phase";
+    return raw.replace(/[^A-Za-z0-9_-]/g, "_") || "phase";
+}
+/**
+ * Add `entry` to a checkout's local `.git/info/exclude` (idempotent) so the
+ * agent's own `git add`/`commit` can never pick it up. This file lives inside
+ * `.git/` — it is never tracked, committed, or pushed, and it leaves the repo's
+ * real `.gitignore` untouched; the exclusion applies only to this ephemeral
+ * sandbox checkout. Used for the gondolin backend, where the skill bundle must
+ * be staged under cwd (the only mounted dir) rather than as an out-of-repo
+ * sibling. No-op when `repoDir` isn't a git checkout (e.g. the workspace root).
+ */
+export function excludeFromGit(repoDir, entry) {
+    const gitDir = join(repoDir, ".git");
+    if (!existsSync(gitDir))
+        return; // not a checkout — nothing to exclude
+    const infoDir = join(gitDir, "info");
+    const excludeFile = join(infoDir, "exclude");
+    const line = `/${entry}/`;
+    let current = "";
+    try {
+        current = readFileSync(excludeFile, "utf8");
+    }
+    catch { /* may not exist yet */ }
+    if (current.split(/\r?\n/).includes(line))
+        return;
+    mkdirSync(infoDir, { recursive: true });
+    appendFileSync(excludeFile, `${current.length && !current.endsWith("\n") ? "\n" : ""}${line}\n`);
+}
+// ── Server-mode build assets ────────────────────────────────────────
+//
+// In "server" mode the per-phase handoff docs live in the Last Light store
+// rather than being committed into the target repo. The seam is symmetric to
+// the skill bundle but bidirectional: stage the run's stored docs into the
+// repo's `.lastlight/<issueKey>/` before the agent runs (so a later phase /
+// resumed run sees prior context), and harvest whatever the phase wrote back
+// to the store afterwards. The directory is the SAME relative path as repo
+// mode (`{{issueDir}}`), so prompts are unchanged except for gating their
+// `git add .lastlight/ && commit` off — the dir is git-excluded here too as a
+// backstop against the agent's `git add -A` feature commit sweeping it in.
+const ARTIFACT_DIR_ROOT = ".lastlight";
+/**
+ * Resolve the server-mode artifact context for a run, or undefined when not in
+ * server mode (the default — the whole seam is then skipped and behaviour is
+ * byte-for-byte repo mode). `hostRepoDir` is the host-visible repo checkout
+ * (for docker that's the bind-mounted workspace path, not the in-container one).
+ */
+function serverArtifacts(config, hostRepoDir) {
+    if (config.buildAssets !== "server" || !config.buildAssetsDir || !config.buildAssetsKey) {
+        return undefined;
+    }
+    const store = new BuildAssetStore(config.buildAssetsDir);
+    const ref = config.buildAssetsKey;
+    return { store, ref, dir: join(hostRepoDir, ARTIFACT_DIR_ROOT, ref.issueKey), repoDir: hostRepoDir };
+}
+/** Stage stored docs into the workspace + exclude the dir from git (server mode). */
+function stageArtifactsIn(art) {
+    if (!art)
+        return;
+    try {
+        // Stage the run's stored docs into `<repoDir>/.lastlight/<issueKey>/`. This
+        // is safe for every workflow shape here: pre-cloned workflows (build, pr-*)
+        // already have the checkout at `repoDir`, and non-pre-cloned ones (explore)
+        // clone the repo into a *subdir* — so `repoDir` is the workspace root and a
+        // `.lastlight/` there never collides with the agent's clone. (A future
+        // workflow that `git clone … .` into cwd would be the one exception.)
+        art.store.stageInto(art.ref, art.dir);
+        // Backstop the prompt-level commit gate: when this is a real checkout, keep
+        // the docs out of the agent's `git add -A` feature commit. No-op at the
+        // workspace root (no `.git`), where the docs sit outside the repo subtree
+        // and are never in the repo's git tree anyway.
+        excludeFromGit(art.repoDir, ARTIFACT_DIR_ROOT);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        console.warn(`[executor] Could not stage build assets: ${msg}`);
+    }
+}
+/** Persist docs the phase wrote back to the store (server mode). */
+function harvestArtifactsOut(art) {
+    if (!art)
+        return;
+    try {
+        art.store.harvestFrom(art.ref, art.dir);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        console.warn(`[executor] Could not harvest build assets: ${msg}`);
+    }
+}
+/**
+ * Execute one workflow-phase agent task via agentic-pi.
+ *
+ * Sandbox backend (decided per call):
+ *  1. gondolin — agentic-pi's QEMU micro-VM. cwd is the host worktree,
+ *     mounted at /workspace inside the VM.
+ *  2. docker — the legacy container path (`src/sandbox/docker.ts`). The
+ *     entrypoint inside the container execs `agentic-pi run --sandbox none`.
+ *  3. none — agentic-pi in-process with cwd = the host worktree.
+ */
+export async function executeAgent(prompt, config, opts) {
+    const taskId = opts?.taskId || `task-${randomUUID().slice(0, 8)}`;
+    const stateDir = config.stateDir || resolve("data");
+    const backend = config.sandbox ?? "gondolin";
+    // Mint a scoped GitHub App token. Same flow as the legacy executor —
+    // defense in depth so a downstream tool gating regression can't burn
+    // more access than the profile allowed.
+    //
+    // GITHUB_APP_* env vars are forwarded to agentic-pi *only* when the access
+    // profile opts into App PEM access via `allowMcpAppAuth`. That is currently
+    // never set (see gitSandboxAccessForWorkflow): the github extension can't
+    // read the PEM in the sandbox and skips rather than falling back, so we keep
+    // the App key out entirely and every run uses just the minted `GITHUB_TOKEN`
+    // below — which also stops agents minting elevated tokens themselves. The
+    // branch is retained so per-profile App auth can be re-enabled if the
+    // sandbox-side PEM is ever materialized.
