@wildo-ai/framework-knowledge 1.1.12 → 1.1.13
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.
- package/dist/claude-marketplace/.claude-plugin/marketplace.json +1 -1
- package/dist/claude-marketplace/plugins/wildo/.claude-plugin/plugin.json +1 -1
- package/dist/esm/framework-knowledge.definitions.d.ts +1 -1
- package/dist/esm/framework-knowledge.definitions.d.ts.map +1 -1
- package/dist/esm/framework-knowledge.definitions.js +10 -0
- package/dist/esm/framework-knowledge.definitions.js.map +1 -1
- package/dist/esm/playbooks/bottom_up.foundation/playbook.definition.d.ts +2 -1
- package/dist/esm/playbooks/bottom_up.foundation/playbook.definition.d.ts.map +1 -1
- package/dist/esm/playbooks/bottom_up.foundation/playbook.definition.js +8 -1
- package/dist/esm/playbooks/bottom_up.foundation/playbook.definition.js.map +1 -1
- package/dist/knowledge/assurance-authority-bases/wildo-safe-baseline-v1/external-framework-mapping.json +1 -1
- package/dist/knowledge/manifest.json +7 -7
- package/dist/knowledge/playbooks/bottom_up.foundation/playbook.json +5 -2
- package/dist/knowledge/playbooks/bottom_up.maturity-reading/method.md +9 -0
- package/dist/knowledge/playbooks/bottom_up.maturity-reading/playbook.json +1 -1
- package/dist/knowledge/playbooks/top_down.requirements-definition/method.md +14 -0
- package/dist/knowledge/playbooks/top_down.requirements-definition/playbook.json +1 -1
- package/dist/knowledge/references/application-maturity.config.md +43 -0
- package/dist/knowledge/references/example.infra.environment-config.local.md +2 -2
- package/dist/knowledge/references/example.resource.config-wiring.retention-policy-operations.md +21 -0
- package/dist/knowledge/references/platform.config.infrastructure-host.md +1 -0
- package/dist/knowledge/references/platform.config.platform-key-management.md +5 -2
- package/dist/knowledge/references/requirements.engine-substance.md +38 -0
- package/dist/knowledge/references/requirements.functional-requirement.config.md +11 -0
- package/dist/knowledge/references/requirements.non-functional-requirement.config.md +11 -0
- package/dist/knowledge/references/saas.approval-seals.admin-endpoints.md +20 -0
- package/dist/knowledge/references/saas.backend.flows-actors.function-tool.md +72 -0
- package/dist/knowledge/references/saas.frontend.chart-dashboard-view.md +3 -3
- package/dist/knowledge/references/saas.frontend.chart-dashboard.md +38 -35
- package/dist/knowledge/references/saas.models.chart-dashboard-tile-sizes.md +40 -0
- package/dist/knowledge/references/saas.models.flows-actors.agent-content-guardrails-environment.md +52 -0
- package/dist/knowledge/references/saas.models.organization-dashboard-layout.md +61 -0
- package/dist/knowledge/registries/assurance-authority-bases.json +1 -1
- package/dist/knowledge/registries/framework-topics.json +41 -1
- package/dist/knowledge/registries/playbooks.json +1 -1
- package/dist/knowledge/registries/source-tags.json +1052 -982
- package/dist/knowledge/skills/backend-custom-operation/references/framework-hook-transaction-posture-api.md +11 -0
- package/dist/knowledge/skills/config-saas/references/framework-key-management-api.md +43 -0
- package/dist/knowledge/skills/electronic-signature-deliverables/references/what-makes-a-request-signed.md +20 -3
- package/dist/knowledge/skills/flows-actors-autonomous-agents/SKILL.md +12 -2
- package/dist/knowledge/skills/flows-actors-chat-assistant/SKILL.md +16 -4
- package/dist/knowledge/skills/flows-actors-chat-assistant/references/framework-agent-content-guardrails-api.md +144 -6
- package/dist/knowledge/skills/flows-actors-chat-assistant/references/framework-flows-actors-tool-api.md +72 -0
- package/dist/knowledge/skills/flows-actors-declared-flow/SKILL.md +12 -2
- package/dist/knowledge/skills/flows-actors-llm-agent-invocation/SKILL.md +12 -2
- package/dist/knowledge/skills/frontend-component-catalog/SKILL.md +1 -1
- package/dist/knowledge/skills/frontend-component-catalog/references/framework-component-catalog.md +3 -2
- package/dist/knowledge/skills/frontend-component-catalog/references/framework-component-usage.md +13 -0
- package/dist/knowledge/skills/infra-environment-config/SKILL.md +26 -2
- package/dist/knowledge/skills/infra-environment-config/references/framework-agent-content-guardrails-environment-api.md +67 -0
- package/dist/knowledge/skills/infra-environment-config/references/framework-infrastructure-host-api.md +1 -0
- package/dist/knowledge/skills/infra-environment-config/references/framework-platform-key-management-api.md +5 -2
- package/dist/knowledge/skills/provider-backend-package/SKILL.md +17 -3
- package/dist/knowledge/skills/resource-config-wiring/references/framework-operation-audit-api.md +2 -1
- package/dist/knowledge/skills/resource-config-wiring/references/framework-retention-policy-operations.md +21 -0
- package/dist/knowledge/skills/views-composite-authoring/SKILL.md +1 -1
- package/dist/knowledge/skills/views-composite-authoring/references/framework-chart-dashboard-api.md +166 -39
- package/dist/knowledge/topic-index.json +197 -58
- package/dist/tsconfig.build.tsbuildinfo +1 -1
- package/package.json +4 -4
|
@@ -36,6 +36,17 @@ audit row that must exist before the operation answers), and write why beside it
|
|
|
36
36
|
its own reported reason, not as unreviewed. A registration that declares nothing is treated as unreviewed, and so is an
|
|
37
37
|
operation one of whose registrations declares nothing.
|
|
38
38
|
|
|
39
|
+
When a hook's effect outside the database must come FIRST (a lookup whose answer the write needs, a revocation that
|
|
40
|
+
must precede what you write), it does not belong in a hook that runs inside the transaction. Give the registration a
|
|
41
|
+
`transactionalCore` with a `beforeTransactionalCore` handler: it runs once, outside any transaction, before the core, and
|
|
42
|
+
hands what it learned to PREFIX and REPLACE as `utils.beforeTransactionalCore.continuation`. Re-check inside the
|
|
43
|
+
transaction that the row still holds what that answer was about, and refuse when it does not.
|
|
44
|
+
|
|
45
|
+
On a multi-row verb (`updateMany`, `deleteMany`) whose REPLACE handler writes the rows itself, a transaction must also be
|
|
46
|
+
short: declare `maximumRowsWrittenPerAttempt` (at most 100) beside `RUNS_INSIDE_TRANSACTION` when the handler cannot write
|
|
47
|
+
more by construction (one row by a unique key, a page it bounds itself), and say why. Without it the write stays best
|
|
48
|
+
effort. A multi-row write with no custom implementation needs nothing: the framework runs it 100 rows per transaction.
|
|
49
|
+
|
|
39
50
|
## Import Origins
|
|
40
51
|
|
|
41
52
|
- `@wildo-ai/saas-backend-lib`
|
|
@@ -326,6 +326,16 @@ export const ApprovalSealsAdminEndpointPath = {
|
|
|
326
326
|
* which no copy can be made (`reason` names why, `detail` says more).
|
|
327
327
|
*/
|
|
328
328
|
MANIFESTED_COPY_REBUILD: '/manifested-copy/rebuild',
|
|
329
|
+
/**
|
|
330
|
+
* POST — whether ONE seal's approval is still CURRENT over the record it is about (#2914): `{ "sealId": "…" }` →
|
|
331
|
+
* {@link ApprovalSealCurrencyResult}. The seal's presentation carries the values of the fields its tool declared the
|
|
332
|
+
* approval covers (`sealedRecordFields`); they are compared with the record as it is now. CURRENT right after the
|
|
333
|
+
* approval, NO_LONGER_CURRENT once one of them changed, SUBJECT_GONE when the record is gone, NO_SEALED_FIELD_SET when
|
|
334
|
+
* the seal covers no record field. Names which fields moved, never a value, so it is not recorded on the trail. 404 for
|
|
335
|
+
* an unknown seal (a seal its approver's erasure retained is not read here); 409 for one whose presentation file is
|
|
336
|
+
* unreadable or not the one its row recorded.
