gentle-pi 3.2.1 → 3.3.0

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 (71) hide show
  1. package/docs/gentle-shell.md +27 -14
  2. package/docs/readme-reference.md +41 -7
  3. package/docs/review-integration.md +22 -17
  4. package/extensions/gentle-agents.ts +85 -17
  5. package/extensions/gentle-ai.ts +123 -21
  6. package/extensions/gentle-shell.ts +379 -31
  7. package/extensions/gentle-todo.ts +19 -1
  8. package/lib/agents-view.ts +41 -14
  9. package/lib/agents-widget.ts +84 -13
  10. package/lib/command-palette-catalog.ts +1 -0
  11. package/lib/double-esc-cancel-policy.ts +138 -0
  12. package/lib/inprocess-reviewer.ts +260 -0
  13. package/lib/native-review-cli.ts +14 -0
  14. package/lib/odd-runtime-delegation-gate.ts +88 -0
  15. package/lib/review-host-relay.ts +246 -171
  16. package/lib/review-integration-v2.ts +110 -26
  17. package/lib/shell-bar.ts +150 -75
  18. package/lib/shell-card.ts +19 -9
  19. package/lib/shell-changes-view.ts +43 -5
  20. package/lib/shell-changes.ts +92 -5
  21. package/lib/shell-hover.ts +39 -0
  22. package/lib/shell-prompt.ts +10 -1
  23. package/lib/shell-sidebar-layout.ts +111 -15
  24. package/lib/shell-sidebar.ts +16 -0
  25. package/lib/shell-todo.ts +7 -1
  26. package/lib/shell-usage-view.ts +98 -10
  27. package/package.json +1 -1
  28. package/runtime/native-review-cli.mjs +14 -0
  29. package/runtime/review-integration-v2.mjs +110 -26
  30. package/scripts/gentle-ai-installer.mjs +10 -10
  31. package/scripts/maintainer/provider-relay-matrix.mjs +118 -47
  32. package/scripts/verify-package-files.mjs +3 -3
  33. package/tests/agents-grouping.test.ts +75 -18
  34. package/tests/agents-view.test.ts +28 -18
  35. package/tests/agents-widget.test.ts +100 -12
  36. package/tests/command-palette.test.ts +1 -0
  37. package/tests/devbinary/pi-host-relay.devtest.ts +176 -138
  38. package/tests/double-esc-cancel-policy.test.ts +194 -0
  39. package/tests/gentle-agents.test.ts +528 -5
  40. package/tests/gentle-ai-binary.test.ts +1 -1
  41. package/tests/gentle-ai-installer.test.ts +47 -47
  42. package/tests/gentle-ai.test.ts +69 -5
  43. package/tests/gentle-shell.test.ts +795 -23
  44. package/tests/gentle-todo.test.ts +17 -4
  45. package/tests/inprocess-reviewer.test.ts +368 -0
  46. package/tests/maintainer/provider-relay.maintest.ts +101 -143
  47. package/tests/native-review-capability-contract.test.ts +19 -1
  48. package/tests/odd-runtime-delegation-gate.test.ts +212 -0
  49. package/tests/orchestrator-rdd-ownership.test.ts +3 -3
  50. package/tests/package-manifest.test.ts +6 -6
  51. package/tests/review-host-relay-routing.test.ts +77 -0
  52. package/tests/review-host-relay.test.ts +277 -300
  53. package/tests/review-integration-v2-forward.test.ts +61 -0
  54. package/tests/review-integration-v2.test.ts +116 -1
  55. package/tests/review-relay-transport-agent.test.ts +22 -24
  56. package/tests/runtime-harness.mjs +11 -0
  57. package/tests/session-changes-shell.test.ts +27 -0
  58. package/tests/session-worktree-registry.test.ts +41 -0
  59. package/tests/shell-bar.test.ts +192 -124
  60. package/tests/shell-card.test.ts +5 -3
  61. package/tests/shell-changes-view.test.ts +47 -0
  62. package/tests/shell-changes.test.ts +177 -0
  63. package/tests/shell-hover.test.ts +19 -0
  64. package/tests/shell-prompt.test.ts +20 -0
  65. package/tests/shell-sidebar-fullscreen.test.ts +59 -0
  66. package/tests/shell-sidebar-layout.test.ts +243 -5
  67. package/tests/shell-sidebar.test.ts +25 -1
  68. package/tests/shell-todo.test.ts +36 -0
  69. package/tests/shell-usage-view.test.ts +120 -1
  70. package/lib/opaque-pi-reviewer-adapter.ts +0 -404
  71. package/tests/opaque-pi-reviewer-adapter.test.ts +0 -410
@@ -4,9 +4,9 @@ import { tmpdir } from "node:os";
4
4
  import { join } from "node:path";
5
5
  import test from "node:test";
6
6
  import { createNodeExecFileAdapter } from "../lib/native-review-cli.ts";
7
+ import { INPROCESS_REVIEWER_FAILURE, type InProcessReviewerOutcome, type InProcessReviewerRequest, type InProcessReviewerDeps, runInProcessReviewer } from "../lib/inprocess-reviewer.ts";
7
8
  import {
8
9
  REVIEW_HOST_RELAY_FAILURE,
9
- REVIEW_HOST_RELAY_PI_ARGV,
10
10
  REVIEW_HOST_RELAY_PI_TIMEOUT_ENV,
11
11
  REVIEW_HOST_RELAY_PI_TIMEOUT_FLOOR_MS,
12
12
  REVIEW_HOST_RELAY_PI_TIMEOUT_MAX_MS,
@@ -27,14 +27,26 @@ import {
27
27
  type ReviewHostRelayPreparedResult,
28
28
  type ReviewHostRelayRequest,
29
29
  } from "../lib/review-host-relay.ts";
30
+ import * as reviewHostRelayModule from "../lib/review-host-relay.ts";
30
31
  import { GENTLE_PI_REVIEW_RELAY_CONTRACT, GENTLE_PI_REVIEW_RELAY_CONTRACT_ENV } from "../lib/review-relay-contract.ts";
31
32
  import { decodeReviewNextTransitionV3, type ReviewCaptureSubmissionV1, type ReviewCollectInputV3 } from "../lib/review-integration-v2.ts";
32
33
 
34
+ // gentle-pi#311 P2: the reviewer transport this file exercises no longer
35
+ // spawns a pi child. `runReviewHostRelaySlot`/`prepareReviewHostRelaySlot`
36
+ // take an injectable `runReviewer` (defaulting to the real
37
+ // `runInProcessReviewer`), so every test below fakes that seam instead of a
38
+ // fake `pi` executable on disk. `lib/inprocess-reviewer.ts`'s own tests own
39
+ // the completion's internal behavior (model resolution, auth, thinking
40
+ // mapping, text extraction); this file owns the relay's mapping of that
41
+ // completion's outcome onto a typed `ReviewHostRelayError`, plus everything
42
+ // unrelated to the reviewer transport (materialize, submission, admission,
43
+ // slot detection) which stayed exactly as it was.
44
+
33
45
  // ---------------------------------------------------------------------------
34
- // Fake binaries. Following the repo's fake-executable idiom (shell/git
35
- // wrappers in related review tests), these are
36
- // shebang scripts written into a scratch directory; node scripts are used so
37
- // binary-unsafe bytes survive verbatim.
46
+ // Fake gentle-ai binary. Following the repo's fake-executable idiom (shell/git
47
+ // wrappers in related review tests), this is a shebang script written into a
48
+ // scratch directory; a node script is used so binary-unsafe bytes survive
49
+ // verbatim.
38
50
  // ---------------------------------------------------------------------------
39
51
 
40
52
  const FAKE_GENTLE_AI = `#!/usr/bin/env node
@@ -94,52 +106,10 @@ process.stderr.write("unexpected fake gentle-ai invocation\\n");
94
106
  process.exit(9);
95
107
  `;
96
108
 