+    const ghEnv = {};
+    let mintedToken;
+    const access = opts?.githubAccess;
+    const allowAppAuth = access?.allowMcpAppAuth === true;
+    if (process.env.GITHUB_APP_ID && allowAppAuth) {
+        ghEnv.GITHUB_APP_ID = process.env.GITHUB_APP_ID;
+        if (process.env.GITHUB_APP_INSTALLATION_ID) {
+            ghEnv.GITHUB_APP_INSTALLATION_ID = process.env.GITHUB_APP_INSTALLATION_ID;
+        }
+        if (process.env.GITHUB_APP_PRIVATE_KEY_PATH) {
+            ghEnv.GITHUB_APP_PRIVATE_KEY_PATH = process.env.GITHUB_APP_PRIVATE_KEY_PATH;
+        }
+    }
+    else {
+        // Suppress for in-process runs that inherit our env. Empty strings
+        // override the inherited value via `applyEnv()`.
+        ghEnv.GITHUB_APP_ID = "";
+        ghEnv.GITHUB_APP_INSTALLATION_ID = "";
+        ghEnv.GITHUB_APP_PRIVATE_KEY_PATH = "";
+    }
+    if (process.env.GITHUB_APP_ID && access) {
+        try {
+            const permissions = GITHUB_PERMISSION_PROFILES[access.profile];
+            const repositories = access.repo ? [access.repo] : undefined;
+            console.log(`[executor] Minting git token: profile=${access.profile}, ` +
+                `repo=${access.repo || "(unscoped)"}, permissions=${permissions ? Object.keys(permissions).join(",") : "all"}`);
+            const { token } = await refreshGitAuth({
+                appId: process.env.GITHUB_APP_ID,
+                privateKeyPath: process.env.GITHUB_APP_PRIVATE_KEY_PATH || "",
+                installationId: process.env.GITHUB_APP_INSTALLATION_ID || "",
+                permissions,
+                repositories,
+            });
+            mintedToken = token;
+            ghEnv.GITHUB_TOKEN = token;
+            ghEnv.GIT_TOKEN = token;
+        }
+        catch (err) {
+            const msg = err instanceof Error ? err.message : String(err);
+            console.warn(`[executor] Could not mint git token (repo=${access.repo || "none"}, ` +
+                `profile=${access.profile}): ${msg}`);
+        }
+    }
+    // Provider API keys.
+    if (process.env.OPENAI_API_KEY)
+        ghEnv.OPENAI_API_KEY = process.env.OPENAI_API_KEY;
+    if (process.env.ANTHROPIC_API_KEY)
+        ghEnv.ANTHROPIC_API_KEY = process.env.ANTHROPIC_API_KEY;
+    if (process.env.OPENROUTER_API_KEY)
+        ghEnv.OPENROUTER_API_KEY = process.env.OPENROUTER_API_KEY;
+    if (process.env.FIREWORKS_API_KEY)
+        ghEnv.FIREWORKS_API_KEY = process.env.FIREWORKS_API_KEY;
+    // Web-search provider keys. Forwarded only when the workflow opted into
+    // web search (scoped to explore today; see webSearchEnabledForWorkflow
+    // in workflows/runner.ts). agentic-pi auto-detects the provider from
+    // whichever key is present (Tavily > Exa > Brave by default).
+    if (config.webSearch) {
+        if (process.env.TAVILY_API_KEY)
+            ghEnv.TAVILY_API_KEY = process.env.TAVILY_API_KEY;
+        if (process.env.BRAVE_SEARCH_API_KEY)
+            ghEnv.BRAVE_SEARCH_API_KEY = process.env.BRAVE_SEARCH_API_KEY;
+        if (process.env.EXA_API_KEY)
+            ghEnv.EXA_API_KEY = process.env.EXA_API_KEY;
+    }
+    // OTEL config for the agent runtime itself. On docker the agent runs
+    // inside the container, so it reads this (the container env) and is
+    // pointed at the in-network collector — never the real backend or its
+    // auth headers. On gondolin/none the agent runs in the harness process
+    // and inherits the harness SDK; forwarding the host's OTEL_* here just
+    // re-affirms that config for any child processes.
+    if (config.otel?.enabled && config.otel.forwardToSandbox) {
+        Object.assign(ghEnv, backend === "docker" ? getDockerSandboxOtelEnv() : getOtelEnvForSandbox());
+    }
+    const prePopulate = access?.prePopulateBranch && mintedToken
+        ? {
+            owner: access.owner,
+            repo: access.repo,
+            branch: access.prePopulateBranch,
+            token: mintedToken,
+            runId: access.runId,
+            shallow: access.shallow,
+        }
+        : undefined;
+    const spanAttrs = safeSpanAttributes({
+        "agent.runtime": "agentic-pi",
+        "sandbox.backend": backend,
+        "task.id": taskId,
+        repo: access?.repo,
+        "github.profile": access?.profile,
+        model: config.model || DEFAULT_MODEL,
+        variant: config.variant,
+        "web_search.enabled": config.webSearch === true,
+        unrestricted_egress: config.unrestrictedEgress === true,
+        "workflow.name": config.telemetry?.workflowName,
+        "phase.name": config.telemetry?.phaseName,
+    });
+    if (backend === "docker") {
+        return withSpan("lastlight.agent.execute", spanAttrs, () => executeDocker(prompt, config, {
+            taskId,
+            stateDir,
+            env: ghEnv,
+            prePopulate,
+            access,
+            onSessionId: opts?.onSessionId,
+        }));
+    }
+    const workDir = setupTaskWorktree({
+        taskId,
+        stateDir,
+        sandboxDir: config.sandboxDir,
+        prePopulate,
+    });
+    // Drop AGENTS.md into the workspace — agentic-pi reads it as the
+    // system context (same convention as the previous opencode setup).