|
|
337
|
+
*/
|
|
338
|
+
CURRENCY: '/currency',
|
|
329
339
|
} as const;
|
|
330
340
|
|
|
331
341
|
/**
|
|
@@ -411,6 +421,16 @@ export interface ApprovalSealRetryResult {
|
|
|
411
421
|
readonly sealed: boolean;
|
|
412
422
|
}
|
|
413
423
|
|
|
424
|
+
/** `POST /currency`'s answer (#2914). Names no person and no value. */
|
|
425
|
+
export interface ApprovalSealCurrencyResult {
|
|
426
|
+
readonly sealId: string;
|
|
427
|
+
readonly standing: ApprovalSealSubjectCurrency;
|
|
428
|
+
/** The record compared, as the seal names it; absent for NO_SEALED_FIELD_SET. */
|
|
429
|
+
readonly subject?: { readonly resourceType: string; readonly resourceId: string };
|
|
430
|
+
/** Each sealed field's currency (CURRENT / CHANGED / ABSENT); absent unless compared. */
|
|
431
|
+
readonly fields?: readonly { readonly key: string; readonly currency: HumanDecision_PresentedFieldCurrency }[];
|
|
432
|
+
}
|
|
433
|
+
|
|
414
434
|
/** `POST /manifested-copy/rebuild`'s answer. */
|
|
415
435
|
export interface ApprovalSealManifestedCopyRebuildResult {
|
|
416
436
|
readonly sealId: string;
|
|
@@ -479,6 +499,29 @@ changed byte and the seal does not verify. A file the engine cannot hash (a remo
|
|
|
479
499
|
keeps the bytes, a missing file, a file of another organization) is shown and sealed as not bound, never
|
|
480
500
|
bound to an empty hash.
|
|
481
501
|
|
|
502
|
+
When the person approves what a RECORD says (a month's figures signed off, an invoice's amounts), declare
|
|
503
|
+
the fields the approval covers, `sealedRecordFields` on the function tool, beside
|
|
504
|
+
`argumentsResource: { resourceType, rowIdArgument }`. The card then shows those fields of the row the call
|
|
505
|
+
names, read when the card is built, and the seal commits to them inside its presentation, so you can ask
|
|
506
|
+
later whether the approval still holds: `wildo seal currency <sealId>` (the operator door
|
|
507
|
+
`POST /internal/admin/approval-seals/currency`) answers CURRENT, NO_LONGER_CURRENT with the fields that
|
|
508
|
+
moved (never their values), SUBJECT_GONE, or NO_SEALED_FIELD_SET; offline,
|
|
509
|
+
`wildo seal verify <bundle.json> --subject-record <record.json>` compares the same values with the record
|
|
510
|
+
you give, and reports a change without failing the seal. Name only what the approval is about, never a
|
|
511
|
+
status, a timestamp or the column the approved call writes the seal's id into: those change when the
|
|
512
|
+
approval is recorded, and would make every approval stale at once. A card read before one of the values
|
|
513
|
+
changed is refused as stale. In your own code, `assessApprovalSealCurrency` (from
|
|
514
|
+
`@wildo-ai/saas-backend-lib`) and `approvalSealSubjectCurrencyOf` (from `@wildo-ai/saas-models`) give the
|
|
515
|
+
same answer. A record of another organization than the decision's is shown and sealed with no values, and
|
|
516
|
+
so is an answer clicked on a chat card: no chat card shows the record, so its seal claims none of it.
|
|
517
|
+
|
|
518
|
+
When the same decision is taken again on a record after the record was reopened (signed off, reopened,
|
|
519
|
+
signed off again), it is a NEW decision, and its seal names no predecessor unless you declare
|
|
520
|
+
`approvalReplacesEarlierApprovalOfRow: true` on the tool: its seal then supersedes the latest seal of that
|
|
521
|
+
tool on that record (`supersedesSealId`, and `supersedes` in the signed claims), as a decision answered
|
|
522
|
+
again supersedes its own earlier answer, and the trail records the earlier one `approval_seal_superseded`.
|
|
523
|
+
An approval made through another tool on the same row is never superseded.
|
|
524
|
+
|
|
482
525
|
When the approved file is a PDF, the engine also makes its **manifested copy** as it seals: a separate
|
|
483
526
|
file holding the original byte for byte, followed by a last page naming the approver, the date and time
|
|
484
527
|
of the decision, the meaning and outcome and the seal's identity, with the seal's JWS attached
|
|
@@ -40,8 +40,23 @@ signer or a lower level ends it FAILED.
|
|
|
40
40
|
|
|
41
41
|
A SIGNED request names two files of its own, written in the same move: the signed PDF (a vendor-signed copy of the
|
|
42
42
|
original) and the validation's evidence as JSON (each signature, its chain, the check date, and every CRL and OCSP
|
|
43
|
-
response used). They go with the request when its record goes, and the record's legal holds keep them.
|
|
44
|
-
|
|
43
|
+
response used). They go with the request when its record goes, and the record's legal holds keep them. Where your
|
|
44
|
+
application scans files, the signed PDF is linked to its request once its scan ends; the poll does it, so nothing is lost
|
|
45
|
+
while the scan runs. The audit trail records each sending, each move (with the hashes and the signers), each late event
|
|
46
|
+
and each cancellation.
|
|
47
|
+
|
|
48
|
+
## Watching it
|
|
49
|
+
|
|
50
|
+
The poll's batch output counts what each run sent, moved, linked and left pending. The metrics answer whether signing is
|
|
51
|
+
healthy without reading the audit rows:
|
|
52
|
+
|
|
53
|
+
| Metric | Labels | Worth an alert when |
|
|
54
|
+
|---|---|---|
|
|
55
|
+
| `signature_request_sent_total` | `outcome` (`sent` or the refusal), `required_level` | a refusal other than `already_open` keeps appearing |
|
|
56
|
+
| `signature_request_moved_total` | `provider`, `to` | `to="failed"` appears (a signer or level that can never be met) |
|
|
57
|
+
| `signature_request_evidence_refused_total` | `provider`, `failure` | any `failure` but `rejected`: the engine cannot validate or keep evidence |
|
|
58
|
+
| `signature_request_vendor_not_read_total` | `provider` | it rises steadily: the vendor is down |
|
|
59
|
+
| `signature_request_evidence_link_total` | `outcome` | `outcome="lost"`: a SIGNED request names evidence nothing keeps |
|
|
45
60
|
|
|
46
61
|
## Notifications
|
|
47
62
|
|
|
@@ -58,4 +73,6 @@ poll re-reads that request within five minutes.
|
|
|
58
73
|
it delegated to, and current).
|
|
59
74
|
- OCSP is read when the signature carries it, never fetched; the CRL is fetched instead.
|
|
60
75
|
- The appearance a signature field draws is not inspected.
|
|
61
|
-
-
|
|
76
|
+
- Signers are invited, and their certificate's e-mail compared, at the address in your database. Someone who can change
|
|
77
|
+
a signer's address can receive their invitation. Requiring identity proofing that does not rely on that data (an
|
|
78
|
+
identity document or eID checked by the vendor, proved by the certificate it issues) is designed and not yet available.
|
|
@@ -83,8 +83,18 @@ A screening or moderation declaration with no content-safety provider makes the
|
|
|
83
83
|
fails closed, by design, and it does so in EVERY environment that cannot serve it, so a provider and its
|
|
84
84
|
secrets authored for `local` only leave the agent refusing every turn in staging and production. The
|
|
85
85
|
backend warns at startup, one line per such agent, and `wildo local doctor` lists every authored
|
|
86
|
-
environment that cannot serve it.
