@pome-sh/cli 0.5.0 → 0.7.0

package/dist/src/cli/{scenarios.js → tasks.js}

@@ -1,9 +1,9 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * `pome scenarios` — browse the bundled task library and optionally
+ * `pome tasks` — browse the bundled task library and optionally
  * copy a twin's runnable tasks into the current project.
  *
- * Discovery only — no network. Source files live under `scenarios/` in
+ * Discovery only — no network. Source files live under `tasks/` in
  * the published tarball; `resolvePackageRoot` locates them whether the
  * CLI was started from `dist/` (published) or `src/` (dev).
  */
@@ -11,8 +11,8 @@ import { existsSync } from "node:fs";
 import { copyFile, mkdir } from "node:fs/promises";
 import { join, resolve } from "node:path";
 import { resolvePackageRoot } from "./resolve-package-root.js";
-import { SCENARIO_TWINS, findTwin, runnableScenarios, } from "./scenarios-catalog.js";
-const DEFAULT_DEST_DIR = "scenarios";
+import { TASK_TWINS, findTwin, runnableTasks, } from "./tasks-catalog.js";
+const DEFAULT_DEST_DIR = "tasks";
 function useColor() {
     return Boolean(process.stdout.isTTY && !process.env.NO_COLOR);
 }
@@ -22,10 +22,10 @@ function dim(s) {
 function bold(s) {
     return useColor() ? `\x1b[1m${s}\x1b[0m` : s;
 }
-export async function runScenariosCommand(twinArg, opts) {
+export async function runTasksCommand(twinArg, opts) {
     if (!twinArg) {
         if (opts.copy || opts.force || opts.dest) {
-            console.error("Specify a twin to copy from, e.g. `pome scenarios github --copy`.");
+            console.error("Specify a twin to copy from, e.g. `pome tasks github --copy`.");
             process.exitCode = 2;
             return;
         }
@@ -34,57 +34,57 @@ export async function runScenariosCommand(twinArg, opts) {
     }
     const twin = findTwin(twinArg);
     if (!twin) {
-        const available = SCENARIO_TWINS.map((t) => t.id).join(", ");
-        console.error(`Unknown twin "${twinArg}". Available: ${available}. Run \`pome scenarios\` for the index.`);
+        const available = TASK_TWINS.map((t) => t.id).join(", ");
+        console.error(`Unknown twin "${twinArg}". Available: ${available}. Run \`pome tasks\` for the index.`);
         process.exitCode = 2;
         return;
     }
     if (opts.copy) {
-        await copyTwinScenarios(twin, {
+        await copyTwinTasks(twin, {
             destDir: opts.dest ?? DEFAULT_DEST_DIR,
             force: Boolean(opts.force),
         });
         return;
     }
-    printTwinScenarios(twin);
+    printTwinTasks(twin);
 }
 function printTwinIndex() {
     console.log(bold("Pome tasks"));
     console.log(dim("Bundled task library, grouped by twin."));
     console.log("");
-    for (const twin of SCENARIO_TWINS) {
-        const count = runnableScenarios(twin).length;
+    for (const twin of TASK_TWINS) {
+        const count = runnableTasks(twin).length;
         console.log(`  ${bold(twin.id)} ${dim(`(${count} tasks)`)} — ${twin.label}`);
         console.log(`    ${dim(twin.description)}`);
     }
     console.log("");
-    console.log(dim("Run `pome scenarios <twin>` to list tasks, or add `--copy` to drop them into ./scenarios/."));
+    console.log(dim("Run `pome tasks <twin>` to list tasks, or add `--copy` to drop them into ./tasks/."));
 }
-function printTwinScenarios(twin) {
-    const runnable = runnableScenarios(twin);
+function printTwinTasks(twin) {
