@genroc/eval-node 0.0.0-edge.ea911c5 → 0.0.0-edge.eadd43a

package/README.md

@@ -90,6 +90,23 @@ queue worker rather than the HTTP sidecar it used to be:
   that genroc had to be able to reach. A worker only needs outbound access, so it can live
   anywhere — behind NAT, in another trust zone.
 
+The cost of that inversion is that genroc cannot reach a running worker, so anything it needs
+to say has to ride the renewal the worker already makes. **Renewing is therefore mandatory,
+not an optimisation** — a worker that stops renewing is indistinguishable from one that died.
+The response answers per token:
+
+| list | what it means | what this worker does |
+|---|---|---|
+| `renewed` | still yours | carry on |
+| `lost` | already someone else's | stop; do **not** release — that would bump the new holder's claim |
+| `cancelled` | still yours, nobody wants it | abort the script and release the claim |
+
+`cancelled` is how an operator's `genctl cancel` reaches work already running: the evaluation
+is aborted through an `AbortSignal`, its realm is terminated, and no outcome is submitted —
+the process is stopping, so there is nothing left to answer. `renew_before_ms` on the claim
+says how long the worker may wait before renewing again, so the interval is a value it reads
+rather than one it guesses.
+
 ## The task input
 
 The `input` of the external task IS the evaluation request:
@@ -169,7 +186,7 @@ no retry policy to write.
 
 - id: script_failed
   switch:
-    - case: 'error.data.name == "LimitExceeded"'
+    - case: 'last_error.data.name == "LimitExceeded"'
       raise: { code: limit_exceeded, message: "the script rejected the amount" }
     - raise: { code: script_failed, message: "the script failed" }
 ```

package/dist/eval.js

@@ -20,13 +20,31 @@ class RealmFault extends Error {
         this.name = "RealmFault";
     }
 }
-export async function evaluate(req) {
+/** Thrown when `signal` aborts: the work was cancelled server-side, so there is no outcome to
+ *  report and worker.ts must release rather than answer. Distinct from RealmFault because the
+ *  runner did not fault -- nothing is wrong, the answer is simply no longer wanted. */
+export class Cancelled extends Error {
+    constructor() {
+        super("cancelled");
+        this.name = "Cancelled";
+    }
+}
+export async function evaluate(req, signal) {
     const budget = typeof req.timeout_ms === "number" ? req.timeout_ms : DEFAULT_TIMEOUT_MS;
     const worker = new Worker(REALM_URL);
     let timer;
+    let onAbort;
     try {
         return await new Promise((resolve, reject) => {
             timer = setTimeout(() => resolve(timedOut(budget)), budget);
+            // The abort is wired to the same promise as the budget, so both settle through the one
+            // finally below -- which is what guarantees the thread is gone before either returns.
+            if (signal) {
+                if (signal.aborted)
+                    reject(new Cancelled());
+                onAbort = () => reject(new Cancelled());
+                signal.addEventListener("abort", onAbort, { once: true });
+            }
             worker.once("message", (reply) => resolve(reply));
             // A script may end its own realm (`process.exit()`), which is not a throw and would
             // otherwise present as a hang until the budget expired. Our own terminate() raises
@@ -38,6 +56,8 @@ export async function evaluate(req) {
     }
     finally {
         clearTimeout(timer);
+        if (signal && onAbort)
+            signal.removeEventListener("abort", onAbort);
         // Awaited, and the whole point: on the timeout path a thread is still burning a core, and
         // resolving before it is gone would report an evaluation the machine is still running.
         await worker.terminate();

package/dist/import.js

@@ -7,9 +7,10 @@
 // code; "build" typechecks and bundles. A separate types hook would mean a second `tsc`
 // over the same project.