97
- const FAKE_PI = `#!/usr/bin/env node
98
- const fs = require("node:fs");
99
- const path = require("node:path");
100
- const argv = process.argv.slice(2);
101
- if (process.env.RELAY_FAKE_PI_LOG) fs.appendFileSync(process.env.RELAY_FAKE_PI_LOG, JSON.stringify({ argv, cwd: process.cwd(), entries: fs.readdirSync(process.cwd()), contract: process.env.${GENTLE_PI_REVIEW_RELAY_CONTRACT_ENV} ?? null }) + "\\n");
102
- const chunks = [];
103
- process.stdin.on("data", (chunk) => chunks.push(chunk));
104
- process.stdin.on("end", () => {
105
- const stdin = Buffer.concat(chunks);
106
- if (process.env.RELAY_FAKE_PI_STDIN_CAPTURE) fs.writeFileSync(process.env.RELAY_FAKE_PI_STDIN_CAPTURE, stdin);
107
- const finish = () => {
108
- const mode = process.env.RELAY_FAKE_PI_MODE || "ok";
109
- if (process.env.RELAY_FAKE_PI_BREAK_CLEANUP === "true") fs.chmodSync(path.dirname(process.cwd()), 0o500);
110
- if (mode === "ok") { process.stdout.write(Buffer.from(process.env.RELAY_FAKE_PI_OUTPUT_B64 || "", "base64")); process.exit(0); }
111
- if (mode === "empty") process.exit(0);
112
- if (mode === "hang") { setTimeout(() => process.exit(0), 10_000); return; }
113
- process.stderr.write("pi exploded\\n");
114
- process.exit(4);
115
- };
116
- const barrierDirectory = process.env.RELAY_FAKE_PI_BARRIER_DIR;
117
- if (barrierDirectory === undefined) { finish(); return; }
118
- fs.mkdirSync(barrierDirectory, { recursive: true });
119
- fs.writeFileSync(path.join(barrierDirectory, String(process.pid)), "started");
120
- const expected = Number(process.env.RELAY_FAKE_PI_BARRIER_COUNT || "4");
121
- const deadline = Date.now() + 1_000;
122
- const releaseFile = process.env.RELAY_FAKE_PI_RELEASE_FILE;
123
- const waitForGroup = () => {
124
- if (fs.readdirSync(barrierDirectory).length < expected) {
125
- if (Date.now() >= deadline) { process.stderr.write("reviewer barrier timed out\\n"); process.exit(5); return; }
126
- setTimeout(waitForGroup, 10);
127
- return;
128
- }
129
- if (releaseFile !== undefined && !fs.existsSync(releaseFile)) { setTimeout(waitForGroup, 10); return; }
130
- finish();
131
- };
132
- waitForGroup();
133
- });
134
- `;
135
-
136
109
  interface RelayHarness {
137
110
  directory: string;
138
111
  gentleAi: string;
139
- pi: string;
140
112
  logPath: string;
141
- piLogPath: string;
142
- stdinCapturePath: string;
143
113
  submitCapturePath: string;
144
114
  targetCwd: string;
145
115
  environment: NodeJS.ProcessEnv;
@@ -149,32 +119,25 @@ function harness(t: test.TestContext, overrides: Record<string, string> = {}): R
149
119
  const directory = realpathSync(mkdtempSync(join(tmpdir(), "gentle-pi-relay-harness-")));
150
120
  t.after(() => rmSync(directory, { recursive: true, force: true }));
151
121
  const gentleAi = join(directory, "gentle-ai");
152
- const pi = join(directory, "pi");
153
122
  writeFileSync(gentleAi, FAKE_GENTLE_AI);
154
123
  chmodSync(gentleAi, 0o755);
155
- writeFileSync(pi, FAKE_PI);
156
- chmodSync(pi, 0o755);
157
124
  const logPath = join(directory, "gentle-ai.log");
158
- const piLogPath = join(directory, "pi.log");
159
- const stdinCapturePath = join(directory, "pi-stdin.bin");
160
125
  const submitCapturePath = join(directory, "submitted.bin");
161
126
  const targetCwd = join(directory, "target-worktree");
162
127
  mkdirSync(targetCwd);
163
128
  const environment: NodeJS.ProcessEnv = {
164
129
  ...process.env,
165
130
  RELAY_FAKE_LOG: logPath,
166
- RELAY_FAKE_PI_LOG: piLogPath,
167
- RELAY_FAKE_PI_STDIN_CAPTURE: stdinCapturePath,
168
131
  RELAY_FAKE_SUBMIT_CAPTURE: submitCapturePath,
169
132
  ...overrides,
170
133
  };
171
134
  // The relay itself must add the handshake; the base environment never
172
135
  // carries it, so the fake-binary log proves the injection.
173
136
  delete environment[GENTLE_PI_REVIEW_RELAY_CONTRACT_ENV];
174
- return { directory, gentleAi, pi, logPath, piLogPath, stdinCapturePath, submitCapturePath, targetCwd, environment };
137
+ return { directory, gentleAi, logPath, submitCapturePath, targetCwd, environment };
175
138
  }
176
139
 
177
- function readLog(path: string): Array<{ argv: string[]; contract: string | null; cwd?: string; entries?: string[] }> {
140
+ function readLog(path: string): Array<{ argv: string[]; contract: string | null; cwd?: string }> {
178
141
  if (!existsSync(path)) return [];
179
142
  return readFileSync(path, "utf8").split("\n").filter((line) => line.length > 0).map((line) => JSON.parse(line));
180
143
  }
@@ -207,44 +170,61 @@ const SUBMISSION: ReviewCaptureSubmissionV1 = Object.freeze({
207
170
  values: Object.freeze([{ slot: "reviewer_result", domain: "artifact_path_or_stdin", substitutionLocation: BINDING_TOKENS.length }]),
208
171
  });
209
172
 
210
- // Prompt and result bytes deliberately include binary-unsafe content: NUL,
211
- // control bytes, quotes, backslashes, CRLF, multi-byte UTF-8, and bytes that
212
- // are not valid UTF-8 at all. The relay must move them verbatim.
173
+ // Prompt bytes deliberately include binary-unsafe content: NUL, control
174
+ // bytes, quotes, backslashes, CRLF, multi-byte UTF-8, and bytes that are not
175
+ // valid UTF-8 at all. The relay must materialize them verbatim into the
176
+ // completion request.
213
177
  const PROMPT_BYTES = Buffer.concat([
214
178
  Buffer.from('GENTLE_AI_REVIEW_BINDING {"lineage":"review-1d5aadacc600e167"}\n"quotes" \\backslash\r\n\u00e9\u{1F3A9}\n', "utf8"),
215
179
  Buffer.from([0x00, 0x01, 0x07, 0xff, 0xfe, 0x00]),
216
180
  ]);
217
- // The reviewer child runs `pi --mode json` and the transport submits the
218
- // assistant text of that event stream, so the fixture wraps the reviewer
219
- // payload in the minimal event shape the extraction understands.
220
- const PI_REVIEWER_TEXT = `{"subject_hash":"sha256:${"a".repeat(64)}","findings":[]}\n🎉\r\n`;
221
- const piEventStream = (...events: readonly unknown[]): Buffer => Buffer.from(events.map((event) => JSON.stringify(event)).join("\n") + "\n", "utf8");
222
- const assistantEnd = (text: string, reviewerModel?: string) => {
223
- const message: Record<string, unknown> = { role: "assistant", content: [{ type: "text", text }] };
224
- if (reviewerModel !== undefined) message["model"] = reviewerModel;
225
- return { type: "message_end", message };
226
- };
227
- const PI_OUTPUT_BYTES = piEventStream(
228
- { type: "message_start", message: { role: "user", content: [] } },
229
- assistantEnd(PI_REVIEWER_TEXT),
230
- );
231
-
232
- function relayRequest(fixture: RelayHarness, overrides: Record<string, unknown> = {}) {
181
+ // The reviewer's completion text — what used to be the pi event stream's
182
+ // extracted assistant text is now simply the in-process outcome's `text`.
183
+ const REVIEWER_TEXT = `{"subject_hash":"sha256:${"a".repeat(64)}","findings":[]}\n🎉\r\n`;
184
+
185
+ function relayRequest(fixture: RelayHarness, overrides: Record<string, unknown> = {}): ReviewHostRelayRequest {
233
186
  return {
234
187
  captureArgumentTokens: CAPTURE_TOKENS,
235
188
  submission: SUBMISSION,
236
189
  gentleAiExecutable: fixture.gentleAi,
237
- piExecutable: fixture.pi,
238
190
  targetCwd: fixture.targetCwd,
191
+ reviewerRegistry: {
192
+ find: () => ({ id: "gpt-5", name: "GPT-5", api: "openai-responses", provider: "openai", baseUrl: "https://example.invalid", reasoning: true, input: ["text"], cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }, contextWindow: 100_000, maxTokens: 8192 }) as never,
193
+ getApiKeyAndHeaders: async () => ({ ok: true, apiKey: "test-key" }),
194
+ },
195
+ selection: "openai/gpt-5",
196
+ routingKey: "review-reliability",
239
197
  environment: {
240
198
  ...fixture.environment,
241
199
  RELAY_FAKE_PROMPT_B64: PROMPT_BYTES.toString("base64"),
242
- RELAY_FAKE_PI_OUTPUT_B64: PI_OUTPUT_BYTES.toString("base64"),
243
200
  },
244
201
  gentleAiTimeoutMs: 30_000,
245
202
  piTimeoutMs: 30_000,
246
203
  ...overrides,
247
- };
204
+ } as ReviewHostRelayRequest;
205
+ }
206
+
207
+ /** A `runReviewer` fake resolving with canned completion text, capturing every call. */
208
+ function textReviewer(text: string, reviewerModel = "openai/gpt-5") {
209
+ const calls: InProcessReviewerRequest[] = [];
210
+ const runReviewer = (async (request: InProcessReviewerRequest, _deps: InProcessReviewerDeps): Promise<InProcessReviewerOutcome> => {
211
+ calls.push(request);
212
+ return { kind: "text", text, reviewerModel };
213
+ }) as typeof runInProcessReviewer;
214
+ return { runReviewer, calls };
215
+ }
216
+
217
+ /** A `runReviewer` fake resolving with a canned typed refusal. */
218
+ function refusingReviewer(code: (typeof INPROCESS_REVIEWER_FAILURE)[keyof typeof INPROCESS_REVIEWER_FAILURE], message: string, evidence?: Record<string, unknown>, delayMs = 0): typeof runInProcessReviewer {
219
+ return (async () => {
220
+ if (delayMs > 0) await new Promise((resolve) => setTimeout(resolve, delayMs));
221
+ return { kind: "refused", code, message, ...(evidence === undefined ? {} : { evidence }) };
222
+ }) as typeof runInProcessReviewer;
223
+ }
224
+
225
+ /** Runs one relay slot with a safe-by-default fake reviewer; no test may reach the real network unless it explicitly overrides this. */
226
+ function runRelay(fixture: RelayHarness, overrides: Record<string, unknown> = {}, runReviewer: typeof runInProcessReviewer = textReviewer(REVIEWER_TEXT).runReviewer): Promise<{ promptByteLength: number; resultByteLength: number; submission: string }> {
227
+ return runReviewHostRelaySlot(relayRequest(fixture, overrides), runReviewer);
248
228
  }
249
229
 
