@foldrun/core 0.3.0 → 0.4.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (83) hide show
  1. package/README.md +27 -10
  2. package/dist/src/approvals.d.ts +104 -1
  3. package/dist/src/approvals.js +192 -2
  4. package/dist/src/budget.d.ts +32 -0
  5. package/dist/src/budget.js +177 -0
  6. package/dist/src/catalog.js +2 -1
  7. package/dist/src/confine.js +6 -1
  8. package/dist/src/egress.d.ts +19 -2
  9. package/dist/src/egress.js +2 -1
  10. package/dist/src/evals.d.ts +4 -0
  11. package/dist/src/evals.js +20 -11
  12. package/dist/src/flow-lint.d.ts +4 -0
  13. package/dist/src/flow-lint.js +108 -0
  14. package/dist/src/host-env.d.ts +8 -0
  15. package/dist/src/host-env.js +48 -0
  16. package/dist/src/json-schema.d.ts +12 -0
  17. package/dist/src/json-schema.js +202 -0
  18. package/dist/src/library.js +13 -1
  19. package/dist/src/notify.d.ts +79 -1
  20. package/dist/src/notify.js +243 -9
  21. package/dist/src/paths.d.ts +5 -0
  22. package/dist/src/paths.js +32 -0
  23. package/dist/src/platform.d.ts +102 -2
  24. package/dist/src/platform.js +18 -2
  25. package/dist/src/provider-health.d.ts +55 -0
  26. package/dist/src/provider-health.js +232 -0
  27. package/dist/src/run-container.d.ts +57 -4
  28. package/dist/src/run-container.js +147 -41
  29. package/dist/src/runner.d.ts +39 -3
  30. package/dist/src/runner.js +422 -73
  31. package/dist/src/script-tools.d.ts +12 -0
  32. package/dist/src/script-tools.js +10 -2
  33. package/dist/src/secret-health.d.ts +59 -0
  34. package/dist/src/secret-health.js +155 -0
  35. package/dist/src/secrets.d.ts +19 -1
  36. package/dist/src/secrets.js +44 -12
  37. package/dist/src/starter.js +3 -0
  38. package/dist/src/step-exec.d.ts +47 -1
  39. package/dist/src/step-exec.js +155 -60
  40. package/dist/src/store.d.ts +276 -3
  41. package/dist/src/store.js +552 -26
  42. package/dist/src/tar.js +2 -1
  43. package/dist/src/test-mode.d.ts +111 -0
  44. package/dist/src/test-mode.js +285 -0
  45. package/dist/src/tool-test.js +48 -21
  46. package/dist/src/translator.js +2 -1
  47. package/dist/src/trigger-log.d.ts +43 -0
  48. package/dist/src/trigger-log.js +122 -0
  49. package/dist/src/triggers.d.ts +20 -0
  50. package/dist/src/triggers.js +91 -10
  51. package/dist/src/webhook.d.ts +15 -3
  52. package/dist/src/webhook.js +11 -11
  53. package/package.json +4 -4
  54. package/src/approvals.ts +253 -2
  55. package/src/budget.ts +178 -0
  56. package/src/catalog.ts +2 -1
  57. package/src/confine.ts +6 -1
  58. package/src/egress.ts +20 -4
  59. package/src/evals.ts +25 -9
  60. package/src/flow-lint.ts +118 -0
  61. package/src/host-env.ts +51 -0
  62. package/src/json-schema.ts +184 -0
  63. package/src/library.ts +13 -1
  64. package/src/notify.ts +272 -11
  65. package/src/paths.ts +29 -0
  66. package/src/platform.ts +107 -3
  67. package/src/provider-health.ts +274 -0
  68. package/src/run-container.ts +175 -36
  69. package/src/runner.ts +435 -65
  70. package/src/script-tools.ts +18 -2
  71. package/src/secret-health.ts +185 -0
  72. package/src/secrets.ts +44 -10
  73. package/src/starter.ts +3 -0
  74. package/src/step-exec.ts +128 -13
  75. package/src/storage.ts +8 -2
  76. package/src/store.ts +746 -25
  77. package/src/tar.ts +2 -1
  78. package/src/test-mode.ts +311 -0
  79. package/src/tool-test.ts +49 -20
  80. package/src/translator.ts +2 -1
  81. package/src/trigger-log.ts +159 -0
  82. package/src/triggers.ts +92 -10
  83. package/src/webhook.ts +17 -11
@@ -20,6 +20,7 @@ import { spawn } from "node:child_process";
20
20
  import { z } from "zod";
21
21
  import { createSdkMcpServer, tool } from "@anthropic-ai/claude-agent-sdk";
22
22
  import { runInContainer } from "./container.ts";
23
+ import { hostSafeEnv } from "./host-env.ts";
23
24
 
