typeclaw 0.48.0 → 0.48.2

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 (41) hide show
  1. package/package.json +1 -1
  2. package/src/agent/plugin-tools.ts +68 -9
  3. package/src/agent/system-prompt.ts +8 -0
  4. package/src/agent-messenger/config-dir.ts +131 -0
  5. package/src/agent-messenger/host-config-dir.ts +121 -0
  6. package/src/bundled-plugins/guard/policies/managed-config.ts +7 -20
  7. package/src/bundled-plugins/guard/policies/non-workspace-write.ts +7 -20
  8. package/src/bundled-plugins/guard/policies/skill-authoring.ts +8 -20
  9. package/src/bundled-plugins/memory/README.md +2 -0
  10. package/src/bundled-plugins/memory/dreaming.ts +2 -0
  11. package/src/bundled-plugins/security/policies/private-surface-read.ts +10 -23
  12. package/src/cli/channel.ts +15 -16
  13. package/src/cli/hostd.ts +8 -13
  14. package/src/cli/init.ts +21 -7
  15. package/src/cli/model.ts +6 -3
  16. package/src/cli/restart.ts +13 -15
  17. package/src/compose/restart.ts +14 -6
  18. package/src/config/providers.ts +106 -143
  19. package/src/container/agent-operation-lock.ts +85 -0
  20. package/src/container/controller.ts +6 -0
  21. package/src/container/index.ts +8 -0
  22. package/src/container/restart.ts +55 -0
  23. package/src/container/shared.ts +5 -0
  24. package/src/container/start.ts +112 -4
  25. package/src/container/stop.ts +17 -1
  26. package/src/hostd/daemon.ts +1 -1
  27. package/src/hostd/paths.ts +17 -0
  28. package/src/hostd/supervisor.ts +27 -10
  29. package/src/init/dockerfile.ts +6 -5
  30. package/src/init/env-file.ts +42 -18
  31. package/src/init/instagram-auth.ts +55 -31
  32. package/src/init/kakaotalk-auth.ts +1 -5
  33. package/src/init/line-auth.ts +45 -29
  34. package/src/init/models-dev.ts +114 -7
  35. package/src/init/opengateway-catalog.ts +120 -0
  36. package/src/path-safety/real-intended-path.ts +111 -0
  37. package/src/sandbox/canonical-secrets.ts +10 -1
  38. package/src/sandbox/errors.ts +0 -32
  39. package/src/sandbox/privileged-runtime.ts +8 -94
  40. package/src/test-helpers/fake-agent-operation-lock.ts +66 -0
  41. package/typeclaw.schema.json +10 -30
@@ -654,16 +654,12 @@ async function maybePromptCredentialRefresh(
654
654
  return
655
655
  }
656
656
 
657
- const stopped = await controller.stop({ cwd })
658
- if (!stopped.ok) {
659
- console.error(errorLine(`Restart failed during stop: ${stopped.reason}`))
660
- process.exit(1)
661
- }
662
- const started = await controller.start({ cwd, preferredHostPort: config.port, cliEntry: process.argv[1] })
663
- if (!started.ok) {
664
- console.error(errorLine(`Restart failed during start: ${started.reason}`))
657
+ const restarted = await controller.restart({ cwd, preferredHostPort: config.port, cliEntry: process.argv[1] })
658
+ if (!restarted.ok) {
659
+ console.error(errorLine(formatRestartFailure(restarted.reason)))
665
660
  process.exit(1)
666
661
  }
662
+ const started = restarted.start
667
663
  done({
668
664
  title: c.green(`${label} ${verbPast}. Restarted ${started.plan.containerName} on host port ${started.hostPort}.`),
669
665
  hints: [
@@ -2002,16 +1998,12 @@ async function maybePromptRestart(
2002
1998
  return
2003
1999
  }
2004
2000
 
2005
- const stopped = await controller.stop({ cwd })
2006
- if (!stopped.ok) {
2007
- console.error(errorLine(`Restart failed during stop: ${stopped.reason}`))
2008
- process.exit(1)
2009
- }
2010
- const started = await controller.start({ cwd, preferredHostPort: config.port, cliEntry: process.argv[1] })
2011
- if (!started.ok) {
2012
- console.error(errorLine(`Restart failed during start: ${started.reason}`))
2001
+ const restarted = await controller.restart({ cwd, preferredHostPort: config.port, cliEntry: process.argv[1] })
2002
+ if (!restarted.ok) {
2003
+ console.error(errorLine(formatRestartFailure(restarted.reason)))
2013
2004
  process.exit(1)
2014
2005
  }
2006
+ const started = restarted.start
2015
2007
  done({
2016
2008
  title: c.green(
2017
2009
  `${label} channel ${verb}. Restarted ${started.plan.containerName} on host port ${started.hostPort}.`,