|
|
87
|
-
|
|
86
|
+
environment that cannot serve it. On the development machine the classifier is usually a local double: the
|
|
87
|
+
provider's host secret names `127.0.0.1:<port>`, and `wildo local dev` starts the double on that port when the
|
|
88
|
+
application's root `package.json` declares `"@wildo-ai/provider-simulators": "*"` in its `devDependencies` (add it
|
|
89
|
+
in the same task that selects the provider). Without it nothing answers there, and every guarded turn is withheld
|
|
90
|
+
locally as `OUTPUT_NOT_MODERATED`. Two registry providers serve it: `azure-content-safety` screens and moderates
|
|
91
|
+
(its own Azure resource), `openai-moderation` moderates only, free, with an OpenAI key. An environment that is meant to
|
|
92
|
+
run without one (a staging stack with no key for it, while production has one) does not drop the declaration: it
|
|
93
|
+
states `agentContentGuardrails: { withoutClassifier: RUN_UNGUARDED, reason }` in its `wildo.infra.<env>.config.ts`, and
|
|
94
|
+
the agent runs unguarded there, named at startup. Never author that for an environment because a provider's setup is
|
|
95
|
+
pending; it is the operator's decision, and the infra-environment-config skill describes it. Until the application
|
|
96
|
+
selects a provider, declare `personalData`, and write one line at the agent saying which of the other two it will
|
|
97
|
+
declare and why. A guardrail that cannot treat
|
|
88
98
|
anything (an agent with no tools and no conversation, fed only text your code composes) is not declared at
|
|
89
99
|
all: a declaration that does nothing reads as a control. The contract and every refusal are in the
|
|
90
100
|
`flows-actors-chat-assistant` skill, `references/framework-agent-content-guardrails-api.md`.
|
|
@@ -162,7 +162,9 @@ where none can serve it the agent refuses every turn, the backend warns at start
|
|
|
162
162
|
environment. To keep what it
|
|
163
163
|
**writes** within a content policy (no hate, sexual content, violence or self-harm reaching a person, a record or a
|
|
164
164
|
message), declare `contentGuardrails.outputModeration` with the severity to withhold per category. See
|
|
165
|
-
`references/framework-agent-content-guardrails-api.md`.
|
|
165
|
+
`references/framework-agent-content-guardrails-api.md`. When a reply instead reads *"This reply couldn't be checked
|
|
166
|
+
against the content policy"*, the classifier is configured and did not answer (on a development machine, usually its
|
|
167
|
+
local double is not running): that reference's section of the same name diagnoses it from the turn's audit event.
|
|
166
168
|
|
|
167
169
|
Every agent call leaves an **operation record** (`agentOperationRecords`): the model the provider answered with, each tool
|
|
168
170
|
call and how it ended, what the guardrails decided and how the call ended, never what a tool returned. You write
|
|
@@ -213,8 +215,18 @@ A screening or moderation declaration with no content-safety provider makes the
|
|
|
213
215
|
fails closed, by design, and it does so in EVERY environment that cannot serve it, so a provider and its
|
|
214
216
|
secrets authored for `local` only leave the agent refusing every turn in staging and production. The
|
|
215
217
|
backend warns at startup, one line per such agent, and `wildo local doctor` lists every authored
|
|
216
|
-
environment that cannot serve it.
|
|
217
|
-
|
|
218
|
+
environment that cannot serve it. On the development machine the classifier is usually a local double: the
|
|
219
|
+
provider's host secret names `127.0.0.1:<port>`, and `wildo local dev` starts the double on that port when the
|
|
220
|
+
application's root `package.json` declares `"@wildo-ai/provider-simulators": "*"` in its `devDependencies` (add it
|
|
221
|
+
in the same task that selects the provider). Without it nothing answers there, and every guarded turn is withheld
|
|
222
|
+
locally as `OUTPUT_NOT_MODERATED`. Two registry providers serve it: `azure-content-safety` screens and moderates
|
|
223
|
+
(its own Azure resource), `openai-moderation` moderates only, free, with an OpenAI key. An environment that is meant to
|
|
224
|
+
run without one (a staging stack with no key for it, while production has one) does not drop the declaration: it
|
|
225
|
+
states `agentContentGuardrails: { withoutClassifier: RUN_UNGUARDED, reason }` in its `wildo.infra.<env>.config.ts`, and
|
|
226
|
+
the agent runs unguarded there, named at startup. Never author that for an environment because a provider's setup is
|
|
227
|
+
pending; it is the operator's decision, and the infra-environment-config skill describes it. Until the application
|
|
228
|
+
selects a provider, declare `personalData`, and write one line at the agent saying which of the other two it will
|
|
229
|
+
declare and why. A guardrail that cannot treat
|
|
218
230
|
anything (an agent with no tools and no conversation, fed only text your code composes) is not declared at
|
|
219
231
|
all: a declaration that does nothing reads as a control. The contract and every refusal are in the
|
|
220
232
|
`flows-actors-chat-assistant` skill, `references/framework-agent-content-guardrails-api.md`.
|
|
@@ -576,7 +588,7 @@ export const moduleBackend_FlowsActorsRegistry: {
|
|
|
576
588
|
- `references/framework-present-records-tool-api.md` — `presentRecordsTool()`: the agent shows its person records as summary cards in the conversation, kept as a reference and fetched with the viewer's session
|
|
577
589
|
- `references/framework-retrieval-tool-api.md` — `retrievalTool()`: the agent searches your `.ragSource()` fields as its person, cites the record and field of each passage, and reads `outcome` before answering
|
|
578
590
|
- `references/framework-ask-person-tool-api.md` — `askPersonTool()`: the agent pauses its turn to ask its person a structured question, answered through a form, and gets typed answers back
|
|
579
|
-
- `references/framework-agent-content-guardrails-api.md` — `contentGuardrails.personalData` (what the model is sent in place of a personal field's value, who reads the real values back, the `personalDataOutput` every tool declares), `contentGuardrails.promptInjection` (screening what others wrote before the model reads it), `contentGuardrails.outputModeration` (judging what the model writes before anyone reads, stores or sends it)
|
|
591
|
+
- `references/framework-agent-content-guardrails-api.md` — `contentGuardrails.personalData` (what the model is sent in place of a personal field's value, who reads the real values back, the `personalDataOutput` every tool declares), `contentGuardrails.promptInjection` (screening what others wrote before the model reads it), `contentGuardrails.outputModeration` (judging what the model writes before anyone reads, stores or sends it), `argumentsResource` on every write tool (what a guarded write may store), and how to diagnose a reply withheld because the classifier did not answer (`OUTPUT_NOT_MODERATED`, the local double `wildo local dev` starts)
|
|
580
592
|
- `references/framework-agent-suspension-api.md` — an administrator's runtime off switch for one agent: who may suspend and resume it (an organization's administrator, or the application's super-administrator for everybody), what a suspended agent's caller and conversation see, what is audited, and what a suspension does not stop
|
|
581
593
|
- `references/framework-agent-operation-record-api.md` — the `agentOperationRecords` row every agent call leaves: the model the provider reported, each tool call and how it ended, the guardrail verdicts and the end, never a tool's result; how to read it per run and as counts through the evidence door, and its retention
|
|
582
594
|
- `references/framework-agent-behavior-reports-api.md` — a person reports an agent's reply from the conversation (a closed category, an optional comment; the report points at the reply and never copies it), the organization's administrators triage, resolve or dismiss it against declared deadlines, and how it is audited, erased and measured
|
|
@@ -107,9 +107,9 @@ contentGuardrails: { promptInjection: { screening: FlowsActors_PromptInjectionSc
|
|
|
107
107
|
naming where the content came from, and routed to the organization's SIEM.
|
|
108
108
|
4. A failed tool tells the model only the kind of failure: an error's text can carry what a third party wrote.
|
|
109
109
|
5. The classifier is the provider selected for `EngineCapability.AI_CONTENT_SAFETY`. Without one, or where provider calls
|
|
110
|
-
are not recorded, the agent refuses to run: running unscreened is what the declaration prevents. The
|
|
111
|
-
Azure AI Content Safety's Prompt Shields: `wildo registry add wildo/providers/azure-content-safety`, then
|
|
112
|
-
own skill.
|
|
110
|
+
are not recorded, the agent refuses to run: running unscreened is what the declaration prevents. The provider that
|
|
111
|
+
screens is Azure AI Content Safety's Prompt Shields: `wildo registry add wildo/providers/azure-content-safety`, then
|
|
112
|
+
follow its own skill. (`openai-moderation` moderates output only and does not screen.)