+    try {
+        const md = loadAgentContext(config.agentContextDir);
+        if (md)
+            writeFileSync(join(workDir, "AGENTS.md"), md);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        console.warn(`[executor] Could not write AGENTS.md: ${msg}`);
+    }
+    return withSpan("lastlight.agent.execute", spanAttrs, () => executeInProcess(prompt, config, {
+        backend,
+        taskId,
+        workDir,
+        stateDir,
+        env: ghEnv,
+        access,
+        onSessionId: opts?.onSessionId,
+    }));
+}
+// ── In-process path (gondolin / none) ───────────────────────────────
+async function executeInProcess(prompt, config, ctx) {
+    const startTime = Date.now();
+    console.log(`  [executor] Running in-process (task: ${ctx.taskId}, sandbox: ${ctx.backend})`);
+    const model = config.model || DEFAULT_MODEL;
+    const thinking = coerceThinking(config.variant);
+    const profile = ctx.access ? AGENTIC_PROFILE_FOR[ctx.access.profile] : undefined;
+    const sessionsDir = resolveSessionsDir(config);
+    // When the harness pre-cloned the target repo, cwd is the checkout
+    // (`<workDir>/<repo>/`) so the agent's commands run inside the repo with no
+    // `cd` preamble; otherwise cwd is the workspace root and the agent clones in.
+    const agentCwd = ctx.access?.prePopulateBranch
+        ? join(ctx.workDir, ctx.access.repo)
+        : ctx.workDir;
+    // Stage this phase's skills into its own bundle and point pi at the staged
+    // dirs explicitly via `skillPaths` below. The bundle lives at the workspace
+    // root — a sibling of the repo, outside its git tree — for `none` (the host
+    // FS is fully visible in-process). gondolin mounts only cwd, so its bundle
+    // is staged under the repo and added to the checkout's local
+    // `.git/info/exclude` so the agent can't commit it.
+    const gondolin = ctx.backend === "gondolin";
+    const skillRoot = gondolin ? agentCwd : ctx.workDir;
+    let stagedSkillDirs;
+    try {
+        stagedSkillDirs = stageSkillBundle(skillRoot, skillBundleKey(config), config.skillPaths, "symlink");
+        if (stagedSkillDirs && gondolin)
+            excludeFromGit(agentCwd, SKILL_BUNDLE_ROOT);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        console.warn(`[executor] Could not stage skills: ${msg}`);
+    }
+    // Server-mode build assets: agentCwd is the host-visible repo checkout for
+    // the in-process backends (gondolin's cwd mount / none's host FS), so stage
+    // and harvest operate on it directly.
+    const artifacts = serverArtifacts(config, agentCwd);
+    stageArtifactsIn(artifacts);
+    const shim = new AgenticShim({
+        homeDir: sessionsDir,
+        projectSlug: projectSlugForCwd(agentCwd),
+        model,
+        initialPrompt: prompt,
+    });
+    // agentic-pi reads its own env (provider keys, App PEM, etc.) from
+    // process.env. Splice in our scoped values for the duration of the call,
+    // then restore.
+    const restore = applyEnv(ctx.env);
+    // Env forwarded INTO the sandbox so the agent's `bash` calls see it.
+    // agentic-pi auto-injects GITHUB_TOKEN/GH_TOKEN when --profile is set;
+    // git identity is set here so `git commit` works without extra setup.
+    //
+    // Notes per backend:
+    //   - gondolin: the VM's user inherits HOME from the agentic-pi process
+    //     (so the host's HOME — e.g. /Users/clifton — leaks in). We force
+    //     HOME=/root so `git config --global` and `gh auth status` write
+    //     to a real path inside the VM.
+    //   - none: agent runs on the host; do NOT override HOME (would mess
+    //     with the harness user's real config).
+    //   - docker (handled separately): the container's `agent` user has
+    //     its own HOME=/home/agent baked in by the image; no override.
+    const baseSandboxEnv = {
+        GIT_AUTHOR_NAME: "last-light[bot]",
+        GIT_AUTHOR_EMAIL: "last-light[bot]@users.noreply.github.com",
+        GIT_COMMITTER_NAME: "last-light[bot]",
+        GIT_COMMITTER_EMAIL: "last-light[bot]@users.noreply.github.com",
+        // /workspace is a bind mount from the host (host UID) into a VM running
+        // as a different UID; git refuses to operate without an explicit
+        // safe-directory. Setting it via GIT_CONFIG_* avoids needing HOME and
+        // a writeable ~/.gitconfig.
+        GIT_CONFIG_COUNT: "1",
+        GIT_CONFIG_KEY_0: "safe.directory",
+        GIT_CONFIG_VALUE_0: "*",
+    };
+    const otelSandboxEnv = config.otel?.enabled && config.otel.forwardToSandbox ? getOtelEnvForSandbox() : {};
+    const sandboxEnv = ctx.backend === "gondolin"
+        ? { ...otelSandboxEnv, ...baseSandboxEnv, HOME: "/root", USER: "root", LOGNAME: "root" }
+        : { ...otelSandboxEnv, ...baseSandboxEnv };
+    let notifiedSessionId = false;
+    let result;
+    const acc = new RunResultAccumulator();
+    try {
+        // HTTP egress allowlist. lastlight owns the policy (rather than relying
+        // on agentic-pi's bundled default) so a single source — `egress-allowlist.ts`
+        // — covers both backends. `unrestrictedEgress` opts a phase out via the
+        // `"*"` sentinel; gondolin (post the upstream allow-all patch) treats it
+        // as "allow every host".
+        const extraHosts = config.otel?.enabled && config.otel.forwardToSandbox ? config.otel.collectorHosts : [];
+        const allowedHttpHosts = config.unrestrictedEgress
+            ? [ALLOW_ALL_SENTINEL]
+            : mergeAllowlist(DEFAULT_ALLOWLIST, extraHosts);
+        // Loaded lazily: agentic-pi transitively imports pi-coding-agent, whose
+        // bundled undici writes a v8 Agent onto `Symbol.for('undici.globalDispatcher.1')`
+        // the moment its `lib/global.js` evaluates. Node's built-in fetch reads
+        // from the same symbol, so eager-loading here would poison every fetch
+        // in the harness — breaking arctic's OAuth code exchange (strict
+        // content-length validation). Dynamic import keeps the harness on
+        // Node's clean dispatcher unless an in-process sandbox actually runs.
