@gaunt-sloth/core 2.0.0-alpha.35 → 2.0.0-alpha.37

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 (95) hide show
  1. package/dist/config/configDiscovery.d.ts +79 -0
  2. package/dist/config/configDiscovery.js +80 -0
  3. package/dist/config/configDiscovery.js.map +1 -0
  4. package/dist/config/loader.d.ts +6 -5
  5. package/dist/config/loader.js +229 -92
  6. package/dist/config/loader.js.map +1 -1
  7. package/dist/config/schema.d.ts +3 -2
  8. package/dist/config/schema.js +57 -9
  9. package/dist/config/schema.js.map +1 -1
  10. package/dist/config/shell-policy.d.ts +58 -26
  11. package/dist/config/shell-policy.js +60 -27
  12. package/dist/config/shell-policy.js.map +1 -1
  13. package/dist/config/tool-descriptions.d.ts +6 -6
  14. package/dist/config/tool-descriptions.js +7 -7
  15. package/dist/config/tool-descriptions.js.map +1 -1
  16. package/dist/config/types.d.ts +34 -31
  17. package/dist/config/types.js.map +1 -1
  18. package/dist/config.d.ts +3 -0
  19. package/dist/config.js +3 -0
  20. package/dist/config.js.map +1 -1
  21. package/dist/core/GthAbstractAgent.d.ts +11 -13
  22. package/dist/core/GthAbstractAgent.js +9 -11
  23. package/dist/core/GthAbstractAgent.js.map +1 -1
  24. package/dist/core/GthAgentRunner.d.ts +17 -33
  25. package/dist/core/GthAgentRunner.js +31 -41
  26. package/dist/core/GthAgentRunner.js.map +1 -1
  27. package/dist/core/GthLangChainAgent.js +25 -34
  28. package/dist/core/GthLangChainAgent.js.map +1 -1
  29. package/dist/core/approvals/toolAnnotationSources.d.ts +4 -4
  30. package/dist/core/approvals/toolAnnotationSources.js +8 -8
  31. package/dist/core/debugCapture.d.ts +2 -3
  32. package/dist/core/debugCapture.js +1 -2
  33. package/dist/core/debugCapture.js.map +1 -1
  34. package/dist/core/gthLeanAgentFactory.d.ts +4 -4
  35. package/dist/core/gthLeanAgentFactory.js +4 -4
  36. package/dist/core/launchBanner.d.ts +10 -0
  37. package/dist/core/launchBanner.js +6 -1
  38. package/dist/core/launchBanner.js.map +1 -1
  39. package/dist/core/reasoningBlocks.d.ts +7 -2
  40. package/dist/core/reasoningBlocks.js +7 -2
  41. package/dist/core/reasoningBlocks.js.map +1 -1
  42. package/dist/core/shell/ShellCommandFailedError.d.ts +3 -4
  43. package/dist/core/shell/ShellCommandFailedError.js +3 -4
  44. package/dist/core/shell/ShellCommandFailedError.js.map +1 -1
  45. package/dist/core/shell/approvalStop.d.ts +68 -29
  46. package/dist/core/shell/approvalStop.js +178 -28
  47. package/dist/core/shell/approvalStop.js.map +1 -1
  48. package/dist/core/shell/hardline.d.ts +22 -0
  49. package/dist/core/shell/hardline.js +81 -0
  50. package/dist/core/shell/hardline.js.map +1 -1
  51. package/dist/core/shell/negotiation.d.ts +81 -2
  52. package/dist/core/shell/negotiation.js +144 -11
  53. package/dist/core/shell/negotiation.js.map +1 -1
  54. package/dist/core/shell/rater.d.ts +7 -69
  55. package/dist/core/shell/rater.js +5 -24
  56. package/dist/core/shell/rater.js.map +1 -1
  57. package/dist/core/shell/raterVocabulary.d.ts +119 -0
  58. package/dist/core/shell/raterVocabulary.js +114 -0
  59. package/dist/core/shell/raterVocabulary.js.map +1 -0
  60. package/dist/core/types.d.ts +17 -3
  61. package/dist/providers/anthropic.js +12 -0
  62. package/dist/providers/anthropic.js.map +1 -1
  63. package/dist/providers/configurationPassthrough.d.ts +107 -0
  64. package/dist/providers/configurationPassthrough.js +148 -0
  65. package/dist/providers/configurationPassthrough.js.map +1 -0
  66. package/dist/providers/geminiThinking.d.ts +8 -0
  67. package/dist/providers/geminiThinking.js +33 -13
  68. package/dist/providers/geminiThinking.js.map +1 -1
  69. package/dist/providers/google-genai.js +11 -0
  70. package/dist/providers/google-genai.js.map +1 -1
  71. package/dist/providers/groq.js +12 -0
  72. package/dist/providers/groq.js.map +1 -1
  73. package/dist/providers/ollama.js +12 -0
  74. package/dist/providers/ollama.js.map +1 -1
  75. package/dist/providers/openrouter.d.ts +23 -0
  76. package/dist/providers/openrouter.js +71 -4
  77. package/dist/providers/openrouter.js.map +1 -1
  78. package/dist/providers/vertexai.js +12 -0
  79. package/dist/providers/vertexai.js.map +1 -1
  80. package/dist/providers/xai.js +20 -0
  81. package/dist/providers/xai.js.map +1 -1
  82. package/dist/runtime/conversation.d.ts +2 -2
  83. package/dist/runtime/conversation.js +19 -4
  84. package/dist/runtime/conversation.js.map +1 -1
  85. package/dist/runtime/singleShot.d.ts +2 -3
  86. package/dist/runtime/singleShot.js +22 -7
  87. package/dist/runtime/singleShot.js.map +1 -1
  88. package/dist/utils/systemPromptNotes.d.ts +17 -20
  89. package/dist/utils/systemPromptNotes.js +22 -24
  90. package/dist/utils/systemPromptNotes.js.map +1 -1
  91. package/dist/utils/untrustedText.d.ts +24 -4
  92. package/dist/utils/untrustedText.js +25 -4
  93. package/dist/utils/untrustedText.js.map +1 -1
  94. package/package.json +4 -4
