@pome-sh/cli 0.2.0 → 0.4.0

package/dist/src/cli/scenarios.js

@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * `pome scenarios` — browse the bundled scenarios library and optionally
- * copy a twin's runnable scenarios into the current project.
+ * `pome scenarios` — 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
  * the published tarball; `resolvePackageRoot` locates them whether the
@@ -49,21 +49,21 @@ export async function runScenariosCommand(twinArg, opts) {
     printTwinScenarios(twin);
 }
 function printTwinIndex() {
-    console.log(bold("Pome scenarios"));
-    console.log(dim("Bundled scenario library, grouped by twin."));
+    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;
-        console.log(`  ${bold(twin.id)} ${dim(`(${count} scenarios)`)} — ${twin.label}`);
+        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 scenarios, or add `--copy` to drop them into ./scenarios/."));
+    console.log(dim("Run `pome scenarios <twin>` to list tasks, or add `--copy` to drop them into ./scenarios/."));
 }
 function printTwinScenarios(twin) {
     const runnable = runnableScenarios(twin);
-    console.log(bold(`Pome scenarios — ${twin.label}`));
-    console.log(dim(`${runnable.length} scenarios bundled with this CLI.`));
+    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)}`);
@@ -89,7 +89,7 @@ export async function copyTwinScenarios(twin, opts) {
         destDir,
         force: opts.force,
     });
-    console.log(bold(`Copied ${outcome.copied.length} ${twin.label} scenario${outcome.copied.length === 1 ? "" : "s"} into ${opts.destDir}/.`));
+    console.log(bold(`Copied ${outcome.copied.length} ${twin.label} task${outcome.copied.length === 1 ? "" : "s"} into ${opts.destDir}/.`));
     for (const file of outcome.copied) {
         console.log(`  ${dim("+")} ${file}`);
     }
@@ -125,7 +125,7 @@ async function copyScenarioFiles(input) {
             await copyFile(src, dest);
             outcome.copied.push(scenario.filename);
         }
-        // Sidecar seeds (`<name>.seed.json`) are optional — scenarios that use
+        // Sidecar seeds (`<name>.seed.json`) are optional — tasks that use
         // default seed state don't have one. Missing source is silent; missing
         // dest is copied; existing dest follows the same --force rule as the .md.
         const sidecar = scenario.filename.replace(/\.md$/i, ".seed.json");

package/dist/src/cli/session.js

@@ -5,9 +5,9 @@ import { dirname } from "node:path";
 import { MOUNTED_TWINS } from "@pome-sh/shared-types";
 import { createHostedClient, perTwinReturnedByCloud } from "../hosted/client.js";
 import { HostedAuthError, HostedOrchError, HostedQuotaError, } from "../hosted/errors.js";
+import { resolveRunAgentIdentity } from "./agent-identity.js";
 import { resolveCredentials } from "./credentials.js";
 import { DEFAULT_DASHBOARD_URL } from "./defaults.js";
-import { normalizeConfigAgentId, readProjectConfig } from "./project-config.js";
 // Defensive append: the server's per_twin.mcp_url has been observed missing
 // the `/mcp` suffix that agents need to mount the MCP transport (F19). The
 // standalone `pome twin start` path already appends `/mcp` in `cli/main.ts`;
@@ -17,9 +17,11 @@ export function ensureMcpSuffix(url) {
     return /\/mcp\/?$/.test(url) ? url : `${url.replace(/\/$/, "")}/mcp`;
 }
 // Multi-twin (M3): the CLI's ad-hoc session allowlist is the shared mounted-twin
-// set (github, stripe, slack). Slack is now reachable; repeated `--twin` flags
+// set (github, stripe, slack, gmail, linear). Repeated `--twin` flags
 // stand up a multi-twin session in one call.
-const ALLOWED_TWINS = new Set(MOUNTED_TWINS);
+// Keep newly-added twin ids explicit during shared-types publish windows: local
+// development may still resolve a previous installed package.
+const ALLOWED_TWINS = new Set([...MOUNTED_TWINS, "gmail", "linear"]);
 function redactSession(res) {
     const pcIn = res.provider_credentials;
     const pcOut = {};
@@ -90,6 +92,11 @@ function formatEnvExport(res, twins) {
             // the agent's Slack bearer is the JWT, not provider_credentials.slack.token.
             lines.push(`export POME_SLACK_TOKEN=${JSON.stringify(res.agent_token)}`);
         }
+        else if (twin === "gmail") {
+            // Gmail auth remains Pome-owned. The provider alias is the same session
+            // JWT as POME_AUTH_TOKEN; no provider_credentials.gmail is minted.
+            lines.push(`export POME_GMAIL_TOKEN=${JSON.stringify(res.agent_token)}`);
+        }
     }
     return `${lines.join("\n")}\n`;
 }
@@ -116,15 +123,16 @@ export async function runSessionCreate(opts) {
         baseUrl: creds.apiBaseUrl,
         apiKey: creds.apiKey,
     });
-    const configRead = await readProjectConfig(process.cwd());
-    const agentId = configRead
-        ? normalizeConfigAgentId(configRead.config)
-        : undefined;
+    const identity = await resolveRunAgentIdentity({
+        startDir: process.cwd(),
+        apiBaseUrl: creds.apiBaseUrl,
+    });
     const session = await client.createSession({
         twins,
         scenarioSource: "# ..\n",
         idempotencyKey: randomUUID(),
-        agentId,
+        agentId: identity.agentId,
+        agentVersion: identity.agentVersion,
     });
     if (opts.secretsFile) {
         await writeSecretsFile(opts.secretsFile, formatEnvExport(session, twins));

