Repository navigation
Correct native hosted routing documentation for Hermes, Deep Agents and Pi #13341
Description
Parent epic: #13338
Documentation defect
Public documentation still describes Hermes, Deep Agents and Pi custom endpoints as always using the global shared inference.local route. PR #13201 introduced native hosted custom provider bindings. Those unconditional claims now misdescribe endpoint selection and credential configuration.
The post-merge documentation workflow, run 38866879815, rejected these claims in review job 116861682770 and report artifact 12031454922. The affected pages are byte-identical between merged commit bdf2172e47538adf94ce5e8b07aeea6010a58dfd and inspected main snapshot 486ff49e9fe58cedd959c1454cb25c875291480c.
Owning pages
- Set up an Anthropic-compatible endpoint: Hermes and Deep Agents sections prescribe shared-route addresses and environment names without accounting for native hosted selection.
- How inference routing works: the custom endpoint and agent descriptions retain shared-route generalizations.
- Pi quickstart: the custom/local routing sentence is too broad.
- Pi support reference: the support matrix and credential explanation repeat that generalization.
Minimal plan and delivery owner
Use the existing managed documentation catch-up PR #13308, currently draft at e3043763b4636543d88ceb9042834f3f489497d5, as the delivery owner or coordinate with its successor. Its inspected routing patch does not resolve these native custom claims, and the other three pages are absent from that patch. Do not create a competing catch-up branch or edit the workflow-owned branch while it is being generated.
- Trace the documented variants to the current native provider selection, agent config and credential owners.
- Replace universal shared-route statements with the actual hosted-native versus retained shared/local behavior for each agent. Do not generalize all local inference as shared; qualifying DCode local inference already has a native route.
- Correct examples, matrix entries and credential explanations together. Keep the agent-specific Anthropic/OpenAI API compatibility rules intact.
- Run the documented docs build/link checks and obtain an independent documentation review before publication.
Acceptance criteria
- Each affected agent's hosted custom endpoint and credential behavior agrees with the source and accepted product scope.
- Legacy/shared-route examples remain correct for the modes that use them.
- Host credential custody, canonical sandbox placeholders, API restrictions and Pi activation/support status remain explicit.
- The accepted durable private endpoint binding policy is preserved: removing an ambient allowlist entry is not documented as revocation of an already verified binding.
- The three accepted live combination gaps remain limitations. Documentation does not claim successful Hermes native Anthropic-to-OpenAI, fresh custom OpenAI onboarding, or Pi credential-rewrite real-request qualification without evidence.
- Generated variants and public pages are consistent, and the independent review no longer reports the stale routing claims.
This issue needs documentation validation, not a new cloud run or a product/runtime change. It records a confirmed content defect, not a new integration proposal.
Activity
docs(inference): correct hosted custom routing guidance (NVIDIA#13609)