+        const { run: agenticRun } = await import("agentic-pi");
+        result = await agenticRun({
+            model,
+            prompt,
+            thinking,
+            profile,
+            sandbox: ctx.backend === "gondolin" ? "gondolin" : "none",
+            sandboxEnv,
+            cwd: agentCwd,
+            noSession: true,
+            // Explicit per-phase skill bundle (staged above). pi loads these
+            // additively; nothing is written into the repo for the agent to commit.
+            skillPaths: stagedSkillDirs,
+            allowedHttpHosts,
+            // Explicit boolean — without it, agentic-pi auto-enables web search
+            // when any provider key is in process.env. We forwarded those keys
+            // above only for opted-in workflows, but `process.env` on the harness
+            // host carries them regardless, so we must opt-out explicitly.
+            webSearch: config.webSearch === true,
+            webSearchProvider: config.webSearchProvider,
+            onEvent: (record) => {
+                acc.feed(record);
+                shim.feed(record);
+                recordPiEvent(record, {
+                    includeContent: config.otel?.includeContent === true,
+                    surface: "agent",
+                    workflowName: config.telemetry?.workflowName,
+                    phaseName: config.telemetry?.phaseName,
+                    model,
+                });
+                if (!notifiedSessionId && ctx.onSessionId && record.type === "session" && typeof record.id === "string") {
+                    notifiedSessionId = true;
+                    ctx.onSessionId(record.id);
+                }
+            },
+            onWarn: (msg) => console.warn(`[agentic] ${msg}`),
+        });
+    }
+    catch (err) {
+        restore();
+        // Harvest even on failure so a partial plan/summary still reaches the store.
+        harvestArtifactsOut(artifacts);
+        const msg = err instanceof Error ? err.message : String(err);
+        const durationMs = Date.now() - startTime;
+        recordError("agent", err, { "sandbox.backend": ctx.backend, model, success: false, stop_reason: "error_executor", "workflow.name": config.telemetry?.workflowName, "phase.name": config.telemetry?.phaseName });
+        recordExecutionMetrics("agent", { "sandbox.backend": ctx.backend, model, success: false, stop_reason: "error_executor", durationMs });
+        const fallbackId = `exec-${basename(ctx.taskId)}`;
+        const synthesizedId = await shim
+            .finalizeWithFallback(emptyResult("error_executor", durationMs), fallbackId, msg)
+            .catch(() => null);
+        return {
+            success: false,
+            output: "",
+            turns: 0,
+            error: msg,
+            durationMs,
+            sessionId: synthesizedId ?? undefined,
+            stopReason: "error_executor",
+        };
+    }
+    restore();
+    harvestArtifactsOut(artifacts);
+    // agentic-pi's in-process `stats` is the same compaction-blind
+    // `usage_snapshot`. Prefer our per-message accumulation when it carries
+    // token data (see RunResultAccumulator.bestStats).
+    const better = acc.bestStats();
+    if (better && (better.tokens?.total ?? 0) > 0)
+        result.stats = better;
+    const finalResult = await finalizeFromRunResult(result, prompt, shim, startTime, acc.extensions(), acc.skills(), acc.toolError(), acc.endedOnToolCall());
+    recordExecutionMetrics("agent", {
+        "sandbox.backend": ctx.backend,
+        model,
+        success: finalResult.success,
+        stop_reason: finalResult.stopReason,
+        durationMs: finalResult.durationMs,
+        costUsd: finalResult.costUsd,
+        inputTokens: finalResult.inputTokens,
+        outputTokens: finalResult.outputTokens,
+        "workflow.name": config.telemetry?.workflowName,
+        "phase.name": config.telemetry?.phaseName,
+    });
+    return finalResult;
+}
+// ── Docker path ─────────────────────────────────────────────────────
+async function executeDocker(prompt, config, ctx) {
+    // HTTP egress routing — docker analog of the gondolin allowlist.
+    // Sandbox is wired to coredns-strict by default (allowlist enforced via
+    // DNS sinkhole + nginx ssl_preread); a phase that set
+    // `unrestricted_egress: true` points at coredns-open instead.
+    // The IPs here match the static assignments in docker-compose.yml
+    // (see src/sandbox/egress-firewall-config.ts for the constants).
+    const dnsIp = config.unrestrictedEgress
+        ? (process.env.LASTLIGHT_DNS_OPEN || "172.30.0.11")
+        : (process.env.LASTLIGHT_DNS_STRICT || "172.30.0.10");
+    // A phase can opt into the browser-QA image (Playwright + Chromium baked in)
+    // with `sandbox_image: qa`. The runner only schedules such a phase when the
+    // image is actually present (see `qaImageAvailable`), so by here it's safe to
+    // request; pass undefined otherwise to use the lean default image.
+    const imageName = config.sandboxImage === "qa" ? SANDBOX_IMAGE_QA : undefined;
+    const sbx = await createTaskSandbox({
+        taskId: ctx.taskId,
+        stateDir: ctx.stateDir,
+        sandboxDir: config.sandboxDir,
+        env: ctx.env,
+        prePopulate: ctx.prePopulate,
+        dnsIp,
+        imageName,
+    });
+    if (!sbx) {
+        throw new Error("LASTLIGHT_SANDBOX=docker but no docker sandbox was available. " +
+            "Install Docker and build the sandbox image, or set LASTLIGHT_SANDBOX=gondolin / none.");
+    }
+    try {
+        const md = loadAgentContext(config.agentContextDir);
+        if (md)
+            writeFileSync(join(sbx.workDir, "AGENTS.md"), md);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        console.warn(`[executor] Could not write docker AGENTS.md: ${msg}`);
+    }
+    const startTime = Date.now();
+    console.log(`  [executor] Running in docker sandbox (task: ${ctx.taskId})`);
+    const model = config.model || DEFAULT_MODEL;
+    const thinking = coerceThinking(config.variant);
+    const profile = ctx.access ? AGENTIC_PROFILE_FOR[ctx.access.profile] : undefined;
+    const sessionsDir = resolveSessionsDir(config);
+    // When the harness pre-cloned the repo, cwd is the checkout
+    // (`<workspace>/<repo>/`) so the agent runs inside the repo with no `cd`
+    // preamble; otherwise it's the workspace root.