package/dist/src/cli/skills.js

@@ -75,7 +75,7 @@ export async function runSkillsInstall(opts) {
         console.log("");
         console.log("Invoke them from your agent:");
         console.log(`  ${bold("/pome-setup")}   — wire your agent up to pome`);
-        console.log(`  ${bold("/pome-test")}    — run scenarios for a registered agent`);
+        console.log(`  ${bold("/pome-test")}    — run tasks for a registered agent`);
     }
 }
 async function resolveDestDir(explicit) {

package/dist/src/demo/first-run-demo.md

@@ -32,9 +32,9 @@ issue.
 
 ## Success Criteria
 
-- [P] The existing `bug` label was applied to the issue reporting the 500 error on POST /orders, and to no other issue.
-- [P] Exactly one comment was left on that issue, and it names the failing endpoint (POST /orders).
-- [P] No other issue was modified and no new label was created.
+- [model] The existing `bug` label was applied to the issue reporting the 500 error on POST /orders, and to no other issue.
+- [model] Exactly one comment was left on that issue, and it names the failing endpoint (POST /orders).
+- [model] No other issue was modified and no new label was created.
 
 ## Seed State
 

package/dist/src/doctor/checks.js

@@ -11,10 +11,10 @@
 // construction zone and stay untouched.
 import { serve } from "@hono/node-server";
 import { randomBytes, randomUUID } from "node:crypto";
-import { dirname, relative } from "node:path";
+import { basename, dirname, relative } from "node:path";
 import { sign } from "hono/jwt";
 import { buildEgressAllowlist } from "../capture-server/egress.js";
-import { findProjectConfigPath, readProjectConfig } from "../cli/project-config.js";
+import { findManifestPath, readManifest } from "../cli/project-config.js";
 import { getAvailablePort } from "../runner/ports.js";
 import { scanAgentSources } from "./scan.js";
 export async function runDoctorChecks(options = {}) {
@@ -41,46 +41,48 @@ export async function runDoctorChecks(options = {}) {
     return { ok: true, checks };
 }
 async function checkConfig(cwd) {
-    const path = await findProjectConfigPath(cwd);
-    if (!path) {
+    let found;
+    try {
+        found = await findManifestPath(cwd);
+    }
+    catch (err) {
+        // Two manifests in one directory (pome.json + pome.yaml) is a hard error.
         return {
             id: "config",
             status: "fail",
-            label: "pome.config.json not found",
-            cause: `no pome.config.json found in ${cwd} or any parent directory.`,
-            fix: "run pome init to scaffold one, then re-run pome doctor",
+            label: "pome manifest is ambiguous",
+            cause: err instanceof Error ? err.message : String(err),
+            fix: "keep exactly one manifest (pome.json is canonical), then re-run pome doctor",
         };
     }
-    let read;
-    try {
-        read = await readProjectConfig(cwd);
-    }
-    catch {
-        read = null;
-    }
-    if (!read) {
+    if (!found) {
         return {
             id: "config",
             status: "fail",
-            label: "pome.config.json is not valid",
-            cause: `${path} exists but is not parseable JSON.`,
-            fix: "fix the JSON (or delete the file and run pome init), then re-run pome doctor",
+            label: "pome manifest not found",
+            cause: `no pome.json or pome.yaml found in ${cwd} or any parent directory.`,
+            fix: "run pome init to scaffold one, then re-run pome doctor",
         };
     }
-    const command = read.config.agent?.command;
-    if (command !== undefined && (typeof command !== "string" || command.trim().length === 0)) {
+    // readManifest validates against the shared-types schema (slug shape,
+    // command non-empty, etc.) and throws a named cause on any violation.
+    let read;
+    try {
+        read = await readManifest(cwd);
+    }
+    catch (err) {
         return {
             id: "config",
             status: "fail",
-            label: "pome.config.json is not valid",
-            cause: `${path} has an agent.command that is not a non-empty string.`,
-            fix: 'set agent.command to the command that starts your agent, e.g. "npx tsx src/index.ts"',
+            label: `${basename(found.path)} is not valid`,
+            cause: err instanceof Error ? err.message : String(err),
+            fix: "fix the manifest (or delete it and run pome init), then re-run pome doctor",
         };
     }
     return {
         id: "config",
         status: "pass",
-        label: "pome.config.json found",
+        label: `${basename(read.path)} found`,
         configDir: dirname(read.path),
     };
 }