250
230
  function reviewerGroupRequests(fixture: RelayHarness): ReviewHostRelayRequest[] {
@@ -252,15 +232,14 @@ function reviewerGroupRequests(fixture: RelayHarness): ReviewHostRelayRequest[]
252
232
  captureArgumentTokens: CAPTURE_TOKENS.map((token) => token === "--lens=review-reliability"
253
233
  ? `--lens=${lens}`
254
234
  : token === "--order=0" ? `--order=${order}` : token),
235
+ routingKey: lens,
255
236
  environment: {
256
237
  ...fixture.environment,
257
238
  RELAY_FAKE_PROMPT_B64: PROMPT_BYTES.toString("base64"),
258
- RELAY_FAKE_PI_OUTPUT_B64: PI_OUTPUT_BYTES.toString("base64"),
259
239
  },
260
240
  }));
261
241
  }
262
242
 
263
-
264
243
  async function rejectsWithRelayError(promise: Promise<unknown>, kind: string, stage: string): Promise<ReviewHostRelayError> {
265
244
  let caught: ReviewHostRelayError | undefined;
266
245
  await assert.rejects(promise, (error: unknown) => {
@@ -299,28 +278,36 @@ test("relay contract constants are the compiled gentle-ai handshake values", ()
299
278
  assert.equal(GENTLE_PI_REVIEW_RELAY_CONTRACT, "gentle-pi.review-relay/v1");
300
279
  });
301
280
 
281
+ test("no pi child launcher is exported: the lens path completes in-process", () => {
282
+ const relayExports = reviewHostRelayModule as unknown as Record<string, unknown>;
283
+ assert.equal("REVIEW_HOST_RELAY_PI_ARGV" in relayExports, false, "the pinned pi argv export must be gone");
284
+ assert.equal("resolveReviewHostRelayExtensionPaths" in relayExports, false, "the extension allowlist reader must be gone");
285
+ });
286
+
302
287
  // ---------------------------------------------------------------------------
303
288
  // Happy path
304
289
  // ---------------------------------------------------------------------------
305
290
 
306
- test("relay happy path moves prompt bytes verbatim and the pi event stream's assistant text through a fresh empty scratch pi subprocess", async (t) => {
291
+ test("relay happy path moves prompt bytes verbatim into the completion and submits its text through the exact provider submission token", async (t) => {
307
292
  const fixture = harness(t);
308
- const result = await runReviewHostRelaySlot(relayRequest(fixture));
293
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
294
+ const result = await runRelay(fixture, {}, runReviewer);
309
295
 
310
296
  assert.equal(result.promptByteLength, PROMPT_BYTES.length);
311
- assert.equal(result.resultByteLength, Buffer.byteLength(PI_REVIEWER_TEXT));
297
+ assert.equal(result.resultByteLength, Buffer.byteLength(REVIEWER_TEXT));
312
298
  assert.equal(JSON.parse(result.submission).admission_decision, "completed");
313
299
 
314
- // Prompt bytes reached pi stdin verbatim.
315
- assert.deepEqual(readFileSync(fixture.stdinCapturePath), PROMPT_BYTES);
316
- // Submission --input file bytes are EXACTLY the reviewer's assistant text.
317
- assert.deepEqual(readFileSync(fixture.submitCapturePath), Buffer.from(PI_REVIEWER_TEXT, "utf8"));
300
+ // Prompt bytes reached the completion request verbatim.
301
+ assert.equal(calls.length, 1);
302
+ assert.ok(calls[0]!.prompt.equals(PROMPT_BYTES));
303
+ // Submission --input file bytes are EXACTLY the reviewer's completion text.
304
+ assert.deepEqual(readFileSync(fixture.submitCapturePath), Buffer.from(REVIEWER_TEXT, "utf8"));
318
305
 
319
306
  const gentleAiCalls = readLog(fixture.logPath);
320
307
  assert.equal(gentleAiCalls.length, 2);
321
308
  // (a) exact provider tokens, verbatim, in provider order.
322
309
  assert.deepEqual(gentleAiCalls[0]!.argv, ["review", "capture-result", ...CAPTURE_TOKENS]);
323
- // (d) the provider-owned submission form, verbatim: its exact operation
310
+ // (b) the provider-owned submission form, verbatim: its exact operation
324
311
  // and argument tokens with only the artifact path substituted into the
325
312
  // declared {{value}} slot. No agent/materialize, nothing synthesized.
326
313
  assert.deepEqual(gentleAiCalls[1]!.argv.slice(0, 2 + BINDING_TOKENS.length), ["review", SUBMISSION.operationToken, ...BINDING_TOKENS]);
@@ -334,83 +321,74 @@ test("relay happy path moves prompt bytes verbatim and the pi event stream's ass
334
321
  // environment carried none.
335
322
  assert.deepEqual(gentleAiCalls.map((call) => call.contract), [GENTLE_PI_REVIEW_RELAY_CONTRACT, GENTLE_PI_REVIEW_RELAY_CONTRACT]);
336
323
  assert.deepEqual(gentleAiCalls.map((call) => call.cwd), [fixture.targetCwd, fixture.targetCwd]);
337
-
338
- const piCalls = readLog(fixture.piLogPath);
339
- assert.equal(piCalls.length, 1);
340
- // The pi environment stays untouched: no relay handshake is injected.
341
- assert.equal(piCalls[0]!.contract, null);
342
- // Fresh EMPTY scratch cwd, removed after the run.
343
- assert.deepEqual(piCalls[0]!.entries, []);
344
- assert.notEqual(piCalls[0]!.cwd, process.cwd());
345
- assert.equal(existsSync(piCalls[0]!.cwd!), false);
346
324
  });
347
325
 
348
- test("the pi lockdown argv is pinned exactly with no default model or provider selection", async (t) => {
326
+ // gentle-pi#311 P3: a v9 provider role slot (refuter, targeted validator)
327
+ // reaches this same relay through the exact same request shape a lens
328
+ // materialize slot uses — the only difference is the operation name the
329
+ // provider's own submission descriptor names. The materialize invocation
330
+ // must follow that name, never a hardcoded "capture-result".
331
+ test("the materialize invocation follows the provider's own submission operation, not a hardcoded capture-result", async (t) => {
349
332
  const fixture = harness(t);
350
- await runReviewHostRelaySlot(relayRequest(fixture));
351
- const piCalls = readLog(fixture.piLogPath);
352
- const expected = [
353
- "--print",
354
- "--mode", "json",
355
- "--no-session",
356
- "--no-tools",
357
- "--no-extensions",
358
- "--no-skills",
359
- "--no-prompt-templates",
360
- "--no-themes",
361
- "--no-context-files",
362
- "--no-approve",
333
+ const bindingTokens = [
334
+ "--lineage=review-1d5aadacc600e167",
335
+ `--expected-revision=sha256:${"c".repeat(64)}`,
336
+ `--target=sha256:${"d".repeat(64)}`,
337
+ `--repository-context=rctx1_${"e".repeat(64)}`,
363
338
  ];
364
- assert.deepEqual([...REVIEW_HOST_RELAY_PI_ARGV], expected);
365
- assert.deepEqual(piCalls[0]!.argv, expected);
366
- // The default launch carries no selection; a caller-owned selection may only
367
- // ride the validated forwarding path, never the pinned argv.
368
- assert.equal(piCalls[0]!.argv.some((token) => token.startsWith("--model") || token.startsWith("--provider") || token.startsWith("--profile")), false);
339
+ const refuterCaptureTokens = [...bindingTokens, "--agent=pi", "--materialize=true"];
340
+ const refuterSubmission: ReviewCaptureSubmissionV1 = {
341
+ operationToken: "capture-refuter",
342
+ argumentTokens: [...bindingTokens, "--agent=pi", "--input={{value}}"],
343
+ values: [{ slot: "provider_refuter", domain: "artifact_path_or_stdin", substitutionLocation: bindingTokens.length + 1 }],
344
+ };
345
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
346
+ const result = await runRelay(fixture, { captureArgumentTokens: refuterCaptureTokens, submission: refuterSubmission, routingKey: "review-refuter" }, runReviewer);
347
+
348
+ assert.equal(JSON.parse(result.submission).admission_decision, "completed");
349
+ assert.equal(calls.length, 1);
350
+ const gentleAiCalls = readLog(fixture.logPath);
351
+ assert.equal(gentleAiCalls.length, 2);
352
+ assert.deepEqual(gentleAiCalls[0]!.argv, ["review", "capture-refuter", ...refuterCaptureTokens]);
353
+ assert.deepEqual(gentleAiCalls[1]!.argv.slice(0, 2 + bindingTokens.length), ["review", "capture-refuter", ...bindingTokens]);
354
+ assert.equal(gentleAiCalls[1]!.argv[2 + bindingTokens.length], "--agent=pi");
355
+ const submitted = gentleAiCalls[1]!.argv.at(-1)!;
356
+ assert.match(submitted, /^--input=\S+$/);
369
357
  });
370
358
 