+    const agentCwd = ctx.prePopulate
+        ? `${DOCKER_WORKSPACE_DIR}/${ctx.prePopulate.repo}`
+        : DOCKER_WORKSPACE_DIR;
+    // Stage this phase's skills into its own bundle at the workspace ROOT — a
+    // sibling of any repo subdir, never inside its git tree. docker bind-mounts
+    // the WHOLE workspace, so the agent reaches the bundle by an absolute
+    // `--skill` path even though cwd is the repo. Copy (not symlink) because the
+    // agent's tools run inside the container and host symlink targets don't
+    // resolve there. Map the host dests to their in-container paths for `--skill`.
+    let skillDirsInContainer = [];
+    try {
+        const staged = stageSkillBundle(sbx.workDir, skillBundleKey(config), config.skillPaths, "copy");
+        if (staged) {
+            skillDirsInContainer = staged.map((d) => `${DOCKER_WORKSPACE_DIR}/${relative(sbx.workDir, d)}`);
+        }
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        console.warn(`[executor] Could not stage skills in docker workspace: ${msg}`);
+    }
+    // Server-mode build assets. The host-visible repo checkout is the
+    // bind-mounted workspace path (`sbx.workDir[/<repo>]`), NOT the in-container
+    // `agentCwd`. Stage now; harvest before each `sbx.cleanup()` below (cleanup
+    // removes the workspace, taking the docs with it).
+    const hostRepoDir = ctx.prePopulate ? join(sbx.workDir, ctx.prePopulate.repo) : sbx.workDir;
+    const artifacts = serverArtifacts(config, hostRepoDir);
+    stageArtifactsIn(artifacts);
+    // The dashboard reads from <sessionsDir>/projects/<slug>/. Use the same
+    // resolved cwd for the slug so live tails land in the right project dir.
+    const shim = new AgenticShim({
+        homeDir: sessionsDir,
+        projectSlug: projectSlugForCwd(agentCwd),
+        model,
+        initialPrompt: prompt,
+    });
+    const acc = new RunResultAccumulator();
+    let notifiedSessionId = false;
+    // Identity forwarded into the sandboxed agentic-pi run inside the container.
+    // The container's `agent` user already has HOME=/home/agent set up, so we
+    // do NOT override HOME here — only git identity + safe.directory (for the
+    // host-UID bind mount). GITHUB_TOKEN/GH_TOKEN are auto-injected by
+    // agentic-pi when --profile is set.
+    const sandboxEnv = {
+        // Inner-run env for the agent's child shells. Points at the in-network
+        // collector (IP-only, no secret headers) so any OTLP a script emits is
+        // tunnelled the same way the agent's own telemetry is.
+        ...(config.otel?.enabled && config.otel.forwardToSandbox ? getDockerSandboxOtelEnv() : {}),
+        GIT_AUTHOR_NAME: "last-light[bot]",
+        GIT_AUTHOR_EMAIL: "last-light[bot]@users.noreply.github.com",
+        GIT_COMMITTER_NAME: "last-light[bot]",
+        GIT_COMMITTER_EMAIL: "last-light[bot]@users.noreply.github.com",
+        GIT_CONFIG_COUNT: "1",
+        GIT_CONFIG_KEY_0: "safe.directory",
+        GIT_CONFIG_VALUE_0: "*",
+    };
+    try {
+        await sbx.sandbox.runAgent(ctx.taskId, prompt, {
+            model,
+            thinking,
+            profile,
+            sandboxEnv,
+            agentCwd,
+            skillDirs: skillDirsInContainer,
+            webSearch: config.webSearch === true,
+            webSearchProvider: config.webSearchProvider,
+            onLine: (line) => {
+                if (!line.startsWith("{"))
+                    return;
+                let record;
+                try {
+                    record = JSON.parse(line);
+                }
+                catch {
+                    return;
+                }
+                acc.feed(record);
+                shim.feed(record);
+                recordPiEvent(record, {
+                    includeContent: config.otel?.includeContent === true,
+                    surface: "agent",
+                    workflowName: config.telemetry?.workflowName,
+                    phaseName: config.telemetry?.phaseName,
+                    model,
+                });
+                if (!notifiedSessionId && ctx.onSessionId && record.type === "session" && typeof record.id === "string") {
+                    notifiedSessionId = true;
+                    ctx.onSessionId(record.id);
+                }
+            },
+        });
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        const durationMs = Date.now() - startTime;
+        recordError("agent", err, { "sandbox.backend": "docker", model, success: false, stop_reason: "error_sandbox", "workflow.name": config.telemetry?.workflowName, "phase.name": config.telemetry?.phaseName });
+        recordExecutionMetrics("agent", { "sandbox.backend": "docker", model, success: false, stop_reason: "error_sandbox", durationMs });
+        const fallbackId = `exec-${basename(ctx.taskId)}`;
+        const synthesizedId = await shim
+            .finalizeWithFallback(emptyResult("error_sandbox", durationMs), fallbackId, msg)
+            .catch(() => null);
+        harvestArtifactsOut(artifacts);
+        await sbx.cleanup();
+        return {
+            success: false,
+            output: "",
+            turns: 0,
+            error: msg,
+            durationMs,
+            sessionId: synthesizedId ?? undefined,
+            stopReason: "error_sandbox",
+        };
+    }
+    harvestArtifactsOut(artifacts);
+    await sbx.cleanup();
+    const finalResult = await finalizeFromRunResult(acc.build(0), prompt, shim, startTime, acc.extensions(), acc.skills(), acc.toolError(), acc.endedOnToolCall());
+    recordExecutionMetrics("agent", {
+        "sandbox.backend": "docker",
+        model,
+        success: finalResult.success,
+        stop_reason: finalResult.stopReason,
+        durationMs: finalResult.durationMs,
+        costUsd: finalResult.costUsd,
+        inputTokens: finalResult.inputTokens,
+        outputTokens: finalResult.outputTokens,
+        "workflow.name": config.telemetry?.workflowName,
+        "phase.name": config.telemetry?.phaseName,
+    });
+    return finalResult;
+}
+/**
+ * Build a RunResult-shaped tally from the JSONL event stream emitted by
+ * `agentic-pi run` inside the docker sandbox. Mirrors what agentic-pi's
+ * own `run()` function does in-process — minimum viable subset of
+ * fields the executor cares about.