  95. package/schema/gsloth-config.schema.json +0 -1
@@ -10,9 +10,10 @@ import { displayDebug, displayError, displayInfo, displayWarning, setConsoleLeve
10
10
  import { findApprovalsGrammarIssues, findApprovalsRaterProfiles, findDeprecatedConfigIssues, findUnknownTopLevelKeys, formatConfigValidationError, formatDeprecatedConfigIssues, isRecordConfig, rawGthConfigSchema, unresolvedRaterProfileMessage, validateRawGthConfig, } from '#src/config/schema.js';
11
11
  import { parseJsonc } from '#src/config/jsonc.js';
12
12
  import { isMissingProviderKeyError, MissingProviderKeyError } from '#src/config/providerKeys.js';
13
+ import { ConfigDiscoveryError, isConfigDiscoveryError } from '#src/config/configDiscovery.js';
13
14
  import { getGslothConfigReadPath, importExternalFile } from '#src/utils/fileUtils.js';
14
15
  import { getGlobalGslothConfigReadPath } from '#src/utils/globalConfigUtils.js';
15
- import { env, error, exit, getCurrentWorkDir, isStdoutTTY, isTTY, setProjectDir, setUseColour, } from '#src/utils/systemUtils.js';
16
+ import { env, getCurrentWorkDir, isStdoutTTY, isTTY, setProjectDir, setUseColour, } from '#src/utils/systemUtils.js';
16
17
  import { resolveUseColour } from '#src/config/colour.js';
17
18
  import { resolveUseMouse } from '#src/config/mouse.js';
18
19
  import { existsSync, readFileSync } from 'node:fs';
@@ -28,7 +29,7 @@ import { DEFAULT_CONFIG } from '#src/config/defaults.js';
28
29
  * 1. Deprecated-shape reject (GS2-28): a removed pre-2.0 shape — a top-level command key
29
30
  * or a deprecated `*Provider*` name (root + per-command), detected by
30
31
  * {@link findDeprecatedConfigIssues} — is a HARD error naming the canonical replacement +
31
- * migration path, then exit. 2.0 dropped back-compat coercion, so these fail rather than
32
+ * migration path. 2.0 dropped back-compat coercion, so these fail rather than
32
33
  * remap. Runs FIRST so a deprecated name never merely surfaces as an unknown-key warning.
33
34
  * 1b. Approvals rule-grammar reject (EXT-71): a bare string in `allow`/`deny`/`escalate`, or a
34
35
  * configured `mcpServers` key named `*`, detected by {@link findApprovalsGrammarIssues} — a
@@ -36,13 +37,25 @@ import { DEFAULT_CONFIG } from '#src/config/defaults.js';
36
37
  * user wrote) is the only one they see.
37
38
  * 2. Unknown top-level keys: warn (do NOT fail) so likely typos are surfaced while
38
39
  * forward-compatible / extension keys still pass through untouched.
39
- * 3. Schema parse: on a genuine type mismatch on a known field, emit a friendly,
40
- * path-scoped error and exit (matching the loader's existing invalid-config
41
- * behaviour). Validation is shape-only — the loose schema preserves unknown keys,
40
+ * 3. Schema parse: on a genuine type mismatch on a known field, a friendly, path-scoped
41
+ * error. Validation is shape-only the loose schema preserves unknown keys,
42
42
  * so the original `raw` is returned unchanged on success.
43
43
  *
44
+ * CFG-36 / CFG-47 — every hard failure above RAISES a {@link ConfigDiscoveryError} rather than
45
+ * printing and calling `exit(1)`, and so does every other "config present and unusable" site in this
46
+ * file. Config loading is a library operation: the caller chooses the exit code (the CLI's top-level
47
+ * guard prints the message and exits 1; `gth eval` classifies it as a harness error and exits 2).
48
+ * **This file no longer calls `exit` at all** — that is the invariant, and it is easier to keep than
49
+ * a list of which sites do. Because these throw, every `catch` between here and a top level must
50
+ * re-raise them rather than fall through to another format or treat the layer as absent — see the
51
+ * {@link isConfigDiscoveryError} re-raises in {@link loadGlobalRawConfig}, {@link initConfig},
52
+ * {@link tryModuleConfig} and {@link tryJsonConfig}. Swallowing one would silently downgrade a hard
53
+ * config error to a different (or absent) config, which is the false-green this change exists to
54
+ * prevent.
55
+ *
44
56
  * @param raw The freshly loaded config layer (read-only here).
45
57
  * @param sourceLabel Human-readable source name for messages (e.g. the filename).
58
+ * @throws ConfigDiscoveryError when the layer carries a hard configuration error.
46
59
  */
47
60
  function validateRawConfigLayer(raw, sourceLabel) {
48
61
  // Only an object config can carry deprecated/unknown keys; a null/array/primitive config skips
@@ -51,11 +64,7 @@ function validateRawConfigLayer(raw, sourceLabel) {
51
64
  if (isRecordConfig(raw)) {
52
65
  const deprecatedIssues = findDeprecatedConfigIssues(raw);
53
66
  if (deprecatedIssues.length > 0) {
54
- displayError(`Invalid configuration in ${sourceLabel}:\n${formatDeprecatedConfigIssues(deprecatedIssues)}`);
55
- exit(1);
56
- // Unreachable past exit(1) in production; keeps the mocked-exit test path from falling
57
- // through into the schema parse below.