+    const runnable = runnableTasks(twin);
     console.log(bold(`Pome tasks — ${twin.label}`));
     console.log(dim(`${runnable.length} tasks bundled with this CLI.`));
     console.log("");
-    for (const scenario of runnable) {
-        console.log(`  ${bold(scenario.filename)}`);
-        console.log(`    ${dim(scenario.title)} — ${scenario.summary}`);
+    for (const task of runnable) {
+        console.log(`  ${bold(task.filename)}`);
+        console.log(`    ${dim(task.title)} — ${task.summary}`);
     }
     console.log("");
-    console.log(dim(`Copy locally: \`pome scenarios ${twin.id} --copy\` (or \`--copy --dest <dir>\`).`));
-    console.log(dim(`Run one: \`pome run scenarios/${runnable[0]?.filename ?? "01-bug-happy-path.md"}\`.`));
+    console.log(dim(`Copy locally: \`pome tasks ${twin.id} --copy\` (or \`--copy --dest <dir>\`).`));
+    console.log(dim(`Run one: \`pome run tasks/${runnable[0]?.filename ?? "01-bug-happy-path.md"}\`.`));
 }
-export async function copyTwinScenarios(twin, opts) {
+export async function copyTwinTasks(twin, opts) {
     const root = resolvePackageRoot(import.meta.url);
     if (!root) {
         console.error("Could not locate the installed pome package (package.json not found).");
         process.exitCode = 2;
         return;
     }
-    const sourceDir = join(root, "scenarios");
+    const sourceDir = join(root, "tasks");
     const destDir = resolve(process.cwd(), opts.destDir);
     await mkdir(destDir, { recursive: true });
-    const outcome = await copyScenarioFiles({
-        scenarios: runnableScenarios(twin),
+    const outcome = await copyTaskFiles({
+        tasks: runnableTasks(twin),
         sourceDir,
         destDir,
         force: opts.force,
@@ -109,26 +109,26 @@ export async function copyTwinScenarios(twin, opts) {
         console.log(dim(`Next: \`pome run ${opts.destDir}/${first}\`.`));
     }
 }
-async function copyScenarioFiles(input) {
+async function copyTaskFiles(input) {
     const outcome = { copied: [], skipped: [], missingSources: [] };
-    for (const scenario of input.scenarios) {
-        const src = join(input.sourceDir, scenario.filename);
-        const dest = join(input.destDir, scenario.filename);
+    for (const task of input.tasks) {
+        const src = join(input.sourceDir, task.filename);
+        const dest = join(input.destDir, task.filename);
         if (!existsSync(src)) {
-            outcome.missingSources.push(scenario.filename);
+            outcome.missingSources.push(task.filename);
             continue;
         }
         if (existsSync(dest) && !input.force) {
-            outcome.skipped.push(scenario.filename);
+            outcome.skipped.push(task.filename);
         }
         else {
             await copyFile(src, dest);
-            outcome.copied.push(scenario.filename);
+            outcome.copied.push(task.filename);
         }
         // Sidecar seeds (`<name>.seed.json`) are optional — tasks that use
         // default seed state don't have one. Missing source is silent; missing
         // dest is copied; existing dest follows the same --force rule as the .md.
-        const sidecar = scenario.filename.replace(/\.md$/i, ".seed.json");
+        const sidecar = task.filename.replace(/\.md$/i, ".seed.json");
         const sidecarSrc = join(input.sourceDir, sidecar);
         const sidecarDest = join(input.destDir, sidecar);
         if (!existsSync(sidecarSrc))

