[Feature] AI credentials pool (instance-level named credentials)

[HCH-hash]

Summary

Introduce named, instance-level AI credentials (API keys, Claude Max tokens, Gemini keys, etc.) that can be selected when creating or configuring an agent, instead of requiring a per-agent secret every time.

Problem

Today each agent stores its own key via Agent.LLMAPIKeySecret. Teams running many agents (especially ACP agents on a shared Max subscription or shared API keys) must repeat the same credential setup, which is error-prone, hard to rotate, and duplicates secrets.

Proposed Direction

• Add a named credentials store at the instance (or workspace) level, integrated with the existing secrets store and permission model.
• At agent creation/edit time, allow selecting a credential from a dropdown (gated by a permission such as credentials.use).
• Keep per-agent overrides possible for the minority of cases that need a unique key.

Scope

• Product behavior
• Repository structure
• Architecture
• Developer experience
• Documentation
• Other

Additional Context

• Builds on the existing secrets store rather than inventing a parallel system.
• Especially valuable for ACP agents sharing a single Claude Max / Codex / Gemini subscription.
• Trade-off: slightly more complex permission surface, but removes a large amount of operational repetition.

[pullfrog]

  ^{｜ Build this ➔ ｜ Make a plan ➔}