@@ -2022,3 +2014,10 @@ async function maybePromptRestart(
2022
2014
  ],
2023
2015
  })
2024
2016
  }
2017
+
2018
+ function formatRestartFailure(reason: string): string {
2019
+ if (reason.startsWith('stop failed: ')) return `Restart failed during stop: ${reason.slice('stop failed: '.length)}`
2020
+ if (reason.startsWith('start failed: '))
2021
+ return `Restart failed during start: ${reason.slice('start failed: '.length)}`
2022
+ return `Restart failed: ${reason}`
2023
+ }
package/src/cli/hostd.ts CHANGED
@@ -5,9 +5,8 @@ import {
5
5
  resolveController,
6
6
  resolveHostPort,
7
7
  resolveTuiToken,
8
- type StartOptions,
9
- type StartResult,
10
- type StopResult,
8
+ type RestartOptions,
9
+ type RestartResult,
11
10
  } from '@/container'
12
11
  import {
13
12
  applyCredentialRotation,
@@ -116,17 +115,15 @@ export const hostdCommand = defineCommand({
116
115
 
117
116
  export type HostdRestartDeps = {
118
117
  validateConfig: (cwd: string) => ValidateConfigResult
119
- stop: (opts: { cwd: string }) => Promise<StopResult>
120
118
  loadConfigSync: (cwd: string) => Config
121
- start: (opts: StartOptions) => Promise<StartResult>
119
+ restart: (opts: RestartOptions) => Promise<RestartResult>
122
120
  }
123
121
 
124
122
  const controller = resolveController()
125
123
  const defaultRestartDeps: HostdRestartDeps = {
126
124
  validateConfig,
127
- stop: (opts) => controller.stop(opts),
128
125
  loadConfigSync,
129
- start: (opts) => controller.start(opts),
126
+ restart: (opts) => controller.restart(opts),
130
127
  }
131
128
 
132
129
  export function buildHostdRestart(
@@ -134,7 +131,7 @@ export function buildHostdRestart(
134
131
  deps: HostdRestartDeps = defaultRestartDeps,
135
132
  daemonVersion?: string,
136
133
  ): SupervisorRestart {
137
- return async ({ containerName, cwd, build = false, currentHostDaemon }) => {
134
+ return async ({ containerName, cwd, build = false, currentHostDaemon, operationLease }) => {
138
135
  const drift = await detectSourceDrift(cliEntry, daemonVersion)
139
136
  if (drift) return { ok: false, reason: drift }
140
137
 
@@ -142,19 +139,17 @@ export function buildHostdRestart(
142
139
  if (!validated.ok) {
143
140
  return { ok: false, reason: `invalid config for ${containerName}: ${validated.reason}` }
144
141
  }
145
- const stopResult = await deps.stop({ cwd })
146
- if (!stopResult.ok) return { ok: false, reason: `stop failed: ${stopResult.reason}` }
147
-
148
142
  const cfg = deps.loadConfigSync(cwd)
149
- const startResult = await deps.start({
143
+ const restartResult = await deps.restart({
150
144
  cwd,
151
145
  preferredHostPort: cfg.port,
152
146
  forceBuild: build,
153
147
  cliEntry,
154
148
  reuseCurrentHostDaemon: true,
149
+ operationLease,
155
150
  ...(currentHostDaemon ? { currentHostDaemon } : {}),
156
151
  })
157
- if (!startResult.ok) return { ok: false, reason: `start failed: ${startResult.reason}` }
152
+ if (!restartResult.ok) return restartResult
158
153
  return { ok: true }
159
154
  }
160
155
  }
package/src/cli/init.ts CHANGED
@@ -64,7 +64,12 @@ import {
64
64
  type WizardCheckpointStore,
65
65
  } from '@/init/checkpoint'
66
66
  import { runKakaotalkBootstrap } from '@/init/kakaotalk-auth'
67
- import { customModelMetaFromOption, fetchModelOptions, type ModelOption } from '@/init/models-dev'
67
+ import {
68
+ customModelMetaFromOption,
69
+ fetchModelOptions,
70
+ type FetchModelsResult,
71
+ type ModelOption,
72
+ } from '@/init/models-dev'
68
73
  import { makeOAuthLoginRunner, type OAuthLoginResult } from '@/init/oauth-login'
69
74
  import { detectInitProgress } from '@/init/progress'
70
75
  import { runTeamsBootstrap } from '@/init/teams-auth'
@@ -1223,14 +1228,23 @@ function oauthDiscoveryRef(providerId: KnownProviderId): KnownModelRef {
1223
1228
 
1224
1229
  async function loadCatalog(): Promise<NonNullable<WizardState['catalog']>> {
1225
1230
  const s = spinner()
1226
- s.start('Loading model catalog from models.dev...')