package/dist/src/demo/agent.js

@@ -4,12 +4,12 @@
 // calls served by pome's anonymous demo gateway (FDRS-637).
 //
 // It is spawned by `pome demo` as `pome demo-agent`, a child of the REAL
-// capture path (runScenario): the runner injects the standard POME_* twin
+// capture path (runTask): the runner injects the standard POME_* twin
 // contract plus the demo-specific vars below, and HTTP(S)_PROXY pointing at
 // the capture-server so the gateway calls land in events.jsonl as genuine
 // LlmCallEvent rows.
 //
-// Env contract (beyond runScenario's standard POME_* set):
+// Env contract (beyond runTask's standard POME_* set):
 //   POME_DEMO_LLM_URL    — {POME_API_BASE}/v1/demo/sessions/{sid}/llm
 //   POME_DEMO_TOKEN      — the trial session's demo_token (Bearer)
 //   POME_DEMO_TASK_NAME  — server-allowlisted task name ("first-run-demo")

package/dist/src/demo/runDemo.d.ts

@@ -1,4 +1,4 @@
-import { runScenario, type RunScenarioOptions } from "../runner/runScenario.js";
+import { runTask, type RunTaskOptions } from "../runner/runTask.js";
 import { type HostedClient } from "../hosted/client.js";
 import { mintDemoSessions, type DemoSession } from "./mint.js";
 import { type TrialVerdict } from "./render.js";
@@ -13,8 +13,8 @@ export interface RunDemoOptions {
     out?: (line: string) => void;
     /** Agent command override. Default re-invokes this binary: `<node> <main.js> demo-agent`. */
     agentCommand?: string;
-    captureServerCommand?: RunScenarioOptions["captureServerCommand"];
-    runScenarioFn?: typeof runScenario;
+    captureServerCommand?: RunTaskOptions["captureServerCommand"];
+    runTaskFn?: typeof runTask;
     mintFn?: typeof mintDemoSessions;
     /** Per-session upload/finalize client factory (bearer demo_token). */
     trialClientFactory?: (session: DemoSession) => DemoTrialClient;

package/dist/src/demo/runDemo.js

@@ -6,7 +6,7 @@
 //   2. Mint ALL k demo sessions upfront (POST /v1/demo/sessions, one shared
 //      grp_ id) — each 15-min TTL comfortably covers the run.
 //   3. Per trial: run the bundled agent through the REAL capture path
-//      (runScenario: in-process twin + capture-server child + POME_* env),
+//      (runTask: in-process twin + capture-server child + POME_* env),
 //      upload the captured blobs with the trial's demo_token as Bearer, then
 //      POST /finalize (explicit 60s timeout) — the terminal verdict comes
 //      from that cloud evaluation. The CLI never scores locally.
@@ -23,9 +23,9 @@
 import { readFile } from "node:fs/promises";
 import { join } from "node:path";
 import { getAvailablePort } from "../runner/ports.js";
-import { runScenario } from "../runner/runScenario.js";
+import { runTask } from "../runner/runTask.js";
 import { bootTwin } from "../twin/twinHarness.js";
-import { parseScenarioFile } from "../scenario/parseScenario.js";
+import { parseTaskFile } from "../task/parseTask.js";
 import { createHostedClient } from "../hosted/client.js";
 import { scoreFromFinalizeResponse, uploadRunBlobs, } from "../hosted/uploadAndFinalize.js";
 import { outcomeOf, scoreStatus } from "../hosted/evalResultView.js";
@@ -42,7 +42,7 @@ export async function runDemo(options) {
     const out = options.out ?? ((line) => console.error(line));
     const trials = options.trials ?? 5;
     const artifactsDir = options.artifactsDir ?? "runs";
-    const runScenarioFn = options.runScenarioFn ?? runScenario;
+    const runTaskFn = options.runTaskFn ?? runTask;
     const mintFn = options.mintFn ?? mintDemoSessions;
     const trialClientFactory = options.trialClientFactory ??
         ((session) => createHostedClient({
@@ -53,7 +53,7 @@ export async function runDemo(options) {
             finalizeTimeoutMs: DEMO_FINALIZE_TIMEOUT_MS,
         }));
     const groupId = newGroupId();
-    const scenarioPath = demoTaskPath();
+    const taskPath = demoTaskPath();
     for (const line of reassuranceBox())
         out(line);
     out("");
@@ -79,9 +79,9 @@ export async function runDemo(options) {
     }
     // Warm-up boot: validates the packaged seed and gives the design's
     // "spinning up github twin … ready" line an honest measurement. Each trial
-    // then boots its own isolated twin inside runScenario.
+    // then boots its own isolated twin inside runTask.
     if (options.skipTwinWarmup !== true) {
-        const scenario = await parseScenarioFile(scenarioPath);
+        const scenario = await parseTaskFile(taskPath);
         const warmupStart = Date.now();
         const port = await getAvailablePort();
         const harness = await bootTwin({
@@ -113,11 +113,11 @@ export async function runDemo(options) {
             verdict = await runOneTrial({
                 trialNumber,
                 session,
-                scenarioPath,
+                taskPath,
                 agentCommand,
                 artifactsDir,
                 apiBase: options.apiBase,
-                runScenarioFn,
+                runTaskFn,
                 client,
                 captureServerCommand: options.captureServerCommand,
                 out,
@@ -180,8 +180,8 @@ export async function runDemo(options) {
 }
 async function runOneTrial(input) {
     const startedAt = Date.now();
-    const result = await input.runScenarioFn({
-        scenarioPath: input.scenarioPath,
+    const result = await input.runTaskFn({
+        taskPath: input.taskPath,
         agentCommand: input.agentCommand,
         artifactsDir: input.artifactsDir,
         captureServerCommand: input.captureServerCommand,
@@ -238,9 +238,9 @@ async function runOneTrial(input) {
         agentModel: "demo-gateway",
         agentSdk: null,
         criteria: [],
-        scenarioName: DEMO_TASK_NAME,
-        scenarioHash: "",
-        scenarioPrompt: "",
+        taskName: DEMO_TASK_NAME,
+        taskHash: "",
+        taskPrompt: "",
         expectedBehavior: "",
         traceStorageKey: uploaded.eventsKey ?? undefined,
         stateInitialStorageKey: uploaded.stateInitialKey ?? undefined,

