Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

fix: stop advertising an output cap the Codex wire cannot enforce

Alfredo Lopeze65c7e19a07f
diff --git a/charts/clerum-crds/crds/workflowrecipe.yaml b/charts/clerum-crds/crds/workflowrecipe.yaml
index d5f1362b7..741b54c2f 100644
--- a/charts/clerum-crds/crds/workflowrecipe.yaml
+++ b/charts/clerum-crds/crds/workflowrecipe.yaml
@@ -477,6 +477,14 @@ spec:
maxOutputTokens:
type: integer
minimum: 1
+ description: >-
+ DEPRECATED — accepted for
+ compatibility but no longer enforced.
+ The Codex ChatGPT wire discards
+ max_output_tokens, so no layer capped
+ the response by token count.
+ Scheduled for removal in a future
+ version.
clientNotifications:
type: object
required: [allowedEventTypes]
@@ -1711,6 +1719,12 @@ spec:
maxOutputTokens:
type: integer
minimum: 1
+ description: >-
+ DEPRECATED — accepted for compatibility but no
+ longer enforced. The Codex ChatGPT wire discards
+ max_output_tokens, so no layer capped the response
+ by token count. Scheduled for removal in a future
+ version.
maxRequestsPerRun:
type: integer
minimum: 1
diff --git a/control-api/package-lock.json b/control-api/package-lock.json
index d8dff2505..79765c78a 100644
--- a/control-api/package-lock.json
+++ b/control-api/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-control-api",
- "version": "0.1.536",
+ "version": "0.1.537",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-control-api",
- "version": "0.1.536",
+ "version": "0.1.537",
"dependencies": {
"@clerum/display-field": "file:../packages/display-field",
"@clerum/image-policy": "file:../packages/image-policy",
diff --git a/control-api/package.json b/control-api/package.json
index 6cb79190f..53e51c3af 100644
--- a/control-api/package.json
+++ b/control-api/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-control-api",
- "version": "0.1.536",
+ "version": "0.1.537",
"description": "Internal Clerum control plane REST API + UI for managing CRDs and secrets",
"main": "dist/main.js",
"engines": {
diff --git a/control-api/src/routes/admin/pluginWorkloadSdk.ts b/control-api/src/routes/admin/pluginWorkloadSdk.ts
index de166e0b4..833f20cf3 100644
--- a/control-api/src/routes/admin/pluginWorkloadSdk.ts
+++ b/control-api/src/routes/admin/pluginWorkloadSdk.ts
@@ -147,15 +147,22 @@ function parseModelArray(value: unknown, field: string, res: Response): string[]
// Issue #348: per-run quota keys are deprecated. They remain accepted on the
// wire for compatibility and are validated with the same shape rules, but they
// are stripped before persistence — only per-minute/platform limits are stored.
+//
+// J8: `maxOutputTokens` joins them. It was advertised as a policy ceiling but
+// bound nothing: the Codex ChatGPT wire rejects `max_output_tokens`, so no
+// layer ever capped the response by token count. The only real bound is the
+// 300s wall clock, and the budget reservation is admission control that denies
+// the NEXT call, after the overspend already happened. Persisting the key kept
+// a promise the platform could not keep.
const DEPRECATED_QUOTA_LIMIT_KEYS: ReadonlyArray<keyof PluginWorkloadSdkQuotaLimits> = [
'maxRequestsPerRun',
'maxNotificationsPerRun',
+ 'maxOutputTokens',
]
const ACTIVE_QUOTA_LIMIT_KEYS: ReadonlyArray<keyof PluginWorkloadSdkQuotaLimits> = [
'maxInvocationsPerMinute',
'maxNotificationsPerMinute',
- 'maxOutputTokens',
]
function parseQuotaLimits(value: unknown, res: Response): PluginWorkloadSdkQuotaLimits | null {
diff --git a/control-api/src/routes/mcp-host/plugin-workload-sdk.routes.ts b/control-api/src/routes/mcp-host/plugin-workload-sdk.routes.ts
index 1f75d5bd2..d2cabdf47 100644
--- a/control-api/src/routes/mcp-host/plugin-workload-sdk.routes.ts
+++ b/control-api/src/routes/mcp-host/plugin-workload-sdk.routes.ts
@@ -611,7 +611,6 @@ async function handlePromptBridgeAuthorization(
attemptGeneration: result.value.attemptGeneration,
policyRevision: result.value.policyRevision,
policyHash: result.value.policyHash,
- maxOutputTokens: result.value.maxOutputTokens,
})
}
diff --git a/control-api/src/services/pluginWorkloadSdkAuthorizer.ts b/control-api/src/services/pluginWorkloadSdkAuthorizer.ts
index ff5151c7e..be07334b5 100644
--- a/control-api/src/services/pluginWorkloadSdkAuthorizer.ts
+++ b/control-api/src/services/pluginWorkloadSdkAuthorizer.ts
@@ -128,7 +128,6 @@ export interface AuthorizedPromptBridge {
attemptGeneration: number
policyRevision: number
policyHash: string
- maxOutputTokens: number | null
}
type PromptTargetResolution =
@@ -398,7 +397,6 @@ async function authorizePromptBridgeInner(
attemptGeneration: recorded.invocation.attemptGeneration,
policyRevision: grant.policyRevision,
policyHash: resolvedPolicyHash,
- maxOutputTokens: grant.quotaLimits.maxOutputTokens ?? null,
},
}
}
@@ -496,7 +494,6 @@ async function authorizePromptBridgeInner(
attemptGeneration: recorded.invocation.attemptGeneration,
policyRevision: grant.policyRevision,
policyHash: resolvedPolicyHash,
- maxOutputTokens: grant.quotaLimits.maxOutputTokens ?? null,
},
}
}
@@ -599,6 +596,20 @@ export async function reissuePromptBridgeCredentialTicket(
return deny('target_not_allowed', 'target is not in the current promptBridge policy')
}
+ const reservationOnly =
+ isLlmProviderId(target.provider) && PROVIDER_AUTH_MODE[target.provider] === 'oauth-broker'
+ // RP-539-009: an oauth-broker target is redeemable only through the broker
+ // binding the WRC installs for the grant's default target. Any other broker
+ // target would be reserved here and then die in execution with no credential
+ // to redeem, after the reservation already charged the attempt budget. Deny
+ // before reserving, and fail closed when the grant has no default target.