1227
- const { options, source, warning } = await fetchModelOptions()
1228
- if (source === 'curated') {
1229
- s.stop(`Using built-in catalog (models.dev unavailable: ${warning ?? 'unknown'})`)
1231
+ s.start('Loading model catalogs...')
1232
+ const result = await fetchModelOptions()
1233
+ const unavailable = catalogFallbackLabels(result)
1234
+ if (unavailable.length > 0) {
1235
+ s.stop(`Loaded model catalog with fallback for ${unavailable.join(', ')}.`)
1230
1236
  } else {
1231
- s.stop('Loaded model catalog.')
1237
+ s.stop('Loaded model catalogs.')
1232
1238
  }
1233
- return warning !== undefined ? { options, source, warning } : { options, source }
1239
+ return result
1240
+ }
1241
+
1242
+ function catalogFallbackLabels(result: FetchModelsResult): string[] {
1243
+ const labels: string[] = []
1244
+ if (result.sources.modelsDev === 'unavailable') labels.push('native providers')
1245
+ if (result.sources.openGatewayCatalog === 'unavailable') labels.push('OpenGateway models')
1246
+ else if (result.sources.openGatewayPrices === 'unavailable') labels.push('OpenGateway prices')
1247
+ return labels
1234
1248
  }
1235
1249
 
1236
1250
  async function pickVendor(
package/src/cli/model.ts CHANGED
@@ -541,15 +541,18 @@ function describeRef(ref: string): string {
541
541
  }
542
542
 
543
543
  async function printAvailableRefs(): Promise<void> {
544
- const { options, source, warning } = await fetchModelOptions()
544
+ const { options, sources, warnings } = await fetchModelOptions()
545
545
  if (options.length === 0) {
546
546
  console.log(c.dim('No models registered.'))
547
547
  return
548
548
  }
549
549
  console.log(c.dim('Use `typeclaw model set <profile> <ref>` to apply.'))
550
- if (source === 'curated' && warning !== undefined) {
551
- console.log(c.dim(`Using built-in catalog (models.dev unavailable: ${warning}).`))
550
+ if (sources.modelsDev === 'unavailable') console.log(c.dim('Using built-in models for native providers.'))
551
+ if (sources.openGatewayCatalog === 'unavailable') console.log(c.dim('Using the built-in OpenGateway anchor.'))
552
+ else if (sources.openGatewayPrices === 'unavailable') {
553
+ console.log(c.dim('OpenGateway models loaded without live pricing.'))
552
554
  }
555
+ if (warnings.length > 0) console.log(c.dim(`Catalog warnings: ${warnings.join('; ')}`))
553
556
  for (const providerId of Object.keys(KNOWN_PROVIDERS) as KnownProviderId[]) {
554
557
  const providerOptions = options.filter((option) => option.providerId === providerId)
555
558
  if (providerOptions.length === 0) continue
@@ -74,28 +74,26 @@ export const restartCommand = defineCommand({
74
74
 
75
75
  const stopSpin = spinner()
76
76
  stopSpin.start('Stopping container...')
77
- const stopped = await controller.stop({ cwd })
78
- if (!stopped.ok) {
79
- stopSpin.error(stopped.reason)
80
- process.exit(1)
81
- }
82
- stopSpin.stop(stopped.running ? `Stopped ${c.cyan(stopped.containerName)}.` : 'Already stopped.')