24
25
  // A script's only clock is the `timeout:` its own tool.md declares. There
25
26
  // was a two-minute default and a ten-minute cap here; both were the
@@ -55,6 +56,14 @@ export interface ScriptSpec {
55
56
  * default, `materialised`, is a program that reads values from its
56
57
  * environment and needs them real — a browser seeding cookies. */
57
58
  secrets?: "proxied" | "materialised";
59
+ /** `outward: true` — this program makes something happen outside (sends,
60
+ * posts, orders). On a test run a step that grants it is handed no real
61
+ * secret at all, whatever their names. */
62
+ outward?: boolean;
63
+ /** `test_mode: allow` — a reader that happens to use a send-capable
64
+ * secret's name (a GBP reader on a GBP token): on a test run the step
65
+ * still gets the real values. An `outward` tool in the same step wins. */
66
+ testMode?: "allow";
58
67
  }
59
68
 
60
69
  export function parseScripts(raw: unknown): ScriptSpec[] {
@@ -82,6 +91,8 @@ export function parseScripts(raw: unknown): ScriptSpec[] {
82
91
  args,
83
92
  interpreter: typeof e.interpreter === "string" ? e.interpreter : undefined,
84
93
  secrets: e.secrets === "proxied" ? "proxied" : undefined,
94
+ outward: e.outward === true ? true : undefined,
95
+ testMode: e.test_mode === "allow" ? "allow" : undefined,
85
96
  timeout: Number.isFinite(timeout) && timeout > 0 ? timeout : undefined,
86
97
  });
87
98
  }
@@ -128,7 +139,7 @@ function containerInterpreter(abs: string): string {
128
139
  return byExt[path.extname(abs).toLowerCase()] ?? abs;
129
140
  }
130
141
 
131
- function runScript(
142
+ export function runScript(
132
143
  cwd: string,
133
144
  spec: ScriptSpec,
134
145
  values: Record<string, string>,
@@ -229,9 +240,14 @@ function runScript(
229
240
  return;
230
241
  }
231
242
 