+ if (reservationOnly && target.targetRef !== grant.defaultTargetRef) {
+ return deny(
+ 'target_not_allowed',
+ 'oauth-broker fallback is authorized only for the default target of this grant'
+ )
+ }
+
const providerAttempt = await reservePluginWorkloadSdkProviderAttempt({
invocationId: invocation.id,
recipeNamespace: params.claims.recipeNamespace,
@@ -613,8 +624,6 @@ export async function reissuePromptBridgeCredentialTicket(
)
}
- const reservationOnly =
- isLlmProviderId(target.provider) && PROVIDER_AUTH_MODE[target.provider] === 'oauth-broker'
if (reservationOnly) {
return {
ok: true,
diff --git a/control-api/src/services/pluginWorkloadSdkDb.ts b/control-api/src/services/pluginWorkloadSdkDb.ts
index 0fcf3fa38..bf8e6d311 100644
--- a/control-api/src/services/pluginWorkloadSdkDb.ts
+++ b/control-api/src/services/pluginWorkloadSdkDb.ts
@@ -167,10 +167,13 @@ export function getPromptBridgeAttemptLeaseSeconds(): number {
const IDEMPOTENCY_TTL_HOURS = 24
export interface PluginWorkloadSdkQuotaLimits {
+ /** Deprecated (issue #348) — accepted and validated on the wire, never persisted. */
maxRequestsPerRun?: number
+ /** Deprecated (issue #348) — accepted and validated on the wire, never persisted. */
maxNotificationsPerRun?: number
maxInvocationsPerMinute?: number
maxNotificationsPerMinute?: number
+ /** Deprecated (J8) — accepted and validated on the wire, never persisted. */
maxOutputTokens?: number
}
diff --git a/control-api/src/services/pluginWorkloadSdkFinalization.ts b/control-api/src/services/pluginWorkloadSdkFinalization.ts
index 0423a021e..fbc5d281d 100644
--- a/control-api/src/services/pluginWorkloadSdkFinalization.ts
+++ b/control-api/src/services/pluginWorkloadSdkFinalization.ts
@@ -421,6 +421,12 @@ export async function finalizePromptBridgeInTransaction(
if (!providerAttempt) {
throw new PromptBridgeFinalizationError('not_found', 'provider attempt not found', 404)
}
+ // N-19: shape before binding, matching `replayExistingOutcome`. A non-oauth
+ // finalization that omits `target.credentialSlot` is a malformed request, and
+ // both paths must classify it as 400 `invalid_request`. Running the binding
+ // comparison first would report the same body as 403 `binding_mismatch` here
+ // and 400 on replay, so the same defect got two different verdicts.
+ assertPersistedAuthModeCredentialRules(String(providerAttempt.provider), input)
if (
String(providerAttempt.invocation_id) !== input.invocationId ||
String(providerAttempt.recipe_namespace) !== input.recipeNamespace ||
@@ -438,7 +444,6 @@ export async function finalizePromptBridgeInTransaction(
403
)
}
- assertPersistedAuthModeCredentialRules(String(providerAttempt.provider), input)
const currentInvocationStatus = String(invocation.status)
const currentReceiptStatus = String(receipt.status)
@@ -608,8 +613,12 @@ export async function finalizePromptBridgeInTransaction(
// RP-539-003: a successful failover finalizes only its winner, so the
// attempts it displaced would never reach this endpoint. They are terminal
// and immutable by the reservation fence, so they can be discovered and
- // settled here without trusting the host to declare them. Index 1 has no
- // predecessors, so the query is skipped rather than run over an empty range.