83
-
84
- const startSpin = spinner()
85
- startSpin.start('Starting container...')
86
- const started = await controller.start({
77
+ let startSpin: ReturnType<typeof spinner> | undefined
78
+ const restarted = await controller.restart({
87
79
  cwd,
88
80
  preferredHostPort: Number(args.port),
89
81
  forceBuild: args.build,
90
82
  cliEntry: process.argv[1],
83
+ onStopped: (stopped) => {
84
+ stopSpin.stop(stopped.running ? `Stopped ${c.cyan(stopped.containerName)}.` : 'Already stopped.')
85
+ startSpin = spinner()
86
+ startSpin.start('Starting container...')
87
+ },
91
88
  })
92
- if (!started.ok) {
93
- startSpin.error(started.reason)
89
+ if (!restarted.ok) {
90
+ if (startSpin === undefined) stopSpin.error(restarted.reason)
91
+ else startSpin.error(restarted.reason)
94
92
  process.exit(1)
95
93
  }
96
- startSpin.stop('Started.')
94
+ startSpin?.stop('Started.')
97
95
 
98
- reportConfigWarnings(started.dockerfileWarnings)
99
- console.log(renderStartSuccess(started))
96
+ reportConfigWarnings(restarted.start.dockerfileWarnings)
97
+ console.log(renderStartSuccess(restarted.start))
100
98
  },
101
99
  })
@@ -58,12 +58,20 @@ async function runOne(
58
58
  if (!validated.ok) return { name, ok: false, reason: validated.reason }
59
59
  try {
60
60
  const controller = resolveController()
61
- const stopped = await controller.stop({ cwd })
62
- if (!stopped.ok) return { name, ok: false, reason: stopped.reason }
63
- onStopped()
64
- const started = await controller.start({ cwd, preferredHostPort, forceBuild, cliEntry })
65
- if (!started.ok) return { name, ok: false, reason: started.reason }
66
- return { name, ok: true, data: { stop: stopped, start: started }, warnings: validated.warnings }
61
+ const restarted = await controller.restart({
62
+ cwd,
63
+ preferredHostPort,
64
+ forceBuild,
65
+ cliEntry,
66
+ onStopped,
67
+ })
68
+ if (!restarted.ok) return { name, ok: false, reason: restarted.reason }
69
+ return {
70
+ name,
71
+ ok: true,
72
+ data: { stop: restarted.stop, start: restarted.start },
73
+ warnings: validated.warnings,
74
+ }
67
75
  } catch (error) {
68
76
  return { name, ok: false, reason: error instanceof Error ? error.message : String(error) }
69
77
  }
@@ -781,10 +781,14 @@ export const KNOWN_PROVIDERS = {
781
781
  // the console, the OpenAI-SDK example, and the agent docs all use bare `/v1`.
782
782
  //
783
783
  // Model lineup spans two families on the same Console API + key:
784
- // * Closed chat models from the official agent API reference
785
- // (console.upstage.ai/api/docs/for-agents) as of 2026-03-23:
786
- // solar-pro3 (102B MoE / 12B active, 128K, flagship), solar-pro2
787
- // (31B, 65K), solar-mini (10.7B, 32K, cheap/fast).
784
+ // * Closed chat models. solar-pro4 (512K context / 128K output, released
785
+ // 2026-08-06, current flagship) is per the live model page
786
+ // console.upstage.ai/docs/models/solar-pro4; the rest are from the agent
787
+ // API reference (console.upstage.ai/api/docs/for-agents) as of
788
+ // 2026-03-23: solar-pro3 (102B MoE / 12B active, 128K), solar-pro2
789
+ // (31B, 65K), solar-mini (10.7B, 32K, cheap/fast). Note that the
790
+ // for-agents dump is stale (self-dated 2026-03-07) and omits pro4 — its
791
+ // absence there is not evidence the model is unavailable.
788
792
  // * The open-weight Solar Open family: solar-open2 (Solar Open 2), a 102B
789
793
  // MoE / 12B active model with a 128K context. It launched on the Console
790
794
  // API for the Solar Agent Partner program (Stage 1, from 2026-07-17) —
@@ -799,6 +803,7 @@ export const KNOWN_PROVIDERS = {
799
803
  // versioned release.
800
804
  //
801
805
  // Costs are USD per 1M tokens from upstage.ai/pricing/api (VAT-exclusive).