371
359
  test("preparation snapshots mutable submission tokens and values before materialization", async (t) => {
372
360
  const fixture = harness(t);
373
- const barrierDirectory = join(fixture.directory, "submission-snapshot-barrier");
374
- const releaseFile = join(fixture.directory, "submission-snapshot-release");
375
361
  const submissionTokens = [...SUBMISSION.argumentTokens];
376
362
  const submissionValues = SUBMISSION.values.map((value) => ({ ...value }));
377
363
  const submission: ReviewCaptureSubmissionV1 = { ...SUBMISSION, argumentTokens: submissionTokens, values: submissionValues };
378
- const prepared = prepareReviewHostRelaySlot(relayRequest(fixture, {
379
- submission,
380
- environment: {
381
- ...fixture.environment,
382
- RELAY_FAKE_PROMPT_B64: PROMPT_BYTES.toString("base64"),
383
- RELAY_FAKE_PI_OUTPUT_B64: PI_OUTPUT_BYTES.toString("base64"),
384
- RELAY_FAKE_PI_BARRIER_DIR: barrierDirectory,
385
- RELAY_FAKE_PI_BARRIER_COUNT: "1",
386
- RELAY_FAKE_PI_RELEASE_FILE: releaseFile,
387
- },
388
- }));
389
- await waitFor(() => readLog(fixture.piLogPath).length === 1, "reviewer did not reach the snapshot barrier");
364
+ let releaseCompletion: (() => void) | undefined;
365
+ const gate = new Promise<void>((resolve) => { releaseCompletion = resolve; });
366
+ let reached = false;
367
+ const runReviewer = (async () => {
368
+ reached = true;
369
+ await gate;
370
+ return { kind: "text", text: REVIEWER_TEXT, reviewerModel: "openai/gpt-5" } as InProcessReviewerOutcome;
371
+ }) as typeof runInProcessReviewer;
372
+ const prepared = prepareReviewHostRelaySlot(relayRequest(fixture, { submission }), runReviewer);
373
+ await waitFor(() => reached, "the completion was never reached");
390
374
  submissionTokens[submissionTokens.length - 1] = "--input=mutated";
391
375
  submissionValues[0]!.slot = "mutated";
392
- writeFileSync(releaseFile, "release");
376
+ releaseCompletion!();
393
377
 
394
378
  const result = await prepared;
395
379
  assert.deepEqual(result.request.submission, SUBMISSION);
396
380
  await submitReviewHostRelayPreparedResult(result);
397
- assert.deepEqual(readFileSync(fixture.submitCapturePath), Buffer.from(PI_REVIEWER_TEXT, "utf8"));
381
+ assert.deepEqual(readFileSync(fixture.submitCapturePath), Buffer.from(REVIEWER_TEXT, "utf8"));
398
382
  });
399
383
 
400
384
  test("preparation keeps reviewer bytes private through deferred submission", async (t) => {
401
385
  const fixture = harness(t);
402
- const reviewerBytes = Buffer.from(PI_OUTPUT_BYTES);
403
- const prepared = await prepareReviewHostRelaySlot(relayRequest(fixture), async () => ({
404
- stdout: reviewerBytes,
405
- promptByteLength: PROMPT_BYTES.length,
406
- stdoutByteLength: reviewerBytes.length,
407
- }));
408
- reviewerBytes.fill(0);
386
+ const prepared = await prepareReviewHostRelaySlot(relayRequest(fixture), textReviewer(REVIEWER_TEXT).runReviewer);
409
387
  const mutablePrepared = prepared as unknown as { resultBytes?: Buffer };
410
388
  mutablePrepared.resultBytes?.fill(0);
411
389
  assert.throws(() => { mutablePrepared.resultBytes = Buffer.from("fabricated"); }, TypeError);
412
390
  await submitReviewHostRelayPreparedResult(prepared);
413
- assert.deepEqual(readFileSync(fixture.submitCapturePath), PI_OUTPUT_BYTES);
391
+ assert.deepEqual(readFileSync(fixture.submitCapturePath), Buffer.from(REVIEWER_TEXT, "utf8"));
414
392
 
415
393
  const fabricated = { ...prepared, resultBytes: Buffer.from("fabricated") } as unknown as ReviewHostRelayPreparedResult;
416
394
  await assert.rejects(submitReviewHostRelayPreparedResult(fabricated), /recognized prepared result/);
@@ -420,7 +398,7 @@ test("preparation keeps reviewer bytes private through deferred submission", asy
420
398
  test("the supplied AbortSignal stays live through deferred submission", async (t) => {
421
399
  const fixture = harness(t, { RELAY_FAKE_SUBMIT_DELAY_MS: "1000" });
422
400
  const controller = new AbortController();
423
- const prepared = await prepareReviewHostRelaySlot(relayRequest(fixture, { signal: controller.signal }));
401
+ const prepared = await prepareReviewHostRelaySlot(relayRequest(fixture, { signal: controller.signal }), textReviewer(REVIEWER_TEXT).runReviewer);
424
402
  const submission = submitReviewHostRelayPreparedResult(prepared);
425
403
  await waitFor(() => readLog(fixture.logPath).length === 2, "submission did not start");
426
404
  controller.abort();
@@ -429,21 +407,22 @@ test("the supplied AbortSignal stays live through deferred submission", async (t
429
407
 
430
408
  test("four reviewers cross a shared barrier before any result submission can begin", async (t) => {
431
409
  const fixture = harness(t);
432
- const barrierDirectory = join(fixture.directory, "reviewer-barrier");
433
- const requests = reviewerGroupRequests(fixture).map((request) => ({
434
- ...request,
435
- environment: {
436
- ...request.environment,
437
- RELAY_FAKE_PI_BARRIER_DIR: barrierDirectory,
438
- RELAY_FAKE_PI_BARRIER_COUNT: String(REVIEWER_GROUP_LENSES.length),
439
- },
440
- }));
441
-
442
- const prepared = await runReviewHostRelayReviewerGroup(requests);
410
+ const requests = reviewerGroupRequests(fixture);
411
+ let started = 0;
412
+ const releases: Array<() => void> = [];
413
+ const runReviewer = (async () => {
414
+ started += 1;
415
+ await new Promise<void>((resolve) => releases.push(resolve));
416
+ return { kind: "text", text: REVIEWER_TEXT, reviewerModel: "openai/gpt-5" } as InProcessReviewerOutcome;
417
+ }) as typeof runInProcessReviewer;
418
+
419
+ const group = runReviewHostRelayReviewerGroup(requests, (request) => prepareReviewHostRelaySlot(request, runReviewer));
420
+ await waitFor(() => started === REVIEWER_GROUP_LENSES.length, "every reviewer must start before any completes");
421
+ for (const release of releases) release();
422
+ const prepared = await group;
443
423
 
444
424
  assert.equal(prepared.length, REVIEWER_GROUP_LENSES.length);
445
- assert.ok(prepared.every((result) => result.resultByteLength === Buffer.byteLength(PI_REVIEWER_TEXT)));
446
- assert.equal(readLog(fixture.piLogPath).length, REVIEWER_GROUP_LENSES.length);
425
+ assert.ok(prepared.every((result) => result.resultByteLength === Buffer.byteLength(REVIEWER_TEXT)));
447
426
  assert.equal(readLog(fixture.logPath).length, REVIEWER_GROUP_LENSES.length, "preparation materializes only; it does not submit");
448
427
  });
449
428
 