+ // settled here without trusting the host to declare them.
+ //
+ // The index-1 skip is NOT a defensive optimization: `attempt_index < 1`
+ // matches nothing by construction, because attempt_index starts at 1 and
+ // UNIQUE (invocation_id, attempt_generation, attempt_index) forbids gaps.
+ // The query is skipped because it is provably empty, not because it might be.
if (input.providerAttemptIndex > 1) {
await settlePriorProviderAttemptFloors(db, {
invocationId: input.invocationId,
diff --git a/control-api/test/routes.adminPluginWorkloadSdk.test.ts b/control-api/test/routes.adminPluginWorkloadSdk.test.ts
index f7188af8a..0a182f459 100644
--- a/control-api/test/routes.adminPluginWorkloadSdk.test.ts
+++ b/control-api/test/routes.adminPluginWorkloadSdk.test.ts
@@ -531,10 +531,14 @@ describe('routes/admin/pluginWorkloadSdk — grants', () => {
)
})
- it('accepts deprecated per-run keys, strips them, keeps active keys', async () => {
+ it('accepts deprecated quota keys, strips them, keeps active keys', async () => {
// Issue #348 (plan 2.6/2.7): deprecated per-run keys stay shape-validated
// (malformed still 400s, pinned above) but are dropped on write, while the
- // active per-minute/token keys persist untouched.
+ // active per-minute keys persist untouched.
+ //
+ // J8: maxOutputTokens joins the stripped set. It was persisted and
+ // advertised as a policy ceiling, but the Codex ChatGPT wire discards
+ // max_output_tokens, so nothing ever capped the response by token count.
vi.mocked(sdkDb.upsertGrant).mockResolvedValue({ id: 'g1' } as never)
const res = await request(buildApp())
.post('/admin/plugin-workload-sdk/grants')
@@ -543,6 +547,7 @@ describe('routes/admin/pluginWorkloadSdk — grants', () => {
quotaLimits: {
maxRequestsPerRun: 5,
maxNotificationsPerRun: 7,
+ maxOutputTokens: 4096,
maxInvocationsPerMinute: 30,
},
})
@@ -556,6 +561,15 @@ describe('routes/admin/pluginWorkloadSdk — grants', () => {
)
})
+ it('still shape-validates the deprecated maxOutputTokens on the wire (J8)', async () => {
+ // Deprecating the key must not loosen the boundary: a malformed value is
+ // still a 400, exactly as it was while the key was persisted.
+ const res = await request(buildApp())
+ .post('/admin/plugin-workload-sdk/grants')
+ .send({ ...validGrantBody, quotaLimits: { maxOutputTokens: 0 } })
+ expect(res.status).toBe(400)
+ })
+
it('rejects promptBridge targets outside the provider allowlist', async () => {
// Allowlist for provider `zai` enables only glm-4.7; the request asks for a
// model that is not enabled → 400 model_not_allowed listing the offenders.
diff --git a/control-api/test/routes.mcp-host.plugin-workload-sdk.test.ts b/control-api/test/routes.mcp-host.plugin-workload-sdk.test.ts
index 16a32f055..9f376b1b9 100644
--- a/control-api/test/routes.mcp-host.plugin-workload-sdk.test.ts
+++ b/control-api/test/routes.mcp-host.plugin-workload-sdk.test.ts
@@ -162,7 +162,6 @@ beforeEach(() => {
attemptGeneration: 1,
policyRevision: 1,
policyHash: 'policy-hash',
- maxOutputTokens: null,
},
})
vi.mocked(authorizer.reissuePromptBridgeCredentialTicket).mockResolvedValue({
@@ -534,7 +533,6 @@ describe('POST /mcp-host/plugin-workload-sdk/prompt-bridge/v2', () => {
],
policyRevision: 1,
policyHash: 'policy-hash',
- maxOutputTokens: null,
},
})
const res = await request(buildApp())
diff --git a/control-api/test/services.pluginWorkloadSdkAuthorizer.test.ts b/control-api/test/services.pluginWorkloadSdkAuthorizer.test.ts
index a9ad1a214..0b314e973 100644
--- a/control-api/test/services.pluginWorkloadSdkAuthorizer.test.ts
+++ b/control-api/test/services.pluginWorkloadSdkAuthorizer.test.ts
@@ -599,8 +599,12 @@ describe('authorizePromptBridge', () => {
})
expect(result).toMatchObject({
ok: true,
- value: { invocationId: 'inv-1', replay: false, maxOutputTokens: 2048 },
+ value: { invocationId: 'inv-1', replay: false },
})
+ if (!result.ok) throw new Error('expected the authorization to succeed')
+ // J8: a legacy grant may still carry maxOutputTokens, but the authorization
+ // must not re-advertise it as a policy ceiling — nothing enforced it.