|
|
113
113
|
|
|
114
114
|
Not screened, deliberately: the message of a person driving their own conversation, and the answer they give to the
|
|
115
115
|
agent's question (`askPersonTool`), which are their own words on their own authority.
|
|
@@ -140,8 +140,15 @@ contentGuardrails: {
|
|
|
140
140
|
4. Each withholding is recorded on the audit trail (`agent_output_withheld`, `agent_output_moderation_failed`) with the
|
|
141
141
|
categories and severities, never the text. What the provider generated is still metered.
|
|
142
142
|
5. The classifier is the provider selected for `EngineCapability.AI_CONTENT_SAFETY`, and it must assess every category
|
|
143
|
-
your policy names
|
|
144
|
-
|
|
143
|
+
your policy names. Two registry providers moderate output, and both assess all four:
|
|
144
|
+
|
|
145
|
+
| provider | grades severity | cost | also screens prompt injection |
|
|
146
|
+
|---|---|---|---|
|
|
147
|
+
| `azure-content-safety` | LOW / MEDIUM / HIGH | the application's Azure resource (a free tier exists) | yes |
|
|
148
|
+
| `openai-moderation` | no: a flag reads as HIGH, so every threshold withholds what OpenAI flags | free, an OpenAI API key | no |
|
|
149
|
+
|
|
150
|
+
Without one, or where provider calls are not recorded, the agent refuses to run. Under `personalData` the classifier
|
|
151
|
+
receives the placeholders, never the values.
|
|
145
152
|
|
|
146
153
|
A call the person APPROVED is not judged again when it runs: its input was judged when the turn paused for approval.
|
|
147
154
|
|
|
@@ -171,10 +178,141 @@ Two places say so before a person's turn does:
|
|
|
171
178
|
dev companion is running, and otherwise passes saying the agents were not probed.
|
|
172
179
|
|
|
173
180
|
The remedy is the same in both: in each environment named, enable `EngineCapability.AI_CONTENT_SAFETY`, give the
|
|
174
|
-
backend scope a provider that serves it, and author that provider's secrets; or remove the declaration from the agent
|
|
181
|
+
backend scope a provider that serves it, and author that provider's secrets; or remove the declaration from the agent;
|
|
182
|
+
or, deliberately, declare that the environment runs its agents unguarded (next section).
|
|
175
183
|
An organization's policy that switches screening or moderation on is refused when it is saved unless a classifier can
|
|
176
184
|
serve it then; if the environment later loses its classifier, the startup count above is where that shows.
|
|
177
185
|
|
|
186
|
+
## Running an environment without a classifier, on purpose
|
|
187
|
+
|
|
188
|
+
Some environments are meant to run without a classifier: a staging stack nobody pays a vendor for, or a production
|
|
189
|
+
launched before one is bought. Not a good practice, and a common one. Do not delete the declaration from the agent for
|
|
190
|
+
that, and do not look for a classifier that answers "safe" to everything (there is none, on purpose: an agent that looks
|
|
191
|
+
guarded and is not is worse than one that refuses). STATE it, in that environment's
|
|
192
|
+
`infrastructure/<env>/wildo.infra.<env>.config.ts`:
|
|
193
|
+
|
|
194
|
+
```ts
|
|
195
|
+
agentContentGuardrails: {
|
|
196
|
+
withoutClassifier: AgentContentGuardrailsWithoutClassifier.RUN_UNGUARDED,
|
|
197
|
+
reason: 'Staging has no Azure Content Safety key; production authors one.',
|
|
198
|
+
},
|
|
199
|
+
```
|
|
200
|
+
|
|
201
|
+
## Import Origins
|
|
202
|
+
|
|
203
|
+
- `@wildo-ai/saas-models`
|
|
204
|
+
|
|
205
|
+
## Source Excerpts
|
|
206
|
+
|
|
207
|
+
### Origin Package: `@wildo-ai/saas-models`
|
|
208
|
+
|
|
209
|
+
```typescript
|
|
210
|
+
*
|
|
211
|
+
* What an environment does with an agent's classifier-backed content guardrails (`contentGuardrails.promptInjection`,
|
|
212
|
+
* `contentGuardrails.outputModeration`) when NO classifier for the guardrail is present there: its configuration declares
|
|
213
|
+
* no provider for `EngineCapability.AI_CONTENT_SAFETY`, the provider it selects does not offer that guardrail, or that
|
|
214
|
+
* provider's credentials are not authored in this environment's `secrets.json → providerEnv`. Imported
|
|
215
|
+
* from `@wildo-ai/saas-models`; authored as `agentContentGuardrails` in `infrastructure/<env>/wildo.infra.<env>.config.ts`.
|
|
216
|
+
*/
|
|
217
|
+
export enum AgentContentGuardrailsWithoutClassifier {
|
|
218
|
+
/** The default, and what an environment that authors nothing gets: the agent refuses every turn. */
|
|
219
|
+
REFUSE = 'refuse',
|
|
220
|
+
/**
|
|
221
|
+
* The agent runs without each guardrail no classifier is present for here (its tool results unscreened, or its output
|
|
222
|
+
* unmoderated). A classifier that is there and cannot be used is never waived: a declared provider that did not load or
|
|
223
|
+
* a category it does not assess still refuses the turn, and a key that is set and refused, or an outage, still
|
|
224
|
+
* withholds the output. Named at startup, and listed by launch-blockers outside a local environment.
|
|
225
|
+
*/
|
|
226
|
+
RUN_UNGUARDED = 'run_unguarded',
|
|
227
|
+
}
|
|
228
|
+
|
|
229
|
+
/**
|
|
230
|
+
* The environment's declaration. `reason` is required with `RUN_UNGUARDED`: it is printed beside every agent the backend
|
|
231
|
+
* leaves unguarded, so the next reader knows it was decided, and by whom or why.
|
|
232
|
+
*
|
|
233
|
+
* ```ts
|
|
234
|
+
* agentContentGuardrails: {
|
|
235
|
+
* withoutClassifier: AgentContentGuardrailsWithoutClassifier.RUN_UNGUARDED,
|
|
236
|
+
* reason: 'Staging has no Azure Content Safety key; production authors one.',
|
|
237
|
+
* },
|
|
238
|
+
* ```
|
|
239
|
+
*/
|
|
240
|
+
export const EnvironmentAgentContentGuardrailsSchema = z.discriminatedUnion('withoutClassifier', [
|
|
241
|
+
z.strictObject({
|
|
242
|
+
withoutClassifier: z.literal(AgentContentGuardrailsWithoutClassifier.REFUSE),
|
|
243
|
+
}),
|
|
244
|
+
z.strictObject({
|
|
245
|
+
withoutClassifier: z.literal(AgentContentGuardrailsWithoutClassifier.RUN_UNGUARDED),
|
|
246
|
+
reason: z.string().trim().min(10, 'Say why this environment runs its agents unguarded (at least 10 characters).').max(500),
|
|
247
|
+
}),
|
|
248
|
+
]);
|
|
249
|
+
export type EnvironmentAgentContentGuardrails = z.output<typeof EnvironmentAgentContentGuardrailsSchema>;
|
|
250
|
+
```
|
|
251
|
+
|
|
252
|
+
Then run `wildo config sync`. What changes in that environment only:
|
|
253
|
+
|
|
254
|
+
1. A guardrail for which NO classifier is present there runs without it: none is declared for `AI_CONTENT_SAFETY`, the selected one does not offer that guardrail
|
|
255
|
+
(`openai-moderation` does not screen), or its credentials (its key, or the host slot it reads) are not authored in
|
|
256
|
+
this environment's `secrets.json → providerEnv`.
|
|
257
|
+
Its tool results and inbound deliveries reach the model unscreened, or its output leaves unmoderated. A classifier
|
|
258
|
+
that offers moderation but not screening (`openai-moderation`) waives only screening. The classifier is declared once
|
|
259
|
+
for the application in `wildo.saas.config.ts`, so leaving its credentials out of one environment is how that
|
|
260
|
+
environment runs without it while another has it.
|
|
261
|
+
2. A classifier that IS there and cannot be used is NOT waived: that is a failure. A provider declared and not loaded,
|
|
262
|
+
or a category it does not assess, refuses the turn; a key that is set and refused, or the vendor down, withholds the
|
|
263
|
+
output.