58
- return raw;
67
+ throw new ConfigDiscoveryError(`Invalid configuration in ${sourceLabel}:\n${formatDeprecatedConfigIssues(deprecatedIssues)}`, { sourceLabel });
59
68
  }
60
69
  // EXT-71 — the rule-grammar errors that must be seen BEFORE the schema parse: a bare string in
61
70
  // a rule list (whose message shows the object form of that same string) and a reserved `*`
@@ -63,9 +72,7 @@ function validateRawConfigLayer(raw, sourceLabel) {
63
72
  // validate` too, so the validator can never green-light a config a real run refuses.
64
73
  const grammarIssues = findApprovalsGrammarIssues(raw);
65
74
  if (grammarIssues.length > 0) {
66
- displayError(`Invalid configuration in ${sourceLabel}:\n${formatDeprecatedConfigIssues(grammarIssues)}`);
67
- exit(1);
68
- return raw;
75
+ throw new ConfigDiscoveryError(`Invalid configuration in ${sourceLabel}:\n${formatDeprecatedConfigIssues(grammarIssues)}`, { sourceLabel });
69
76
  }
70
77
  const unknownKeys = findUnknownTopLevelKeys(raw);
71
78
  if (unknownKeys.length > 0) {
@@ -76,11 +83,7 @@ function validateRawConfigLayer(raw, sourceLabel) {
76
83
  }
77
84
  const result = rawGthConfigSchema.safeParse(raw);
78
85
  if (!result.success) {
79
- displayError(`Invalid configuration in ${sourceLabel}:\n${formatConfigValidationError(result.error)}`);
80
- exit(1);
81
- // Unreachable past exit(1) in production; in specs exit() is mocked, so returning here keeps
82
- // a shape-invalid config from falling through into the profile check below.
83
- return raw;
86
+ throw new ConfigDiscoveryError(`Invalid configuration in ${sourceLabel}:\n${formatConfigValidationError(result.error)}`, { sourceLabel });
84
87
  }
85
88
  // CFG-26 — `approvals.rater` STRICT resolution (GS2-62): a named profile that does not
86
89
  // resolve to a real profile config is a hard error, never a silent fallback to the main model.
@@ -89,10 +92,8 @@ function validateRawConfigLayer(raw, sourceLabel) {
89
92
  if (isRecordConfig(raw)) {
90
93
  for (const ref of findApprovalsRaterProfiles(raw)) {
91
94
  if (!resolveIdentityProfileConfigPath(ref.profile)) {
92
- displayError(`Invalid configuration in ${sourceLabel}:\n` +
93
- ` - ${ref.path}: ${unresolvedRaterProfileMessage(ref)}`);
94
- exit(1);
95
- return raw;
95
+ throw new ConfigDiscoveryError(`Invalid configuration in ${sourceLabel}:\n` +
96
+ ` - ${ref.path}: ${unresolvedRaterProfileMessage(ref)}`, { sourceLabel, identityProfile: ref.profile });
96
97
  }
97
98
  }
98
99
  }
@@ -240,11 +241,12 @@ const RAW_CONFIG_VALIDATION_OPTIONS = {
240
241
  * config happens to be present / a global config exists," a distinction the loader's fall-through
241
242
  * deliberately blurs.
242
243
  *
243
- * PURE PREDICATE — never throws, never calls `exit` (contrast the loader's interactive-CLI
244
- * `exit(1)` safety net in {@link initConfig}). So batch/eval code (e.g. `gth eval --judge <profile>`
245
- * and BATCH-12's identity matrix) can pre-check an explicitly-requested profile and raise its OWN
246
- * catchable error / graceful exit code instead of dying on an uncatchable `process.exit`. A
247
- * blank/whitespace-only name counts as "no profile" → `undefined`.
244
+ * PURE PREDICATE — never throws, never calls `exit`, so it can be asked the question without
245
+ * committing to an outcome. {@link initConfig} uses it to enforce that an explicitly-named profile
246
+ * really exists (raising a catchable {@link ConfigDiscoveryError} when it does not), and callers
247
+ * that want to CLASSIFY rather than fail BATCH-12's identity matrix checks every declared identity
248
+ * up front so one message can name them all — ask it directly. A blank/whitespace-only name counts
249
+ * as "no profile" → `undefined`.
248
250
  *
249
251
  * @param identityProfile The explicitly-requested identity profile name.
250
252
  * @returns The resolved profile config path, or `undefined` when the profile has no config.
@@ -265,6 +267,31 @@ export function resolveIdentityProfileConfigPath(identityProfile) {
265
267
  }
266
268
  return undefined;
267
269
  }
270
+ /**
271
+ * CFG-36 — the ONE statement of "an explicitly-named identity profile that does not exist", shared
272
+ * by the run path ({@link initConfig}, which raises) and the read path ({@link validateConfig},
273
+ * which records a not-ok layer). Single-sourced deliberately: GS2-29's invariant is that
274
+ * `gth config validate` can never green-light a config a real run refuses, and two copies of this
275
+ * rule is exactly how that invariant rots.
276
+ *
277
+ * Returns the offending profile name, or `undefined` when there is nothing to complain about —
278
+ * no profile named (a blank/whitespace name counts as none, keeping the CFG-8 no-profile path
279
+ * untouched), an explicit `--config` path (which names the file to load and bypasses discovery),
280
+ * or a profile that resolves to its own config.