 import { spawn } from "node:child_process";
+import { existsSync } from "node:fs";
 import { access, mkdir, writeFile } from "node:fs/promises";
 import { builtinModules } from "node:module";
-import { dirname, join, relative } from "node:path";
+import { dirname, join, relative, resolve } from "node:path";
 import { fileURLToPath } from "node:url";
 import commonjs from "@rollup/plugin-commonjs";
 import json from "@rollup/plugin-json";
@@ -112,21 +113,14 @@ function deref(s, defs) {
     }
     return cur;
 }
-/** The manifest's `input` is the type of the whole ACTION input — for /eval that is
- *  `{code, input, timeout_ms, …}`, and only its `input` field is bound as the script's
- *  parameter. genroc cannot know that; this file owns the evaluator's wire contract, so
- *  the navigation belongs here rather than in genctl. */
-function scriptInput(site, defs) {
-    const action = deref(site.input, defs);
-    const bound = action?.properties?.input;
-    return bound ?? site.input;
-}
 /** Emits one named type per reachable $def rather than inlining: a task output may
  *  reference itself (specs/recursive-type-inference.md) and inlining would not terminate. */
-function declarations(site, defs) {
+function declarations(at) {
+    const { site, where } = at;
+    const defs = where.$defs ?? {};
     const used = new Set();
-    const input = tsType(scriptInput(site, defs), used);
-    const output = tsType(site.output, used);
+    const input = tsType(site.types?.Input, used);
+    const output = tsType(site.types?.Output, used);
     const emitted = [];
     const done = new Set();
     while (true) {
@@ -139,7 +133,7 @@ function declarations(site, defs) {
     }
     return [
         "// Generated by genroc. Do not edit - regenerate with `genctl types`.",
-        `// ${site.process}${site.task ? ` / ${site.task}` : ""}  (${site.pointer})`,
+        `// ${where.name}  (${address(site.pointer)})`,
         "",
         ...emitted,
         emitted.length ? "" : "",
@@ -167,7 +161,7 @@ async function nearestTsconfig(from, root) {
             return null;
     }
 }
-async function typecheck(root, sites) {
+async function typecheck(sites) {
     // NOT `.genroc`: that is the project config FILE, and a directory of the same name cannot
     // coexist with it. The suffix is what keeps the scratch area out of its way.
     const dir = join(root, ".genroc-cache");
@@ -175,13 +169,13 @@ async function typecheck(root, sites) {
     // One tsc per distinct base config: `extends` takes a single base, so merging two would
     // check each script under the other author's options.
     const groups = new Map();
-    for (const site of sites) {
-        const base = (await nearestTsconfig(dirname(site.path), root)) ?? "";
+    for (const at of sites) {
+        const base = (await nearestTsconfig(dirname(at.file), root)) ?? "";
         const group = groups.get(base);
         if (group)
-            group.push(site);
+            group.push(at);
         else
-            groups.set(base, [site]);
+            groups.set(base, [at]);
     }
     let n = 0;
     for (const [base, group] of groups) {
@@ -203,8 +197,8 @@ async function typecheck(root, sites) {
                 ...(base ? {} : { types: [] }),
             },
             files: group.flatMap((s) => [
-                relative(dir, s.path),
-                relative(dir, typesPathFor(s.path)),
+                relative(dir, s.file),
+                relative(dir, typesPathFor(s.file)),
             ]),
             // `files` overrides the base's, but a base `include` survives beside it and would
             // drag the author's whole tree in, to be checked under the worker lib.
@@ -287,15 +281,16 @@ function tsResolve(configPath) {
 /** Bundles to a self-contained ES module, which is what the evaluator imports: the default
  *  export it calls is the author's own, so nothing wraps or rewrites the code between the two.
  *  Bundling is entirely the importer's job, so a definition version pins its code forever. */
-async function bundle(site, root) {
+async function bundle(at) {
+    const site = at.site;
     // Builtins are EXTERNALISED as imports the realm resolves natively. Anything else
     // unresolved is a REFUSAL, not an external: rollup's default is to leave it as an import
     // of a module that will not be there, which bundles clean and fails at runtime.
     const built = await rollup({
-        input: site.path,
+        input: at.file,
         external: (id) => BUILTIN.has(id),
         plugins: [
-            tsResolve(await nearestTsconfig(dirname(site.path), root)),
+            tsResolve(await nearestTsconfig(dirname(at.file), root)),
             nodeResolve({ extensions: [".ts", ".tsx", ".mjs", ".js", ".json"] }),
             commonjs(),
             // A `.json` import is a data file inlined at build time, which the previous bundler
@@ -305,10 +300,10 @@ async function bundle(site, root) {
         ],
         onwarn(warning) {
             if (warning.code === "UNRESOLVED_IMPORT") {
-                die(`${site.path}: cannot resolve ${warning.exporter ?? "an import"} — is it installed?`);
+                die(`${at.file}: cannot resolve ${warning.exporter ?? "an import"} — is it installed?`);
             }
         },
-    }).catch((e) => die(`${site.path}: ${e instanceof Error ? e.message : String(e)}`));
+    }).catch((e) => die(`${at.file}: ${e instanceof Error ? e.message : String(e)}`));
     const { output } = await built.generate({
         format: "es",
         inlineDynamicImports: true,
@@ -317,7 +312,7 @@ async function bundle(site, root) {
     // Refused here rather than in the realm: the evaluator can only report it against a running
     // instance, and the file it names is on this machine.
     if (!output[0].exports.includes("default")) {
-        die(`${site.path}: a script must \`export default\` the function to run`);
+        die(`${at.file}: a script must \`export default\` the function to run`);
     }
     return output[0].code;
 }
@@ -330,32 +325,73 @@ const stdin = await new Promise((resolve, reject) => {
     process.stdin.on("error", reject);
 });
 const manifest = JSON.parse(stdin);
-if (!manifest || !Array.isArray(manifest.sites))
+// genctl runs a resolver with the project root as its cwd, so the manifest need not say so.