+ expect(result.value).not.toHaveProperty('maxOutputTokens')
expect(sdkDb.insertInvocation).toHaveBeenCalledOnce()
// Issue #348: the per-run leg is inert — no period quota is consumed even
// though the grant still carries a legacy maxRequestsPerRun value.
@@ -883,6 +887,66 @@ describe('reissuePromptBridgeCredentialTicket', () => {
expect(sdkDb.markPluginWorkloadSdkProviderAttemptStatus).not.toHaveBeenCalled()
})
+ it('denies an oauth-broker fallback that is not the default target (RP-539-009)', async () => {
+ // The WRC installs the broker binding for the grant's default target only.
+ // A non-default Codex fallback is authorizable but not executable: without
+ // this guard it reserved an attempt, burned budget, and died in execution.
+ const policy = grant({
+ promptTargets: [
+ {
+ targetRef: 'primary-zai',
+ provider: 'zai',
+ model: 'glm-4.7',
+ credentialSlot: 'zai-api-key',
+ },
+ {
+ targetRef: 'codex-fallback',
+ provider: 'codex-subscription',
+ model: 'gpt-5.1',
+ credentialSlot: '',
+ },
+ ],
+ defaultTargetRef: 'primary-zai',
+ })
+ const policyHash = sdkDb.hashPromptTargetPolicy(policy)
+ vi.mocked(sdkDb.findGrant).mockResolvedValue(policy)
+ vi.mocked(sdkDb.getInvocationById).mockResolvedValue(
+ invocation({
+ promptAuthorization: {
+ policyRevision: 1,
+ policyHash,
+ authorizedTargetRefs: ['primary-zai', 'codex-fallback'],
+ },
+ })
+ )
+ vi.mocked(sdkDb.getPluginWorkloadSdkAttemptReceipt).mockResolvedValue({
+ invocationId: 'inv-1',
+ recipeNamespace: NS,
+ recipeName: RECIPE,
+ attemptGeneration: 1,
+ method: 'promptBridge',
+ targetRefs: ['primary-zai', 'codex-fallback'],
+ policyRevision: 1,
+ policyHash,
+ status: 'in_progress',
+ startedAt: '2026-06-09T00:00:00.000Z',
+ leaseExpiresAt: '2026-06-09T00:01:00.000Z',
+ completedAt: null,
+ })
+
+ const result = await reissuePromptBridgeCredentialTicket({
+ claims: claims(),
+ invocationId: 'inv-1',
+ targetRef: 'codex-fallback',
+ attemptGeneration: 1,
+ })
+
+ expect(result).toMatchObject({ ok: false, error: 'target_not_allowed' })
+ // The denial precedes the reservation, so no attempt index is consumed.
+ expect(sdkDb.reservePluginWorkloadSdkProviderAttempt).not.toHaveBeenCalled()
+ expect(sdkDb.registerPluginWorkloadSdkCredentialTicketJti).not.toHaveBeenCalled()
+ })
+
it('fails closed when policy changes or the invocation is terminal', async () => {
const policy = twoTargetGrant()
authorizedSuffix.policyHash = sdkDb.hashPromptTargetPolicy(policy)
diff --git a/control-api/test/services.pluginWorkloadSdkFinalization.test.ts b/control-api/test/services.pluginWorkloadSdkFinalization.test.ts
index bec9015d9..b506e9b99 100644
--- a/control-api/test/services.pluginWorkloadSdkFinalization.test.ts
+++ b/control-api/test/services.pluginWorkloadSdkFinalization.test.ts
@@ -1172,8 +1172,41 @@ describe('Plugin Workload SDK promptBridge finalization', () => {
) as never
)
).rejects.toMatchObject({
- code: 'binding_mismatch',
- httpStatus: 403,
+ // N-19: the persisted `openai` attempt still governs, so the missing
+ // credentialSlot is a malformed body (400), not a binding mismatch (403).
+ // `replayExistingOutcome` already classified this exact body as 400; the
+ // fresh path used to answer 403 only because the binding comparison ran
+ // before the persisted-auth-mode assert.
+ code: 'invalid_request',
+ httpStatus: 400,
+ })
+ })
+
+ it('classifies the same malformed body identically on the replay path (N-19)', async () => {
+ await expect(
+ finalizePromptBridgeInTransaction(
+ {
+ ...inputBase,
+ target: {
+ ...inputBase.target,
+ provider: 'codex-subscription',
+ credentialSlot: '',
+ },
+ status: 'complete',
+ usage: {
+ llmSecretName: '',
+ callerRef: 'scanner',
+ fallbackUsed: false,
+ attemptCount: 1,
+ inputTokens: 1,
+ outputTokens: 1,
+ },
+ },
+ dbWithRows([], spendRow({ outcome: 'exact', input_tokens: 1, output_tokens: 1 })) as never
+ )
+ ).rejects.toMatchObject({
+ code: 'invalid_request',
+ httpStatus: 400,
})
})
diff --git a/control-ui/app/plugin-workload-sdk/page.tsx b/control-ui/app/plugin-workload-sdk/page.tsx
index 114e0b684..2b05120c1 100644
--- a/control-ui/app/plugin-workload-sdk/page.tsx
+++ b/control-ui/app/plugin-workload-sdk/page.tsx
@@ -678,13 +678,14 @@ function GrantFormModal({
setError('')
// Edit mode: the upsert is a full-column overwrite, and the form does not expose
// the quota fields. Round-trip every quota key from the original grant except the
- // deprecated per-run keys (no longer enforced, issue #348) so editing does not
- // silently wipe values configured outside this form.