281
+ */
282
+ function findUnresolvedExplicitProfile(commandLineConfigOverrides) {
283
+ const profile = commandLineConfigOverrides.identityProfile?.trim();
284
+ if (!profile || commandLineConfigOverrides.customConfigPath) {
285
+ return undefined;
286
+ }
287
+ return resolveIdentityProfileConfigPath(profile) ? undefined : profile;
288
+ }
289
+ /** The message both paths report for {@link findUnresolvedExplicitProfile}'s failure. */
290
+ function identityProfileNotFoundMessage(profile) {
291
+ return (`identity profile "${profile}" not found: no config file in ` +
292
+ `${GSLOTH_DIR}/${GSLOTH_SETTINGS_DIR}/${profile}/ ` +
293
+ `(checked ${PROJECT_CONFIG_FORMATS.join(', ')})`);
294
+ }
268
295
  /**
269
296
  * Loads the global gsloth config (if present) from the global `~/.gsloth` folder.
270
297
  *
@@ -292,6 +319,13 @@ export async function loadGlobalRawConfig() {
292
319
  return validateRawConfigLayer(parsed, `${filename} (global)`);
293
320
  }
294
321
  catch (e) {
322
+ // CFG-36 — this catch exists to treat an UNREADABLE global as absent. A global that read
323
+ // fine and is MALFORMED is a hard configuration error (it used to `exit(1)` from inside
324
+ // the validator); swallowing it here would silently downgrade that to "ignoring it" and
325
+ // run under a different config — the exact false-green the throw was introduced to avoid.
326
+ if (isConfigDiscoveryError(e)) {
327
+ throw e;
328
+ }
295
329
  displayDebug(e instanceof Error ? e : String(e));
296
330
  displayWarning(`Failed to read global config from ${jsonPath}, ignoring it.`);
297
331
  return undefined;
@@ -308,6 +342,10 @@ export async function loadGlobalRawConfig() {
308
342
  return validateRawConfigLayer(configured, `${filename} (global)`);
309
343
  }
310
344
  catch (e) {
345
+ // CFG-36 — see the JSON branch above: a malformed global is a hard error, not an absent one.
346
+ if (isConfigDiscoveryError(e)) {
347
+ throw e;
348
+ }
311
349
  displayDebug(e instanceof Error ? e : String(e));
312
350
  displayWarning(`Failed to read global config from ${modulePath}, ignoring it.`);
313
351
  return undefined;
@@ -339,7 +377,8 @@ const MAX_EXTENDS_CHAIN_DEPTH = 50;
339
377
  * cycle, a missing base, an over-deep chain, or an unreadable base. It carries the SAME clear,
340
378
  * user-facing message the run path prints, so the two consumers can translate one shared failure
341
379
  * into their own convention WITHOUT the traversal being forked or the checks duplicated:
342
- * - the run path ({@link resolveConfigExtends}) → `displayError` + `exit(1)` (hard-fail a run),
380
+ * - the run path ({@link resolveConfigExtends}) → re-raised as a {@link ConfigDiscoveryError}, so
381
+ * the caller classifies it and chooses the exit code (CFG-36),
343
382
  * - the read path ({@link validateConfig}) → a `not-ok` layer with this message (collect, never
344
383
  * `exit`), so `gth config validate` mirrors what a real run would hit (GS2-29 invariant).
345
384
  */
@@ -383,15 +422,28 @@ export async function resolveConfigExtends(rawConfig, profileLabel) {
383
422
  return await composeExtends(rawConfig, profileLabel);
384
423
  }
385
424
  catch (e) {
386
- // GS2-73 — the traversal now RAISES a {@link ConfigExtendsError} rather than exiting inline, so
387
- // the read side ({@link validateConfig}) can report the same failure without terminating. On the
388
- // RUN path we translate it back to the original behaviour: print the clear message and exit(1).
425
+ // GS2-73 — the traversal RAISES a {@link ConfigExtendsError} rather than exiting inline, so the
426
+ // read side ({@link validateConfig}) can report the same failure without terminating.
427
+ //
428
+ // CFG-36 — the RUN path re-raises it as a {@link ConfigDiscoveryError} instead of printing and
429
+ // calling `exit(1)`. A profile whose `extends` base is missing (or forms a cycle) is a bad
430
+ // profile exactly as a profile with no config at all is, and both must be classifiable by the
431
+ // caller: exiting here from inside a library collapses `gth eval`'s harness-error (exit 2) and
432
+ // product-failure (exit 1) contract onto the same code, which is the collapse this node exists
433
+ // to remove. The CLI's top-level guard prints the same message and exits 1, so what a person at
434
+ // a terminal sees is unchanged.
435
+ //
436
+ // The message already names the profile, the base and the cycle; the optional detail fields are
437
+ // deliberately left unset rather than guessed at. `identityProfile` means "the profile that did
438
+ // not resolve", which for a missing base referenced from another profile is the BASE, not the
439
+ // `profileLabel` in hand here — a wrong value would be worse than an absent one.
440
+ //
441
+ // CFG-47 — the original travels as `cause`. Re-raising by message alone made the
442
+ // {@link ConfigExtendsError} (and any stack under it) unrecoverable from the wrapper, so a
443
+ // consumer that wanted the underlying failure had only the rendered string. The in-repo sibling
444
+ // ({@link MissingProviderKeyError} in {@link tryJsonConfig}) already carries it.