+const root = process.cwd();
+if (!manifest || !Array.isArray(manifest.processes))
     die("stdin is not a genroc resolver manifest");
+/** A pointer as the address it is, so a generated comment or an error can be pasted into
+ *  `genctl schema`. A key no identifier can spell is bracketed, as the grammar spells it. */
+function address(pointer) {
+    return pointer
+        .map((seg) => typeof seg === "string" && /^[A-Za-z_][A-Za-z0-9_]*$/.test(seg)
+        ? `.${seg}`
+        : `[${JSON.stringify(seg)}]`)
+        .join("")
+        .replace(/^\./, "");
+}
+const located = manifest.processes.flatMap((where) => where.sites.map((site) => ({
+    site,
+    where,
+    file: resolve(where.dir, site.argument),
+})));
+// genctl is agnostic about what a script is for; the contract that an evaluation request carries
+// its module in `code` is THIS resolver's, so it is the one that checks a directive landed there.
+// `child` is the shape the scaffold uses — a call to a process that forwards to the evaluator —
+// and `external` is the same request made directly.
+for (const at of located) {
+    // Two halves, from the two places that carry them: the action's KIND is a field, and WHERE in
+    // it the directive sits is the pointer. A child call is the shape the scaffold generates — to a
+    // process that forwards to the evaluator — and an external task is the same request made
+    // directly.
+    const kind = at.site.level === "action" ? (at.site.action ?? "") : "";
+    const slot = at.site.pointer.slice(-2).join(".");
+    if ((kind !== "child" && kind !== "external") || slot !== "input.code") {
+        die(`${join(at.where.dir, at.where.file)}: ${address(at.site.pointer)}: an evaluated script ` +
+            "belongs in the `code` field of a child or external task's input, and this is " +
+            `${kind ? `a ${kind} task's ` : ""}\`${slot}\`.`);
+    }
+}
+// genctl no longer stats the argument — it does not know it is a file — so the resolver that
+// does is the one that must say when it is not there.
+for (const at of located) {
+    if (!existsSync(at.file)) {
+        die(`${join(at.where.dir, at.where.file)}: ${address(at.site.pointer)}: ` +
+            `"$import: ${at.site.argument}" names no file (looked at ${at.file})`);
+    }
+}
 // One script at two sites with different input types is a refusal, not a union: the union
 // is sound and would typecheck a body that is wrong at one of the sites.
 const byPath = new Map();
-for (const site of manifest.sites) {
-    const seen = byPath.get(site.path);
-    const defsOf = (x) => (manifest.schemas[x.process]?.$defs ?? {});
+for (const at of located) {
+    const seen = byPath.get(at.file);
     if (seen &&
-        JSON.stringify(scriptInput(seen, defsOf(seen))) !==
-            JSON.stringify(scriptInput(site, defsOf(site)))) {
-        die(`${site.path} is imported at ${seen.pointer} and ${site.pointer} with different input types.\n` +
+        JSON.stringify(seen.site.types) !== JSON.stringify(at.site.types)) {
+        die(`${at.file} is imported at ${address(seen.site.pointer)} and ` +
+            `${address(at.site.pointer)} with different input types.\n` +
             "Split it into two scripts, or make the two call sites pass the same shape.");
     }
-    byPath.set(site.path, site);
+    byPath.set(at.file, at);
 }
-for (const site of byPath.values()) {
-    const defs = (manifest.schemas[site.process]?.$defs ?? {});
-    await write(typesPathFor(site.path), declarations(site, defs));
+for (const at of byPath.values()) {
+    await write(typesPathFor(at.file), declarations(at));
 }
 if (manifest.mode === "types") {
+    console.log(JSON.stringify(manifest, null, 2));
     process.exit(0);
 }
-await typecheck(manifest.root, [...byPath.values()]);
+await typecheck([...byPath.values()]);
 const code = [];
-for (const site of manifest.sites) {
-    code.push(await bundle(site, manifest.root));
+for (const at of located) {
+    code.push(await bundle(at));
 }
 process.stdout.write(JSON.stringify({ code }));

package/dist/worker.js

@@ -5,7 +5,7 @@
 //
 // See README.md for the contract, and specs/external-task-queue.md for the queue itself.