243
+ // Not process.env: on a platform that is every tenant's root key and
244
+ // the datastore credentials. The child starts from the allowlisted
245
+ // base and gets what the agent declared, by name — see host-env.ts.
246
+ // In a run container `env` already carries the container's whole
247
+ // environment, which is the boundary, so nothing is lost there.
232
248
  const child = spawn(cmd, args, {
233
249
  cwd,
234
- env: { ...process.env, ...env },
250
+ env: { ...hostSafeEnv(), ...env },
235
251
  timeout: timeoutFor(spec),
236
252
  });
237
253
  let out = "";
@@ -0,0 +1,185 @@
1
+ // Does this credential still work?
2
+ //
3
+ // The vault knows a secret's name, its scope and its shape. It knows nothing
4
+ // about whether the thing on the other end still accepts it, and that is the
5
+ // only question anyone actually has. A key that has quietly stopped working
6
+ // looks identical to a key nobody has used yet, and both look identical to a
7
+ // key that is fine — three very different situations rendered the same way.
8
+ //
9
+ // Two facts, recorded where they are already known and nowhere else:
10
+ //
11
+ // when it was last used, and against which host
12
+ // what the other end said the last time — accepted, or refused
13
+ //
14
+ // The egress proxy sees both for every http tool and every model call,
15
+ // because filling the credential is its whole job. Notifications and the
16
+ // provider check report their own. Nothing new is instrumented, and a secret
17
+ // materialised into a sandbox reports nothing, which the page says plainly
18
+ // rather than guessing.
19
+ //
20
+ // Never the value, never a hash of the value, never the request body. A
21
+ // name, a host, a status and a time.
22
+
23
+ import fs from "node:fs";
24
+ import path from "node:path";
25
+ import crypto from "node:crypto";
26
+ import { accountDir } from "./store.ts";
27
+
28
+ /** How a credential is identified here. A name alone is not enough: the
29
+ * account and each workspace can hold a secret by the same name, and they
30
+ * are different credentials with different health. */
31
+ export function healthKey(name: string, scope: "account" | "workspace", workspace?: string | null): string {
32
+ return scope === "workspace" && workspace ? `workspace:${workspace}:${name}` : `account:${name}`;
33
+ }
34
+
35
+ export interface SecretUse {
36
+ /** ISO time of the most recent use. */
37
+ at: string;
38
+ /** The host it was sent to, so "which of my four keys is this" has an
39
+ * answer without opening a run. */
40
+ host: string;
41
+ /** The response status, when there was one. */
42
+ status: number | null;
43
+ /** Refused means the credential itself was rejected: 401, 402 or 403.
44
+ * A 500 is the other end failing, which says nothing about the key. */
45
+ outcome: "accepted" | "refused" | "error";
46
+ }
47
+
48
+ export interface SecretRecord {
49
+ last: SecretUse;
50
+ /** Consecutive refusals, newest first — one is a fluke, five is a dead
51
+ * key. Reset by any acceptance. */
52
+ refusals: number;
53
+ /** The last acceptance, kept through a run of refusals so a page can say
54
+ * "worked until Tuesday", which is the sentence that dates the breakage. */
55
+ lastAccepted: string | null;
56
+ }
57
+
58
+ type HealthFile = Record<string, SecretRecord>;
59
+
60
+ const healthFile = (tenant: string) => path.join(accountDir(tenant), "secret-health.json");
61
+
62
+ function read(tenant: string): HealthFile {
63
+ try {
64
+ return JSON.parse(fs.readFileSync(healthFile(tenant), "utf8")) as HealthFile;
65
+ } catch {
66
+ return {};
67
+ }
68
+ }
69
+
70
+ function write(tenant: string, data: HealthFile): void {
71
+ const file = healthFile(tenant);
72
+ if (!fs.existsSync(path.dirname(file))) return;
73
+ // pid alone is not unique across pods on a shared volume.
74
+ const tmp = `${file}.${process.pid}.${crypto.randomBytes(4).toString("hex")}`;
75
+ fs.writeFileSync(tmp, JSON.stringify(data, null, 2));
76
+ fs.renameSync(tmp, file);
77
+ }
78
+
79
+ // Writes are queued and coalesced off the caller's path. The egress proxy
80
+ // notes a use per secret per request, before it streams the response body,
81
+ // and a synchronous read-modify-write of a JSON file there is latency on
82
+ // every proxied call — "never blocks" has to be true, not aspirational.
83
+ // One writer per process; the queue drains in order.
84
+ const pending = new Map<string, { name: string; use: { host: string; status: number | null; at?: string } }[]>();
85
+ let draining = false;
86
+
87
+ function drain(): void {
88
+ if (draining) return;
89
+ draining = true;
90
+ setImmediate(() => {
91
+ try {
92
+ for (const [tenant, uses] of pending) {
93
+ pending.delete(tenant);
94
+ const data = read(tenant);
95
+ for (const { name, use } of uses) apply(data, name, use);
96
+ write(tenant, data);
97
+ }
98
+ } catch {
99
+ // best effort, always
100
+ } finally {
101
+ draining = false;
102
+ if (pending.size) drain();
103
+ }
104
+ });
105
+ }
106
+
107
+ function apply(data: HealthFile, name: string, use: { host: string; status: number | null; at?: string }): void {
108
+ const outcome = outcomeFor(use.status);
109
+ const at = use.at ?? new Date().toISOString();
110
+ const prev = data[name];
111
+ data[name] = {
112
+ last: { at, host: use.host, status: use.status, outcome },
113
+ // Only a refusal counts against a key. An error at the other end is
114
+ // the other end's problem and must not accumulate into a false alarm.
115
+ refusals: outcome === "refused" ? (prev?.refusals ?? 0) + 1 : outcome === "accepted" ? 0 : prev?.refusals ?? 0,
116
+ lastAccepted: outcome === "accepted" ? at : prev?.lastAccepted ?? null,
117
+ };
118
+ }
119
+
120
+ /** 401, 402 and 403 are the credential being refused. Everything else is
121
+ * the request, the model or the provider — none of which is the key's
122
+ * fault, and calling them a credential failure would send people to
123
+ * rotate a key that was never the problem. */
124
+ export function outcomeFor(status: number | null): SecretUse["outcome"] {
125
+ if (status === null) return "error";
126
+ if (status === 401 || status === 402 || status === 403) return "refused";
127
+ if (status >= 200 && status < 400) return "accepted";
128
+ return "error";
129
+ }
130
+
131
+ /**
132
+ * Record one use. `key` is a health key (see healthKey), never a bare name.
133
+ * Never throws and never blocks: the write is queued and happens off the
134
+ * caller's path, because this runs on every proxied request.
135
+ */
136
+ export function noteSecretUse(
137
+ tenant: string,
138
+ key: string,
139
+ use: { host: string; status: number | null; at?: string },
140
+ ): void {
141
+ if (!key) return;
142
+ const list = pending.get(tenant) ?? [];
143
+ list.push({ name: key, use: { ...use, at: use.at ?? new Date().toISOString() } });
144
+ pending.set(tenant, list);
145
+ drain();
146
+ }
147
+
148
+ /** Wait for queued writes to land — for tests, and for a process that is
149
+ * about to exit. */
150
+ export function flushSecretHealth(): Promise<void> {
151
+ return new Promise((resolve) => {
152
+ const check = () => (pending.size === 0 && !draining ? resolve() : setImmediate(check));
153
+ check();
154
+ });
155
+ }
156
+
157
+ /** Everything known about this account's credentials, by name. Names the
158
+ * vault holds but nothing has used are simply absent, which is the honest
159
+ * answer: unused is not the same as broken. */
160
+ export function secretHealth(tenant: string): HealthFile {
161
+ return read(tenant);
162
+ }
163
+
164
+ /** The credentials worth saying something about: refused the last time they
165
+ * were used. Sorted worst first — most consecutive refusals, then oldest
166
+ * last-success, which is the order someone would work through them. */
167
+ export function failingSecrets(tenant: string): { name: string; record: SecretRecord }[] {
168
+ return Object.entries(read(tenant))
169
+ .filter(([, r]) => r.last.outcome === "refused")
170
+ .map(([name, record]) => ({ name, record }))
171
+ .sort((a, b) => b.record.refusals - a.record.refusals || (a.record.lastAccepted ?? "").localeCompare(b.record.lastAccepted ?? ""));
172
+ }
173
+
174
+ /** Drop what is known about a secret — for a delete, and for a rotation,
175
+ * where the old key's refusals say nothing about the new one. */
176
+ export function forgetSecretHealth(tenant: string, name: string): void {
177
+ try {
178
+ const data = read(tenant);
179
+ if (!(name in data)) return;
180
+ delete data[name];
181
+ write(tenant, data);
182
+ } catch {
183
+ // best effort
184
+ }
185
+ }
package/src/secrets.ts CHANGED
@@ -24,25 +24,52 @@ import path from "node:path";
24
24
  import { dataRoot } from "./paths.ts";
25
25
  import { platform } from "./platform.ts";
26
26
 