389
445
  if (e instanceof ConfigExtendsError) {
390
- displayError(e.message);
391
- exit(1);
392
- // Unreachable past exit(1) in production; LOAD-BEARING under the specs' mocked exit so the
393
- // caller never observes a silently-composed config after a cycle/missing base.
394
- throw new Error('Unexpected error occurred.');
446
+ throw new ConfigDiscoveryError(e.message, {}, { cause: e });
395
447
  }
396
448
  throw e;
397
449
  }
@@ -400,7 +452,7 @@ export async function resolveConfigExtends(rawConfig, profileLabel) {
400
452
  * GS2-73 — seed the `extends` chain from the selected profile's own name and run the throwing
401
453
  * traversal ({@link resolveExtendsChain}). Shared by BOTH consumers so the walk and its
402
454
  * cycle/missing-base checks live in ONE place: the run-path {@link resolveConfigExtends} (which
403
- * translates a {@link ConfigExtendsError} into `displayError` + `exit`) and the read-path
455
+ * re-raises a {@link ConfigExtendsError} as a {@link ConfigDiscoveryError}) and the read-path
404
456
  * {@link validateConfig} (which records it as a not-ok layer). Propagates the typed error to its
405
457
  * caller; the caller owns the reporting convention.
406
458
  */
@@ -554,11 +606,12 @@ async function readProjectConfiguredTui(commandLineConfigOverrides) {
554
606
  // profile whose `tui` is inherited while the run composes it, which is precisely the
555
607
  // reader-vs-run divergence this seam exists to prevent.
556
608
  //
557
- // Deliberately NOT inside the try above: on a cycle / missing base / malformed base,
558
- // `resolveConfigExtends` reports the problem and calls `exit(1)`, which ends the process — a
559
- // catch could not suppress that, and the sentinel throw past it is load-bearing under the specs'
560
- // mocked `exit` (the convention used throughout this file), so swallowing it would let a test
561
- // observe a silent fall-through production can never reach.
609
+ // Deliberately NOT inside the try above: a cycle, a missing base and a MALFORMED base all raise a
610
+ // {@link ConfigDiscoveryError} that must reach the top level to be reported. The try above exists
611
+ // to treat an unreadable config as "no configured tui"; extending it over these would swallow a
612
+ // hard configuration error and let the surface be selected from a config the run itself refuses
613
+ // to load. This reader runs BEFORE the TUI/readline choice, so the error surfaces here rather than
614
+ // through `createTuiSession` — either way it is reported once, by the CLI's top-level guard.
562
615
  if (typeof raw.extends === 'string') {
563
616
  raw = await resolveConfigExtends(raw, commandLineConfigOverrides.identityProfile);
564
617
  }
@@ -571,7 +624,13 @@ async function readProjectConfiguredTui(commandLineConfigOverrides) {
571
624
  export async function initConfig(commandLineConfigOverrides) {
572
625
  if (commandLineConfigOverrides.customConfigPath &&
573
626
  !existsSync(commandLineConfigOverrides.customConfigPath)) {
574
- throw new Error(`Provided manual config "${commandLineConfigOverrides.customConfigPath}" does not exist`);
627
+ // CFG-47 a {@link ConfigDiscoveryError}, not a plain `Error`. An explicitly named `-c` path
628
+ // that is not there is the same "this configuration cannot be used" failure as a named profile
629
+ // with no config, and it reaches the same two consumers: the CLI's top-level guard (prints the
630
+ // message, exits 1) and `gth eval` (harness error, exit 2). As a plain Error it was invisible to
631
+ // BOTH — it fell through the guard to the crash handler, so `gth -c <missing path> code` printed
632
+ // a false "TUI unavailable … falling back to the readline session" and then a crash snapshot.
633
+ throw new ConfigDiscoveryError(`Provided manual config "${commandLineConfigOverrides.customConfigPath}" does not exist`, { sourceLabel: commandLineConfigOverrides.customConfigPath });
575
634
  }
576
635
  // Clear the project root BEFORE discovery. Discovery and detection must resolve against cwd,
577
636
  // and the up-tree walk itself goes through getGslothConfigReadPath -> getProjectDir(); clearing
@@ -583,6 +642,26 @@ export async function initConfig(commandLineConfigOverrides) {
583
642
  // from cwd to the stop boundary (see findProjectConfigPath). Detection and loading share this
584
643
  // resolver, and the discovered dir becomes the base for the per-format cascade below.
585
644
  const discovered = findProjectConfigPath(commandLineConfigOverrides);
645
+ // GS2-62 / CFG-36 — an EXPLICITLY named identity profile (`-i <name>` / eval `--judge <name>`)
646
+ // must resolve to its OWN config; it must never silently run under some OTHER config. That is a
647
+ // false-green trap: `gth -i typo …` would run under the wrong model while appearing to use the
648
+ // named profile — in an authorization/eval context, hiding a real misconfiguration.
649
+ //
650
+ // Checked with the STRICT resolver, NOT with `discovered`. `findProjectConfigPath` deliberately
651
+ // falls back to a plain `<dir>/<config>` when the named profile has no config of its own (see its
652
+ // note, and the "Case C" spec in config.uptree.spec.ts), so gating on `!discovered` only catches
653
+ // the case where NO config exists anywhere — a project that has a plain config would sail past it
654
+ // and load that instead. `resolveIdentityProfileConfigPath` never falls through, which is what
655
+ // makes this check see the case the discovery gate cannot.
656
+ //
657
+ // The rule itself lives in findUnresolvedExplicitProfile, shared with `validateConfig` so the
658
+ // validator can never green-light a profile a run refuses (GS2-29).