 import { readFileSync } from "node:fs";
-import { evaluate } from "./eval.js";
+import { Cancelled, evaluate } from "./eval.js";
 const SERVER = (process.env.GENROC_SERVER ?? "http://localhost:8448").replace(/\/$/, "");
 const WORKER_ID = process.env.WORKER_ID ?? `evaluator-${process.pid}`;
 // The credential, when the server runs with --auth token. A worker needs exactly the `worker`
@@ -185,7 +185,7 @@ async function answer(token, outcome) {
     console.error(`genroc refused the outcome for ${token}: ${JSON.stringify(data)}`);
     await release(token);
 }
-async function run(job) {
+async function run(job, signal) {
     let resolved;
     try {
         resolved = await resolveObjects(job);
@@ -206,9 +206,17 @@ async function run(job) {
     }
     let result;
     try {
-        result = await evaluate(req);
+        result = await evaluate(req, signal);
     }
     catch (err) {
+        // Cancelled is not a fault: the process was stopped while this ran. The claim still goes
+        // back — the row is terminal, so nothing re-claims it, and the release is what stops it
+        // waiting out a lease nobody is serving.
+        if (err instanceof Cancelled) {
+            console.log(`cancelled: ${job.token}`);
+            await release(job.token);
+            return;
+        }
         // The RUNNER faulted, not the script — the one class where a retry can help. There is no
         // error code for it on purpose: releasing the claim is how a queue spells "retryable", and
         // it puts the task in front of a different worker instead of burning the definition's
@@ -245,11 +253,22 @@ async function renewLoop() {
             tokens,
             lease_ms: LEASE_MS,
         });
-        // A short count means a claim lapsed and was taken over. Nothing to do about it — the work
-        // continues and its answer will be refused — but say so, because it is the signal that
-        // LEASE_MS is too short for what these scripts actually take.
-        if (ok && data?.renewed < tokens.length) {
-            console.error(`renewed ${data.renewed}/${tokens.length} claims; a lease lapsed under load`);
+        if (!ok)
+            continue;
+        // Cancelled: the process was stopped, so the evaluation is abandoned and the claim handed
+        // back. Aborting first is the point of the list — waiting for the script to finish would
+        // keep burning a core on work an operator has already stopped. The release rides run()'s
+        // own catch, so nothing is released twice.
+        for (const token of (data?.cancelled ?? [])) {
+            inFlight.get(token)?.abort.abort();
+        }
+        // Lost: someone else holds this claim now. The work continues and its answer will be
+        // refused — that is not fixable here — but say so, because it is the signal that LEASE_MS
+        // is too short for what these scripts actually take. Deliberately NOT released: the claim
+        // is the new holder's, and releasing would bump the epoch out from under it.
+        const lost = (data?.lost ?? []);
+        if (lost.length) {
+            console.error(`lost ${lost.length}/${tokens.length} claims; a lease lapsed under load`);
         }
     }
 }
@@ -258,8 +277,9 @@ async function pollLoop() {
         const free = CONCURRENCY - inFlight.size;
         const jobs = free > 0 ? await claim(free) : [];
         for (const job of jobs) {
-            inFlight.set(job.token, job);
-            void run(job).finally(() => inFlight.delete(job.token));
+            const abort = new AbortController();
+            inFlight.set(job.token, { job, abort });
+            void run(job, abort.signal).finally(() => inFlight.delete(job.token));
         }
         // Only idle when there was nothing to take: a full queue should be drained at the speed the
         // realms allow, not at the poll interval.

package/eval.ts

@@ -57,14 +57,32 @@ class RealmFault extends Error {
   }
 }
 
-export async function evaluate(req: EvalRequest): Promise<EvalResult> {
+/** Thrown when `signal` aborts: the work was cancelled server-side, so there is no outcome to
+ *  report and worker.ts must release rather than answer. Distinct from RealmFault because the
+ *  runner did not fault -- nothing is wrong, the answer is simply no longer wanted. */
+export class Cancelled extends Error {
+  constructor() {
+    super("cancelled");
+    this.name = "Cancelled";
+  }
+}
+
+export async function evaluate(req: EvalRequest, signal?: AbortSignal): Promise<EvalResult> {
   const budget = typeof req.timeout_ms === "number" ? req.timeout_ms : DEFAULT_TIMEOUT_MS;
 
   const worker = new Worker(REALM_URL);
   let timer: ReturnType<typeof setTimeout> | undefined;
+  let onAbort: (() => void) | undefined;
   try {
     return await new Promise<EvalResult>((resolve, reject) => {
       timer = setTimeout(() => resolve(timedOut(budget)), budget);
+      // The abort is wired to the same promise as the budget, so both settle through the one
+      // finally below -- which is what guarantees the thread is gone before either returns.