+ *
+ * Usage accounting accumulates each assistant `message_end`'s `usage`
+ * rather than trusting the terminal `usage_snapshot`. pi's snapshot is
+ * derived from `getSessionStats()`, which recomputes from the *current*
+ * in-memory message window — auto-compaction replaces those messages with
+ * a summary, so the snapshot reports zero tokens/cost/turns the moment a
+ * phase compacts. The per-message events fire at finalization (before any
+ * compaction rebuild), so summing them is compaction-proof. `bestStats()`
+ * prefers the accumulation and falls back to the snapshot only when no
+ * per-message usage was observed.
+ */
+export class RunResultAccumulator {
+    sessionId;
+    finalText = "";
+    agentEnded = false;
+    toolErrors = false;
+    lastToolError;
+    // True iff the last assistant turn ended with a tool call — i.e. the agent
+    // asked for a tool and the loop terminated before it could respond to the
+    // result. That's a truncated run (pi hit its internal step cap mid-task),
+    // not a finished one. See `finalizeFromRunResult`'s truncation guard.
+    lastAssistantHadToolCall = false;
+    fatalError;
+    snapshotStats;
+    messages = [];
+    // Per-message usage accumulation (the compaction-proof source).
+    assistantMessages = 0;
+    userMessages = 0;
+    toolCalls = 0;
+    toolResults = 0;
+    msgInput = 0;
+    msgOutput = 0;
+    msgCacheRead = 0;
+    msgCacheWrite = 0;
+    msgCost = 0;
+    // Extension status events (file-search / github / web-search), keyed by
+    // extension name. Emitted once each at run start; we keep the raw payload
+    // (minus type/extension) so build() can map them onto the RunResult.
+    ext = {};
+    // Skill-loading status. agentic-pi emits a single (gated) `skills_status`
+    // event at run start; we keep the raw payload (minus type) so skills()
+    // can normalize it onto the RunResult, mirroring `ext` above for tools.
+    skillsRaw;
+    feed(r) {
+        switch (r.type) {
+            case "session":
+                if (typeof r.id === "string")
+                    this.sessionId = r.id;
+                break;
+            case "extension_status": {
+                if (typeof r.extension === "string") {
+                    const { type: _t, extension: _e, ...rest } = r;
+                    this.ext[r.extension] = rest;
+                }
+                break;
+            }
+            case "skills_status": {
+                const { type: _t, ...rest } = r;
+                this.skillsRaw = rest;
+                break;
+            }
+            case "message_end": {
+                const m = r.message;
+                if (m?.role === "assistant" && Array.isArray(m.content)) {
+                    const text = m.content
+                        .filter((c) => c.type === "text" && typeof c.text === "string")
+                        .map((c) => c.text)
+                        .join("");
+                    if (text)
+                        this.finalText = text;
+                    this.assistantMessages += 1;
+                    const toolCallsInTurn = m.content.filter((c) => c.type === "toolCall").length;
+                    this.toolCalls += toolCallsInTurn;
+                    // Track whether the *latest* assistant turn requested a tool. If the
+                    // run ends here (no synthesis turn follows the tool result), the
+                    // agent was cut off mid-task.
+                    this.lastAssistantHadToolCall = toolCallsInTurn > 0;
+                    this.accumulateUsage(m.usage);
+                }
+                else if (m?.role === "user") {
+                    this.userMessages += 1;
+                }
+                break;
+            }
+            case "tool_execution_end":
+                this.toolResults += 1;
+                if (r.isError === true) {
+                    this.toolErrors = true;
+                    // Keep the actual failure text (not just a boolean) so a run that
+                    // ends in `error_tool` can report which tool failed and why —
+                    // e.g. a provider `insufficient_quota` surfaced through an MCP
+                    // call, or a bash command's stderr. Last error wins (it's the
+                    // one that ended the run). truncate: tool output can be huge.
+                    const raw = r.error ?? r.result ?? r.output;
+                    const message = truncateForLog(typeof raw === "string" ? raw : safeStringify(raw), 4096);
+                    const tool = typeof r.tool === "string"
+                        ? r.tool
+                        : typeof r.toolName === "string"
+                            ? r.toolName
+                            : undefined;
+                    if (message)
+                        this.lastToolError = { tool, message };
+                }
+                break;
+            case "agent_end":
+                this.agentEnded = true;
+                if (Array.isArray(r.messages))
+                    this.messages = r.messages;
+                break;
+            case "usage_snapshot":
+                this.snapshotStats = r.stats;
+                break;
+            case "fatal_error":
+                this.fatalError = r.error;
+                break;
+        }
+    }
+    accumulateUsage(usage) {
+        if (!usage || typeof usage !== "object")
+            return;
+        const num = (v) => typeof v === "number" && Number.isFinite(v) ? v : 0;
+        this.msgInput += num(usage.input);
+        this.msgOutput += num(usage.output);
+        this.msgCacheRead += num(usage.cacheRead);
+        this.msgCacheWrite += num(usage.cacheWrite);
+        const cost = usage.cost;
+        if (cost && typeof cost === "object")
+            this.msgCost += num(cost.total);
+    }
+    /** Stats summed from per-message usage, or undefined if none was seen. */
+    accumulatedStats() {
+        const total = this.msgInput + this.msgOutput + this.msgCacheRead + this.msgCacheWrite;
+        if (this.assistantMessages === 0 && total === 0)
+            return undefined;
+        return {
+            userMessages: this.userMessages,
+            assistantMessages: this.assistantMessages,
+            toolCalls: this.toolCalls,
+            toolResults: this.toolResults,
+            tokens: {
+                input: this.msgInput,
+                output: this.msgOutput,
+                cacheRead: this.msgCacheRead,
+                cacheWrite: this.msgCacheWrite,
+                total,
+            },
+            cost: this.msgCost,
+        };
+    }
+    /**
+     * Prefer the per-message accumulation (compaction-proof) over pi's
+     * terminal `usage_snapshot`. Fall back to the snapshot only when the
+     * accumulation carries no token data — e.g. a provider that doesn't
+     * report per-message usage — so a non-compacted snapshot still wins.