659
+ const explicitProfile = findUnresolvedExplicitProfile(commandLineConfigOverrides);
660
+ if (explicitProfile) {
661
+ throw new ConfigDiscoveryError(identityProfileNotFoundMessage(explicitProfile), {
662
+ identityProfile: explicitProfile,
663
+ });
664
+ }
586
665
  const baseDir = discovered?.dir ?? getCurrentWorkDir();
587
666
  // Set the project root for post-config, project-relative artifact resolution (guidelines,
588
667
  // prompts, .gsloth-settings, outputs). up-tree and --config both set it here; a global-only /
@@ -602,24 +681,10 @@ export async function initConfig(commandLineConfigOverrides) {
602
681
  // standalone global config (loaded alone) before erroring. Project config still takes
603
682
  // precedence: this branch only runs when there is no project file to apply the global under.
604
683
  if (!discovered) {
605
- // GS2-62 an EXPLICITLY named identity profile (`-i <name>` / eval `--judge <name>`) that
606
- // discovered no project config must NOT silently fall back to the global config. Doing so is a
607
- // false-green trap: `gth -i typo …` would run under the GLOBAL model while appearing to use the
608
- // named profile (in an authorization/eval context that hides a real misconfiguration). Fail
609
- // loudly instead. Gated on a non-empty identityProfile, so the CFG-8 no-profile global fallback
610
- // just below is UNTOUCHED — a run with no profile still loads the global exactly as before.
611
- const explicitProfile = commandLineConfigOverrides.identityProfile?.trim();
612
- if (explicitProfile) {
613
- displayError(`identity profile "${explicitProfile}" not found: no config file in ` +
614
- `${GSLOTH_DIR}/${GSLOTH_SETTINGS_DIR}/${explicitProfile}/ ` +
615
- `(checked ${PROJECT_CONFIG_FORMATS.join(', ')})`);
616
- exit(1);
617
- // Unreachable past exit(1) in production. In specs exit() is mocked to a no-op, so this throw
618
- // is LOAD-BEARING: without it execution would fall through into loadGlobalRawConfig() below
619
- // and the test would observe the global silently loaded — masking the very regression this
620
- // guard fixes. Matches the loader's existing post-exit sentinel-throw pattern.
621
- throw new Error('Unexpected error occurred.');
622
- }
684
+ // The explicitly-named-profile guard that used to sit here now runs BEFORE discovery is
685
+ // consulted at all (see above), because gating it on `!discovered` missed the case where a
686
+ // plain project config exists. A run with no profile reaches the global fallback below exactly
687
+ // as before.
623
688
  const globalRawConfig = await loadGlobalRawConfig();
624
689
  if (globalRawConfig) {
625
690
  if (globalRawConfig.llm &&
@@ -628,10 +693,11 @@ export async function initConfig(commandLineConfigOverrides) {
628
693
  // Route the global config through the same path the project JSON uses.
629
694
  return await tryJsonConfig(globalRawConfig, commandLineConfigOverrides);
630
695
  }
631
- displayError('Global configuration found but it is not in valid format. Should at least define llm.type');
632
- exit(1);
633
- // Unreachable past exit(1) in production; keeps TS happy and prevents test exit.
634
- throw new Error('Unexpected error occurred.');
696
+ // CFG-47 a global config that read fine and does not define `llm.type` is "config present
697
+ // and unusable", the same class CFG-36 converted: raise it, let the caller choose the exit
698
+ // code. The message is the one this branch has always printed, so the CLI's top-level guard
699
+ // reproduces the previous output exactly.
700
+ throw new ConfigDiscoveryError('Global configuration found but it is not in valid format. Should at least define llm.type', { sourceLabel: `${USER_PROJECT_CONFIG_JSON} (global)` });
635
701
  }
636
702
  }
637
703
  // Try loading the JSON/JSONC config file first (GS2-69 — an explicit `-c foo.jsonc` takes
@@ -656,12 +722,12 @@ export async function initConfig(commandLineConfigOverrides) {
656
722
  return await tryJsonConfig(jsonConfig, commandLineConfigOverrides);
657
723
  }
658
724
  else {
659
- error(`${jsonConfigPath} is not in valid format. Should at least define llm.type`);
660
- exit(1);
661
- // noinspection ExceptionCaughtLocallyJS
662
- // This throw is unreachable due to exit(1) above, but satisfies TS type analysis and prevents tests from exiting
725
+ // CFG-47 same class, same treatment. Deliberately a ConfigDiscoveryError and not a plain
726
+ // one: the catch below re-raises this class and swallows everything else into the next
727
+ // FORMAT, so a plain throw here would silently fall through to the module loader and end in
728
+ // the terminal "No configuration file found" hiding the real, clearly-worded problem.
663
729
  // noinspection ExceptionCaughtLocallyJS
664
- throw new Error('Unexpected error occurred.');
730
+ throw new ConfigDiscoveryError(`${jsonConfigPath} is not in valid format. Should at least define llm.type`, { sourceLabel: jsonConfigName });
665
731
  }
666
732
  }
667
733
  catch (e) {
@@ -669,10 +735,17 @@ export async function initConfig(commandLineConfigOverrides) {
669
735
  // failure. This catch exists to move on to the next config FORMAT when the JSON layer could
670
736
  // not be read; a config that read fine and named a provider we have no key for is not a
671
737
  // read failure, and falling through would end in the terminal "No configuration file found"
672
- // exitthe opposite of catchable, and a misleading message besides.
738
+ // failurewhich blames a missing config for what is actually a missing key, burying the
739
+ // clearly-worded message this branch already has.