+      if (signal) {
+        if (signal.aborted) reject(new Cancelled());
+        onAbort = () => reject(new Cancelled());
+        signal.addEventListener("abort", onAbort, { once: true });
+      }
       worker.once("message", (reply: WorkerReply) => resolve(reply));
       // A script may end its own realm (`process.exit()`), which is not a throw and would
       // otherwise present as a hang until the budget expired. Our own terminate() raises
@@ -75,6 +93,7 @@ export async function evaluate(req: EvalRequest): Promise<EvalResult> {
     });
   } finally {
     clearTimeout(timer);
+    if (signal && onAbort) signal.removeEventListener("abort", onAbort);
     // Awaited, and the whole point: on the timeout path a thread is still burning a core, and
     // resolving before it is gone would report an evaluation the machine is still running.
     await worker.terminate();

package/import.ts

@@ -8,9 +8,10 @@
 // over the same project.
 
 import { spawn } from "node:child_process";
+import { existsSync } from "node:fs";
 import { access, mkdir, writeFile } from "node:fs/promises";
 import { builtinModules } from "node:module";
-import { dirname, join, relative } from "node:path";
+import { dirname, join, relative, resolve } from "node:path";
 import { fileURLToPath } from "node:url";
 
 import commonjs from "@rollup/plugin-commonjs";
@@ -22,20 +23,37 @@ import ts from "typescript";
 type Schema = Record<string, any>;
 
 type Site = {
-  resolver: string;
-  process: string;
+  // Which namespace the directive sits in: `process`, `task` or `action`.
+  level: "process" | "task" | "action";
   task?: string;
-  pointer: string;
-  path: string;
-  input?: Schema;
-  output?: Schema;
+  // What the site IS, at action level: the action's type, and the process a child call names.
+  action?: string;
+  child?: string;
+  // The slot as a path of keys and indices, so nothing has to unescape a JSON Pointer.
+  pointer: (string | number)[];
+  // Everything after `$<resolver>:`, verbatim. genctl does not read it as a path — that is
+  // this resolver's reading of it.
+  argument: string;
+  // What .genroc asked genctl to type, keyed by the name it chose. This resolver reads Input
+  // and Output; which ADDRESS each came from is the config's business, not ours.
+  types?: Record<string, Schema>;
+};
+
+type ManifestProcess = {
+  name: string;
+  // The definition's own location, split because a relative argument is relative to the
+  // DIRECTORY — joining is ours to do, and this is the base.
+  dir: string;
+  file: string;
+  sites: Site[];
+  // Only what this process's fragments reach — a `$ref` in `types` points here.
+  $defs?: Record<string, Schema>;
 };
 
 type Manifest = {
   mode: "types" | "build";
   root: string;
-  schemas: Record<string, Schema>;
-  sites: Site[];
+  processes: ManifestProcess[];
 };
 
 function die(message: string): never {
@@ -150,25 +168,14 @@ function deref(
   return cur;
 }
 
-/** The manifest's `input` is the type of the whole ACTION input — for /eval that is
- *  `{code, input, timeout_ms, …}`, and only its `input` field is bound as the script's
- *  parameter. genroc cannot know that; this file owns the evaluator's wire contract, so
- *  the navigation belongs here rather than in genctl. */
-function scriptInput(
-  site: Site,
-  defs: Record<string, Schema>,
-): Schema | undefined {
-  const action = deref(site.input, defs);
-  const bound = action?.properties?.input;
-  return bound ?? site.input;
-}
-
 /** Emits one named type per reachable $def rather than inlining: a task output may
  *  reference itself (specs/recursive-type-inference.md) and inlining would not terminate. */
-function declarations(site: Site, defs: Record<string, Schema>): string {
+function declarations(at: Located): string {
+  const { site, where } = at;
+  const defs = where.$defs ?? {};
   const used = new Set<string>();
-  const input = tsType(scriptInput(site, defs), used);
-  const output = tsType(site.output, used);
+  const input = tsType(site.types?.Input, used);
+  const output = tsType(site.types?.Output, used);
 
   const emitted: string[] = [];
   const done = new Set<string>();
@@ -182,7 +189,7 @@ function declarations(site: Site, defs: Record<string, Schema>): string {
 
   return [
     "// Generated by genroc. Do not edit - regenerate with `genctl types`.",
-    `// ${site.process}${site.task ? ` / ${site.task}` : ""}  (${site.pointer})`,
+    `// ${where.name}  (${address(site.pointer)})`,
     "",
     ...emitted,
     emitted.length ? "" : "",
@@ -215,7 +222,7 @@ async function nearestTsconfig(
   }
 }
 
-async function typecheck(root: string, sites: Site[]): Promise<void> {
+async function typecheck(sites: Located[]): Promise<void> {
   // NOT `.genroc`: that is the project config FILE, and a directory of the same name cannot
   // coexist with it. The suffix is what keeps the scratch area out of its way.