package/dist/src/doctor/checks.js

@@ -93,7 +93,7 @@ async function checkTwinReachable(_configDir) {
     const fail = (cause) => ({
         id: "twin",
         status: "fail",
-        label: "twin not reachable",
+        label: "local twin check failed",
         cause,
         fix: "npm install (twin dependencies), then re-run pome doctor — if it persists, file an issue with the error above",
     });
@@ -130,7 +130,16 @@ async function checkTwinReachable(_configDir) {
         if (!sessionHealth.ok) {
             return fail(`the twin is up but rejected the session route (${sessionHealth.status} at /s/<sid>/_pome/health) — the auth contract the agent relies on is broken.`);
         }
-        return { id: "twin", status: "pass", label: "twin reachable", detail: "github · local" };
+        // "boots locally", not "reachable": this check brings a throwaway twin up
+        // on an ephemeral port and tears it down — it does NOT prove a twin is
+        // listening now. Overstating that (the F-906 cold walk read "reachable" as
+        // "a twin is up") is exactly what this wording avoids.
+        return {
+            id: "twin",
+            status: "pass",
+            label: "twin boots locally",
+            detail: "github · health + session ok",
+        };
     }
     catch (err) {
         return fail(`could not reach the locally served twin: ${firstLine(err)}`);

package/dist/src/doctor/render.d.ts

@@ -4,5 +4,14 @@ export interface RenderDoctorReportOptions {
      *  moment-03 default; `pome install` passes moment 02's
      *  "verifying the wiring …" (FDRS-642). */
     header?: string;
+    /** F-906 — when the report passes, append a caveat that a green check means
+     *  the wiring is right, not that the examinee runs cleanly. `pome doctor`
+     *  never launches the agent; and a `pome run` preflight probe launches it
+     *  with POME_PREFLIGHT=1, which most scaffolds honour by exiting before
+     *  their real work path — so a bug on that skipped path (as in F-900)
+     *  surfaces only on a full trial run. Opt-in so only `pome doctor` shows it
+     *  — the `run`/`install` consumers print the report as a gate, not as a
+     *  self-diagnosis. */
+    passNote?: boolean;
 }
 export declare function renderDoctorReport(report: DoctorReport, options?: RenderDoctorReportOptions): string[];

