fix(agent): hide Anthropic-only assistant surfaces where the provider cannot run them (#2204) (#2343)

* fix(agent): hide Anthropic-only assistant surfaces where the provider cannot run them

Self-hosted deployments on an OpenAI-compatible provider (or with no AI
configured) still showed every entry point into the tool-loop runtime
behind /api/agent/invoke, which answers 503 there. The capability lived
server-side only (getAiStatus().assistantAvailable); no UI could read it.

Hand the flag to the client through CompanyContext (useAssistantAvailable,
beside the paid-capability gate) and gate each entry point that opens
AgentChat: the bookkeeping page's "Skapa med assistent" and "Med
assistenten", the inbox workspace's "Fråga assistenten" doors, /chat/intake
and /chat/new?intent=. The floating trigger falls back to general help (the
single-call console runs on any provider) instead of hiding, and AgentChat
itself never fires an invoke without the runtime, so a resumed thread or a
forgotten entry point shows a notice instead of a 503. The Hem checklist's
"Anslut till Claude" step renders only where the assistant runs on Claude
and the mcp-server extension is on. Provider-agnostic AI (ask console,
categorization, extraction) and the server-side 503 are unchanged; on
hosted the flag is true and nothing changes.

Closes #2204

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SaJfqNi4VmsG8FMKq99G6

* test(ai): use a placeholder that cannot match an Anthropic key shape

The new direct-Anthropic status test assigned a string in the exact
format of a live API key, which trips secret scanners on every run. The
config only reads presence, so any non-empty string exercises the path.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

---------

Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
Jakob Wennberg
2026-09-06 18:54:47 +02:00
committed by GitHub
co-authored by Claude Fable 5.1 Jakob Wennberg
parent 272d19b287
commit 9cb1d105e3
16 changed files with 220 additions and 30 deletions
+23
View File
@@ -49,6 +49,16 @@ interface CompanyContextValue {
isSandbox: boolean
/** PAID capability keys the active company currently holds (entitled + enabled). */
capabilities: CapabilityKey[]
/**
* Whether this deployment can run the in-app assistant, i.e. the tool-loop
* runtime behind /api/agent/invoke: getAiStatus().assistantAvailable, false
* without AI credentials or on an OpenAI-compatible endpoint (#2204).
* Deployment-level, not per company. Gates ONLY the surfaces that open
* that runtime; provider-agnostic AI (the single-call ask console,
* categorization, document extraction) never reads it. The route's 503
* stays the real enforcement.
*/
assistantAvailable: boolean
/**
* Trial expiry while the trial is the company's only source of paid access;
* null when paying/comped or after the trial lapsed. Drives the countdown
@@ -111,3 +121,16 @@ export function useCapability(key: CapabilityKey): boolean {
if (!ctx) return true
return ctx.capabilities.includes(key)
}
/**
* Whether the deployment can run the in-app assistant (see
* CompanyContextValue.assistantAvailable). Read it at every entry point that
* opens AgentChat / /api/agent/invoke; the single-call console (general.help)
* runs on any provider and does not need it. Fail-open outside a
* CompanyProvider for the same reason as useCapability: the server answers 503.
*/
export function useAssistantAvailable(): boolean {
const ctx = useContext(CompanyContext)
if (!ctx) return true
return ctx.assistantAvailable
}