|
|
264
|
+
3. The backend's startup log names every such agent as `RUNS UNGUARDED`, with your `reason`, on every boot.
|
|
265
|
+
4. Outside a local environment, `wildo context launch-blockers` lists the environment until a classifier is selected
|
|
266
|
+
there or the declaration is removed.
|
|
267
|
+
5. An organization's administrator still cannot switch screening or moderation ON there: the write is refused, because
|
|
268
|
+
nothing could provide it.
|
|
269
|
+
|
|
270
|
+
The declaration on the agent stays, so the environment that selects a classifier enforces it with no edit. An
|
|
271
|
+
application that wants no guardrail anywhere simply declares none on its agents.
|
|
272
|
+
|
|
273
|
+
## "This reply couldn't be checked against the content policy"
|
|
274
|
+
|
|
275
|
+
That is what a person reads when a reply, or a tool call the model wrote, was withheld as `OUTPUT_NOT_MODERATED`: a
|
|
276
|
+
classifier IS configured, and it did not give a verdict for this turn. Both checks above pass in that case (they read
|
|
277
|
+
configuration, not whether the classifier answers), so diagnose it from what the turn recorded:
|
|
278
|
+
|
|
279
|
+
1. Find the `agent_output_moderation_failed` audit event for the turn (or the backend's
|
|
280
|
+
`An agent's output was withheld by its output moderation` WARN with the same `executionId`). Its `failureReason`
|
|
281
|
+
names the cause and its `outputKind` what was being judged (`text`, `tool_input`, …). Never the text itself.
|
|
282
|
+
2. Read the cause:
|
|
283
|
+
|
|
284
|
+
| `failureReason` says | Cause | Remedy |
|
|
285
|
+
|---|---|---|
|
|
286
|
+
| `was unreachable … ECONNREFUSED` | Nothing answers at the provider's host. On a development machine, the local double is not running | See "On a development machine" below |
|
|
287
|
+
| the provider answered `401` / `403` | The key in `providerEnv` is not the one the classifier (or the double) accepts | Re-author the key slot with `wildo secrets set` |
|
|
288
|
+
| a spend ceiling | The application's screening allowance is used up | Raise the ceiling or wait for its period (`references/framework-model-spend-limits-api.md`) |
|
|
289
|
+
| a category left out | The classifier does not assess a category the policy names | Drop the category from the policy, or select a classifier that assesses it |
|
|
290
|
+
|
|
291
|
+
Once the cause is fixed the person can send the message again, after checking what the turn already did: no reply was
|
|
292
|
+
stored and the withheld output's tool call (with the other calls of its step) did not run, but a tool called in an
|
|
293
|
+
EARLIER step of the same turn did. The turn's trace lists those calls; resending repeats a write that already landed.
|
|
294
|
+
|
|
295
|
+
### On a development machine
|
|
296
|
+
|
|
297
|
+
A provider whose LOCAL host secret names this machine (`AZURE_CONTENT_SAFETY_HOST=127.0.0.1:<port>`, in
|
|
298
|
+
`infrastructure/local/secrets/secrets.json → providerEnv`) is served by a local double, not by the vendor.
|
|
299
|
+
`wildo local dev` starts that double itself, on that port and accepting that key, before the application's processes,
|
|
300
|
+
and restarts it if it stops: look for `Starting the azure-content-safety double` in its output. It needs the doubles'
|
|
301
|
+
package: add `"@wildo-ai/provider-simulators": "*"` to the `devDependencies` of the application's root `package.json`,
|
|
302
|
+
then run `wildo config sync` and `pnpm install`. Without it, `wildo local dev` warns that the double cannot be started,
|
|
303
|
+
and every guarded agent's turn is withheld until something answers on the port.
|
|
304
|
+
|
|
305
|
+
If the stack is already running and the double has died, restart `wildo local dev`. Or run the double alone in another
|
|
306
|
+
terminal, with the port and key from the two secrets:
|
|
307
|
+
|
|
308
|
+
```bash
|
|
309
|
+
CONTENT_SAFETY_MOCK_PORT=<port> CONTENT_SAFETY_MOCK_API_KEY=<key> pnpm exec wildo-provider-simulators-content-safety
|
|
310
|
+
```
|
|
311
|
+
|
|
312
|
+
The double judges every text SAFE unless it carries a harm marker (`WILDO-SIMULATED-HARM-Violence` reaches the top
|
|
313
|
+
severity in Violence), which is how a test makes it withhold. It proves the guardrails are wired, not that real content
|
|
314
|
+
is moderated well: that needs the real classifier in a deployed environment.
|
|
315
|
+
|
|
178
316
|
## An organization can only tighten them
|
|
179
317
|
|
|
180
318
|
Each organization can make the guardrails of the agents that work on its records STRICTER than they declare, never
|
|
@@ -434,6 +434,42 @@ export interface FunctionTool {
|
|
|
434
434
|
* has no top-level string argument of that name. A tool without it is unaffected.
|
|
435
435
|
*/
|
|
436
436
|
deliverableFileArgument?: string;
|
|
437
|
+
/**
|
|
438
|
+
* The fields of the row the call acts on that the approval COVERS (#2914): the record's SEALED FIELD SET. Declare it
|
|
439
|
+
* on an approval-gated tool whose approval is a statement about a record's values (a month's figures signed off, an
|
|
440
|
+
* invoice's amounts approved), so that whether the approval still holds can be answered later:
|
|
441
|
+
*
|
|
442
|
+
* - the decision's card shows these values of the row named by `argumentsResource.rowIdArgument`, read as the system
|
|
443
|
+
* when the card is built, so the person approves what they read; a card read before one of them changed is refused
|
|
444
|
+
* as stale;
|
|
445
|
+
* - the approval's seal commits to them inside its presentation, and the currency check
|
|
446
|
+
* (`approvalSealSubjectCurrencyOf`, the operator door `POST /internal/admin/approval-seals/currency`, and
|
|
447
|
+
* `wildo seal verify --subject-record`) compares them with the record as it is: CURRENT right after the approval,
|
|
448
|
+
* NO_LONGER_CURRENT once one of them changes.
|
|
449
|
+
*
|
|
450
|
+
* Name only what the approval is ABOUT. Never a lifecycle or approval field (a status, `updatedAt`, the column the
|
|
451
|
+
* approved call writes the seal's id or the signer into): the approved call's own write would otherwise make its own
|
|
452
|
+
* seal look stale. A row of another organization than the decision's, or one that cannot be read, is shown and sealed
|
|
453
|
+
* with no record values, and so is an answer given on a chat card: no notice card shows the row, so its SENT seal
|
|
454
|
+
* claims none of it (its currency reads NO_SEALED_FIELD_SET).
|
|
455
|
+
*
|
|
456
|
+
* Registration refuses it on a tool that is not approval-gated, on a question tool, without
|
|
457
|
+
* `argumentsResource.rowIdArgument`, and when it names no field, a field twice, or a field the resource's main schema
|
|
458
|
+
* does not declare (checked when the resource is registered).
|
|
459
|
+
*/
|
|
460
|
+
sealedRecordFields?: readonly string[];
|
|
461
|
+
/**
|
|
462
|
+
* An approval of this tool on a row REPLACES the latest earlier approval of this tool on the same row (#2914): its seal
|
|
463
|
+
* names that earlier seal as the one it supersedes, as a decision answered again supersedes its own earlier seal. For a
|
|
464
|
+
* decision that is taken again on the same record after the record was reopened (a month signed off, reopened and
|
|
465
|
+
* signed off again): the two are different decisions, and without this the second seal names no predecessor, so a
|
|
466
|
+
* verifier sees two approvals of one record and nothing saying which one stands.
|
|
467
|
+
*
|
|
468
|
+
* Only the latest seal of the same source, the same record (`argumentsResource.rowIdArgument`) and the same tool is
|
|
469
|
+
* superseded, never an approval made through another tool on that row; a decision answered again still supersedes its
|
|
470
|
+
* own earlier answer first. Requires an approval gate and `argumentsResource.rowIdArgument`.