@@ -511,106 +490,112 @@ test("partial reviewer failures wait for a later pending transport and report th
511
490
  // Fail-closed legs — a typed transport error and NO submission.
512
491
  // ---------------------------------------------------------------------------
513
492
 
514
- test("materialize nonzero exit fails closed with a typed error and never launches pi or submits", async (t) => {
493
+ test("materialize nonzero exit fails closed with a typed error and never runs the completion or submits", async (t) => {
515
494
  const fixture = harness(t, { RELAY_FAKE_MATERIALIZE_MODE: "fail" });
516
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture)), REVIEW_HOST_RELAY_FAILURE.MATERIALIZE_FAILED, "materialize");
495
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
496
+ const error = await rejectsWithRelayError(runRelay(fixture, {}, runReviewer), REVIEW_HOST_RELAY_FAILURE.MATERIALIZE_FAILED, "materialize");
517
497
  assert.equal(error.exitCode, 3);
518
498
  assert.equal(error.mutationOutcome, "none");
519
499
  assert.equal(readLog(fixture.logPath).length, 1);
520
- assert.equal(readLog(fixture.piLogPath).length, 0);
500
+ assert.equal(calls.length, 0, "no completion may run after a materialize failure");
521
501
  assert.equal(existsSync(fixture.submitCapturePath), false);
522
502
  });
523
503
 
524
- test("an empty materialized prompt fails closed before pi launches", async (t) => {
504
+ test("an empty materialized prompt fails closed before the completion runs", async (t) => {
525
505
  const fixture = harness(t, { RELAY_FAKE_MATERIALIZE_MODE: "empty" });
526
- await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture)), REVIEW_HOST_RELAY_FAILURE.EMPTY_PROMPT, "materialize");
527
- assert.equal(readLog(fixture.piLogPath).length, 0);
506
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
507
+ await rejectsWithRelayError(runRelay(fixture, {}, runReviewer), REVIEW_HOST_RELAY_FAILURE.EMPTY_PROMPT, "materialize");
508
+ assert.equal(calls.length, 0);
528
509
  assert.equal(existsSync(fixture.submitCapturePath), false);
529
510
  });
530
511
 
531
- test("pi nonzero exit fails closed with a typed error and no submission", async (t) => {
532
- const fixture = harness(t, { RELAY_FAKE_PI_MODE: "fail" });
533
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture)), REVIEW_HOST_RELAY_FAILURE.PI_FAILED, "pi");
534
- assert.equal(error.exitCode, 4);
535
- assert.equal(error.mutationOutcome, "none");
536
- assert.equal(readLog(fixture.logPath).length, 1, "no submission invocation after pi failure");
537
- assert.equal(existsSync(fixture.submitCapturePath), false);
538
- });
512
+ // ---------------------------------------------------------------------------
513
+ // Every in-process reviewer refusal code maps to its typed relay code.
514
+ // ---------------------------------------------------------------------------
539
515
 
540
- test("empty pi stdout fails closed with a typed error and no submission", async (t) => {
541
- const fixture = harness(t, { RELAY_FAKE_PI_MODE: "empty" });
542
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture)), REVIEW_HOST_RELAY_FAILURE.PI_EMPTY_OUTPUT, "pi");
543
- assert.equal(readLog(fixture.logPath).length, 1);
544
- assert.equal(existsSync(fixture.submitCapturePath), false);
545
- // #1156: the envelope carries what the child's stream revealed, not just a
546
- // bare kind code.
547
- assert.match(error.message, /no assistant text|no output|not a pi event stream/);
516
+ test("every in-process reviewer refusal maps to its typed relay failure code, carrying evidence when the outcome had any", async (t) => {
517
+ const cases: Array<{ code: (typeof INPROCESS_REVIEWER_FAILURE)[keyof typeof INPROCESS_REVIEWER_FAILURE]; kind: string; evidence?: Record<string, unknown> }> = [
518
+ { code: INPROCESS_REVIEWER_FAILURE.SELECTION_INVALID, kind: REVIEW_HOST_RELAY_FAILURE.REVIEWER_CONFIG_INVALID },
519
+ { code: INPROCESS_REVIEWER_FAILURE.MODEL_NOT_FOUND, kind: REVIEW_HOST_RELAY_FAILURE.REVIEWER_MODEL_NOT_FOUND },
520
+ { code: INPROCESS_REVIEWER_FAILURE.AUTH_UNAVAILABLE, kind: REVIEW_HOST_RELAY_FAILURE.REVIEWER_AUTH_UNAVAILABLE },
521
+ { code: INPROCESS_REVIEWER_FAILURE.THINKING_INVALID, kind: REVIEW_HOST_RELAY_FAILURE.REVIEWER_THINKING_INVALID },
522
+ { code: INPROCESS_REVIEWER_FAILURE.TOOL_CALL_ATTEMPTED, kind: REVIEW_HOST_RELAY_FAILURE.REVIEWER_TOOL_CALL },
523
+ { code: INPROCESS_REVIEWER_FAILURE.EMPTY_OUTPUT, kind: REVIEW_HOST_RELAY_FAILURE.REVIEWER_EMPTY_OUTPUT, evidence: { stopReason: "stop" } },
524
+ { code: INPROCESS_REVIEWER_FAILURE.OUTPUT_TOO_LARGE, kind: REVIEW_HOST_RELAY_FAILURE.REVIEWER_OUTPUT_TOO_LARGE },
525
+ { code: INPROCESS_REVIEWER_FAILURE.PROVIDER_FAILED, kind: REVIEW_HOST_RELAY_FAILURE.PI_FAILED },
526
+ ];
527
+ for (const { code, kind, evidence } of cases) {
528
+ const fixture = harness(t);
529
+ const message = `refused: ${code}`;
530
+ const error = await rejectsWithRelayError(runRelay(fixture, {}, refusingReviewer(code, message, evidence)), kind, "pi");
531
+ assert.equal(error.message, message);
532
+ assert.equal(error.mutationOutcome, "none");
533
+ assert.equal(existsSync(fixture.submitCapturePath), false, `${code} must never reach submission`);
534
+ if (evidence !== undefined) assert.deepEqual(error.reviewerEvidence, evidence, `${code} must carry its outcome evidence`);
535
+ }
548
536
  });
549
537
 
550
- test("a reviewer run that only attempted a tool call fails typed with the child's own evidence", async (t) => {
538
+ test("a reviewer run that only attempted a tool call fails typed and never submits", async (t) => {
551
539
  const fixture = harness(t);
552
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture, {
553
- environment: {
554
- ...fixture.environment,
555
- RELAY_FAKE_PROMPT_B64: PROMPT_BYTES.toString("base64"),
556
- RELAY_FAKE_PI_OUTPUT_B64: piEventStream(
557
- { type: "message_end", message: { role: "assistant", content: [{ type: "toolCall", id: "call_1", name: "bash" }] } },
558
- assistantEnd("", "nan/deepseek-v4-flash"),
559
- ).toString("base64"),
560
- },
561
- })), REVIEW_HOST_RELAY_FAILURE.PI_EMPTY_OUTPUT, "pi");
540
+ const error = await rejectsWithRelayError(
541
+ runRelay(fixture, {}, refusingReviewer(INPROCESS_REVIEWER_FAILURE.TOOL_CALL_ATTEMPTED, "Reviewer attempted a tool call for review-reliability; the in-process reviewer completion must answer in text only.")),
542
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_TOOL_CALL,
543
+ "pi",
544
+ );
562
545
  assert.equal(existsSync(fixture.submitCapturePath), false);
563
- assert.match(error.message, /no assistant text/);
564
- assert.ok((error as unknown as { reviewerEvidence?: { stdoutKind?: string } }).reviewerEvidence !== undefined, "the failure report must carry the reviewer evidence");
546
+ assert.match(error.message, /tool call/);
565
547
  });
566
548
 
567
- test("the relay forwards the caller-owned reviewer selection and extension allowlist to the child argv", async (t) => {
549
+ test("the relay forwards the caller-owned selection and thinking level to the completion request", async (t) => {
568
550
  const fixture = harness(t);
569
- const adapterPath = join(fixture.directory, "auth-adapter.ts");
570
- writeFileSync(adapterPath, "export default () => {};");
571
- await runReviewHostRelaySlot(relayRequest(fixture, {
572
- reviewerModel: "minimax/MiniMax-M3",
573
- reviewerExtensionPaths: [adapterPath],
574
- }));
575
- const calls = readLog(fixture.piLogPath);
551
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
552
+ await runRelay(fixture, { selection: "minimax/MiniMax-M3", thinking: "high", routingKey: "review-risk" }, runReviewer);
576
553
  assert.equal(calls.length, 1);
577
- assert.deepEqual(calls[0]!.argv.slice(-4), ["--model", "minimax/MiniMax-M3", "-e", adapterPath]);
554
+ assert.equal(calls[0]!.selection, "minimax/MiniMax-M3");
555
+ assert.equal(calls[0]!.thinking, "high");
556
+ assert.equal(calls[0]!.routingKey, "review-risk");
578
557
  });
579
558
 