27
+ /**
28
+ * GCM's auth tag, pinned to its full 16 bytes at every cipher call. Without
29
+ * `authTagLength`, Node accepts whatever length `setAuthTag` is handed — 4, 8,
30
+ * 12, 13, 14, 15 or 16 bytes — so a stored record whose tag had been cut to
31
+ * four bytes would still "verify" on 32 bits of authentication instead of
32
+ * 128. The tag comes from the record on disk; the record is the thing an
33
+ * attacker with file access would edit. One option makes a short tag an
34
+ * error rather than a weaker check.
35
+ */
36
+ const GCM = { authTagLength: 16 } as const;
37
+
27
38
  // The account's own data key, when the platform holds one; the install key
28
39
  // otherwise. See platform.ts — locally there is no account key.
29
40
  const tenantKey = (tenant: string) => platform.tenantKey(tenant);
30
41
  import crypto from "node:crypto";
31
42
  import { assertSafeName } from "./store.ts";
43
+ import { trimSlashes } from "./paths.ts";
32
44
 
33
45
  const keyFile = () => path.join(dataRoot(), ".secret-key");
34
46
 
35
47
  /** Which store a secret came from. */
36
48
  export type SecretScope = "account" | "workspace";
37
49
 
38
- function masterKey(): Buffer {
50
+ /**
51
+ * The install's root secret, as text: FOLDRUN_SECRET_KEY when set, else the
52
+ * contents of `.secret-key` under the data root — created here, the first
53
+ * time anything asks, so there is never a moment when a derived token has
54
+ * nothing real to hang off. The vault, hook tokens, approval links and git
55
+ * push secrets all derive from this one value; rotating it invalidates every
56
+ * one of them together. Exported for webhook.ts, which used to fall back to
57
+ * a fixed dev string until the vault happened to create this file — so hook
58
+ * URLs were computable by anyone who read the source, and then changed the
59
+ * first time a secret was stored.
60
+ */
61
+ export function installKeyMaterial(): string {
39
62
  const fromEnv = process.env.FOLDRUN_SECRET_KEY;
40
- if (fromEnv) return crypto.createHash("sha256").update(fromEnv).digest();
63
+ if (fromEnv) return fromEnv;
41
64
  fs.mkdirSync(dataRoot(), { recursive: true });
42
65
  if (!fs.existsSync(keyFile())) {
43
66
  fs.writeFileSync(keyFile(), crypto.randomBytes(32).toString("hex"), { mode: 0o600 });
44
67
  }
45
- return crypto.createHash("sha256").update(fs.readFileSync(keyFile(), "utf8")).digest();
68
+ return fs.readFileSync(keyFile(), "utf8");
69
+ }
70
+
71
+ function masterKey(): Buffer {
72
+ return crypto.createHash("sha256").update(installKeyMaterial()).digest();
46
73
  }
47
74
 