+ // deprecated ones (per-run: issue #348; maxOutputTokens: J8 — no longer enforced)
+ // so editing does not silently wipe values configured outside this form.
const quotaLimits: PluginWorkloadSdkGrantInput['quotaLimits'] = {}
if (grant) {
const {
maxRequestsPerRun: _d1,
maxNotificationsPerRun: _d2,
+ maxOutputTokens: _d3,
...keptQuotaLimits
} = grant.quotaLimits
Object.assign(quotaLimits, keptQuotaLimits)
diff --git a/control-ui/lib/api.ts b/control-ui/lib/api.ts
index 522130379..7b26a6c74 100644
--- a/control-ui/lib/api.ts
+++ b/control-ui/lib/api.ts
@@ -3658,6 +3658,11 @@ export type PluginWorkloadSdkQuotaLimits = {
maxNotificationsPerRun?: number
maxInvocationsPerMinute?: number
maxNotificationsPerMinute?: number
+ /**
+ * @deprecated No longer enforced (J8); the Codex ChatGPT wire discards
+ * max_output_tokens, so no layer capped the response by token count. Still
+ * returned on legacy grants; never sent by this UI.
+ */
maxOutputTokens?: number
}
@@ -3717,10 +3722,13 @@ export type PluginWorkloadSdkGrantInput = {
allowedTargetRefs?: string[]
allowedUserRefs?: string[]
allowedCallers?: string[]
- // Input type omits the deprecated per-run keys (issue #348): this UI never
- // sends them and the server strips them on write. The response
- // PluginWorkloadSdkQuotaLimits still carries them for legacy grants.
- quotaLimits?: Omit<PluginWorkloadSdkQuotaLimits, 'maxRequestsPerRun' | 'maxNotificationsPerRun'>
+ // Input type omits the deprecated keys (per-run: issue #348; maxOutputTokens:
+ // J8): this UI never sends them and the server strips them on write. The
+ // response PluginWorkloadSdkQuotaLimits still carries them for legacy grants.
+ quotaLimits?: Omit<
+ PluginWorkloadSdkQuotaLimits,
+ 'maxRequestsPerRun' | 'maxNotificationsPerRun' | 'maxOutputTokens'
+ >
modelPolicies?: Record<string, PluginWorkloadSdkModelPolicy>
promptTargets?: PluginWorkloadSdkPromptTarget[]
defaultTargetRef?: string
diff --git a/control-ui/package-lock.json b/control-ui/package-lock.json
index afb8dde75..bc520064e 100644
--- a/control-ui/package-lock.json
+++ b/control-ui/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-control-ui",
- "version": "0.1.486",
+ "version": "0.1.487",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-control-ui",
- "version": "0.1.486",
+ "version": "0.1.487",
"dependencies": {
"@clerum/display-field": "file:../packages/display-field",
"@clerum/llm-providers": "file:../packages/llm-providers",
diff --git a/control-ui/package.json b/control-ui/package.json
index 9434e6613..6b85b2738 100644
--- a/control-ui/package.json
+++ b/control-ui/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-control-ui",
- "version": "0.1.486",
+ "version": "0.1.487",
"private": true,
"scripts": {
"dev": "NEXT_IGNORE_INCORRECT_LOCKFILE=1 next dev",
diff --git a/docs/crds/workflowrecipe.md b/docs/crds/workflowrecipe.md
index 753bc1e64..1698a18a8 100644
--- a/docs/crds/workflowrecipe.md
+++ b/docs/crds/workflowrecipe.md
@@ -1374,7 +1374,7 @@ Declaring this block forces an always-on `mcp-host`. Runtime enforcement is addi
| Field | Type | Required | Description |
| -------------------------------------------- | -------- | --------- | --------------------------------------------------------------------------- |
| `promptBridge.allowedModels` | string[] | No | Max 32. |
-| `promptBridge.maxOutputTokens` | integer | No | Min 1. |
+| `promptBridge.maxOutputTokens` | integer | No | Min 1. **Deprecated:** accepted but ignored; the Codex ChatGPT wire discards `max_output_tokens`, so nothing capped the response by token count. Removal planned. |