580
- test("a malformed reviewer selection or a relative extension path is refused typed before anything launches", async (t) => {
559
+ test("a missing model registry is refused typed before materialize ever runs", async (t) => {
581
560
  const fixture = harness(t);
582
- for (const broken of [
583
- { reviewerModel: "minimax / MiniMax M3" },
584
- { reviewerModel: "" },
585
- { reviewerExtensionPaths: ["relative/auth.ts"] },
586
- { reviewerExtensionPaths: [join(fixture.directory, "missing-adapter.ts")] },
587
- ]) {
588
- let caught: unknown;
589
- try {
590
- await runReviewHostRelaySlot(relayRequest(fixture, broken));
591
- } catch (error) {
592
- caught = error;
593
- }
594
- assert.ok(caught instanceof ReviewHostRelayError, `expected a typed relay error for ${JSON.stringify(broken)}`);
595
- const error = caught as ReviewHostRelayError;
596
- assert.equal(error.kind, "reviewer-config-invalid");
597
- assert.equal(error.stage, "pi");
598
- assert.equal(readLog(fixture.piLogPath).length, 0, "no reviewer may launch on a broken selection");
599
- assert.equal(readLog(fixture.logPath).length, 0, "no materialization may run on a broken selection");
600
- }
561
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
562
+ const error = await rejectsWithRelayError(runRelay(fixture, { reviewerRegistry: undefined }, runReviewer), REVIEW_HOST_RELAY_FAILURE.REVIEWER_CONFIG_INVALID, "pi");
563
+ assert.equal(calls.length, 0, "no completion may run on a broken configuration");
564
+ assert.equal(readLog(fixture.logPath).length, 0, "no materialization may run without a model registry");
565
+ assert.match(error.message, /model registry/);
601
566
  });
602
567
 
603
- test("pi timeout fails closed with a typed error and no submission", async (t) => {
604
- const fixture = harness(t, { RELAY_FAKE_PI_MODE: "hang" });
605
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture, { piTimeoutMs: 300 })), REVIEW_HOST_RELAY_FAILURE.PI_TIMED_OUT, "pi");
606
- assert.equal(error.timedOut, true);
607
- assert.equal(readLog(fixture.logPath).length, 1);
568
+ test("a routing entry with no configured model is refused typed, naming the routing key, before materialize ever runs", async (t) => {
569
+ const fixture = harness(t);
570
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
571
+ const error = await rejectsWithRelayError(runRelay(fixture, { selection: undefined, routingKey: "review-risk" }, runReviewer), REVIEW_HOST_RELAY_FAILURE.REVIEWER_CONFIG_INVALID, "pi");
572
+ assert.equal(calls.length, 0);
573
+ assert.equal(readLog(fixture.logPath).length, 0, "no materialization may run without a configured model");
574
+ assert.match(error.message, /review-risk/);
575
+ });
576
+
577
+ test("a tool call's cancelled signal maps to REVIEWER_ABORTED and forwards the caller's own signal to the completion", async (t) => {
578
+ const fixture = harness(t);
579
+ const controller = new AbortController();
580
+ let receivedSignal: AbortSignal | undefined;
581
+ // The signal aborts once the completion actually starts (materialize must
582
+ // already have succeeded with a live signal); this proves the same signal
583
+ // object reaches the completion request, not just that an early abort
584
+ // short-circuits materialize too.
585
+ const runReviewer = (async (request: InProcessReviewerRequest) => {
586
+ receivedSignal = request.signal;
587
+ controller.abort();
588
+ return { kind: "refused", code: INPROCESS_REVIEWER_FAILURE.ABORTED, message: "Reviewer completion for review-reliability was aborted by the caller." } as InProcessReviewerOutcome;
589
+ }) as typeof runInProcessReviewer;
590
+ const error = await rejectsWithRelayError(runRelay(fixture, { signal: controller.signal }, runReviewer), REVIEW_HOST_RELAY_FAILURE.REVIEWER_ABORTED, "pi");
591
+ assert.equal(receivedSignal, controller.signal, "the relay must forward the tool call's own signal into the completion request");
592
+ assert.match(error.message, /aborted/);
608
593
  assert.equal(existsSync(fixture.submitCapturePath), false);
609
594
  });
610
595
 
611
596
  // ---------------------------------------------------------------------------
612
597
  // gentle-pi#367 — the reviewer bound is reachable from production, and a
613
- // reviewer killed by it says so with both measurements.
598
+ // timed-out completion says so with both measurements.
614
599
  // ---------------------------------------------------------------------------
615
600
 
616
601
  test("the reviewer bound scales with materialized prompt bytes instead of one fixed number", () => {
@@ -644,50 +629,26 @@ test("the reviewer bound honours the environment override and ignores malformed
644
629
  }
645
630
  });
646
631
 
647
- test("the production relay path resolves the reviewer bound from the environment, with no injected timeout", async (t) => {
648
- // The regression: piTimeoutMs was reachable only through the test seam, so
649
- // production always ran against the fixed 600s bound. This request injects
650
- // no timeout at all — the override must reach the real spawn.
651
- const fixture = harness(t, { RELAY_FAKE_PI_MODE: "hang", [REVIEW_HOST_RELAY_PI_TIMEOUT_ENV]: "300" });
632
+ test("the production relay path resolves the reviewer bound from the environment and forwards it to the completion, with no injected timeout", async (t) => {
633
+ // The regression this guards: piTimeoutMs was reachable only through the
634
+ // test seam, so production always ran against a fixed bound. This request
635
+ // injects no timeout at all — the override must reach the completion.
636
+ const fixture = harness(t, { [REVIEW_HOST_RELAY_PI_TIMEOUT_ENV]: "300000" });
637
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
638
+ await runRelay(fixture, { piTimeoutMs: undefined }, runReviewer);
639
+ assert.equal(calls[0]!.timeoutMs, 300_000);
640
+ });
641
+
642
+ test("a timed-out reviewer completion reports elapsed, limit, and what to change instead of an opaque transport failure", async (t) => {
643
+ const fixture = harness(t);
652
644
  const error = await rejectsWithRelayError(
653
- runReviewHostRelaySlot(relayRequest(fixture, { piTimeoutMs: undefined })),
645
+ runRelay(fixture, { piTimeoutMs: 300 }, refusingReviewer(INPROCESS_REVIEWER_FAILURE.TIMED_OUT, "Reviewer completion for review-reliability exceeded its 300ms bound.", undefined, 20)),
654
646
  REVIEW_HOST_RELAY_FAILURE.PI_TIMED_OUT,
655
647
  "pi",
656
648
  );
657
649
  assert.equal(error.timeoutMs, 300);
650
+ assert.ok(error.elapsedMs !== null && error.elapsedMs >= 15, `elapsed ${error.elapsedMs} must record the real wall time`);
658
651
  assert.equal(error.timedOut, true);
659
- assert.equal(readLog(fixture.logPath).length, 1, "no submission after a reviewer timeout");
660
- assert.equal(existsSync(fixture.submitCapturePath), false);
661
- });
662
-
663
- test("a relay Pi timeout keeps its typed mapping and timing evidence when opaque scratch cleanup also fails", async (t) => {
664
- const fixture = harness(t, { RELAY_FAKE_PI_MODE: "hang", RELAY_FAKE_PI_BREAK_CLEANUP: "true" });
665
- const scratchParent = realpathSync(mkdtempSync(join(tmpdir(), "gentle-pi-relay-pi-primary-failure-")));
666
- const originalTmpdir = process.env.TMPDIR;
667
- process.env.TMPDIR = scratchParent;
668
- try {
669
- const error = await rejectsWithRelayError(
670
- runReviewHostRelaySlot(relayRequest(fixture, { piTimeoutMs: 300 })),
671
- REVIEW_HOST_RELAY_FAILURE.PI_TIMED_OUT,
672
- "pi",
673
- );
674
- assert.equal(error.timedOut, true);
675
- assert.equal(error.timeoutMs, 300);
676
- assert.ok(error.elapsedMs !== null && error.elapsedMs >= 250);
677
- } finally {
678
- if (originalTmpdir === undefined) delete process.env.TMPDIR;
679
- else process.env.TMPDIR = originalTmpdir;
680
- chmodSync(scratchParent, 0o700);
681
- rmSync(scratchParent, { recursive: true, force: true });
682
- }
683
- });
684
-
685
- test("a killed reviewer reports elapsed, limit, and what to change instead of an opaque transport failure", async (t) => {
686
- const fixture = harness(t, { RELAY_FAKE_PI_MODE: "hang" });
687
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture, { piTimeoutMs: 300 })), REVIEW_HOST_RELAY_FAILURE.PI_TIMED_OUT, "pi");
688
- assert.equal(error.timeoutMs, 300);
689
- assert.ok(error.elapsedMs !== null && error.elapsedMs >= 250, `elapsed ${error.elapsedMs} must record the real wall time`);
690
- assert.ok(error.elapsedMs! < 10_000, "the reviewer was killed at the bound, not left to finish");
691
652
  assert.match(error.message, /exceeded the relay bound/);
692
653
  assert.match(error.message, new RegExp(String(error.elapsedMs)));
693
654
  assert.match(error.message, /limit for a \d+-byte materialized prompt/);
@@ -695,18 +656,21 @@ test("a killed reviewer reports elapsed, limit, and what to change instead of an
695
656
  assert.equal(error.mutationOutcome, "none");
696
657
  });
697
658
 