806
+ // solar-pro4 is $0.30 in / $1.20 out with a $0.06 "Input(Cached)" rate.
802
807
  // solar-pro3/pro2 publish an "Input(Cached)" cache-read rate ($0.015); no
803
808
  // cache-write surcharge is published, so cacheWrite is 0. solar-mini
804
809
  // publishes a single flat $0.15 rate with no cache line, so cacheRead and
@@ -819,18 +824,37 @@ export const KNOWN_PROVIDERS = {
819
824
  // definitions are both undocumented, so supportsUsageInStreaming and
820
825
  // supportsStrictMode are false (both conservative — pi-ai defaults them to
821
826
  // true, and Upstage only documents `strict` inside response_format.json_schema,
822
- // never on tool defs). `reasoning_effort` IS documented for solar-pro3/pro2
823
- // (and the solar-open2 chat template), so supportsReasoningEffort stays true
824
- // there; solar-mini ignores it, so it is false and reasoning is off.
827
+ // never on tool defs). `reasoning_effort` IS documented for solar-pro4,
828
+ // solar-open2, solar-pro3 and solar-pro2, so supportsReasoningEffort stays
829
+ // true there; solar-mini rejects it, so it is false and reasoning is off.
825
830
  //
826
- // Upstage documents reasoning_effort values minimal/low/medium/high only. pi
827
- // adds an `xhigh` level that typeclaw's attention escalation can select
828
- // (src/agent/attention-escalation.ts), and pi-ai passes the pi level straight
829
- // through as reasoning_effort when no map is set — which would send the
830
- // unsupported `xhigh`. `thinkingLevelMap` clamps xhigh -> high (and maps off ->
831
- // null so nothing is sent) on the reasoning models so every emitted
832
- // reasoning_effort is a value Upstage accepts. solar-mini needs no map (it
833
- // never sends reasoning_effort).
831
+ // Reasoning splits the lineup into two groups with OPPOSITE defaults, per
832
+ // Upstage's reasoning table (console.upstage.ai/docs/capabilities/generate/
833
+ // reasoning). pi adds an `xhigh` level that typeclaw's attention escalation
834
+ // can select (src/agent/attention-escalation.ts), and pi-ai passes the pi
835
+ // level straight through as reasoning_effort when no map is set, so each
836
+ // group needs a different `thinkingLevelMap`:
837
+ //
838
+ // * solar-pro4 and solar-open2 — omitted reasoning_effort means reasoning
839
+ // is ON; `none`/`minimal` turn it OFF; low/medium/high/xhigh/max turn it
840
+ // on. So `off` maps to the string 'none', NOT null: pi-ai emits a string
841
+ // off value verbatim and omits the field only when it is null, and
842
+ // omitting it here would silently leave reasoning enabled (and billed) on
843
+ // every non-reasoning turn. `xhigh` is documented, so it passes through
844
+ // unclamped. Upstage's `max` sits above xhigh, but pi has no level above
845
+ // xhigh to reach it, so `max` is intentionally unreachable.
846
+ // * solar-pro3 and solar-pro2 — omitted means reasoning is OFF, `minimal`
847
+ // and `low` ALSO turn it off, and only `medium`/`high` turn it on. So
848
+ // `off` maps to null (nothing sent) and xhigh clamps to high.
849
+ //
850
+ // Note the trap in both rows: Upstage's `minimal` is an OFF value, but pi's
851
+ // `minimal` is an enabled level meaning "reason a little". Forwarding it
852
+ // verbatim would silently disable reasoning on a request that asked for it.
853
+ // Each group therefore floors pi's enabled levels at its own lowest
854
+ // reasoning-ON value — `low` for pro4/open2, `medium` for pro3/pro2 (which is
855
+ // why pi's `low` also maps to `medium` there). Every level pi can select thus
856
+ // emits a value Upstage treats as reasoning-on. solar-mini needs no map: it
857
+ // does not accept reasoning_effort at all (any value returns HTTP 400).