package/dist/src/doctor/render.js

@@ -28,5 +28,12 @@ export function renderDoctorReport(report, options = {}) {
         lines.push("until this passes, your agent would hit production.");
         lines.push("pome will not run trials against a live API.");
     }
+    else if (options.passNote) {
+        lines.push("");
+        lines.push("note: a green check means the wiring is right, not that your agent runs");
+        lines.push("cleanly — pome doctor never launches it, and a run's preflight probe exits");
+        lines.push("it early (POME_PREFLIGHT=1), so a bug on the code path they skip surfaces");
+        lines.push("only on a full trial run.");
+    }
     return lines;
 }

package/dist/src/fix-prompt/prompt.d.ts

@@ -1,12 +1,12 @@
 import type { RecorderEvent } from "../types/shared.js";
-import type { Scenario } from "../scenario/scenarioSchema.js";
+import type { Task } from "../task/taskSchema.js";
 import type { VerdictArtifact } from "../hosted/evalResultCache.js";
 export declare const FIX_PROMPT_TEMPLATE_VERSION = "v1";
 export declare const FIX_PROMPT_SYSTEM_PROMPT: string;
 export declare function escapeTagContent(text: string): string;
 export interface FixPromptContext {
     events: RecorderEvent[];
-    scenario: Scenario;
+    scenario: Task;
 }
 export declare function buildFixUserPrompt(ctx: FixPromptContext): string;
 export interface TrialFixInput {
@@ -22,7 +22,7 @@ export interface GroupFixPromptContext {
     groupId: string | null;
     /** Parsed task file when it still resolves. Null degrades the prompt to
      *  the verdict-embedded criteria (the file may have moved since the run). */
-    scenario: Scenario | null;
+    scenario: Task | null;
     /** Completed trials of the run set (verdict.json present), run order. */
     trials: TrialFixInput[];
 }

package/dist/src/fix-prompt/prompt.js

@@ -89,13 +89,13 @@ function renderCriteria(criteria) {
 export function buildFixUserPrompt(ctx) {
     const criteria = redactSecrets(renderCriteria(ctx.scenario.criteria));
     const trace = renderEvents(ctx.events.map((event) => redactEvent(event)));
-    const scenarioTitle = redactSecrets(ctx.scenario.title);
-    const scenarioPrompt = redactSecrets(ctx.scenario.prompt);
+    const taskTitle = redactSecrets(ctx.scenario.title);
+    const taskPrompt = redactSecrets(ctx.scenario.prompt);
     return `## Task
-${scenarioTitle}
+${taskTitle}
 
 ## Task prompt (what the agent was told to do)
-${scenarioPrompt}
+${taskPrompt}
 
 ## Criteria the run had to satisfy
 ${escapeTagContent(criteria)}

package/dist/src/hosted/client.d.ts

@@ -35,7 +35,7 @@ export interface HostedClientConfig {
     authScheme?: "x-api-key" | "bearer";
 }
 export interface CreateSessionInput {
-    scenarioSource: string;
+    taskSource: string;
     twins: string[];
     /** Omit for one-off sessions; reuse only when intentionally replaying a create after network uncertainty. */
     idempotencyKey?: string;
@@ -85,8 +85,8 @@ export interface FetchOnTwinInput {
     agentToken: string;
 }
 export interface SubmitResultInput {
-    scenarioName: string;
-    scenarioHash: string;
+    taskName: string;
+    taskHash: string;
     durationMs: number;
     agentModel: string;
     satisfactionScore: number;
@@ -139,9 +139,9 @@ export interface FinalizeInput {
     agentModel: string;
     agentSdk?: string | null;
     criteria: CriterionDefWire[];
-    scenarioName: string;
-    scenarioHash: string;
-    scenarioPrompt: string;
+    taskName: string;
+    taskHash: string;
+    taskPrompt: string;
     expectedBehavior: string;
     traceStorageKey?: string;
     stateInitialStorageKey?: string;