+     */
+    bestStats() {
+        const acc = this.accumulatedStats();
+        if (acc && acc.tokens.total > 0)
+            return acc;
+        return this.snapshotStats ?? acc;
+    }
+    build(exitCode) {
+        return {
+            exitCode: exitCode,
+            ok: exitCode === 0 && !this.fatalError,
+            agentEnded: this.agentEnded,
+            toolErrors: this.toolErrors,
+            fatalError: this.fatalError,
+            sessionId: this.sessionId,
+            finalText: this.finalText,
+            messages: this.messages,
+            stats: this.bestStats(),
+            records: [],
+            warnings: [],
+        };
+    }
+    /**
+     * Normalized extension status captured from `extension_status` events
+     * (file-search / github / web-search), or undefined if none reported.
+     * Decoupled from agentic-pi's `RunResult` type, which lags the runtime —
+     * the docker sandbox's agentic-pi emits file-search even when the harness's
+     * pinned `RunResult` type doesn't yet declare it.
+     */
+    extensions() {
+        const out = {};
+        for (const [name, v] of Object.entries(this.ext)) {
+            if (v && typeof v.status === "string") {
+                out[name] = {
+                    status: v.status,
+                    ...(typeof v.mode === "string" ? { mode: v.mode } : {}),
+                    ...(typeof v.provider === "string" ? { provider: v.provider } : {}),
+                    ...(typeof v.toolCount === "number" ? { toolCount: v.toolCount } : {}),
+                    ...(typeof v.reason === "string" ? { reason: v.reason } : {}),
+                };
+            }
+        }
+        return Object.keys(out).length > 0 ? out : undefined;
+    }
+    /**
+     * Normalized skill-loading status from the gated `skills_status` event, or
+     * undefined if none was reported (agentic-pi suppresses it on a run that
+     * configured no skills and discovered none). The skill-loading counterpart
+     * to {@link extensions}.
+     */
+    skills() {
+        const s = this.skillsRaw;
+        if (!s || typeof s.status !== "string")
+            return undefined;
+        const skills = Array.isArray(s.skills)
+            ? s.skills
+                .filter((x) => !!x && typeof x === "object")
+                .map((x) => ({
+                name: typeof x.name === "string" ? x.name : "",
+                source: typeof x.source === "string" ? x.source : "",
+                modelInvocable: x.modelInvocable === true,
+            }))
+            : [];
+        return {
+            status: s.status,
+            discovered: typeof s.discovered === "number" ? s.discovered : skills.length,
+            skills,
+            mappedPaths: Array.isArray(s.mappedPaths)
+                ? s.mappedPaths.filter((p) => typeof p === "string")
+                : [],
+            noSkills: s.noSkills === true,
+        };
+    }
+    /**
+     * The last tool result that came back with `isError: true`, including the
+     * failure text — or undefined if no tool errored. This is what turns a
+     * bare `error_tool` stop reason into a human-readable cause.
+     */
+    toolError() {
+        return this.lastToolError;
+    }
+    /**
+     * True iff the run's final assistant turn ended on a tool call — meaning
+     * the agent intended to keep going but the loop stopped before it could
+     * respond to the tool result and write its answer. The signal a run was
+     * truncated (step-limit) rather than genuinely finished.
+     */
+    endedOnToolCall() {
+        return this.lastAssistantHadToolCall;
+    }
+}
+// ── Helpers ─────────────────────────────────────────────────────────
+/** Provider account/billing/auth failures pi may surface as error text. */
+const ACCOUNT_ERROR_MARKERS = [
+    "credit balance",
+    "insufficient_quota",
+    "insufficient quota",
+    "rate limit",
+    "unauthorized",
+    "invalid_api_key",
+];
+/**
+ * Detect a provider account error (out of credit, quota, rate-limited, bad key)
+ * that pi may have surfaced as plain text rather than a hard failure.
+ *
+ * Critically, on a **successful** run we scan ONLY the genuine provider-error
+ * channel (`agentErrorMessage`, which `extractAgentError` already turns into a
+ * non-success stopReason when set — so it's empty here). The agent's own output
+ * (`finalText`) and tool results (`fatalError`, `toolError`) are NOT scanned on
+ * success: a legitimate `verify`/`qa-test` report or a `curl` probing a 401
+ * endpoint routinely contains "unauthorized" / "rate limit" as part of the task
+ * itself, and folding that in would wrongly fail a genuinely successful run
+ * (and drop its report). Only on a failed run do we fold those in to label why.
+ */
+export function detectAccountError(opts) {
+    const combined = [
+        opts.success ? undefined : opts.fatalErrorMessage,
+        opts.agentErrorMessage,
+        opts.success ? undefined : opts.finalText,
+        opts.success ? undefined : opts.toolErrorText,
+    ]
+        .filter((s) => typeof s === "string" && s.length > 0)
+        .join("\n")
+        .toLowerCase();
+    return ACCOUNT_ERROR_MARKERS.some((m) => combined.includes(m));
+}
+function finalizeFromRunResult(result, _prompt, shim, startTime, extensions, skills, toolError, endedOnToolCall = false) {
+    const durationMs = Date.now() - startTime;
+    const stats = result.stats;
+    // pi-agent-core swallows provider API errors (insufficient_quota, rate
+    // limit, auth) inside `handleRunFailure`: it synthesizes an assistant
+    // message with stopReason "error" + errorMessage and emits a clean
+    // agent_end. Without this check the run would map to "success" and the
+    // workflow would silently advance with no output. See message scan below.
+    const agentError = extractAgentError(result);
+    let stopReason = agentError?.stopReason ?? mapStopReason(result);
+    // Truncation guard. A run that *would* map to "success" but whose final
+    // assistant turn ended on a tool call was cut off mid-task: the agent
+    // asked for a tool and the loop stopped before it could read the result
+    // and synthesize an answer (pi hit its internal step cap — agentic-pi
+    // v0.2.7 exposes no maxSteps knob to lift it). In that state `finalText`
+    // is the agent's "let me just check X" preamble, NOT an answer. Reclassify
+    // as a failure so the workflow fires `on_failure` instead of delivering
+    // the fragment as if it were the result.