834
858
  upstage: {
835
859
  id: 'upstage',
836
860
  name: 'Upstage (Solar)',
@@ -839,6 +863,35 @@ export const KNOWN_PROVIDERS = {
839
863
  apiKeyEnv: 'UPSTAGE_API_KEY',
840
864
  oauthProviderId: null,
841
865
  models: {
866
+ 'solar-pro4': {
867
+ id: 'solar-pro4',
868
+ name: 'Solar Pro 4',
869
+ api: 'openai-completions',
870
+ provider: 'upstage',
871
+ baseUrl: 'https://api.upstage.ai/v1',
872
+ reasoning: true,
873
+ // pro4 and open2 reason by DEFAULT — see the reasoning-group note above.
874
+ thinkingLevelMap: {
875
+ off: 'none',
876
+ minimal: 'low',
877
+ low: 'low',
878
+ medium: 'medium',
879
+ high: 'high',
880
+ xhigh: 'xhigh',
881
+ },
882
+ input: ['text'],
883
+ cost: { input: 0.3, output: 1.2, cacheRead: 0.06, cacheWrite: 0 },
884
+ contextWindow: 524288,
885
+ maxTokens: 131072,
886
+ compat: {
887
+ supportsStore: false,
888
+ supportsDeveloperRole: false,
889
+ supportsReasoningEffort: true,
890
+ supportsUsageInStreaming: false,
891
+ supportsStrictMode: false,
892
+ maxTokensField: 'max_tokens',
893
+ },
894
+ },
842
895
  'solar-open2': {
843
896
  id: 'solar-open2',
844
897
  name: 'Solar Open 2',
@@ -846,7 +899,14 @@ export const KNOWN_PROVIDERS = {
846
899
  provider: 'upstage',
847
900
  baseUrl: 'https://api.upstage.ai/v1',
848
901
  reasoning: true,
849
- thinkingLevelMap: { off: null, minimal: 'minimal', low: 'low', medium: 'medium', high: 'high', xhigh: 'high' },
902
+ thinkingLevelMap: {
903
+ off: 'none',
904
+ minimal: 'low',
905
+ low: 'low',
906
+ medium: 'medium',
907
+ high: 'high',
908
+ xhigh: 'xhigh',
909
+ },
850
910
  input: ['text'],
851
911
  cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
852
912
  contextWindow: 128000,
@@ -867,7 +927,14 @@ export const KNOWN_PROVIDERS = {
867
927
  provider: 'upstage',
868
928
  baseUrl: 'https://api.upstage.ai/v1',
869
929
  reasoning: true,
870
- thinkingLevelMap: { off: null, minimal: 'minimal', low: 'low', medium: 'medium', high: 'high', xhigh: 'high' },
930
+ thinkingLevelMap: {
931
+ off: null,
932
+ minimal: 'medium',
933
+ low: 'medium',
934
+ medium: 'medium',
935
+ high: 'high',
936
+ xhigh: 'high',
937
+ },
871
938
  input: ['text'],
872
939
  cost: { input: 0.15, output: 0.6, cacheRead: 0.015, cacheWrite: 0 },
873
940
  contextWindow: 128000,
@@ -888,7 +955,14 @@ export const KNOWN_PROVIDERS = {
888
955
  provider: 'upstage',
889
956
  baseUrl: 'https://api.upstage.ai/v1',
890
957
  reasoning: true,
891
- thinkingLevelMap: { off: null, minimal: 'minimal', low: 'low', medium: 'medium', high: 'high', xhigh: 'high' },
958
+ thinkingLevelMap: {
959
+ off: null,
960
+ minimal: 'medium',
961
+ low: 'medium',
962
+ medium: 'medium',
963
+ high: 'high',
964
+ xhigh: 'high',
965
+ },
892
966
  input: ['text'],
893
967
  cost: { input: 0.15, output: 0.6, cacheRead: 0.015, cacheWrite: 0 },
894
968
  contextWindow: 65000,
@@ -1048,25 +1122,13 @@ export const KNOWN_PROVIDERS = {
1048
1122
  // `providerForModelRef()` matches on the registered provider prefix, the same
1049
1123
  // mechanism that already handles the slash-heavy Fireworks router ids.
1050
1124
  //
1051
- // Curation rule (why only six of the ~56 live chat models): every curated
1052
- // entry mirrors a model THIS registry already prices natively, so the numbers
1053
- // below are copied from a sibling entry rather than invented. Models with no
1054
- // in-repo price (google/*, qwen/*) and models whose native reasoning contract
1055
- // is NOT plain OpenAI `reasoning_effort` (deepseek's `thinking:{type}`,
1056
- // z-ai/qwen's `enable_thinking`) are deliberately left out: pi-ai picks
1057
- // thinkingFormat from the baseUrl, and `apis.opengateway.ai` resolves to
1058
- // "openai", so routing those here would silently send the wrong reasoning
1059
- // field. They all remain reachable as custom `opengateway/...` refs — curation
1060
- // only drives the picker, JSON-schema enum, and autocomplete.