package/dist/src/hosted/client.js

@@ -390,7 +390,7 @@ export function createHostedClient(config) {
         async createSession(input) {
             const body = {
                 twins: input.twins,
-                scenario_source: Buffer.from(input.scenarioSource, "utf8").toString("base64"),
+                scenario_source: Buffer.from(input.taskSource, "utf8").toString("base64"),
             };
             if (input.idempotencyKey) {
                 body.idempotency_key = input.idempotencyKey;
@@ -536,9 +536,9 @@ export function createHostedClient(config) {
                 agent_model: input.agentModel,
                 agent_sdk: normalizeAgentSdk(input.agentSdk),
                 criteria: input.criteria,
-                scenario_name: input.scenarioName,
-                scenario_hash: input.scenarioHash,
-                scenario_prompt: input.scenarioPrompt,
+                scenario_name: input.taskName,
+                scenario_hash: input.taskHash,
+                scenario_prompt: input.taskPrompt,
                 expected_behavior: input.expectedBehavior,
             };
             if (input.traceStorageKey !== undefined) {
@@ -571,8 +571,8 @@ export function createHostedClient(config) {
         },
         async submitResult(sessionId, input) {
             return postJson(`/v1/sessions/${encodeURIComponent(sessionId)}/result`, {
-                scenario_name: input.scenarioName,
-                scenario_hash: input.scenarioHash,
+                scenario_name: input.taskName,
+                scenario_hash: input.taskHash,
                 duration_ms: input.durationMs,
                 agent_model: input.agentModel,
                 satisfaction_score: input.satisfactionScore,

package/dist/src/hosted/errors.js

@@ -13,7 +13,7 @@
 //   - F0-5c: `pome logout && pome run` was returning exit 2 instead of 3
 //     because credential-resolution failure threw a plain Error.
 //   - F18: hosted sub-threshold runs sometimes returned exit 3 because
-//     `runScenarioHosted.ts` mapped any non-zero agent exit to exit 3,
+//     `runTaskHosted.ts` mapped any non-zero agent exit to exit 3,
 //     stealing the auth slot.
 //
 // FDRS-636 — trial groups (`pome run -n k`, k>1) map the WHOLE GROUP to one

package/dist/src/hosted/evalResultCache.d.ts

@@ -38,7 +38,7 @@ export interface RunSet {
     groupId: string | null;
     taskName: string;
     /** The scenario path recorded at run time (first trial's). */
-    scenarioPath: string;
+    taskPath: string;
     /** Trials sorted by finalized_at ascending. */
     trials: TrialVerdict[];
     latestFinalizedAt: string;

package/dist/src/hosted/evalResultCache.js

@@ -136,7 +136,7 @@ export function groupRunSets(trials) {
         sets.push({
             groupId: bucket[0].verdict.group_id,
             taskName: bucket[0].verdict.task_name,
-            scenarioPath: bucket[0].verdict.scenario_path,
+            taskPath: bucket[0].verdict.scenario_path,
             trials: bucket,
             latestFinalizedAt: last.verdict.finalized_at,
             anyFailed: bucket.some((t) => !t.verdict.passed),

package/dist/src/hosted/evalResultView.d.ts

@@ -31,7 +31,7 @@ export type Score = {
 export declare function outcomeOf(result: CriterionResult): CriterionOutcome;
 export type ScoreStatus = "pass" | "fail" | "unevaluated";
 export declare function scoreStatus(score: Score, passThreshold: number): ScoreStatus;
-export declare function scenarioPassed(score: Score, passThreshold: number): boolean;
+export declare function taskPassed(score: Score, passThreshold: number): boolean;
 export declare function markerFor(outcome: CriterionOutcome): string;
 export declare function criterionMarkerLabel(criterion: WireCriterion): string;
 export declare function twinSkipSuffix(result: CriterionResult): string;

package/dist/src/hosted/evalResultView.js

@@ -36,7 +36,7 @@ export function scoreStatus(score, passThreshold) {
         return "unevaluated";
     return score.satisfaction >= passThreshold ? "pass" : "fail";
 }
-export function scenarioPassed(score, passThreshold) {
+export function taskPassed(score, passThreshold) {
     return scoreStatus(score, passThreshold) === "pass";
 }
 // FDRS-591/611 per-criterion marker: ✓ passed, ✗ failed, - skipped, ! errored.