   const dir = join(root, ".genroc-cache");
@@ -223,12 +230,12 @@ async function typecheck(root: string, sites: Site[]): Promise<void> {
 
   // One tsc per distinct base config: `extends` takes a single base, so merging two would
   // check each script under the other author's options.
-  const groups = new Map<string, Site[]>();
-  for (const site of sites) {
-    const base = (await nearestTsconfig(dirname(site.path), root)) ?? "";
+  const groups = new Map<string, Located[]>();
+  for (const at of sites) {
+    const base = (await nearestTsconfig(dirname(at.file), root)) ?? "";
     const group = groups.get(base);
-    if (group) group.push(site);
-    else groups.set(base, [site]);
+    if (group) group.push(at);
+    else groups.set(base, [at]);
   }
 
   let n = 0;
@@ -251,8 +258,8 @@ async function typecheck(root: string, sites: Site[]): Promise<void> {
         ...(base ? {} : { types: [] }),
       },
       files: group.flatMap((s) => [
-        relative(dir, s.path),
-        relative(dir, typesPathFor(s.path)),
+        relative(dir, s.file),
+        relative(dir, typesPathFor(s.file)),
       ]),
       // `files` overrides the base's, but a base `include` survives beside it and would
       // drag the author's whole tree in, to be checked under the worker lib.
@@ -351,15 +358,16 @@ function tsResolve(configPath: string | null): Plugin {
 /** Bundles to a self-contained ES module, which is what the evaluator imports: the default
  *  export it calls is the author's own, so nothing wraps or rewrites the code between the two.
  *  Bundling is entirely the importer's job, so a definition version pins its code forever. */
-async function bundle(site: Site, root: string): Promise<string> {
+async function bundle(at: Located): Promise<string> {
+  const site = at.site;
   // Builtins are EXTERNALISED as imports the realm resolves natively. Anything else
   // unresolved is a REFUSAL, not an external: rollup's default is to leave it as an import
   // of a module that will not be there, which bundles clean and fails at runtime.
   const built = await rollup({
-    input: site.path,
+    input: at.file,
     external: (id) => BUILTIN.has(id),
     plugins: [
-      tsResolve(await nearestTsconfig(dirname(site.path), root)),
+      tsResolve(await nearestTsconfig(dirname(at.file), root)),
       nodeResolve({ extensions: [".ts", ".tsx", ".mjs", ".js", ".json"] }),
       commonjs(),
       // A `.json` import is a data file inlined at build time, which the previous bundler
@@ -370,12 +378,12 @@ async function bundle(site: Site, root: string): Promise<string> {
     onwarn(warning) {
       if (warning.code === "UNRESOLVED_IMPORT") {
         die(
-          `${site.path}: cannot resolve ${warning.exporter ?? "an import"} — is it installed?`,
+          `${at.file}: cannot resolve ${warning.exporter ?? "an import"} — is it installed?`,
         );
       }
     },
   }).catch((e: unknown) =>
-    die(`${site.path}: ${e instanceof Error ? e.message : String(e)}`),
+    die(`${at.file}: ${e instanceof Error ? e.message : String(e)}`),
   );
 
   const { output } = await built.generate({
@@ -386,7 +394,7 @@ async function bundle(site: Site, root: string): Promise<string> {
   // Refused here rather than in the realm: the evaluator can only report it against a running
   // instance, and the file it names is on this machine.
   if (!output[0].exports.includes("default")) {
-    die(`${site.path}: a script must \`export default\` the function to run`);
+    die(`${at.file}: a script must \`export default\` the function to run`);
   }
   return output[0].code;
 }
@@ -401,45 +409,99 @@ const stdin: string = await new Promise((resolve, reject) => {
   process.stdin.on("error", reject);
 });
 const manifest = JSON.parse(stdin) as Manifest;
-if (!manifest || !Array.isArray(manifest.sites))
+// genctl runs a resolver with the project root as its cwd, so the manifest need not say so.