1061
- //
1062
- // Costs are UPSTREAM LIST PRICES in USD per 1M tokens, pinned to the sibling
1063
- // entry by a guard test in providers.test.ts. OpenGateway bills upstream usage
1064
- // plus a platform fee and exposes no pricing in /v1/models, so `typeclaw usage`
1065
- // is an estimate that reads LOW against the real invoice. Do NOT "fix" these
1066
- // to zero: zero means "no attributable per-token charge" in this registry
1067
- // (flat-subscription products like the Fireworks Fire Pass router), and
1068
- // OpenGateway is metered. Replace with gateway-billed rates only if
1069
- // OpenGateway ever publishes them.
1125
+ // The live picker gets ids and modalities from OpenGateway's public catalog,
1126
+ // then best-effort joins its rate card. This registry intentionally keeps one
1127
+ // anchor only: `resolveModel()` uses `Object.keys(provider.models)[0]` as the
1128
+ // transport/limits template for every uncurated ref, so entries 2..N buy
1129
+ // nothing. An empty map would make custom refs throw, making one the true
1130
+ // floor. Nano is the cheapest conservative context/limit donor and keeps
1131
+ // `curatedOptions()` useful as the offline fallback.
1070
1132
  //
1071
1133
  // Every model carries an explicit `compat` for the same reason Upstage does:
1072
1134
  // pi-ai auto-detects compatibility from the baseUrl, and `apis.opengateway.ai`
@@ -1077,13 +1139,12 @@ export const KNOWN_PROVIDERS = {
1077
1139
  // defaults it true) — turning it off would break token and cost reporting.
1078
1140
  opengateway: {
1079
1141
  id: 'opengateway',
1080
- name: 'OpenGateway.ai',
1142
+ name: 'OpenGateway',
1081
1143
  baseUrl: 'https://apis.opengateway.ai/v1',
1082
1144
  auth: ['api-key'],
1083
1145
  apiKeyEnv: 'OPENGATEWAY_API_KEY',
1084
1146
  oauthProviderId: null,
1085
1147
  models: {
1086
- // Mirrors openai/gpt-5.4-nano.
1087
1148
  'openai/gpt-5.4-nano': {
1088
1149
  id: 'openai/gpt-5.4-nano',
1089
1150
  name: 'GPT-5.4 nano',
@@ -1099,113 +1160,15 @@ export const KNOWN_PROVIDERS = {
1099
1160
  maxTokensField: 'max_tokens',
1100
1161
  },
1101
1162
  input: ['text', 'image'],
1163
+ // Offline snapshot of OpenGateway's published rate for this model, NOT
1164
+ // the upstream OpenAI list price. Every other model is priced live from
1165
+ // the rate card, so this is the one number that can go stale; it is the
1166
+ // fallback used when the card is unreachable. Re-check it against
1167
+ // https://opengateway.ai/api/model-prices if you touch it.
1102
1168
  cost: { input: 0.2, output: 1.25, cacheRead: 0.02, cacheWrite: 0 },
1103
1169
  contextWindow: 400000,
1104
1170
  maxTokens: 128000,
1105
1171
  },
1106
- // Mirrors openai/gpt-5.5.
1107
- 'openai/gpt-5.5': {
1108
- id: 'openai/gpt-5.5',
1109
- name: 'GPT-5.5',
1110
- api: 'openai-completions',
1111
- provider: 'opengateway',
1112
- baseUrl: 'https://apis.opengateway.ai/v1',
1113
- reasoning: true,
1114
- compat: {
1115
- supportsStore: false,
1116
- supportsDeveloperRole: false,
1117
- supportsReasoningEffort: false,
1118
- supportsStrictMode: false,
1119
- maxTokensField: 'max_tokens',
1120
- },
1121
- input: ['text', 'image'],
1122
- cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 0 },
1123
- contextWindow: 1050000,
1124
- maxTokens: 128000,
1125
- },
1126
- // Mirrors anthropic/claude-sonnet-5. Note the transport differs from the
1127
- // native entry: Anthropic is `anthropic-messages` first-party, but the
1128
- // gateway only speaks OpenAI chat-completions.