package/dist/src/hosted/uploadAndFinalize.d.ts

@@ -81,5 +81,5 @@ export declare function uploadRunBlobs(client: UploadClient, sessionId: string,
 export declare function scoreFromFinalizeResponse(finalized: FinalizeResponse): Score;
 /** Redact a JSONL payload line-by-line before upload. Lines that fail to
  *  parse as JSON are redacted as raw strings. Moved verbatim from
- *  runScenarioHosted.ts (FDRS-656). */
+ *  runTaskHosted.ts (FDRS-656). */
 export declare function redactJsonl(body: string): string;

package/dist/src/hosted/uploadAndFinalize.js

@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: Apache-2.0
 //
 // Shared presigned-upload + finalize-score orchestration for cloud sessions
-// (ADR-013). Extracted verbatim from runScenarioHosted.ts (FDRS-656) so
+// (ADR-013). Extracted verbatim from runTaskHosted.ts (FDRS-656) so
 // `pome eval <run-dir>` reuses the exact same best-effort PUT semantics
 // without duplicating them. Behavior-preserving for the hosted run path:
 // warning text, the 30s PUT timeout, and the null-key fallbacks all match
@@ -218,7 +218,7 @@ export function scoreFromFinalizeResponse(finalized) {
 }
 /** Redact a JSONL payload line-by-line before upload. Lines that fail to
  *  parse as JSON are redacted as raw strings. Moved verbatim from
- *  runScenarioHosted.ts (FDRS-656). */
+ *  runTaskHosted.ts (FDRS-656). */
 export function redactJsonl(body) {
     const lines = body.split("\n");
     const redacted = lines

package/dist/src/recorder/artifacts.d.ts

@@ -1,5 +1,5 @@
 import { toTwinHttpEventRow } from "@pome-sh/sdk/server";
-import type { Scenario } from "../scenario/scenarioSchema.js";
+import type { Task } from "../task/taskSchema.js";
 import type { RecorderEvent } from "@pome-sh/shared-types";
 export declare const toTwinHttpEvent: typeof toTwinHttpEventRow;
 export type RunArtifacts = {
@@ -9,7 +9,7 @@ export type RunArtifacts = {
 export type RunArtifactCoreInput = {
     artifactsDir: string;
     runId: string;
-    scenario: Scenario;
+    scenario: Task;
     startedAt: string;
     completedAt: string;
     stdout: string;

package/dist/src/recorder/inspect.d.ts

@@ -28,9 +28,9 @@ export type TraceHealthLayer = {
 };
 export type TraceHealthInput = {
     events: Event[];
-    scenarioUsesTwin?: boolean;
-    scenarioExpectsLlm?: boolean;
-    scenarioExpectsCas?: boolean;
+    taskUsesTwin?: boolean;
+    taskExpectsLlm?: boolean;
+    taskExpectsCas?: boolean;
 };
 export declare function computeTraceHealth(input: TraceHealthInput): TraceHealthLayer[];
 export declare function renderTraceHealth(layers: TraceHealthLayer[]): string[];