698
- test("a crashed reviewer stays distinguishable from a killed one and still carries its measurements", async (t) => {
699
- const fixture = harness(t, { RELAY_FAKE_PI_MODE: "fail" });
700
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture, { piTimeoutMs: 30_000 })), REVIEW_HOST_RELAY_FAILURE.PI_FAILED, "pi");
659
+ test("a provider-failed completion stays distinguishable from a timed-out one and still carries its measurements", async (t) => {
660
+ const fixture = harness(t);
661
+ const error = await rejectsWithRelayError(
662
+ runRelay(fixture, { piTimeoutMs: 30_000 }, refusingReviewer(INPROCESS_REVIEWER_FAILURE.PROVIDER_FAILED, "Reviewer completion failed for review-reliability: upstream 500")),
663
+ REVIEW_HOST_RELAY_FAILURE.PI_FAILED,
664
+ "pi",
665
+ );
701
666
  assert.equal(error.timedOut, false);
702
- assert.equal(error.exitCode, 4);
703
667
  assert.equal(error.timeoutMs, 30_000);
704
668
  assert.ok(error.elapsedMs !== null && error.elapsedMs >= 0);
705
669
  });
706
670
 
707
671
  test("submission refusal is a typed error whose outcome is unknown pending STATUS", async (t) => {
708
672
  const fixture = harness(t, { RELAY_FAKE_SUBMIT_MODE: "refuse" });
709
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture)), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_REFUSED, "submit");
673
+ const error = await rejectsWithRelayError(runRelay(fixture), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_REFUSED, "submit");
710
674
  assert.equal(error.mutationOutcome, "unknown");
711
675
  assert.match(error.stderr, /capture binding does not match/);
712
676
  assert.equal(readLog(fixture.logPath).length, 2);
@@ -716,21 +680,14 @@ test("submission refusal is a typed error whose outcome is unknown pending STATU
716
680
  // typed [invalid_request] refusal and states that the lens slot was not
717
681
  // consumed. That is a proven non-mutation, not an unknown outcome, and the
718
682
  // refusal text is the only thing that tells the host what to change.
719
- function admittingRequest(fixture: RelayHarness, reviewerOutput: Buffer) {
720
- return relayRequest(fixture, {
721
- environment: {
722
- ...fixture.environment,
723
- RELAY_FAKE_PROMPT_B64: PROMPT_BYTES.toString("base64"),
724
- // The reviewer's payload arrives as the event stream's assistant text.
725
- RELAY_FAKE_PI_OUTPUT_B64: piEventStream(assistantEnd(reviewerOutput.toString("utf8"))).toString("base64"),
726
- },
727
- });
683
+ function admittingRequest(fixture: RelayHarness): ReviewHostRelayRequest {
684
+ return relayRequest(fixture);
728
685
  }
729
686
 
730
687
  test("a reviewer printing garbage is refused at admission as a proven non-mutation carrying Go's refusal", async (t) => {
731
688
  const fixture = harness(t, { RELAY_FAKE_SUBMIT_MODE: "admit", RELAY_FAKE_EXPECTED_SUBJECT: `sha256:${"a".repeat(64)}` });
732
689
  const error = await rejectsWithRelayError(
733
- runReviewHostRelaySlot(admittingRequest(fixture, Buffer.from("not json at all", "utf8"))),
690
+ runReviewHostRelaySlot(admittingRequest(fixture), textReviewer("not json at all").runReviewer),
734
691
  REVIEW_HOST_RELAY_FAILURE.SUBMISSION_REFUSED,
735
692
  "submit",
736
693
  );
@@ -745,9 +702,9 @@ test("a reviewer printing garbage is refused at admission as a proven non-mutati
745
702
 
746
703
  test("a reviewer echoing a different subject is refused at admission as a proven non-mutation carrying Go's continuation", async (t) => {
747
704
  const fixture = harness(t, { RELAY_FAKE_SUBMIT_MODE: "admit", RELAY_FAKE_EXPECTED_SUBJECT: `sha256:${"a".repeat(64)}` });
748
- const wrongSubject = Buffer.from(JSON.stringify({ subject_hash: `sha256:${"0".repeat(64)}`, inspection: { status: "completed", paths: [] }, findings: [], evidence: ["x"] }), "utf8");
705
+ const wrongSubject = JSON.stringify({ subject_hash: `sha256:${"0".repeat(64)}`, inspection: { status: "completed", paths: [] }, findings: [], evidence: ["x"] });
749
706
  const error = await rejectsWithRelayError(
750
- runReviewHostRelaySlot(admittingRequest(fixture, wrongSubject)),
707
+ runReviewHostRelaySlot(admittingRequest(fixture), textReviewer(wrongSubject).runReviewer),
751
708
  REVIEW_HOST_RELAY_FAILURE.SUBMISSION_REFUSED,
752
709
  "submit",
753
710
  );
@@ -762,7 +719,8 @@ test("a reviewer echoing a different subject is refused at admission as a proven
762
719
 
763
720
  test("a submission the fake admits with the expected subject still completes", async (t) => {
764
721
  const fixture = harness(t, { RELAY_FAKE_SUBMIT_MODE: "admit", RELAY_FAKE_EXPECTED_SUBJECT: `sha256:${"a".repeat(64)}` });
765
- const result = await runReviewHostRelaySlot(admittingRequest(fixture, Buffer.from(JSON.stringify({ subject_hash: `sha256:${"a".repeat(64)}`, inspection: { status: "completed", paths: [] }, findings: [], evidence: ["x"] }), "utf8")));
722
+ const expected = JSON.stringify({ subject_hash: `sha256:${"a".repeat(64)}`, inspection: { status: "completed", paths: [] }, findings: [], evidence: ["x"] });
723
+ const result = await runReviewHostRelaySlot(admittingRequest(fixture), textReviewer(expected).runReviewer);
766
724
  assert.equal(JSON.parse(result.submission).admission_decision, "completed");
767
725
  });
768
726
 
@@ -773,7 +731,7 @@ test("submission refusal preserves its primary evidence when result staging clea
773
731
  process.env.TMPDIR = scratchParent;
774
732
  try {
775
733
  const error = await rejectsWithRelayError(
776
- runReviewHostRelaySlot(relayRequest(fixture)),
734
+ runRelay(fixture),
777
735
  REVIEW_HOST_RELAY_FAILURE.SUBMISSION_REFUSED,
778
736
  "submit",
779
737
  );
@@ -791,12 +749,10 @@ test("submission refusal preserves its primary evidence when result staging clea
791
749
  }
792
750
  });
793
751
 
794
- test("relay scratch and staging directories are removed after failures too", async (t) => {
795
- const fixture = harness(t, { RELAY_FAKE_PI_MODE: "fail" });
796
- await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture)), REVIEW_HOST_RELAY_FAILURE.PI_FAILED, "pi");
797
- const piCalls = readLog(fixture.piLogPath);
798
- assert.equal(piCalls.length, 1);
799
- assert.equal(existsSync(piCalls[0]!.cwd!), false);
752
+ test("relay staging directories are removed after a submission failure too", async (t) => {
753
+ const fixture = harness(t, { RELAY_FAKE_SUBMIT_MODE: "refuse" });
754
+ await rejectsWithRelayError(runRelay(fixture), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_REFUSED, "submit");
755
+ assert.equal(readLog(fixture.logPath).length, 2);
800
756
  });
801
757
 
802
758
  test("a result staging cleanup failure remains a typed submit failure", async (t) => {
@@ -806,7 +762,7 @@ test("a result staging cleanup failure remains a typed submit failure", async (t
806
762
  process.env.TMPDIR = scratchParent;
807
763
  try {
808
764
  const error = await rejectsWithRelayError(
809
- runReviewHostRelaySlot(relayRequest(fixture)),
765
+ runRelay(fixture),
810
766
  REVIEW_HOST_RELAY_FAILURE.SUBMISSION_REFUSED,
811
767
  "submit",
812
768
  );
@@ -826,20 +782,22 @@ test("a result staging cleanup failure remains a typed submit failure", async (t
826
782
 