673
740
  if (isMissingProviderKeyError(e)) {
674
741
  throw e;
675
742
  }
743
+ // CFG-36 — same reasoning for a MALFORMED config: the config read fine and is invalid, which
744
+ // is a hard error the user must see. Falling through to the next FORMAT would end in the
745
+ // terminal "No configuration file found" failure, hiding the real (clearly-worded) problem.
746
+ if (isConfigDiscoveryError(e)) {
747
+ throw e;
748
+ }
676
749
  displayDebug(e instanceof Error ? e : String(e));
677
750
  displayError(`Failed to read config from ${jsonConfigName}, will try other formats.`);
678
751
  // Continue to try other formats
@@ -726,15 +799,24 @@ async function tryModuleConfig(format, commandLineConfigOverrides, baseDir) {
726
799
  return await mergeConfig(mergedWithGlobal, commandLineConfigOverrides);
727
800
  }
728
801
  catch (e) {
802
+ // CFG-36 — a config that read fine and is MALFORMED (or names an unresolvable profile) is a
803
+ // hard error, not a reason to try the next format. Re-raise before the fall-through, exactly
804
+ // as the JSON branch in initConfig does.
805
+ if (isConfigDiscoveryError(e)) {
806
+ throw e;
807
+ }
729
808
  displayDebug(e instanceof Error ? e : String(e));
730
809
  if (nextFormat) {
731
810
  displayError(`Failed to read config from ${filename}, will try other formats.`);
732
811
  // Continue to try other formats
733
812
  return await tryModuleConfig(nextFormat, commandLineConfigOverrides, baseDir);
734
813
  }
735
- displayError(`Failed to read config from ${filename}.`);
736
- displayError(`No valid configuration found. Please create a valid configuration file.`);
737
- exit(1);
814
+ // CFG-47 — the last format in the chain failed to read. The config IS present and cannot be
815
+ // used, so this raises like the rest of the class. Both lines are kept in the one message:
816
+ // the first names the file that failed, the second is the advice, and the guard prints them
817
+ // together exactly as the two `displayError` calls did.
818
+ throw new ConfigDiscoveryError(`Failed to read config from ${filename}.\n` +
819
+ `No valid configuration found. Please create a valid configuration file.`, { sourceLabel: filename });
738
820
  }
739
821
  }
740
822
  else if (nextFormat) {
@@ -742,14 +824,19 @@ async function tryModuleConfig(format, commandLineConfigOverrides, baseDir) {
742
824
  return await tryModuleConfig(nextFormat, commandLineConfigOverrides, baseDir);
743
825
  }
744
826
  else {
745
- // No config files found
746
- displayError('No configuration file found. Please create one of: ' +
827
+ // No config files found.
828
+ //
829
+ // CFG-47 — the terminal "nothing to load" exit, raised rather than exited. It is the same class
830
+ // for the same reason as the rest: the caller must classify it. `gth eval` in particular needs
831
+ // "the harness has no config" to be a harness error (exit 2), not the exit 1 that means the SUT
832
+ // ran and failed. The CLI's top-level guard prints this message and exits 1, so a person at a
833
+ // terminal sees exactly what they saw before — and `startSession` still runs the first-run
834
+ // dialog ahead of this on an interactive TTY with no config anywhere (CFG-10), so the ordinary
835
+ // unconfigured path never reaches here.
836
+ throw new ConfigDiscoveryError('No configuration file found. Please create one of: ' +
747
837
  `${USER_PROJECT_CONFIG_JSON}, ${USER_PROJECT_CONFIG_JS}, or ${USER_PROJECT_CONFIG_MJS} ` +
748
838
  'in your project directory.');
749
- exit(1);
750
839
  }
751
- // This throw is unreachable due to exit(1) above, but satisfies TS type analysis and prevents tests from exiting
752
- throw new Error('Unexpected error occurred.');
753
840
  }
754
841
  /**
755
842
  * Process JSON LLM config by creating the appropriate LLM instance
@@ -763,8 +850,12 @@ export async function tryJsonConfig(jsonConfig, commandLineConfigOverrides) {
763
850
  // Get the type of LLM (e.g. 'vertexai', 'anthropic') - this should exist
764
851
  const llmType = jsonConfig.llm.type;
765
852
  if (!llmType) {
766
- displayError('LLM type not specified in config.');
767
- exit(1);
853
+ // CFG-47 same class, raised rather than exited. Note this throw is caught by THIS
854
+ // function's own catch below, which re-raises the class before its "Error processing LLM
855
+ // config" wrapper; without that re-raise the clear message here would be re-worded and then
856
+ // exited on anyway.
857
+ // noinspection ExceptionCaughtLocallyJS
858
+ throw new ConfigDiscoveryError('LLM type not specified in config.');
768
859
  }
769
860
  // Get the configuration for the specific LLM type
770
861
  const llmConfig = jsonConfig.llm;
@@ -791,19 +882,35 @@ export async function tryJsonConfig(jsonConfig, commandLineConfigOverrides) {
791
882
  }
792
883
  }
793
884
  else {
794
- displayWarning(`Config module for ${llmType} does not have processJsonConfig function.`);
795
- exit(1);
885
+ // CFG-47 the node called this site the one that is "genuinely a different case". Measured,
886
+ // it is not: `#src/providers/<type>.js` resolves against the SAME directory that holds five
887
+ // real modules with no `processJsonConfig` export (modelCatalog, modelDiscovery,
888
+ // geminiThinking, geminiSchemaSanitizer, configurationPassthrough), so `llm.type:
889
+ // "modelCatalog"` reaches here from an ordinary config file — verified by resolving that
890
+ // specifier at runtime. It is therefore user-reachable "config present and unusable" like
891
+ // the rest, and is converted with them. Kept on the warning wording it has always had.