+const root = process.cwd();
+if (!manifest || !Array.isArray(manifest.processes))
   die("stdin is not a genroc resolver manifest");
 
+/** Every site with the process it sits in and the file its argument names, in the manifest's own
+ *  order — which is the order `code` must answer in. genctl passes the argument verbatim, so
+ *  joining it to the definition's directory is this resolver's business: only it knows the
+ *  argument is a path at all. */
+type Located = { site: Site; where: ManifestProcess; file: string };
+
+/** A pointer as the address it is, so a generated comment or an error can be pasted into
+ *  `genctl schema`. A key no identifier can spell is bracketed, as the grammar spells it. */
+function address(pointer: (string | number)[]): string {
+  return pointer
+    .map((seg) =>
+      typeof seg === "string" && /^[A-Za-z_][A-Za-z0-9_]*$/.test(seg)
+        ? `.${seg}`
+        : `[${JSON.stringify(seg)}]`,
+    )
+    .join("")
+    .replace(/^\./, "");
+}
+const located: Located[] = manifest.processes.flatMap((where) =>
+  where.sites.map((site) => ({
+    site,
+    where,
+    file: resolve(where.dir, site.argument),
+  })),
+);
+
+// genctl is agnostic about what a script is for; the contract that an evaluation request carries
+// its module in `code` is THIS resolver's, so it is the one that checks a directive landed there.
+// `child` is the shape the scaffold uses — a call to a process that forwards to the evaluator —
+// and `external` is the same request made directly.
+for (const at of located) {
+  // Two halves, from the two places that carry them: the action's KIND is a field, and WHERE in
+  // it the directive sits is the pointer. A child call is the shape the scaffold generates — to a
+  // process that forwards to the evaluator — and an external task is the same request made
+  // directly.
+  const kind = at.site.level === "action" ? (at.site.action ?? "") : "";
+  const slot = at.site.pointer.slice(-2).join(".");
+  if ((kind !== "child" && kind !== "external") || slot !== "input.code") {
+    die(
+      `${join(at.where.dir, at.where.file)}: ${address(at.site.pointer)}: an evaluated script ` +
+        "belongs in the `code` field of a child or external task's input, and this is " +
+        `${kind ? `a ${kind} task's ` : ""}\`${slot}\`.`,
+    );
+  }
+}
+
+// genctl no longer stats the argument — it does not know it is a file — so the resolver that
+// does is the one that must say when it is not there.
+for (const at of located) {
+  if (!existsSync(at.file)) {
+    die(
+      `${join(at.where.dir, at.where.file)}: ${address(at.site.pointer)}: ` +
+        `"$import: ${at.site.argument}" names no file (looked at ${at.file})`,
+    );
+  }
+}
+
 // One script at two sites with different input types is a refusal, not a union: the union
 // is sound and would typecheck a body that is wrong at one of the sites.
-const byPath = new Map<string, Site>();
-for (const site of manifest.sites) {
-  const seen = byPath.get(site.path);
-  const defsOf = (x: Site) =>
-    (manifest.schemas[x.process]?.$defs ?? {}) as Record<string, Schema>;
+const byPath = new Map<string, Located>();
+for (const at of located) {
+  const seen = byPath.get(at.file);
   if (
     seen &&
-    JSON.stringify(scriptInput(seen, defsOf(seen))) !==
-      JSON.stringify(scriptInput(site, defsOf(site)))
+    JSON.stringify(seen.site.types) !== JSON.stringify(at.site.types)
   ) {
     die(
-      `${site.path} is imported at ${seen.pointer} and ${site.pointer} with different input types.\n` +
+      `${at.file} is imported at ${address(seen.site.pointer)} and ` +
+        `${address(at.site.pointer)} with different input types.\n` +
         "Split it into two scripts, or make the two call sites pass the same shape.",
     );
   }
-  byPath.set(site.path, site);
+  byPath.set(at.file, at);
 }
 
-for (const site of byPath.values()) {
-  const defs = (manifest.schemas[site.process]?.$defs ?? {}) as Record<
-    string,
-    Schema
-  >;
-  await write(typesPathFor(site.path), declarations(site, defs));
+for (const at of byPath.values()) {
+  await write(typesPathFor(at.file), declarations(at));
 }
 
 if (manifest.mode === "types") {
+  console.log(JSON.stringify(manifest, null, 2));
   process.exit(0);
 }
 
-await typecheck(manifest.root, [...byPath.values()]);
+await typecheck([...byPath.values()]);
 
 const code: string[] = [];
-for (const site of manifest.sites) {
-  code.push(await bundle(site, manifest.root));
+for (const at of located) {
+  code.push(await bundle(at));
 }
 process.stdout.write(JSON.stringify({ code }));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@genroc/eval-node",
-  "version": "0.0.0-edge.ea911c5",
+  "version": "0.0.0-edge.eadd43a",
   "type": "module",
   "engines": {
     "node": ">=24"

package/worker.ts

@@ -6,7 +6,7 @@
 // See README.md for the contract, and specs/external-task-queue.md for the queue itself.
