$ Credential Flow · AI Notes
How AI was used to build this piece, then the real AI feature living inside it.
## BUILD_PROCESS
Credential Flow was built with Claude Code pairing throughout — scaffolding the FastAPI token client and its in-memory caching logic, drafting the honeypot/rate-limiting/profanity-blocklist write protections shared with this site’s contact form, and iterating on the Relationship View’s live-sync behavior against the real Salesforce org rather than mocked responses.
## RECOMMENDED_NEXT_ACTION
The real AI feature: a short, AI-assisted “recommended next action” suggestion for a given Loan Application, generated from its current Status and how long it has been in that status, by calling the Anthropic API directly. An application sitting in Under Review for several days with no change gets a suggestion that references that duration, not a generic status-only message — and if the Anthropic call fails or times out, the page shows a clean error state for the recommendation without breaking anything else.
Why not Agentforce:Salesforce’s own agentic AI layer isn’t licensed on this free Developer Edition org. Calling Anthropic directly is the honest workaround, not a feature downgrade dressed up — and it proves the same direct-API-integration skill Farpost Dispatch proves from the opposite direction (Apex calling out to Anthropic, this piece’s Python calling out to Salesforce data).