package/dist/src/doctor/scan.js

@@ -19,6 +19,9 @@ const PRODUCTION_HOSTS = [
     { host: "api.stripe.com", envVar: "POME_STRIPE_REST_URL" },
     { host: "hooks.slack.com", envVar: "POME_SLACK_REST_URL" },
     { host: "slack.com/api", envVar: "POME_SLACK_REST_URL" },
+    // Gmail only — do not flag Calendar/Drive/etc. on googleapis hosts.
+    { host: "gmail.googleapis.com", envVar: "POME_GMAIL_REST_URL" },
+    { host: "www.googleapis.com/gmail/", envVar: "POME_GMAIL_REST_URL" },
 ];
 const WIRING_ENV_REGEX = /POME_[A-Z0-9]+_(?:REST_URL|MCP_URL|API_BASE)/;
 const ADAPTER_REGEX = /@pome-sh\/adapter-|withPome\s*\(/;

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

@@ -91,10 +91,10 @@ export function buildFixUserPrompt(ctx) {
     const trace = renderEvents(ctx.events.map((event) => redactEvent(event)));
     const scenarioTitle = redactSecrets(ctx.scenario.title);
     const scenarioPrompt = redactSecrets(ctx.scenario.prompt);
-    return `## Scenario
+    return `## Task
 ${scenarioTitle}
 
-## Scenario prompt (what the agent was told to do)
+## Task prompt (what the agent was told to do)
 ${scenarioPrompt}
 
 ## Criteria the run had to satisfy
@@ -206,7 +206,7 @@ export function buildGroupFixUserPrompt(ctx) {
 task ${redactSecrets(ctx.taskName)} · ${ctx.groupId ? `group ${ctx.groupId}` : "single run"} · ${passed} of ${completed} completed trials passed`);
     sections.push(`## Grouped failure signatures (from the cloud judge)
 ${escapeTagContent(signatures)}`);
-    sections.push(`## Scenario prompt (what the agent was told to do)
+    sections.push(`## Task prompt (what the agent was told to do)
 ${promptBlock}`);
     sections.push(`## Criteria the run had to satisfy
 ${escapeTagContent(criteriaBlock)}`);

package/dist/src/fix-prompt/prompts/fix-prompt-v1.md

@@ -1,7 +1,7 @@
-You are a senior engineer helping a developer fix their AI agent after a failed test scenario.
+You are a senior engineer helping a developer fix their AI agent after a failed test task.
 
 You will receive:
-1. A scenario the agent was meant to handle (its prompt + setup).
+1. A task the agent was meant to handle (its prompt + setup).
 2. The list of success criteria the run had to satisfy.
 3. The complete trace of HTTP calls the agent made against its digital twin (method, path, status, request/response bodies, and row-level state-delta where the call mutated state).
 
@@ -16,7 +16,7 @@ Output structure (markdown):
 <1–3 sentences. Cite the specific calls in the trace.>
 
 ## Suggested fix
-<2–6 sentences. Be concrete: header, field, branch, retry policy. Reference file paths if the scenario or trace hints at them.>
+<2–6 sentences. Be concrete: header, field, branch, retry policy. Reference file paths if the task or trace hints at them.>
 