827
783
  test("an old binary's unknown-flag refusal classifies as relay-unavailable with the exact report", async (t) => {
828
784
  const fixture = harness(t, { RELAY_FAKE_MATERIALIZE_MODE: "unknown-flag" });
829
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture)), REVIEW_HOST_RELAY_FAILURE.RELAY_UNAVAILABLE, "materialize");
785
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
786
+ const error = await rejectsWithRelayError(runRelay(fixture, {}, runReviewer), REVIEW_HOST_RELAY_FAILURE.RELAY_UNAVAILABLE, "materialize");
830
787
  assert.equal(error.message, REVIEW_HOST_RELAY_UNAVAILABLE_MESSAGE);
831
788
  assert.match(error.stderr, /flag provided but not defined: -materialize/);
832
- assert.equal(readLog(fixture.piLogPath).length, 0);
789
+ assert.equal(calls.length, 0);
833
790
  assert.equal(existsSync(fixture.submitCapturePath), false);
834
791
  });
835
792
 
836
793
  test("a handshake refusal surfaces the provider refusal verbatim", async (t) => {
837
794
  const refusal = "review capture-result --agent pi: the active runtime is not eligible for immutable receipt review; supported immutable review runtimes: claude-code, codex, opencode";
838
795
  const fixture = harness(t, { RELAY_FAKE_MATERIALIZE_MODE: "handshake", RELAY_FAKE_HANDSHAKE_STDERR: refusal });
839
- const error = await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture)), REVIEW_HOST_RELAY_FAILURE.HANDSHAKE_REFUSED, "materialize");
796
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
797
+ const error = await rejectsWithRelayError(runRelay(fixture, {}, runReviewer), REVIEW_HOST_RELAY_FAILURE.HANDSHAKE_REFUSED, "materialize");
840
798
  assert.equal(error.message, refusal);
841
799
  assert.equal(error.stderr, refusal);
842
- assert.equal(readLog(fixture.piLogPath).length, 0);
800
+ assert.equal(calls.length, 0);
843
801
  });
844
802
 
845
803
  test("refusal classification distinguishes unknown-flag, handshake, and other", () => {
@@ -854,7 +812,7 @@ test("refusal classification distinguishes unknown-flag, handshake, and other",
854
812
  // deterministic for this exact slot. Generic admission refusals are repairable
855
813
  // by a fresh reviewer and retain the ordinary exact-reoffer behavior.
856
814
  test("the unachievable predicate names only deterministic slot failures", () => {
857
- const killed = new ReviewHostRelayError(REVIEW_HOST_RELAY_FAILURE.PI_TIMED_OUT, "pi", "pi reviewer subprocess exceeded the relay bound", { timedOut: true, elapsedMs: 2_256_004, timeoutMs: 2_256_000 });
815
+ const killed = new ReviewHostRelayError(REVIEW_HOST_RELAY_FAILURE.PI_TIMED_OUT, "pi", "the reviewer completion exceeded the relay bound", { timedOut: true, elapsedMs: 2_256_004, timeoutMs: 2_256_000 });
858
816
  assert.equal(reviewHostRelayUnachievableReason(killed), "relay_transport_bound_exceeded");
859
817
  assert.equal(reviewHostRelayUnachievableDetail(killed), "killed after 2256004ms against a 2256000ms relay bound");
860
818
 
@@ -868,7 +826,24 @@ test("the unachievable predicate names only deterministic slot failures", () =>
868
826
  assert.equal(reviewHostRelayUnachievableReason(unknownOutcome), undefined, "an unknown mutation outcome is never deterministic");
869
827
  assert.equal(reviewHostRelayUnachievableDetail(unknownOutcome), undefined);
870
828
 
871
- for (const kind of [REVIEW_HOST_RELAY_FAILURE.PI_FAILED, REVIEW_HOST_RELAY_FAILURE.PI_LAUNCH_FAILED, REVIEW_HOST_RELAY_FAILURE.PI_EMPTY_OUTPUT, REVIEW_HOST_RELAY_FAILURE.MATERIALIZE_FAILED, REVIEW_HOST_RELAY_FAILURE.EMPTY_PROMPT, REVIEW_HOST_RELAY_FAILURE.RELAY_UNAVAILABLE, REVIEW_HOST_RELAY_FAILURE.HANDSHAKE_REFUSED, REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH]) {
829
+ for (const kind of [
830
+ REVIEW_HOST_RELAY_FAILURE.PI_FAILED,
831
+ REVIEW_HOST_RELAY_FAILURE.PI_LAUNCH_FAILED,
832
+ REVIEW_HOST_RELAY_FAILURE.PI_EMPTY_OUTPUT,
833
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_MODEL_NOT_FOUND,
834
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_AUTH_UNAVAILABLE,
835
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_THINKING_INVALID,
836
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_TOOL_CALL,
837
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_EMPTY_OUTPUT,
838
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_OUTPUT_TOO_LARGE,
839
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_ABORTED,
840
+ REVIEW_HOST_RELAY_FAILURE.REVIEWER_CONFIG_INVALID,
841
+ REVIEW_HOST_RELAY_FAILURE.MATERIALIZE_FAILED,
842
+ REVIEW_HOST_RELAY_FAILURE.EMPTY_PROMPT,
843
+ REVIEW_HOST_RELAY_FAILURE.RELAY_UNAVAILABLE,
844
+ REVIEW_HOST_RELAY_FAILURE.HANDSHAKE_REFUSED,
845
+ REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH,
846
+ ]) {
872
847
  assert.equal(reviewHostRelayUnachievableReason(new ReviewHostRelayError(kind, "pi", "transient")), undefined, `${kind} stays transient`);
873
848
  }
874
849
 
@@ -942,37 +917,39 @@ test("relay input validation rejects empty or malformed inputs before any proces
942
917
 
943
918
  test("a materialize slot without a provider submission fails closed before any process launches", async (t) => {
944
919
  const fixture = harness(t);
920
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
945
921
  const error = await rejectsWithRelayError(
946
- runReviewHostRelaySlot(relayRequest(fixture, { submission: undefined })),
922
+ runRelay(fixture, { submission: undefined }, runReviewer),
947
923
  REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH,
948
924
  "binding",
949
925
  );
950
926
  assert.equal(error.message, REVIEW_HOST_RELAY_SUBMISSION_MISSING_MESSAGE);
951
927
  assert.equal(error.mutationOutcome, "none");
952
928
  assert.equal(readLog(fixture.logPath).length, 0, "no gentle-ai invocation was launched");
953
- assert.equal(readLog(fixture.piLogPath).length, 0, "no pi subprocess was launched");
929
+ assert.equal(calls.length, 0, "no completion was run");
954
930
  assert.equal(existsSync(fixture.submitCapturePath), false);
955
931
  });
956
932
 
957
933
  test("a submission form the relay cannot bind is a typed contract mismatch, never a repaired invocation", async (t) => {
958
934
  const fixture = harness(t);
935
+ const { runReviewer, calls } = textReviewer(REVIEWER_TEXT);
959
936
  const twoValues: ReviewCaptureSubmissionV1 = {
960
937
  ...SUBMISSION,
961
938
  values: [...SUBMISSION.values, { slot: "extra", domain: "artifact_path_or_stdin", substitutionLocation: 0 }],
962
939
  };
963
- await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture, { submission: twoValues })), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH, "binding");
940
+ await rejectsWithRelayError(runRelay(fixture, { submission: twoValues }, runReviewer), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH, "binding");
964
941
  const noSlot: ReviewCaptureSubmissionV1 = {
965
942
  ...SUBMISSION,
966
943
  argumentTokens: [...BINDING_TOKENS, "--input=/etc/somewhere"],
967
944
  };
968
- await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture, { submission: noSlot })), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH, "binding");
945
+ await rejectsWithRelayError(runRelay(fixture, { submission: noSlot }, runReviewer), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH, "binding");
969
946
  const outOfBounds: ReviewCaptureSubmissionV1 = {
970
947
  ...SUBMISSION,
971
948
  values: [{ slot: "reviewer_result", domain: "artifact_path_or_stdin", substitutionLocation: SUBMISSION.argumentTokens.length }],
972
949
  };
973
- await rejectsWithRelayError(runReviewHostRelaySlot(relayRequest(fixture, { submission: outOfBounds })), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH, "binding");
950
+ await rejectsWithRelayError(runRelay(fixture, { submission: outOfBounds }, runReviewer), REVIEW_HOST_RELAY_FAILURE.SUBMISSION_CONTRACT_MISMATCH, "binding");
974
951
  assert.equal(readLog(fixture.logPath).length, 0);
975
- assert.equal(readLog(fixture.piLogPath).length, 0);
952
+ assert.equal(calls.length, 0);
976
953
  });
977
954
 
978
955
  test("resolveReviewHostRelaySubmission returns the provider binding untouched", () => {