| `promptBridge.maxRequestsPerRun` | integer | No | Min 1. **Deprecated:** accepted but ignored; platform per-minute rate limits apply. Removal planned. |
| `promptBridge.maxConcurrentInvocations` | integer | No | Min 1. Runtime default 5. |
| `promptBridge.maxInvocationsPerMinute` | integer | No | Min 1. Platform default 120 when unset (ENV `CONTROL_API_PLUGIN_SDK_PROMPTBRIDGE_PER_MIN`, issue #348). |
diff --git a/docs/testing/optional-playwright-qa-recorder.md b/docs/testing/optional-playwright-qa-recorder.md
index 6cbfbe647..8af773bf0 100644
--- a/docs/testing/optional-playwright-qa-recorder.md
+++ b/docs/testing/optional-playwright-qa-recorder.md
@@ -99,6 +99,7 @@ prompt/notification journey:
```bash
E2E_PLUGIN_SDK_WRITE_CONFIRM=1 \
+E2E_PLUGIN_SDK_EXPECT_PROVIDER=codex-subscription \
CONTROL_UI_BASE_URL=http://127.0.0.1:<random-control-ui-port> \
CONTROL_API_BASE_URL=http://127.0.0.1:<random-control-api-port> \
EXTERNAL_REST_API_BASE_URL=http://127.0.0.1:<random-external-rest-port> \
@@ -107,6 +108,15 @@ KUBECONTEXT=<branch-profile-context> \
make test-e2e-plugin-workload-sdk-desktop
```
+`E2E_PLUGIN_SDK_EXPECT_PROVIDER` is **mandatory and has no default**. It
+declares which provider the run is meant to exercise, and the precondition
+asserts the live recipe declares exactly that one. Without it the spec used to
+accept `openai`, `claude` or `codex-subscription` and skip its Codex assertions
+when the recipe was not Codex — so a run against the OpenAI recipe that
+`seed-e2e-data.sh` provisions by default went green while proving nothing about
+Codex. Declare the provider you intend to evidence; a mismatch now fails the
+precondition instead of quietly narrowing the run.
+
The target sets the opt-in flag only for this spec, requires non-default
branch-owned URLs, and compares them with the exact `ports.env` generated for
the selected branch profile. It runs the actual Electron/WebContentsView
@@ -114,6 +124,32 @@ journey and fails before building if the profile map is missing or belongs to a
different context. It does not start Minikube or create grants; those
prerequisites must already be provided by the T2/T3 profile lane.
+### No-grant guard lane
+
+`plugin-workload-sdk-no-grant-guard.spec.ts` is the regression test for the
+symptom of issue #533: a Codex recipe whose execution binding is missing must
+report `awaiting_policy`, never `validated`. It lives in its own spec and its
+own Make target because it is mutually exclusive with the happy path — that one
+requires the recipe to reach `validated`, this one requires it not to.
+
+```bash
+E2E_PLUGIN_SDK_WRITE_CONFIRM=1 \
+E2E_PLUGIN_SDK_NO_GRANT_RECIPE_NAME=<codex-recipe-without-binding> \
+E2E_PLUGIN_SDK_NO_GRANT_APP_TITLE=<its-app-title> \
+CONTROL_UI_BASE_URL=http://127.0.0.1:<random-control-ui-port> \
+CONTROL_API_BASE_URL=http://127.0.0.1:<random-control-api-port> \
+EXTERNAL_REST_API_BASE_URL=http://127.0.0.1:<random-external-rest-port> \
+RPC_PROXY_BASE_URL=http://127.0.0.1:<random-rpc-port> \
+KUBECONTEXT=<branch-profile-context> \
+make test-e2e-plugin-workload-sdk-desktop-no-grant
+```
+
+Both `E2E_PLUGIN_SDK_NO_GRANT_*` variables are **mandatory**: the spec has no
+default recipe and fails loudly when either is missing, and it refuses to run
+against the happy path's recipe even if pointed at it. **Provisioning a Codex
+recipe with no execution binding is operator work** — the lane will not create
+one, and a lane that cannot find its fixture fails rather than skipping.