1129
- 'anthropic/claude-sonnet-5': {
1130
- id: 'anthropic/claude-sonnet-5',
1131
- name: 'Claude Sonnet 5',
1132
- api: 'openai-completions',
1133
- provider: 'opengateway',
1134
- baseUrl: 'https://apis.opengateway.ai/v1',
1135
- reasoning: true,
1136
- compat: {
1137
- supportsStore: false,
1138
- supportsDeveloperRole: false,
1139
- supportsReasoningEffort: false,
1140
- supportsStrictMode: false,
1141
- maxTokensField: 'max_tokens',
1142
- },
1143
- input: ['text', 'image'],
1144
- cost: { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 },
1145
- contextWindow: 1000000,
1146
- maxTokens: 128000,
1147
- },
1148
- // Mirrors anthropic/claude-opus-4-8.
1149
- 'anthropic/claude-opus-4-8': {
1150
- id: 'anthropic/claude-opus-4-8',
1151
- name: 'Claude Opus 4.8',
1152
- api: 'openai-completions',
1153
- provider: 'opengateway',
1154
- baseUrl: 'https://apis.opengateway.ai/v1',
1155
- reasoning: true,
1156
- compat: {
1157
- supportsStore: false,
1158
- supportsDeveloperRole: false,
1159
- supportsReasoningEffort: false,
1160
- supportsStrictMode: false,
1161
- maxTokensField: 'max_tokens',
1162
- },
1163
- input: ['text', 'image'],
1164
- cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
1165
- contextWindow: 1000000,
1166
- maxTokens: 128000,
1167
- },
1168
- // Mirrors moonshot/kimi-k2.6. The gateway namespaces Moonshot as
1169
- // `moonshotai`, so the id is NOT the native `kimi-k2.6`.
1170
- 'moonshotai/kimi-k2.6': {
1171
- id: 'moonshotai/kimi-k2.6',
1172
- name: 'Kimi K2.6',
1173
- api: 'openai-completions',
1174
- provider: 'opengateway',
1175
- baseUrl: 'https://apis.opengateway.ai/v1',
1176
- reasoning: true,
1177
- compat: {
1178
- supportsStore: false,
1179
- supportsDeveloperRole: false,
1180
- supportsReasoningEffort: false,
1181
- supportsStrictMode: false,
1182
- maxTokensField: 'max_tokens',
1183
- },
1184
- input: ['text', 'image'],
1185
- cost: { input: 0.6, output: 2.5, cacheRead: 0.15, cacheWrite: 0 },
1186
- contextWindow: 256000,
1187
- maxTokens: 64000,
1188
- },
1189
- // Mirrors minimax/MiniMax-M3.
1190
- 'minimax/MiniMax-M3': {
1191
- id: 'minimax/MiniMax-M3',
1192
- name: 'MiniMax M3',
1193
- api: 'openai-completions',
1194
- provider: 'opengateway',
1195
- baseUrl: 'https://apis.opengateway.ai/v1',
1196
- reasoning: true,
1197
- compat: {
1198
- supportsStore: false,
1199
- supportsDeveloperRole: false,
1200
- supportsReasoningEffort: false,
1201
- supportsStrictMode: false,
1202
- maxTokensField: 'max_tokens',
1203
- },
1204
- input: ['text', 'image'],
1205
- cost: { input: 0.3, output: 1.2, cacheRead: 0.06, cacheWrite: 0 },
1206
- contextWindow: 1000000,
1207
- maxTokens: 524288,
1208
- },
1209
1172
  },
1210
1173
  },
1211
1174
  } as const satisfies Record<string, KnownProvider>
@@ -1299,7 +1262,7 @@ export const KNOWN_PROVIDER_VENDORS = {
1299
1262
  // Anthropic should land on the Anthropic row, not the gateway that proxies it.
1300
1263
  opengateway: {
1301
1264
  id: 'opengateway',
1302
- name: 'OpenGateway.ai (multi-vendor gateway)',
1265
+ name: 'OpenGateway',
1303
1266
  providers: ['opengateway'],
1304
1267
  },
1305
1268
  } as const satisfies Record<string, KnownProviderVendor>