48
75
  // `workspace` undefined means the account store.
@@ -135,7 +162,7 @@ export function encryptValue(
135
162
  ): { iv: string; tag: string; data: string; k?: "tenant" } {
136
163
  const { key, mark } = keyFor(tenant);
137
164
  const iv = crypto.randomBytes(12);
138
- const cipher = crypto.createCipheriv("aes-256-gcm", key, iv);
165
+ const cipher = crypto.createCipheriv("aes-256-gcm", key, iv, GCM);
139
166
  const enc = Buffer.concat([cipher.update(value, "utf8"), cipher.final()]);
140
167
  return {
141
168
  iv: iv.toString("base64"),
@@ -180,6 +207,7 @@ function decrypt(tenant: string, rec: StoredSecret | undefined): string | null {
180
207
  "aes-256-gcm",
181
208
  key,
182
209
  Buffer.from(rec.iv, "base64"),
210
+ GCM,
183
211
  );
184
212
  decipher.setAuthTag(Buffer.from(rec.tag, "base64"));
185
213
  return Buffer.concat([
@@ -357,7 +385,7 @@ export function rotateMasterKey(
357
385
  }
358
386
  let plain: string;
359
387
  try {
360
- const d = crypto.createDecipheriv("aes-256-gcm", oldKey, Buffer.from(rec.iv, "base64"));
388
+ const d = crypto.createDecipheriv("aes-256-gcm", oldKey, Buffer.from(rec.iv, "base64"), GCM);
361
389
  d.setAuthTag(Buffer.from(rec.tag, "base64"));
362
390
  plain = Buffer.concat([d.update(Buffer.from(rec.data, "base64")), d.final()]).toString("utf8");
363
391
  } catch {
@@ -366,7 +394,7 @@ export function rotateMasterKey(
366
394
  continue;
367
395
  }
368
396
  const iv = crypto.randomBytes(12);
369
- const c = crypto.createCipheriv("aes-256-gcm", newKey, iv);
397
+ const c = crypto.createCipheriv("aes-256-gcm", newKey, iv, GCM);
370
398
  const enc = Buffer.concat([c.update(plain, "utf8"), c.final()]);
371
399
  next[name] = {
372
400
  iv: iv.toString("base64"),
@@ -498,6 +526,9 @@ async function exchange(config: OAuth2Config, cacheKey: string): Promise<string>
498
526
  */
499
527
  export async function materializeSecrets(
500
528
  env: Record<string, string>,
529
+ /** Whose secrets these are, when the caller knows: lets the platform keep
530
+ * each OAuth connection's health and quote a reconnect link on failure. */
531
+ ctx?: { tenant: string; workspace?: string },
501
532
  ): Promise<Record<string, string>> {
502
533
  const out: Record<string, string> = { ...env };
503
534
  for (const [name, value] of Object.entries(env)) {
@@ -518,8 +549,11 @@ export async function materializeSecrets(
518
549
  const fp = crypto.createHash("sha256").update(grantMaterial).digest("hex").slice(0, 16);
519
550
  out[name] = await exchange(config, `${name}:${config.client_id}:${fp}`);
520
551
  } catch (err) {
521
- throw new Error(`secret ${name}: ${err instanceof Error ? err.message : String(err)}`);
552
+ const message = err instanceof Error ? err.message : String(err);
553
+ const link = ctx ? await platform.noteOAuthRefresh(ctx, name, message).catch(() => undefined) : undefined;
554
+ throw new Error(`secret ${name}: ${message}${link ? ` — reconnect it in one click: ${link}` : ""}`);
522
555
  }
556
+ if (ctx) await platform.noteOAuthRefresh(ctx, name, null).catch(() => undefined);
523
557
  } else if (value.startsWith(SERVICE_ACCOUNT_PREFIX)) {
524
558
  let config: ServiceAccountConfig;
525
559
  try {
@@ -570,7 +604,7 @@ export function setServiceAccountSecret(
570
604
  for (const field of ["token_url", "issuer", "private_key", "scope"] as const) {
571
605
  if (!config[field]?.trim()) throw new Error(`service account secret needs ${field}`);
572
606
  }
573
- if (!/BEGIN [A-Z ]*PRIVATE KEY/.test(config.private_key)) {
607
+ if (!/BEGIN [A-Z ]{0,40}PRIVATE KEY/.test(config.private_key)) {
574
608
  throw new Error("private_key must be a PEM key (from the service account JSON)");
575
609
  }
576
610
  setSecret(tenant, name, SERVICE_ACCOUNT_PREFIX + JSON.stringify(config), workspace, "service-account");
@@ -692,7 +726,7 @@ export function setSshSecret(tenant: string, name: string, config: SshConfig, wo
692
726
  const hasKey = Boolean(config.private_key?.trim());
693
727
  const hasPassword = Boolean(config.password);
694
728
  if (hasKey === hasPassword) throw new Error("ssh secret needs a private key or a password (not both)");
695
- if (hasKey && !/BEGIN [A-Z ]*PRIVATE KEY/.test(config.private_key!)) {
729
+ if (hasKey && !/BEGIN [A-Z ]{0,40}PRIVATE KEY/.test(config.private_key!)) {
696
730
  throw new Error("private_key must be a PEM/OpenSSH private key");
697
731
  }
698
732
  if (hasPassword && /[\n\r]/.test(config.password!)) throw new Error("ssh password cannot contain newlines");
@@ -737,7 +771,7 @@ export function setApiSecret(tenant: string, name: string, config: ApiConfig, wo
737
771
  if (base && !/^https?:\/\//.test(base)) throw new Error("base_url must be http(s)");
738
772
  setSecret(
739
773
  tenant, name,
740
- API_PREFIX + JSON.stringify({ ...(base ? { base_url: base.replace(/\/+$/, "") } : {}), headers: config.headers }),
774
+ API_PREFIX + JSON.stringify({ ...(base ? { base_url: trimSlashes(base) } : {}), headers: config.headers }),
741
775
  workspace, "api",
742
776
  );
743
777
  }
package/src/starter.ts CHANGED
@@ -355,7 +355,10 @@ Indented under a step, any of:
355
355
  | \`preview: draft/*.mdx\` | what the approval box shows — paths under \`storage/\`, globs allowed |
356
356
  | \`delegate: [[a]], [[b]]\` | hand the step to whichever fits, up to 5 |
357
357
  | \`output: json\` | the reply ends with one JSON value; the next step gets it as data, \`each: items\` fans out over it |
358
+ | \`schema: {...}\` | with \`output: json\` — the JSON Schema the value must satisfy; inline, a YAML block under the key, or a file path |
358
359
  | \`each: items\` | run once per element of the previous step's JSON array |
360
+ | \`parallel: 3\` | how many \`each:\` instances run at once — default all |
361
+ | \`max_turns: 20\` | the most model turns before the step is stopped, beside \`budget:\` and \`timeout:\` |
359
362
 