 
 import { readFileSync } from "node:fs";
-import { evaluate, type EvalRequest, type FailureKind } from "./eval.ts";
+import { Cancelled, evaluate, type EvalRequest, type FailureKind } from "./eval.ts";
 
 const SERVER = (process.env.GENROC_SERVER ?? "http://localhost:8448").replace(/\/$/, "");
 const WORKER_ID = process.env.WORKER_ID ?? `evaluator-${process.pid}`;
@@ -135,7 +135,12 @@ function asEvalRequest(input: unknown): EvalRequest | string {
   };
 }
 
-const inFlight = new Map<string, QueueTask>();
+/** A claim this worker is currently serving. The controller is how the renewal loop reaches
+ *  the evaluation: genroc cannot call us, so a cancellation arrives as an answer to our own
+ *  heartbeat and has to be delivered inward from there. */
+type Running = { job: QueueTask; abort: AbortController };
+
+const inFlight = new Map<string, Running>();
 let running = true;
 
 // Whether the last claim reached genroc. A worker polls several times a second, so an
@@ -203,7 +208,7 @@ async function answer(token: string, outcome: Record<string, unknown>): Promise<
   await release(token);
 }
 
-async function run(job: QueueTask): Promise<void> {
+async function run(job: QueueTask, signal: AbortSignal): Promise<void> {
   let resolved: unknown;
   try {
     resolved = await resolveObjects(job);
@@ -224,8 +229,16 @@ async function run(job: QueueTask): Promise<void> {
 
   let result;
   try {
-    result = await evaluate(req);
+    result = await evaluate(req, signal);
   } catch (err) {
+    // Cancelled is not a fault: the process was stopped while this ran. The claim still goes
+    // back — the row is terminal, so nothing re-claims it, and the release is what stops it
+    // waiting out a lease nobody is serving.
+    if (err instanceof Cancelled) {
+      console.log(`cancelled: ${job.token}`);
+      await release(job.token);
+      return;
+    }
     // The RUNNER faulted, not the script — the one class where a retry can help. There is no
     // error code for it on purpose: releasing the claim is how a queue spells "retryable", and
     // it puts the task in front of a different worker instead of burning the definition's
@@ -263,11 +276,22 @@ async function renewLoop(): Promise<void> {
       tokens,
       lease_ms: LEASE_MS,
     });
-    // A short count means a claim lapsed and was taken over. Nothing to do about it — the work
-    // continues and its answer will be refused — but say so, because it is the signal that
-    // LEASE_MS is too short for what these scripts actually take.
-    if (ok && data?.renewed < tokens.length) {
-      console.error(`renewed ${data.renewed}/${tokens.length} claims; a lease lapsed under load`);
+    if (!ok) continue;
+
+    // Cancelled: the process was stopped, so the evaluation is abandoned and the claim handed
+    // back. Aborting first is the point of the list — waiting for the script to finish would
+    // keep burning a core on work an operator has already stopped. The release rides run()'s
+    // own catch, so nothing is released twice.
+    for (const token of (data?.cancelled ?? []) as string[]) {
+      inFlight.get(token)?.abort.abort();
+    }
+    // Lost: someone else holds this claim now. The work continues and its answer will be
+    // refused — that is not fixable here — but say so, because it is the signal that LEASE_MS
+    // is too short for what these scripts actually take. Deliberately NOT released: the claim
+    // is the new holder's, and releasing would bump the epoch out from under it.
+    const lost = (data?.lost ?? []) as string[];
+    if (lost.length) {
+      console.error(`lost ${lost.length}/${tokens.length} claims; a lease lapsed under load`);
     }
   }
 }
@@ -277,8 +301,9 @@ async function pollLoop(): Promise<void> {
     const free = CONCURRENCY - inFlight.size;
     const jobs = free > 0 ? await claim(free) : [];
     for (const job of jobs) {
-      inFlight.set(job.token, job);
-      void run(job).finally(() => inFlight.delete(job.token));
+      const abort = new AbortController();
+      inFlight.set(job.token, { job, abort });
+      void run(job, abort.signal).finally(() => inFlight.delete(job.token));
     }
     // Only idle when there was nothing to take: a full queue should be drained at the speed the
     // realms allow, not at the poll interval.