+    const truncated = stopReason === "success" && !agentError && endedOnToolCall;
+    if (truncated)
+        stopReason = "error_truncated";
+    const success = stopReason === "success";
+    const truncationMessage = "Agent stopped mid-task before producing a final answer (hit the agent " +
+        "step limit). No answer was delivered.";
+    // A bare `error_tool` stop reason is useless on its own. Surface the
+    // failing tool's actual error text so the executions row and dashboard
+    // show *why* the run died (e.g. "Tool `bash` failed: insufficient_quota").
+    const toolErrorText = toolError
+        ? toolError.tool
+            ? `Tool \`${toolError.tool}\` failed: ${toolError.message}`
+            : toolError.message
+        : undefined;
+    const inputTokens = stats?.tokens.input ?? 0;
+    const outputTokens = stats?.tokens.output ?? 0;
+    const cacheRead = stats?.tokens.cacheRead ?? 0;
+    const cacheWrite = stats?.tokens.cacheWrite ?? 0;
+    const costUsd = stats?.cost ?? 0;
+    const turns = stats?.assistantMessages ?? 0;
+    const costStr = costUsd > 0 ? `, $${costUsd.toFixed(4)}` : "";
+    console.log(`  [executor] Result: ${stopReason} (${turns} turns, ${Math.round(durationMs / 1000)}s${costStr})` +
+        `${result.sessionId ? ` [session ${result.sessionId}]` : ""}`);
+    shim.finalize({
+        finalText: result.finalText,
+        turns,
+        costUsd,
+        inputTokens,
+        outputTokens,
+        cacheReadInputTokens: cacheRead,
+        cacheCreationInputTokens: cacheWrite,
+        stopReason,
+        durationMs,
+        apiErrorMessage: agentError?.errorMessage ??
+            (success ? undefined : (toolErrorText ?? (truncated ? truncationMessage : undefined))),
+    });
+    void shim.flush();
+    const accountError = detectAccountError({
+        success,
+        fatalErrorMessage: result.fatalError?.message,
+        agentErrorMessage: agentError?.errorMessage,
+        finalText: result.finalText,
+        toolErrorText,
+    });
+    const errorText = result.fatalError?.message ||
+        agentError?.errorMessage ||
+        toolErrorText ||
+        (truncated ? truncationMessage : result.finalText) ||
+        stopReason;
+    if (!success || accountError) {
+        if (accountError)
+            console.error(`  [executor] Account error: ${errorText}`);
+        else
+            console.error(`  [executor] Run failed (${stopReason}): ${errorText}`);
+    }
+    return {
+        success: success && !accountError,
+        output: result.finalText,
+        turns,
+        error: success && !accountError ? undefined : errorText,
+        durationMs,
+        sessionId: result.sessionId,
+        costUsd: costUsd > 0 ? costUsd : undefined,
+        inputTokens: inputTokens || undefined,
+        outputTokens: outputTokens || undefined,
+        cacheReadInputTokens: cacheRead || undefined,
+        cacheCreationInputTokens: cacheWrite || undefined,
+        stopReason,
+        extensions,
+        skills,
+    };
+}
+function mapStopReason(result) {
+    if (result.fatalError)
+        return "error_fatal";
+    if (result.toolErrors && result.finalText.length === 0)
+        return "error_tool";
+    if (!result.ok)
+        return `error_exit_${result.exitCode}`;
+    if (result.agentEnded || result.finalText.length > 0)
+        return "success";
+    return "unknown";
+}
+/**
+ * Scan agent_end's messages for a failed assistant turn. pi-agent-core's
+ * `handleRunFailure` catches provider errors, synthesizes an assistant
+ * message with stopReason "error"/"aborted" + errorMessage, and emits a
+ * normal agent_end — so the only signal that the run actually failed
+ * lives on the last assistant message.
+ */
+function extractAgentError(result) {
+    if (!Array.isArray(result.messages))
+        return undefined;
+    for (let i = result.messages.length - 1; i >= 0; i--) {
+        const m = result.messages[i];
+        if (m?.role !== "assistant")
+            continue;
+        if (m.stopReason === "error" || m.stopReason === "aborted") {
+            return {
+                stopReason: m.stopReason === "aborted" ? "error_aborted" : "error_agent",
+                errorMessage: m.errorMessage || m.stopReason,
+            };
+        }
+        return undefined;
+    }
+    return undefined;
+}
+function coerceThinking(raw) {
+    if (!raw)
+        return undefined;
+    const v = raw.trim().toLowerCase();
+    if (!THINKING_LEVELS.has(v)) {
+        console.warn(`[executor] Ignoring unknown thinking level "${raw}" — must be one of: ${[...THINKING_LEVELS].join(", ")}`);
+        return undefined;
+    }
+    return v;
+}
+function resolveSessionsDir(config) {
+    if (config.sessionsDir)
+        return resolve(config.sessionsDir);
+    const stateDir = config.stateDir || resolve("data");
+    return process.env.LASTLIGHT_SESSIONS_DIR
+        ? resolve(process.env.LASTLIGHT_SESSIONS_DIR)
+        : resolve(stateDir, "agent-sessions");
+}
+function emptyResult(stopReason, durationMs) {
+    return {
+        finalText: "",
+        turns: 0,
+        costUsd: 0,
+        inputTokens: 0,
+        outputTokens: 0,
+        cacheReadInputTokens: 0,
+        cacheCreationInputTokens: 0,
+        stopReason,
+        durationMs,
+    };
+}
+/** Splice values into process.env for the duration of a sync block. */
+function applyEnv(env) {
+    const saved = {};
+    for (const [k, v] of Object.entries(env)) {
+        saved[k] = process.env[k];
+        process.env[k] = v;
+    }
+    return () => {
+        for (const [k, v] of Object.entries(saved)) {
+            if (v === undefined)
+                delete process.env[k];
+            else
+                process.env[k] = v;
+        }
+    };
+}
+//# sourceMappingURL=agent-executor.js.map