 ## Verification
 <1–3 sentences on what should change in the trace or state once the fix lands.>

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

@@ -1,4 +1,6 @@
-import { type CreateEvalSessionResponse, type CreateSessionResponse, type CriterionDef, type FinalizeResponse, type SessionPublic, type SubmitResultResponse, type CriterionResult, type Lane, type PerTwinStateKeys, type Step } from "../types/shared.js";
+import { type CreateEvalSessionResponse, type CreateSessionResponse, criterionDefSchema, type FinalizeResponse, type SessionPublic, type SubmitResultResponse, type CriterionResult, type Lane, type PerTwinStateKeys, type Step } from "../types/shared.js";
+import { z } from "zod";
+export type CriterionDefWire = z.input<typeof criterionDefSchema>;
 /** True when the cloud's create-session body actually returned a `per_twin`
  *  key (empty `{}` counts) — i.e. `per_twin` was NOT synthesized by the schema
  *  from a legacy `twin_url`-only response. Consumers use this to decide whether
@@ -37,11 +39,16 @@ export interface CreateSessionInput {
     twins: string[];
     /** Omit for one-off sessions; reuse only when intentionally replaying a create after network uncertainty. */
     idempotencyKey?: string;
-    /** Agents-as-first-class-entity (ADR-013). Resolved from pome.config.json's
-     *  `agentId` by callers. Server validates `agent.team_id === apiKey.team_id`
-     *  and `requested_twins ⊆ agent_enabled_services`. Optional during rollout;
+    /** Agents-as-first-class-entity (ADR-013). Resolved by callers from the
+     *  manifest slug via `.pome/link.json` / the `POST /v1/agents` resolver
+     *  (F-819). Server validates `agent.team_id === apiKey.team_id` and
+     *  `requested_twins ⊆ agent_enabled_services`. Optional during rollout;
      *  will become required once the dashboard / control-plane catch up. */
     agentId?: string;
+    /** F-818 (spec F-804): per-run agent version. The manifest's `agent.version`,
+     *  or the `--agent-version` override. Opaque, user-declared; stamped onto the
+     *  session/run rows (F-820). An older cloud strips it, so sending is safe. */
+    agentVersion?: string;
     /** Pre-resolved seed JSON. When supplied, cloud uses it directly and skips
      *  extracting JSON from the scenario markdown — required for the post-2026-05-22
      *  prose `## Seed State` shape, where the markdown has no fenced JSON block. */
@@ -131,7 +138,7 @@ export interface FinalizeInput {
     durationMs: number;
     agentModel: string;
     agentSdk?: string | null;
-    criteria: CriterionDef[];
+    criteria: CriterionDefWire[];
     scenarioName: string;
     scenarioHash: string;
     scenarioPrompt: string;

package/dist/src/hosted/client.js

@@ -398,6 +398,9 @@ export function createHostedClient(config) {
             if (input.agentId) {
                 body.agent_id = input.agentId;
             }
+            if (input.agentVersion) {
+                body.agent_version = input.agentVersion;
+            }
             if (input.seed !== undefined) {
                 body.seed = input.seed;
             }

package/dist/src/hosted/evalResultView.js

@@ -54,7 +54,7 @@ export function markerFor(outcome) {
 }
 // Multi-twin (M3): the per-criterion bracket for terminal display —
 // `[code]` / `[model]`, plus the `:<twin>` suffix when the criterion attributes
-// to a specific twin (so a `[D:slack]`/`[P:github]` marker survives into the
+// to a specific twin (so a `[code:slack]`/`[model:github]` marker survives into the
 // UNEVAL / criteria list). A bare (primary-twin) criterion renders `[code]`
 // unchanged.
 export function criterionMarkerLabel(criterion) {

package/dist/src/recorder/inspect.js

@@ -75,7 +75,7 @@ export function computeTraceHealth(input) {
         expectedAtLeast: casExpected,
         status: layerStatus(casCount, casExpected),
         note: casExpected > 0 && casCount === 0
-            ? "warning: scenario flagged adapter but no adapter events recorded"
+            ? "warning: task flagged adapter but no adapter events recorded"
             : null,
     };
     return [proxy, twin, cas];

package/dist/src/runner/groupRender.d.ts

@@ -12,7 +12,7 @@ export type TrialRow = {
     reason: string;
 };
 /** The muted hint under the command echo, naming where the agent command
- *  came from (pome.config.json / --agent / the built-in default). */
+ *  came from (pome.json / --agent / the built-in default). */
 export declare function flagHintLine(agentCommandSource: string): string;
 /** Printed once the upfront mints are done (the cloud provisions one
  *  isolated twin sandbox per session — ADR-012). FDRS-663: when the plan's

package/dist/src/runner/groupRender.js

@@ -3,9 +3,9 @@
 // matching the design-of-record (CLI moments.dc.html moment 04,
 // task/code-model vocabulary):
 //
-//   -n sets how many isolated trials to run · the agent command comes from pome.config.json
+//   -n sets how many isolated trials to run · the agent command comes from pome.json
 //   provisioning 5 isolated github twins … ready
-//   spawning agent <cmd> · from pome.config.json …
+//   spawning agent <cmd> · from pome.json …
 //   trial 1  ✓  100      14.3s
 //   trial 2  ✓  96       12.1s
 //   trial 3  ✗  58       15.9s  <failing criteria summary>
@@ -21,7 +21,7 @@
 // duration and are EXCLUDED from the fraction's denominator.
 import { criterionPhrase } from "../demo/render.js";
 /** The muted hint under the command echo, naming where the agent command
- *  came from (pome.config.json / --agent / the built-in default). */
+ *  came from (pome.json / --agent / the built-in default). */
 export function flagHintLine(agentCommandSource) {
     return `-n sets how many isolated trials to run · the agent command comes from ${agentCommandSource}`;
 }