360
363
  ## Writing an agent
361
364
 
package/src/step-exec.ts CHANGED
@@ -11,10 +11,13 @@
11
11
  import fs from "node:fs";
12
12
  import path from "node:path";
13
13
  import { query } from "@anthropic-ai/claude-agent-sdk";
14
- import type { McpServerConfig } from "@anthropic-ai/claude-agent-sdk";
14
+ import type { McpServerConfig, SDKMessage } from "@anthropic-ai/claude-agent-sdk";
15
15
  import type { Effort } from "./store.ts";
16
+ import type { TestEffect } from "./test-mode.ts";
16
17
  import { spawn } from "node:child_process";
17
18
  import { checkPaths, checkBash, isFilesystemTool } from "./confine.ts";
19
+ import { hostSafeEnv } from "./host-env.ts";
20
+ import { validateSchema, describeSchemaErrors, looksLikeSchema } from "./json-schema.ts";
18
21
 
19
22
  export interface ExecOutcome {
20
23
  status: "completed" | "failed";
@@ -57,6 +60,9 @@ export interface ExecOptions {
57
60
  * The runner derives it from the flow's `budget:` and what the run has
58
61
  * already spent; absent means no ceiling on this step. */
59
62
  budgetUsd?: number | null;
63
+ /** Which line set the ceiling, for the error that names it — "budget: in
64
+ * the flow file" or "budget: on the <agent> agent". */
65
+ budgetNote?: string;
60
66
  /** Price per token for the model, when the catalogue knows it, so spend
61
67
  * can be counted turn by turn instead of learned at the end. Without it
62
68
  * the step prices its turns from the SDK's own running total. */
@@ -69,15 +75,38 @@ export interface ExecOptions {
69
75
  /** `output: json` — the reply must carry one JSON value; extracting it is
70
76
  * part of finishing the step, and failing to is failing the step. */
71
77
  output?: "json";
78
+ /** `schema:` — what that value must look like; a value that does not fit
79
+ * fails the step naming the field. See json-schema.ts. */
80
+ schema?: Record<string, unknown> | boolean;
81
+ /** `max_turns:` — the most model turns before the step is stopped. The
82
+ * SDK enforces it; the step reads the result's reason and says so. */
83
+ maxTurns?: number;
72
84
  /** false when the caller is already an isolation boundary (a run
73
85
  * container): the SDK's bash sandbox is then redundant and would block
74
86
  * declared network use. Default (undefined/true) keeps it on. */
75
87
  sandboxBash?: boolean;
88
+ /** Has a person asked for the run to stop? Polled every couple of
89
+ * seconds while the model loop runs; true interrupts it. The in-process
90
+ * runner answers from the run record (stopRun writes there from another
91
+ * process); a container has a sandbox that is destroyed instead. */
92
+ stopRequested?: () => boolean;
76
93
  emit: (type: "text" | "tool" | "info" | "error", text: string, extra?: EventExtra) => void;
77
94
  }
78
95
 
96
+ /** How often the stop flag is read while a step runs. */
97
+ export const STOP_POLL_MS = 2000;
98
+ /** After an interrupt is asked for, how long the model loop gets to wind
99
+ * down on its own before the query is aborted outright. */
100
+ export const INTERRUPT_GRACE_MS = 10_000;
101
+
102
+ /** The model loop, as executeStep needs it — the SDK's `query`, or a test's
103
+ * stand-in. Only the shape the loop reads. */
104
+ export type QueryLike = AsyncIterable<unknown> & { interrupt(): Promise<unknown> };
105
+ export type QueryFn = (args: { prompt: string; options: Record<string, unknown> }) => QueryLike;
106
+
79
107
  /** The pairing fields on a tool event — see RunEvent in store.ts. */
80
- export type EventExtra = { call?: string; ms?: number; err?: boolean };
108
+ export type EventExtra = { call?: string; ms?: number; err?: boolean; effect?: TestEffect };
109
+
81
110
 
82
111
  /** Conservative per-token rates for a model nothing else can price — an
83
112
  * unknown id on a gateway with no catalogue. Opus-class, so a ceiling
@@ -145,7 +174,34 @@ export function stepCeiling(budgetUsd: number | null | undefined, spentUsd: numb
145
174
  return Math.max(0, budgetUsd - spentUsd) / Math.max(1, launching);
146
175
  }
147
176
 