892
+ // noinspection ExceptionCaughtLocallyJS
893
+ throw new ConfigDiscoveryError(`Config module for ${llmType} does not have processJsonConfig function.`);
796
894
  }
797
895
  }
798
896
  else {
799
- displayError('No LLM configuration found in config.');
800
- exit(1);
897
+ // noinspection ExceptionCaughtLocallyJS
898
+ throw new ConfigDiscoveryError('No LLM configuration found in config.');
801
899
  }
802
900
  }
803
901
  catch (e) {
902
+ // CFG-47 — the four sites above raise INTO this catch, so it must re-raise the class before its
903
+ // own wrapper reaches them. Without this, a clear "LLM type not specified in config." would be
904
+ // re-worded as "Error processing LLM config: …" and exited on anyway — the local catch would
905
+ // undo the conversion three lines after it happened. Same shape as the re-raises in
906
+ // `initConfig` and `tryModuleConfig`.
907
+ if (isConfigDiscoveryError(e)) {
908
+ throw e;
909
+ }
804
910
  if (e instanceof Error && e.message.includes('Cannot find module')) {
805
- displayError(`LLM type '${jsonConfig.llm.type}' not supported.`);
806
- exit(1);
911
+ // CFG-47 — a configured `llm.type` we have no provider module for is a config error, not a
912
+ // process exit. `{ cause: e }` keeps the resolver's own failure reachable.
913
+ throw new ConfigDiscoveryError(`LLM type '${jsonConfig.llm.type}' not supported.`, {}, { cause: e });
807
914
  }
808
915
  else {
809
916
  const message = `Error processing LLM config: ${e instanceof Error ? e.message : String(e)}`;
@@ -825,12 +932,13 @@ export async function tryJsonConfig(jsonConfig, commandLineConfigOverrides) {
825
932
  if (missingKey) {
826
933
  throw new MissingProviderKeyError(message, missingKey, { cause: e });
827
934
  }
828
- displayError(message);
829
- exit(1);
935
+ // CFG-47 — and when it is NOT a missing key, raise too. This site's immediate neighbour three
936
+ // lines up already threw, so the file disagreed with itself about whether a provider that
937
+ // could not be built terminates the process; it does not. The message is unchanged and
938
+ // `{ cause: e }` keeps the provider's own error reachable, as the missing-key branch does.
939
+ throw new ConfigDiscoveryError(message, {}, { cause: e });
830
940
  }
831
941
  }
832
- // This throw is unreachable due to exit(1) above, but satisfies TS type analysis and prevents tests from exiting
833
- throw new Error('Unexpected error occurred.');
834
942
  }
835
943
  /**
836
944
  * Config array fields whose values ADD UP across merge layers (global → project): the
@@ -1260,6 +1368,31 @@ async function loadGlobalRawConfigUnvalidated() {
1260
1368
  */
1261
1369
  export async function validateConfig(commandLineConfigOverrides) {
1262
1370
  const layers = [];
1371
+ // CFG-36 — mirror the run's STRICT named-profile rule before anything is read. `initConfig`
1372
+ // refuses outright when an explicitly-named profile has no config of its own, so a validator that
1373
+ // walked on would report OK for a config the run rejects — and it would do so in the ordinary
1374
+ // case, because discovery falls back to a plain project config for an unresolved profile. That is
1375
+ // the GS2-29 divergence in its purest form: `gth config validate -i typo` green-lighting a run
1376
+ // that cannot start. Reported as a not-ok layer rather than thrown, because the read side
1377
+ // COLLECTS and never terminates; returned immediately because a run gets no further either.
1378
+ const unresolvedProfile = findUnresolvedExplicitProfile(commandLineConfigOverrides);
1379
+ if (unresolvedProfile) {
1380
+ return {
1381
+ // `found: true` so the caller renders THIS message. `found: false` means "nothing to
1382
+ // validate, run `gth init`", which would both discard the real diagnosis and misdirect a user
1383
+ // whose actual problem is a mistyped `-i`.
1384
+ found: true,
1385
+ ok: false,
1386
+ layers: [
1387
+ {
1388
+ sourceLabel: `${GSLOTH_DIR}/${GSLOTH_SETTINGS_DIR}/${unresolvedProfile}/`,
1389
+ ok: false,
1390
+ warnings: [],
1391
+ errorMessage: identityProfileNotFoundMessage(unresolvedProfile),
1392
+ },
1393
+ ],
1394
+ };
1395
+ }
1263
1396
  // Project layer first (run order): initConfig validates the discovered project config before
1264
1397
  // applying the global base. A parse failure here propagates (surfaced by the caller).
1265
1398
  const discovered = findProjectConfigPath(commandLineConfigOverrides);
@@ -1282,7 +1415,11 @@ export async function validateConfig(commandLineConfigOverrides) {
1282
1415
  await composeExtends(raw, commandLineConfigOverrides.identityProfile);
1283
1416
  }
1284
1417
  catch (e) {
1285
- if (e instanceof ConfigExtendsError) {
1418
+ // CFG-36 a MALFORMED base layer now raises a ConfigDiscoveryError from
1419
+ // `validateRawConfigLayer` instead of exiting the process. Record it as a not-ok layer
1420
+ // alongside the traversal's own failures: the read side collects, it never terminates, so
1421
+ // `gth config validate` reports the broken base rather than dying on it.
1422
+ if (e instanceof ConfigExtendsError || isConfigDiscoveryError(e)) {
1286
1423
  layer.ok = false;
1287
1424
  layer.errorMessage = e.message;
1288
1425
  }