package/dist/src/recorder/inspect.js

@@ -32,7 +32,7 @@ export function computeTraceHealth(input) {
     // never warns). When the scenario manifest declares LLM usage we expect at
     // least one row even with zero observed; absence then surfaces as a warning
     // ("capture-server may not have started").
-    const proxyExpected = input.scenarioExpectsLlm
+    const proxyExpected = input.taskExpectsLlm
         ? 1
         : counts.LlmCallEvent > 0
             ? 1
@@ -48,7 +48,7 @@ export function computeTraceHealth(input) {
     };
     // twin layer — TwinHttpEvent rows. A scenario always declares ≥1 twin in
     // `config.twins`, so when we know the scenario we expect ≥1 row.
-    const twinExpectedBool = input.scenarioUsesTwin ?? counts.TwinHttpEvent > 0;
+    const twinExpectedBool = input.taskUsesTwin ?? counts.TwinHttpEvent > 0;
     const twinExpected = twinExpectedBool ? 1 : 0;
     const twin = {
         name: "twin",
@@ -68,7 +68,7 @@ export function computeTraceHealth(input) {
         counts.SubagentSpawnEvent +
         counts.HookEvent +
         counts.LlmTurnEvent;
-    const casExpected = input.scenarioExpectsCas ? 1 : casCount > 0 ? 1 : 0;
+    const casExpected = input.taskExpectsCas ? 1 : casCount > 0 ? 1 : 0;
     const cas = {
         name: "CAS adapter",
         count: casCount,

package/dist/src/runner/captureServerChild.js

@@ -98,7 +98,7 @@ function defaultBinArgs(eventsOut, allowHosts, egressOut) {
     }
     return [script, "capture-server", "--port", "0", "--events-out", eventsOut, ...egressArgs(allowHosts, egressOut)];
 }
-// Shared by defaultBinArgs and runScenario's test-override path so the two
+// Shared by defaultBinArgs and runTask's test-override path so the two
 // spawn shapes can't drift.
 export function egressArgs(allowHosts, egressOut) {
     const args = [];

package/dist/src/runner/{runScenario.d.ts → runTask.d.ts}

@@ -3,8 +3,8 @@ export type CaptureServerCommand = {
     execPath: string;
     prefixArgs: string[];
 };
-export type RunScenarioOptions = {
-    scenarioPath: string;
+export type RunTaskOptions = {
+    taskPath: string;
     agentCommand: string;
     artifactsDir?: string;
     captureServerCommand?: CaptureServerCommand;
@@ -13,7 +13,7 @@ export type RunScenarioOptions = {
     extraAgentEnv?: Record<string, string>;
     egressExtraHosts?: readonly string[];
 };
-export declare function runScenario(options: RunScenarioOptions): Promise<{
+export declare function runTask(options: RunTaskOptions): Promise<{
     scenario: {
         slug: string;
         title: string;

package/dist/src/runner/{runScenario.js → runTask.js}

@@ -4,7 +4,7 @@ import { randomBytes, randomUUID } from "node:crypto";
 import { mkdir, writeFile } from "node:fs/promises";
 import { join, resolve as resolvePath, sep as pathSep } from "node:path";
 import { sign } from "hono/jwt";
-import { parseScenarioFile, seedStateForTwin } from "../scenario/parseScenario.js";
+import { parseTaskFile, seedStateForTwin } from "../task/parseTask.js";
 import { bootTwin } from "../twin/twinHarness.js";
 import { createRecorder } from "../recorder/recorder.js";
 import { redactSecrets } from "../recorder/redaction.js";
@@ -13,7 +13,7 @@ import { runAgentCommand } from "./agentRunner.js";
 import { egressArgs, spawnCaptureServerChild } from "./captureServerChild.js";
 import { buildEgressAllowlist, readBlockedEgress } from "../capture-server/egress.js";
 import { mergeAdapterSignalsIntoEvents } from "./mergeAdapterSignals.js";
-import { writeRunNoScore } from "./runScenarioCore.js";
+import { writeRunNoScore } from "./runTaskCore.js";
 // Promisify a Hono/node server `close()` so teardown can await in-flight
 // handlers draining before the twin's DB handle is released.
 function closeServer(server) {
@@ -21,8 +21,8 @@ function closeServer(server) {
         server.close((err) => (err ? reject(err) : resolvePromise()));
     });
 }
-export async function runScenario(options) {
-    const scenario = await parseScenarioFile(options.scenarioPath);
+export async function runTask(options) {
+    const scenario = await parseTaskFile(options.taskPath);
     const runId = `run_${randomUUID()}`;
     const artifactsDir = options.artifactsDir ?? "runs";
     const startedAt = new Date().toISOString();