Rework of the connect flow

[tino097]

Rework of the datasource connection experience

1. /connect becomes a single conversational prompt — no numbered menu, no "use existing / create new" branching, no LLM engine lookup
2. Custom datasources have no required fields — users can save with just a name and description, flow never aborts on empty input
3. When credentials arrive via chat (YOLO mode), the connect_new_datasource tool saves them to the vault directly before any scratchpad use

[entelligence-ai-pr-reviews]

EntelligenceAI PR Summary

Simplifies datasource connection UX and enforces optional fields policy for all custom datasources.

• connect.py: Single free-text prompt replaces numbered-menu flow; resolution order is saved slug → registry name → custom fallback
• connect.py / datasources.py: Inline vault.save + restore_namespaced_env + register_secret_vars triples replaced by new save_connection() utility
• custom.py: _force_all_fields_optional() added; required=False enforced at DatasourceField construction, LLM prompt instruction, and serialization; empty-input check changed from not user_answer.strip() to user_answer is None
• datasource_registry.py: Post-construction step sets fields.required = False for all fields (including auth method fields) on custom engines
• tests/test_datasource.py: New TestCustomDatasourceRequiredFieldsPolicy class; non-registry name routing test; cancellation test; fixture iterators updated to remove leading '0' index step
• uv.lock: dill==0.3.8 added as pinned runtime dependency

---

Confidence Score: 4/5 - Mostly Safe

Safe to merge with minor caveats — the PR cleanly consolidates a scattered vault.save + restore_namespaced_env + register_secret_vars pattern into a single save_connection() utility and simplifies the connect flow UX, both of which are clear improvements. No automated review comments were generated and no unresolved pre-existing issues are flagged. The _force_all_fields_optional() addition in custom.py and the slug-resolution order change in connect.py are the two areas worth a human glance, but neither shows evidence of a logic defect from the analysis. Coverage of 4/6 changed files means two files weren't deeply reviewed, which is the only reason this falls short of a full 5.

Key Findings:

• The new save_connection() utility correctly replaces three-step inline patterns (vault.save / restore_namespaced_env / register_secret_vars) across connect.py and datasources.py — centralizing this logic reduces future drift and is the right architectural move.
• _force_all_fields_optional() in custom.py enforcing required=False at DatasourceField construction is a policy enforcement change; if any caller previously relied on required=True validation to gate bad input, that guard is now silently removed — worth confirming no upstream validation is lost.
• The slug → registry name → custom fallback resolution order in connect.py is a behavioral change from a numbered-menu flow; regression risk is low but the happy-path and ambiguous-input paths should be manually verified since automated analysis did not exercise them.
• Two of six changed files were not covered by the automated review, so there is a small blind spot; given the PR scope is a UX/utility refactor rather than a security or data-path change, this is acceptable but noted.

Files requiring special attention

• custom.py
• connect.py
• datasources.py