|
|
471
|
+
*/
|
|
472
|
+
approvalReplacesEarlierApprovalOfRow?: boolean;
|
|
437
473
|
/**
|
|
438
474
|
* The electronic signature the file deliverable needs once the person approved it (#2293): after the approval's seal is
|
|
439
475
|
* SEALED, the engine sends the approved file (`deliverableFileArgument`, bound by its hash) to the vendor selected for
|
|
@@ -1116,6 +1152,42 @@ export function functionToolDeliverableDeclarationProblems(
|
|
|
1116
1152
|
return problems;
|
|
1117
1153
|
}
|
|
1118
1154
|
|
|
1155
|
+
/**
|
|
1156
|
+
* What is wrong with a tool's record-approval declarations (`sealedRecordFields`, `approvalReplacesEarlierApprovalOfRow`,
|
|
1157
|
+
* #2914), one sentence per problem with its remedy; empty when they are sound or absent. `declaredFieldsOf` answers the
|
|
1158
|
+
* fields the resource's main schema declares, or undefined when the resource is not registered (its fields are then not
|
|
1159
|
+
* checked here: a field the row does not hold is shown and sealed as absent).
|
|
1160
|
+
*/
|
|
1161
|
+
export function functionToolRecordApprovalDeclarationProblems(
|
|
1162
|
+
tool: Pick<FunctionTool, 'functionRef' | 'sealedRecordFields' | 'approvalReplacesEarlierApprovalOfRow' | 'argumentsResource' | 'requiresHumanApproval' | 'humanApproval' | 'asksThePerson'>,
|
|
1163
|
+
declaredFieldsOf?: (resourceType: string) => readonly string[] | undefined,
|
|
1164
|
+
): string[] {
|
|
1165
|
+
const fields = tool.sealedRecordFields;
|
|
1166
|
+
const replaces = tool.approvalReplacesEarlierApprovalOfRow === true;
|
|
1167
|
+
if (fields === undefined && !replaces) return [];
|
|
1168
|
+
const problems: string[] = [];
|
|
1169
|
+
const named = `Function tool "${tool.functionRef}" declares ${fields !== undefined ? 'sealedRecordFields' : 'approvalReplacesEarlierApprovalOfRow'}`;
|
|
1170
|
+
if (!tool.humanApproval && !tool.requiresHumanApproval) {
|
|
1171
|
+
problems.push(`${named} and is not approval-gated: it describes what an approval covers, so gate it (humanApproval or requiresHumanApproval).`);
|
|
1172
|
+
}
|
|
1173
|
+
if (tool.asksThePerson) problems.push(`${named} and asks the person: a question approves no record.`);
|
|
1174
|
+
if (!tool.argumentsResource?.rowIdArgument) {
|
|
1175
|
+
problems.push(`${named} without argumentsResource.rowIdArgument: name the argument holding the id of the row the approval is about.`);
|
|
1176
|
+
}
|
|
1177
|
+
if (fields !== undefined) {
|
|
1178
|
+
if (!Array.isArray(fields) || fields.length === 0 || fields.some((field) => typeof field !== 'string' || field.trim().length === 0)) {
|
|
1179
|
+
problems.push(`${named}, which names no field: list the fields of the row the approval covers.`);
|
|
1180
|
+
} else {
|
|
1181
|
+
const duplicated = fields.filter((field, index) => fields.indexOf(field) !== index);
|
|
1182
|
+
if (duplicated.length > 0) problems.push(`${named}, which names ${[...new Set(duplicated)].join(', ')} more than once.`);
|
|
1183
|
+
const declared = tool.argumentsResource ? declaredFieldsOf?.(tool.argumentsResource.resourceType) : undefined;
|
|
1184
|
+
const unknown = declared ? fields.filter((field) => !declared.includes(field)) : [];
|
|
1185
|
+
if (unknown.length > 0) problems.push(`${named}, naming ${unknown.join(', ')}, which ${tool.argumentsResource!.resourceType} does not declare.`);
|
|
1186
|
+
}
|
|
1187
|
+
}
|
|
1188
|
+
return problems;
|
|
1189
|
+
}
|
|
1190
|
+
|
|
1119
1191
|
/** Whether an argument schema is a string once its optional, nullable and default wrappers are removed. */
|
|
1120
1192
|
function isStringArgumentSchema(schema: z.ZodTypeAny): boolean {
|
|
1121
1193
|
let current: z.ZodTypeAny = schema;
|
|
@@ -81,8 +81,18 @@ A screening or moderation declaration with no content-safety provider makes the
|
|
|
81
81
|
fails closed, by design, and it does so in EVERY environment that cannot serve it, so a provider and its
|
|
82
82
|
secrets authored for `local` only leave the agent refusing every turn in staging and production. The
|
|
83
83
|
backend warns at startup, one line per such agent, and `wildo local doctor` lists every authored
|
|
84
|
-
environment that cannot serve it.
|
|
85
|
-
|
|
84
|
+
environment that cannot serve it. On the development machine the classifier is usually a local double: the
|
|
85
|
+
provider's host secret names `127.0.0.1:<port>`, and `wildo local dev` starts the double on that port when the
|
|
86
|
+
application's root `package.json` declares `"@wildo-ai/provider-simulators": "*"` in its `devDependencies` (add it
|
|
87
|
+
in the same task that selects the provider). Without it nothing answers there, and every guarded turn is withheld
|
|
88
|
+
locally as `OUTPUT_NOT_MODERATED`. Two registry providers serve it: `azure-content-safety` screens and moderates
|
|
89
|
+
(its own Azure resource), `openai-moderation` moderates only, free, with an OpenAI key. An environment that is meant to
|
|
90
|
+
run without one (a staging stack with no key for it, while production has one) does not drop the declaration: it
|
|
91
|
+
states `agentContentGuardrails: { withoutClassifier: RUN_UNGUARDED, reason }` in its `wildo.infra.<env>.config.ts`, and
|
|
92
|
+
the agent runs unguarded there, named at startup. Never author that for an environment because a provider's setup is
|
|
93
|
+
pending; it is the operator's decision, and the infra-environment-config skill describes it. Until the application
|
|
94
|
+
selects a provider, declare `personalData`, and write one line at the agent saying which of the other two it will
|
|
95
|
+
declare and why. A guardrail that cannot treat
|
|
86
96
|
anything (an agent with no tools and no conversation, fed only text your code composes) is not declared at
|
|
87
97
|
all: a declaration that does nothing reads as a control. The contract and every refusal are in the
|
|
88
98
|
`flows-actors-chat-assistant` skill, `references/framework-agent-content-guardrails-api.md`.
|
|
@@ -115,8 +115,18 @@ A screening or moderation declaration with no content-safety provider makes the
|
|
|
115
115
|
fails closed, by design, and it does so in EVERY environment that cannot serve it, so a provider and its
|
|
116
116
|
secrets authored for `local` only leave the agent refusing every turn in staging and production. The
|
|
117
117
|
backend warns at startup, one line per such agent, and `wildo local doctor` lists every authored
|
|
118
|
-
environment that cannot serve it.
|
|
119
|
-
|
|
118
|
+
environment that cannot serve it. On the development machine the classifier is usually a local double: the
|
|
119
|
+
provider's host secret names `127.0.0.1:<port>`, and `wildo local dev` starts the double on that port when the
|
|
120
|
+
application's root `package.json` declares `"@wildo-ai/provider-simulators": "*"` in its `devDependencies` (add it
|
|
121
|
+
in the same task that selects the provider). Without it nothing answers there, and every guarded turn is withheld
|
|
122
|
+
locally as `OUTPUT_NOT_MODERATED`. Two registry providers serve it: `azure-content-safety` screens and moderates
|
|
123
|
+
(its own Azure resource), `openai-moderation` moderates only, free, with an OpenAI key. An environment that is meant to
|
|
124
|
+
run without one (a staging stack with no key for it, while production has one) does not drop the declaration: it
|
|
125
|
+
states `agentContentGuardrails: { withoutClassifier: RUN_UNGUARDED, reason }` in its `wildo.infra.<env>.config.ts`, and
|
|
126
|
+
the agent runs unguarded there, named at startup. Never author that for an environment because a provider's setup is
|
|
127
|
+
pending; it is the operator's decision, and the infra-environment-config skill describes it. Until the application
|
|
128
|
+
selects a provider, declare `personalData`, and write one line at the agent saying which of the other two it will
|
|
129
|
+
declare and why. A guardrail that cannot treat
|
|
120
130
|
anything (an agent with no tools and no conversation, fed only text your code composes) is not declared at
|
|
121
131
|
all: a declaration that does nothing reads as a control. The contract and every refusal are in the
|
|
122
132
|
`flows-actors-chat-assistant` skill, `references/framework-agent-content-guardrails-api.md`.