148
- export async function executeStep(opts: ExecOptions): Promise<ExecOutcome> {
177
+ /**
178
+ * The ceiling a step actually runs under, and which line set it.
179
+ *
180
+ * Two caps can apply to one step: the flow's `budget:` gives it an equal
181
+ * share of what the run has left; the agent's own `budget:` is the most that
182
+ * agent may spend in one run, less what its earlier steps in this run cost.
183
+ * The tighter one wins, and the note says which, because "over budget" with
184
+ * no line to go and raise is a message that sends someone to the wrong file.
185
+ */
186
+ export function stepCeilingFor(
187
+ flowShareUsd: number | null,
188
+ agentBudgetUsd: number | null | undefined,
189
+ agentSpentUsd: number,
190
+ agent: string,
191
+ ): { ceilingUsd: number | null; note: string } {
192
+ const flow = { ceilingUsd: flowShareUsd, note: "budget: in the flow file" };
193
+ if (!agentBudgetUsd || agentBudgetUsd <= 0) return flow;
194
+ const left = Math.max(0, agentBudgetUsd - agentSpentUsd);
195
+ if (flowShareUsd !== null && flowShareUsd <= left) return flow;
196
+ return { ceilingUsd: left, note: `budget: on the ${agent} agent` };
197
+ }
198
+
199
+ export async function executeStep(
200
+ opts: ExecOptions,
201
+ /** The model loop. Injected so the clock and the stop can be tested
202
+ * without a model; production passes nothing and gets the SDK. */
203
+ runQuery: QueryFn = query as unknown as QueryFn,
204
+ ): Promise<ExecOutcome> {
149
205
  const { agentDir, workspaceRoot, libraryRoot, emit } = opts;
150
206
  let status: "running" | "completed" | "failed" = "running";
151
207
  let costUsd: number | null = null;
@@ -155,16 +211,23 @@ export async function executeStep(opts: ExecOptions): Promise<ExecOutcome> {
155
211
  fs.mkdirSync(path.join(agentDir, "outputs"), { recursive: true });
156
212
  fs.mkdirSync(path.join(agentDir, "memory"), { recursive: true });
157
213
 
158
- const q = query({
214
+ // The one handle that can end the loop from outside: the timeout and a
215
+ // stop both ask the query to interrupt, and abort it outright if it has
216
+ // not wound down within the grace period.
217
+ const abort = new AbortController();
218
+
219
+ const q = runQuery({
159
220
  prompt: opts.prompt,
160
221
  options: {
161
222
  cwd: agentDir,
223
+ abortController: abort,
162
224
  model: opts.model,
163
225
  // Omitted, not passed as undefined-with-a-default: an unset effort
164
226
  // should mean "whatever this model does normally", which is not a
165
227
  // level we can name — it moves as models ship.
166
228
  ...(opts.effort ? { effort: opts.effort } : {}),
167
229
  systemPrompt: opts.systemPrompt,
230
+ ...(opts.maxTurns ? { maxTurns: opts.maxTurns } : {}),
168
231
  // Restrict the toolset itself, not just approval: an agent that
169
232
  // declares no tools gets none, instead of seeing the full Claude
170
233
  // Code toolset and burning turns on denied calls.
@@ -218,7 +281,34 @@ export async function executeStep(opts: ExecOptions): Promise<ExecOutcome> {
218
281
  },
219
282
  });
220
283
 
221
- const deadline = opts.timeoutSec ? Date.now() + opts.timeoutSec * 1000 : null;
284
+ // Why the loop was ended from outside, if it was. The deadline used to be
285
+ // checked only when a message arrived, so a step waiting on one long tool
286
+ // call — a crawl, a build — sailed past its timeout: with no message there
287
+ // was nothing to check it against. The backstop timer fires regardless,
288
+ // and a stop is read on a clock rather than between groups.
289
+ let ended: "timeout" | "stopped" | null = null;
290
+ const timers: NodeJS.Timeout[] = [];
291
+ const endWith = (why: "timeout" | "stopped") => {
292
+ if (ended) return;
293
+ ended = why;
294
+ void q.interrupt().catch(() => {});
295
+ timers.push(setTimeout(() => abort.abort(), INTERRUPT_GRACE_MS));
296
+ };
297
+ if (opts.timeoutSec) timers.push(setTimeout(() => endWith("timeout"), opts.timeoutSec * 1000));
298
+ if (opts.stopRequested) {
299
+ timers.push(setInterval(() => {
300
+ try {
301
+ if (opts.stopRequested!()) endWith("stopped");
302
+ } catch {
303
+ // an unreadable record is not a stop
304
+ }
305
+ }, STOP_POLL_MS));
306
+ }
307
+ // The clocks stay referenced on purpose: a step hanging in a tool call may
308
+ // be the only thing keeping the loop alive, and an unreferenced timer
309
+ // would let the process exit around it instead of ending it. They are all
310
+ // cleared when the step ends, so nothing outlives the step.
311
+
222
312
  // The hard cap. Spend is counted as each assistant turn arrives — its
223
313
  // usage is in the message — so the step can stop at the ceiling rather
224
314
  // than learn afterwards that it crossed it. Cache traffic is counted as
@@ -231,18 +321,15 @@ export async function executeStep(opts: ExecOptions): Promise<ExecOutcome> {
231
321
  // its call and the trace can say how long each tool ran.
232
322
  const openCalls = new Map<string, { name: string; at: number }>();
233
323
 