+
## Repository isolation
The recorder adds new commands; it does not alter any existing command:
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.json
index b7eb7f138..792528ec0 100644
--- a/mcp-host/package-lock.json
+++ b/mcp-host/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.311",
+ "version": "0.1.312",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-mcp-host",
- "version": "0.1.311",
+ "version": "0.1.312",
"license": "SEE LICENSE IN LICENSE",
"dependencies": {
"@anthropic-ai/sdk": "^0.32.0",
diff --git a/mcp-host/package.json b/mcp-host/package.json
index c9c428555..857b26e11 100644
--- a/mcp-host/package.json
+++ b/mcp-host/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.311",
+ "version": "0.1.312",
"description": "MCP Host - reads Host CRD and provides LLM access via OpenAI/Claude",
"main": "dist/main.js",
"engines": {
diff --git a/mcp-host/src/pluginWorkloadSdk/metrics.test.ts b/mcp-host/src/pluginWorkloadSdk/metrics.test.ts
index c12e1c0ae..559db16df 100644
--- a/mcp-host/src/pluginWorkloadSdk/metrics.test.ts
+++ b/mcp-host/src/pluginWorkloadSdk/metrics.test.ts
@@ -39,7 +39,6 @@ function makeHandler(recipeName: string, complete?: ReturnType<typeof vi.fn>) {
attemptGeneration: 1,
policyRevision: 1,
policyHash: 'a'.repeat(64),
- maxOutputTokens: null,
}),
reportProviderAttemptStatus: vi.fn().mockResolvedValue(undefined),
reportInvocationStatus: vi.fn().mockResolvedValue(undefined),
diff --git a/mcp-host/src/pluginWorkloadSdk/promptBridge/controlApiClient.test.ts b/mcp-host/src/pluginWorkloadSdk/promptBridge/controlApiClient.test.ts
index 9ec3361f7..0ef46fa6f 100644
--- a/mcp-host/src/pluginWorkloadSdk/promptBridge/controlApiClient.test.ts
+++ b/mcp-host/src/pluginWorkloadSdk/promptBridge/controlApiClient.test.ts
@@ -34,7 +34,6 @@ function authorizedBody(invocationId: string) {
attemptGeneration: 1,
policyRevision: 2,
policyHash: 'a'.repeat(64),
- maxOutputTokens: null,
}
}
diff --git a/mcp-host/src/pluginWorkloadSdk/promptBridge/controlApiClient.ts b/mcp-host/src/pluginWorkloadSdk/promptBridge/controlApiClient.ts
index efeff34e2..86d3147d3 100644
--- a/mcp-host/src/pluginWorkloadSdk/promptBridge/controlApiClient.ts
+++ b/mcp-host/src/pluginWorkloadSdk/promptBridge/controlApiClient.ts
@@ -51,7 +51,6 @@ export interface AuthorizePromptBridgeResponse {
attemptGeneration: number
policyRevision: number
policyHash: string
- maxOutputTokens: number | null
}
export interface PluginWorkloadSdkCapabilities {
@@ -186,8 +185,7 @@ function isAuthorizePromptBridgeResponse(v: unknown): v is AuthorizePromptBridge
Number.isInteger(r.policyRevision) &&
(r.policyRevision as number) >= 1 &&
typeof r.policyHash === 'string' &&
- /^[a-f0-9]{64}$/.test(r.policyHash) &&
- (r.maxOutputTokens === null || typeof r.maxOutputTokens === 'number')
+ /^[a-f0-9]{64}$/.test(r.policyHash)
)
}
diff --git a/mcp-host/src/pluginWorkloadSdk/promptBridge/handler.test.ts b/mcp-host/src/pluginWorkloadSdk/promptBridge/handler.test.ts
index f965bd516..0c50bc889 100644
--- a/mcp-host/src/pluginWorkloadSdk/promptBridge/handler.test.ts
+++ b/mcp-host/src/pluginWorkloadSdk/promptBridge/handler.test.ts
@@ -45,7 +45,6 @@ function authorization(overrides: Record<string, unknown> = {}) {
attemptGeneration: 1,
policyRevision: 7,
policyHash: 'a'.repeat(64),
- maxOutputTokens: 2048,
...overrides,
}
}
@@ -183,7 +182,6 @@ describe('PromptBridgeHandler', () => {
expect(complete).toHaveBeenCalledWith(
expect.objectContaining({
invocationId: 'inv-1',
- maxTokens: 2048,
targets: [{ target: primary }, { target: fallback }],
})
)
@@ -208,6 +206,18 @@ describe('PromptBridgeHandler', () => {
)
})
+ it('forwards the workload maxTokens verbatim and applies no grant cap (J8)', async () => {
+ const { handler, complete } = makeDeps()
+ await handler.handle({ ...validBody, maxTokens: 9000 }, 'api')
+ expect(complete).toHaveBeenCalledWith(expect.objectContaining({ maxTokens: 9000 }))
+ })
+
+ it('sends no maxTokens when the workload asks for none (J8)', async () => {
+ const { handler, complete } = makeDeps()
+ await handler.handle(validBody, 'api')
+ expect(complete.mock.calls[0]?.[0]).toMatchObject({ maxTokens: undefined })
+ })
+
it('preserves metering when the logical terminal acknowledgement is unavailable', async () => {
const report = vi.fn().mockRejectedValue(new Error('control-api unavailable'))
const onUsage = vi.fn()
diff --git a/mcp-host/src/pluginWorkloadSdk/promptBridge/handler.ts b/mcp-host/src/pluginWorkloadSdk/promptBridge/handler.ts
index dc333f53f..7e8108dc6 100644
--- a/mcp-host/src/pluginWorkloadSdk/promptBridge/handler.ts
+++ b/mcp-host/src/pluginWorkloadSdk/promptBridge/handler.ts
@@ -429,7 +429,12 @@ export class PromptBridgeHandler {
acknowledgementMode: this.deps.finalizePromptBridge
? 'atomic_terminal_finalization'
: 'per_attempt',
- maxTokens: clampMaxTokens(request.maxTokens, authorized.maxOutputTokens),
+ // J8: the workload's own maxTokens is the only value on this seam. The
+ // grant no longer advertises an output cap, because nothing enforced it:
+ // Codex ChatGPT `/codex/responses` rejects max_output_tokens and the
+ // proxy omits it, so the clamp bounded only the hashed request shape
+ // while the real call ran uncapped.