|
|
@@ -230,7 +230,7 @@ once you have a candidate.
|
|
|
230
230
|
- **App** (11) — `AppLayout_ConsentGate`, `AppLayout_DocumentAcceptanceGate`, `AppLayout_ErrorBoundary`, `AppLayout_Home`, `AppLayout_MainContainer`, `AppLayout_RouterErrorBoundary`, `Files_UploadGrantDrop`, `PublicLayout_LegalNotice`, `PublicLayout_MainContainer`, `PublicLayout_StaticContentContainer`, `PublicLayout_Terms`
|
|
231
231
|
- **Auth** (15) — `Auth_AuthCallback`, `Auth_ChatIdentityLink`, `Auth_ConfirmEmail`, `Auth_EnrollmentRequired`, `Auth_Invitation`, `Auth_MagicLinkVerify`, `Auth_Main`, `Auth_MfaChallenge`, `Auth_OAuthConsent`, `Auth_OAuthLoggedOut`, `Auth_OAuthLogoutConfirm`, `Auth_OneOffSession`, `Auth_OrganizationOnboarding`, `Auth_PasswordLost`, `Auth_Register`
|
|
232
232
|
- **Billing** (6) — `BillingPortalLink`, `InvoiceHistory`, `PlanChangePreview`, `PricingTable`, `SubscriptionStatus`, `UsageDashboard`
|
|
233
|
-
- **Data Display** (
|
|
233
|
+
- **Data Display** (11) — `Avatar`, `Badge`, `DescriptionDetails`, `DescriptionList`, `DescriptionTerm`, `Duration`, `Identifier`, `Kbd`, `NotificationIndicator`, `RecordApprovalSeal`, `Timeline`
|
|
234
234
|
- **Data Display / Formatters** (25) — `DisplayFieldAddress`, `DisplayFieldBigDecimal`, `DisplayFieldBoolean`, `DisplayFieldCode`, `DisplayFieldDate`, `DisplayFieldDateTime`, `DisplayFieldDuration`, `DisplayFieldEnum`, `DisplayFieldFile`, `DisplayFieldFileAudio`, `DisplayFieldFileDocument`, `DisplayFieldFileGallery`, `DisplayFieldFileImage`, `DisplayFieldFileVideo`, `DisplayFieldMoney`, `DisplayFieldNumber`, `DisplayFieldPhone`, `DisplayFieldRangeBigDecimal`, `DisplayFieldRangeDate`, `DisplayFieldRangeDateTime`, `DisplayFieldRangeMoney`, `DisplayFieldRangeNumber`, `DisplayFieldRangeTime`, `DisplayFieldString`, `DisplayFieldTime`
|
|
235
235
|
- **Feedback** (10) — `AppCelebrationOverlay`, `DangerZone`, `EmptyState`, `ErrorDisplay`, `LoadingOverlay`, `LoadingSpinner`, `OperationsConfirmCritical`, `Progress`, `Skeleton`, `Toaster`
|
|
236
236
|
- **Feedback / Features** (4) — `FeatureLimitIndicator`, `FeatureLockBadge`, `FeatureUnavailableDialog`, `FeatureUnavailablePrompt`
|
package/dist/knowledge/skills/frontend-component-catalog/references/framework-component-catalog.md
CHANGED
|
@@ -162,7 +162,7 @@ an application reaches those through configuration, never through JSX.
|
|
|
162
162
|
| `SubscriptionStatus` | Unavailable slices render an explicit unavailable state instead of empty or zero data; independently available billing panels remain usable. Read-only status panel with optional actions to manage billing. | Use on account settings and billing overview routes. | Avoid for users without an active billing subject. |
|
|
163
163
|
| `UsageDashboard` | Unavailable slices render an explicit unavailable state instead of empty or zero data; usage and credit balance availability are independent, so either section can remain usable when the other fails. Dashboard cards/charts fed by billing usage APIs. | Use when plans include metered entitlements or quota warnings. | Avoid when billing is flat-rate without usage signals. |
|
|
164
164
|
|
|
165
|
-
## Data Display (
|
|
165
|
+
## Data Display (11)
|
|
166
166
|
|
|
167
167
|
| Component | Renders | Use when | Avoid when |
|
|
168
168
|
| --- | --- | --- | --- |
|
|
@@ -175,6 +175,7 @@ an application reaches those through configuration, never through JSX.
|
|
|
175
175
|
| `Identifier` | Monospaced, wrapping, receding behind the prose it identifies; the same treatment the field pipeline gives StringDisplayMode.IDENTIFIER. | Use wherever a surface renders an address a machine assigned rather than words a person wrote, including on hand-built rows outside the field pipeline. | Avoid for prose, for a human-authored name, and for a code BLOCK (use the code artifact element, which owns scrolling and copy). |
|
|
176
176
|
| `Kbd` | A small quiet key cap in the monospaced face on the inset surface; the same cap wherever an application names a key. | Use wherever a surface tells a reader what to press: a shortcut hint on a control, a hotkey help list, a keyboard-driven list that says so. | Avoid for a machine address (Identifier) and for inline code (the CODE typography variant): a key cap depicts a control, not a value. |
|
|
177
177
|
| `NotificationIndicator` | A small count bubble, a colored signal dot, or nothing — surfaced on nav items and shell affordances. | Use to surface a notification-badge definition’s live count/dot on a sidebar item or shell control. | Avoid for arbitrary status chips (use `Badge`); this is driven by a resolved notification-badge count. |
|
|
178
|
+
| `RecordApprovalSeal` | A card on a record's READ surface. Its badge names the seal's state (sealed, pending, seal failed, superseded, approver erased, no seal); its facts name the approver from the seal itself, so an approver who has left the organization still shows; a record changed since the approval lists the changed fields by their labels, never their values; a signed seal offers a check of its signature, a download of the deliverable's manifested copy, and a link to the public check page. | Use `RecordApprovalSeal field="<the record's seal id field>"` in the DISPLAY template of a resource that stores an approval seal's id when the approved effect happens (a close storing `signOffApprovalSealId` when it is signed off). | Do not use for a record that stores no seal id: a seal is shown only where the record names it, so the component cannot find one by itself. Do not use for who-decided fields alone (`signedOffByUserId`): show those as fields. Do not use for an electronic signature request sent to a signature vendor; that is the signature request's own status. |
|
|
178
179
|
| `Timeline` | Compound: `Timeline` (the rail, an ordered list), `TimelineEntry` (tone, optional custom marker; under a `selectable` rail, `option` + `selected` make it a pickable grid row — never a listbox option, because an entry carries `TimelineActions` and an option may not contain a control), `TimelineMarker`, `TimelineTime`, `TimelineBody`, `TimelineActions`. Renders in DOM order; ordering, filtering, elapsed clocks and words are the consumer's. | Use for any stream of things that happened in order — runs, dispatches, writes, a flow execution's events, an audit trail — wherever a reader asks what happened, then what. | Avoid for a sequence a person WALKS (the stepper), for an unordered set (a list card), and for a reasoning trace with nested content (the chain-of-thought elements). |
|
|
179
180
|
|
|
180
181
|
## Data Display / Formatters (25)
|
|
@@ -189,7 +190,7 @@ an application reaches those through configuration, never through JSX.