234
- for await (const message of q) {
235
- if (deadline && Date.now() > deadline) {
236
- emit("error", `step exceeded its ${opts.timeoutSec}s timeout`);
237
- status = "failed";
238
- break;
239
- }
324
+ try {
325
+ for await (const message of q as AsyncIterable<SDKMessage>) {
326
+ if (ended) break;
240
327
  if (message.type === "assistant") {
241
328
  const u = (message.message as unknown as { usage?: Record<string, number | undefined> }).usage;
242
329
  if (ceiling && u) {
243
330
  spentUsd += priceTurn(u, opts.price ?? null);
244
331
  if (spentUsd >= ceiling) {
245
- emit("error", `over budget — this step reached $${spentUsd.toFixed(4)} of its $${ceiling.toFixed(4)} ceiling mid-turn and was stopped (budget: in the flow file)`);
332
+ emit("error", `over budget — this step reached $${spentUsd.toFixed(4)} of its $${ceiling.toFixed(4)} ceiling mid-turn and was stopped (${opts.budgetNote ?? "budget: in the flow file"})`);
246
333
  status = "failed";
247
334
  break;
248
335
  }
@@ -275,6 +362,9 @@ export async function executeStep(opts: ExecOptions): Promise<ExecOutcome> {
275
362
  }
276
363
  } else if (message.type === "result") {
277
364
  status = message.subtype === "success" ? "completed" : "failed";
365
+ if (message.subtype === "error_max_turns") {
366
+ emit("error", `stopped after ${opts.maxTurns ?? "its"} turns (max_turns: in the flow file) — the step did not finish`);
367
+ }
278
368
  costUsd = "total_cost_usd" in message ? (message.total_cost_usd ?? null) : null;
279
369
  if ("usage" in message && message.usage) {
280
370
  const u = message.usage as unknown as Record<string, number | undefined>;
@@ -291,6 +381,19 @@ export async function executeStep(opts: ExecOptions): Promise<ExecOutcome> {
291
381
  }
292
382
  }
293
383
  }
384
+ } catch (err) {
385
+ // An abort we asked for is not an error of the step's; anything else is.
386
+ if (!ended) throw err;
387
+ } finally {
388
+ for (const t of timers) clearTimeout(t);
389
+ }
390
+ if (ended === "timeout") {
391
+ emit("error", `timed out after ${opts.timeoutSec}s (timeout: in the flow file) — the step was stopped with the files it wrote so far kept`);
392
+ status = "failed";
393
+ } else if (ended === "stopped") {
394
+ emit("error", "stopped by a person mid-step");
395
+ status = "failed";
396
+ }
294
397
  // The SDK ends a healthy run with a `result` message. A stream that just
295
398
  // stops — subprocess OOM-killed, crashed, or torn down — used to fall
296
399
  // through as "completed", which reported a run that produced nothing as a
@@ -319,6 +422,14 @@ export async function executeStep(opts: ExecOptions): Promise<ExecOutcome> {
319
422
  if (extracted.ok) {
320
423
  data = extracted.value;
321
424
  emit("info", `output: json — ${describeJson(data)}`);
425
+ // The declared shape, checked here beside the extraction: a value
426
+ // that parses but is not what the next step was promised is the same
427
+ // failure as no value, and the message names the field.
428
+ const errors = opts.schema !== undefined && looksLikeSchema(opts.schema) ? validateSchema(data, opts.schema) : [];
429
+ if (errors.length) {
430
+ emit("error", `schema: the value does not fit — ${describeSchemaErrors(errors)}`);
431
+ status = "failed";
432
+ }
322
433
  } else {
323
434
  emit("error", `output: json — ${extracted.reason}`);
324
435
  status = "failed";
@@ -515,9 +626,13 @@ function runVerify(
515
626
  return new Promise((resolve) => {
516
627
  // No clock of the platform's: a verify like `npm run build` takes what
517
628
  // it takes, and the step's own `timeout:` is the bound if the flow set one.
629
+ // The allowlisted host base plus what the caller passed: the step's
630
+ // secrets and identifiers on the in-process path (runner.ts), the
631
+ // container's own environment on the isolated one (run-container.ts).
632
+ // Never process.env whole — see host-env.ts.
518
633
  const child = spawn("bash", ["-lc", command], {
519
634
  cwd: agentDir,
520
- env: { ...process.env, ...env },
635
+ env: { ...hostSafeEnv(), ...env },
521
636
  stdio: ["pipe", "pipe", "pipe"],
522
637
  });
523
638
  // An `output: json` step's data arrives on stdin, so a check can be
package/src/storage.ts CHANGED
@@ -302,8 +302,14 @@ export function s3Config(): S3Config | null {
302
302
  };
303
303
  }
304
304
 
305
- const sha256 = (b: crypto.BinaryLike) => crypto.createHash("sha256").update(b).digest("hex");
306
- const hmac = (key: crypto.BinaryLike, data: string) =>
305
+ // Narrower than crypto.BinaryLike on purpose. BinaryLike also covers views
306
+ // over a SharedArrayBuffer, which .update() has never accepted and which
307
+ // @types/node 26 finally says so about — the old typings let this compile by
308
+ // being vaguer than the runtime. Every caller here passes a string or a
309
+ // Buffer, so this is the honest signature, not a workaround.
310
+ type Hashable = string | Uint8Array;
311
+ const sha256 = (b: Hashable) => crypto.createHash("sha256").update(b).digest("hex");
312
+ const hmac = (key: Hashable, data: string) =>
307
313
  crypto.createHmac("sha256", key).update(data, "utf8").digest();
308
314
 
309
315
  /** RFC 3986. encodeURIComponent leaves !'()* alone and SigV4 does not. */