+ maxTokens: request.maxTokens,
temperature: request.temperature ?? authorized.modelPolicy?.temperature,
timeoutMs: this.deps.promptTimeoutMs,
recipeNamespace: this.deps.recipeNamespace,
@@ -649,18 +654,3 @@ function unexpectedAuthorizationResponse(): PluginWorkloadError {
false
)
}
-
-/**
- * Grant-advertised output cap for providers that honor max_tokens on the wire.
- * Codex ChatGPT `/codex/responses` rejects max_output_tokens; the proxy omits
- * it. This clamp still bounds the host request used for authorize hashing. It
- * is not ChatGPT enforcement.
- */
-function clampMaxTokens(
- requested: number | undefined,
- grantCap: number | null
-): number | undefined {
- if (grantCap === null) return requested
- if (requested === undefined) return grantCap
- return Math.min(requested, grantCap)
-}
diff --git a/mcp-host/src/pluginWorkloadSdk/promptBridge/runtime-http.test.ts b/mcp-host/src/pluginWorkloadSdk/promptBridge/runtime-http.test.ts
index 0a0592cc2..a212a45cd 100644
--- a/mcp-host/src/pluginWorkloadSdk/promptBridge/runtime-http.test.ts
+++ b/mcp-host/src/pluginWorkloadSdk/promptBridge/runtime-http.test.ts
@@ -149,7 +149,6 @@ function makeHarness(scenario: Scenario): RuntimeHarness {
attemptGeneration: 1,
policyRevision: 9,
policyHash: POLICY_HASH,
- maxOutputTokens: 256,
}
})
const controlApiClient = {
diff --git a/mcp-host/src/pluginWorkloadSdk/securityInvariants.test.ts b/mcp-host/src/pluginWorkloadSdk/securityInvariants.test.ts
index 4a9c7d8eb..e3b563a70 100644
--- a/mcp-host/src/pluginWorkloadSdk/securityInvariants.test.ts
+++ b/mcp-host/src/pluginWorkloadSdk/securityInvariants.test.ts
@@ -56,7 +56,6 @@ describe('Plugin Workload SDK — §6.2 response-shape security invariants', ()
attemptGeneration: 1,
policyRevision: 1,
policyHash: 'a'.repeat(64),
- maxOutputTokens: null,
}),
reportProviderAttemptStatus: vi.fn().mockResolvedValue(undefined),
reportInvocationStatus: vi.fn().mockResolvedValue(undefined),
diff --git a/workflow-recipes/package-lock.json b/workflow-recipes/package-lock.json
index ec9ef2bdd..0d2273a9f 100644
--- a/workflow-recipes/package-lock.json
+++ b/workflow-recipes/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "workload-recipes-controller",
- "version": "0.1.328",
+ "version": "0.1.329",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "workload-recipes-controller",
- "version": "0.1.328",
+ "version": "0.1.329",
"license": "SEE LICENSE IN LICENSE",
"dependencies": {
"@aws-sdk/client-s3": "^3.1009.0",
diff --git a/workflow-recipes/package.json b/workflow-recipes/package.json
index a7a77cf17..83f779fea 100644
--- a/workflow-recipes/package.json
+++ b/workflow-recipes/package.json
@@ -1,6 +1,6 @@
{
"name": "workload-recipes-controller",
- "version": "0.1.328",
+ "version": "0.1.329",
"description": "Workload Recipes Controller (WRC) - WorkloadRecipe lifecycle management and MCP server interface",
"main": "dist/main.js",
"engines": {
diff --git a/workflow-recipes/src/types.ts b/workflow-recipes/src/types.ts
index 15293e51e..66dd006ce 100644
--- a/workflow-recipes/src/types.ts
+++ b/workflow-recipes/src/types.ts
@@ -507,6 +507,7 @@ export interface OAuthClientDef {
export interface PluginWorkloadSdkSpec {
promptBridge?: {
allowedModels?: string[]
+ /** @deprecated Accepted for compatibility; no longer enforced (J8). The Codex ChatGPT wire discards max_output_tokens, so no layer capped the response by token count. */
maxOutputTokens?: number
/** @deprecated Accepted for compatibility; no longer enforced (issue #348). Platform per-minute rate limits apply. */
maxRequestsPerRun?: number
@@ -548,6 +549,7 @@ export interface PluginWorkloadSdkSpec {
export interface PluginWorkloadSdkSnippetConfig {
promptBridge?: {
allowedModels?: string[]
+ /** @deprecated Accepted for compatibility; no longer enforced (J8). The Codex ChatGPT wire discards max_output_tokens, so no layer capped the response by token count. */
maxOutputTokens?: number
}
clientNotifications?: {