|
|
|
189
190
|
| `DisplayFieldDateTime` | Timestamp display for point-in-time fields. | Use for audit timestamps, “last updated”, and scheduling instants. | Avoid for wall-time-only values; consider date or time slots. |
|
|
190
191
|
| `DisplayFieldDuration` | Readable duration formatter for interval-backed fields. | Use when the schema represents a time span rather than an instant. | Avoid for single timestamps; use date, datetime, or time slots. |
|
|
191
192
|
| `DisplayFieldEnum` | Read-only enum label resolver for constrained string/enum fields. | Use when displaying picklist values in resource read views or summaries. | Avoid when the raw stored token should remain visible for debugging; use the string slot instead. |
|
|
192
|
-
| `DisplayFieldFile` | File reference display cell with optional actions. | Use for arbitrary file attachments where MIME-specific previews are not required. | Avoid when a dedicated image, video, or audio preview slot is registered for the same shape. |
|
|
193
|
+
| `DisplayFieldFile` | File reference display cell with optional actions. On a record's DISPLAY surface, a single file a person approved by its hash (an approval seal's deliverable, #2913) shows that seal under it (`FileApprovalSeal`): who approved it, when, as what, its signed copy and a check; a file no seal attests shows nothing more, and a SUMMARY never asks. | Use for arbitrary file attachments where MIME-specific previews are not required. | Avoid when a dedicated image, video, or audio preview slot is registered for the same shape. |
|
|
193
194
|
| `DisplayFieldFileAudio` | Read-only audio attachment cell. | Use for podcast clips, voicemails, or other audio blobs. | Avoid when only a download link is required; the generic file slot may suffice. |
|
|
194
195
|
| `DisplayFieldFileDocument` | Read-only document attachment cell. | Use for contracts, PDFs, and office documents. | Avoid for inline text extraction views; those belong in higher-level readers. |
|
|
195
196
|
| `DisplayFieldFileGallery` | Read-only gallery layout for several image files. | Use when the field stores multiple images shown together. | Avoid for a single image; prefer the image file slot. |
|
package/dist/knowledge/skills/frontend-component-catalog/references/framework-component-usage.md
CHANGED
|
@@ -3436,6 +3436,19 @@ Dual-thumb slider for numeric intervals.
|
|
|
3436
3436
|
|
|
3437
3437
|
- Preset implementation ships from `@wildo-ai/presets-components`; runtime resolution uses the component registry in `@wildo-ai/saas-frontend-lib`.
|
|
3438
3438
|
|
|
3439
|
+
## `RecordApprovalSeal`
|
|
3440
|
+
|
|
3441
|
+
The electronic approval seal a record stores, shown where the record is read (#2913): who approved it, when, with which meaning and outcome, whether the approval still describes the record, and the approved deliverable's signed copy.
|
|
3442
|
+
|
|
3443
|
+
**Composition**
|
|
3444
|
+
|
|
3445
|
+
- Preset implementation ships from `@wildo-ai/presets-components`; runtime resolution uses the component registry in `@wildo-ai/saas-frontend-lib`.
|
|
3446
|
+
- `RecordApprovalSeal` reads the record and its resource from the display template context and asks `GET /approval-seals/:sealId/records/:resourceType/:resourceId/:field`, admitted by the record's own READ with the reader's credentials and only for the seal the field holds; it asks again when the record changes.
|
|
3447
|
+
- `ApprovalSeal` (the preset contract) is presentational: it draws the view, the moved fields' labels and the action states it is handed, and decides nothing about access.
|
|
3448
|
+
- `FileApprovalSeal fileId` (drawn by `DisplayFieldFile` on a DISPLAY surface) asks `GET /approval-seals/files/:fileId`, admitted by the READ of the record the file belongs to and only for a seal its own record still names; a 404 draws nothing, since most files are attested by nothing. No currency: that is a statement about the approved record.
|
|
3449
|
+
- `DecisionApprovalSeal decisionRef` asks `GET /approval-seals/decisions/:decisionRef`, which answers only the person who decided; it asks again while the seal is pending (signed seconds after the decision). The inbox draws it under the list for decisions answered there, and a conversation's decision card draws it in its own place once answered. It offers no copy: the deliverable is downloaded where its record is read.
|
|
3450
|
+
- Composed from engine components only: Card, Badge, DescriptionList, Alert, Text, Button, Link, Skeleton.
|
|
3451
|
+
|
|
3439
3452
|
## `ResizablePanelGroup`
|
|
3440
3453
|
|
|
3441
3454
|
Let users resize adjacent panels (IDE-style sidebars, preview panes) with persisted layout optional.
|
|
@@ -487,8 +487,8 @@ export default defineInfraEnvConfig({
|
|
|
487
487
|
* (#2991): its host slot `AZURE_CONTENT_SAFETY_HOST` holds `127.0.0.1:18309` and its key the double's own, both in
|
|
488
488
|
* `secrets.json`, never here. The engine reaches a loopback host only in a development or test environment, so this
|
|
489
489
|
* selection cannot carry the double's address anywhere else. The double is the content-safety simulator of
|
|
490
|
-
* `@wildo-ai/provider-simulators`,
|
|
491
|
-
* answer that their content could not be screened.
|
|
490
|
+
* `@wildo-ai/provider-simulators`, which `wildo local dev` starts on that port with that key; without a service
|
|
491
|
+
* answering there, the guarded agents answer that their content could not be screened or checked.
|
|
492
492
|
*/
|
|
493
493
|
[EngineCapability.AI_CONTENT_SAFETY]: {
|
|
494
494
|
primary: 'azure-content-safety',
|
|
@@ -964,6 +964,29 @@ provider.
|
|
|
964
964
|
|
|
965
965
|
Read `references/framework-platform-key-management-api.md` for the source-backed contract and the policies.
|
|
966
966
|
|
|
967
|
+
## Agents in an environment without a content classifier
|
|
968
|
+
|
|
969
|
+
An agent that declares `contentGuardrails.promptInjection` or `.outputModeration` refuses every turn in an environment
|
|
970
|
+
where no classifier for that guardrail is present: no `AI_CONTENT_SAFETY` provider is declared, the selected one does
|
|
971
|
+
not offer it (`openai-moderation` does not screen), or its credentials are not authored in that environment's
|
|
972
|
+
`secrets.json → providerEnv`. The classifier is declared once for the application, so leaving its key out of staging is
|
|
973
|
+
how staging runs without it while production has it. To run the agent there anyway, on purpose, the environment says
|
|
974
|
+
so, with why:
|
|
975
|
+
|
|
976
|
+
```ts
|
|
977
|
+
agentContentGuardrails: {
|
|
978
|
+
withoutClassifier: AgentContentGuardrailsWithoutClassifier.RUN_UNGUARDED, // from '@wildo-ai/saas-models'
|
|
979
|
+
reason: 'Staging has no Azure Content Safety key; production authors one.',
|
|
980
|
+
},
|
|
981
|
+
```
|
|
982
|
+
|
|
983
|
+
Read `references/framework-agent-content-guardrails-environment-api.md` for the contract. It waives only a guardrail no
|
|
984
|
+
classifier is present for: a classifier that is there and cannot be used (declared and not loaded, a category it does
|
|
985
|
+
not assess) still refuses the turn, and one that fails at call time (a key set and refused, an outage) still withholds
|
|
986
|
+
the output. The backend names each agent it runs unguarded at startup, `wildo local doctor` reports the environment as
|
|
987
|
+
running it unguarded, and outside a local environment `wildo context launch-blockers` lists it. The guardrails
|
|
988
|
+
themselves are the `flows-actors-chat-assistant` skill's.
|
|
989
|
+
|
|
967
990
|
## Running on another application's infrastructure
|
|
968
991
|
|
|
969
992
|
Several applications can share one environment's platform. One application OWNS the
|
|
@@ -1003,6 +1026,7 @@ Read `references/framework-infrastructure-host-api.md` for the source-backed con
|
|
|
1003
1026
|
## References
|
|
1004
1027
|
|
|
1005
1028
|
- `references/framework-infrastructure-host-api.md` — source-backed `infrastructureHost` / hosted descriptor contract
|
|
1029
|
+
- `references/framework-agent-content-guardrails-environment-api.md` — source-backed `agentContentGuardrails`: what the environment does with agents' classifier-backed guardrails when no classifier is selected
|
|
1006
1030
|
- `references/framework-platform-key-management-api.md` — source-backed `platformKeyManagement` contract, the Vault policies and the cloud KMS grants
|
|
1007
1031
|
- `references/framework-application-database-api.md` — source-backed `applicationDatabase` contract
|
|
1008
1032
|
|