…search#17333)

…ousResearch#17332)

Pre-fix, the gate in _anthropic_prompt_cache_policy enabled caching on
api_mode=anthropic_messages only when the model name contained 'claude'.
That worked for third-party gateways serving Claude (LiteLLM proxy,
Zhipu GLM Anthropic-compat) but silently disabled caching for providers
that use the native Anthropic protocol for their *own* model families
(MiniMax-M2.x). PR NousResearch#17333 fixes the symptom by hardcoding 'minimax' /
'minimax-cn' / two specific hostnames.

The issue's 'Scope beyond MiniMax' section explicitly asks for a
'capability-based or provider-allowlist approach' that's future-proof.
This PR delivers exactly that:

1. ProviderConfig.extra carries an opt-in flag
   {anthropic_cache: True, anthropic_cache_hosts: (\u2026)}
   for built-in providers that document cache_control support for
   their own models. Currently set on 'minimax' and 'minimax-cn'.
   Future entrants flip the flag in one place.

2. agent/anthropic_cache_capability.py exposes
   provider_supports_anthropic_cache(provider, base_url,
   user_configured_hosts=\u2026)
   which resolves in this order:
     a. ProviderConfig.extra['anthropic_cache'] for the provider id
        (case-insensitive).
     b. Hostname match against the union of registry-declared hosts
        and operator-configured hosts \u2014 catches provider=='custom'
        setups pointing at MiniMax's URL.

3. New user config knob agent.anthropic_cache_hosts: [\u2026]. Operators
   can opt-in their own private gateway without waiting for an
   upstream patch. Read once in __init__ into
   self._anthropic_cache_user_hosts_cached, threaded through the
   policy lookup. Malformed values are ignored \u2014 a typo never
   *disables* caching, only fails to *enable* it.

4. _anthropic_prompt_cache_policy gains a new branch that fires only
   when api_mode=='anthropic_messages' and the helper returns True.
   Pre-existing native Anthropic / OpenRouter / Claude-on-third-party
   branches are unchanged, so this is purely additive for the
   non-Claude case.

Why this is better than NousResearch#17333
- One brand-name special-case per provider isn't going to age well.
  Capability registry + config opt-in handle MiniMax now and the
  next entrant for free.
- Operator escape hatch (anthropic_cache_hosts) lets users running
  private LiteLLM / vLLM-anthropic deployments enable caching today
  without forking the repo.
- Eliminates the pre-existing 'Claude substring matches non-claude'
  gotcha by encouraging callers to test against curated host/provider
  lists, not free-text model names.
- Single source of truth: built-in MiniMax support reads from the same
  ProviderConfig that already carries auth/base_url \u2014 no parallel
  list to keep in sync.

Tests
- TestCapabilityRegistryAnthropicCache (9 cases): MiniMax via provider
  id, MiniMax via host (custom config), MiniMax-CN both ways, user-
  configured opt-in, case/whitespace normalization, transport gate,
  unlisted gateway stays off, case-insensitive provider id.
- TestCapabilityHelperUnit (5 cases): direct tests of the helper.
- Existing test test_third_party_without_claude_name_does_not_cache
  was using api.minimax.io as the 'unknown' host \u2014 updated to use a
  truly unknown host to reflect the new behavior, which is exactly
  what the issue calls for.

863 passed, 7 skipped on tests/run_agent/ (full sweep, no regressions).
30 passed on the policy file (15 pre-existing + 14 new + 1 updated).

…base_url

PR NousResearch#17425 (merged) enabled prompt caching for MiniMax models on the
anthropic_messages transport, but users still had to manually configure
both api_mode and base_url to actually benefit from it.

This patch makes the defaults ergonomic:

- AIAgent.__init__ now auto-detects provider=minimax / minimax-cn and
  defaults to api_mode=anthropic_messages + the correct /anthropic base_url
  (global or China endpoint respectively).
- .env.example suggests the /anthropic endpoints instead of /v1.
- Explicit base_url or api_mode are preserved when the user sets them.

Tests: 5 new cases covering both providers, explicit overrides, and
prompt-caching flags.

Refs: NousResearch#17332, NousResearch#17333, NousResearch#17425

…base_url

PR NousResearch#17425 (merged) enabled prompt caching for MiniMax models on the
anthropic_messages transport, but users still had to manually configure
both api_mode and base_url to actually benefit from it.

This patch makes the defaults ergonomic:

- AIAgent.__init__ now auto-detects provider=minimax / minimax-cn and
  defaults to api_mode=anthropic_messages + the correct /anthropic base_url
  (global or China endpoint respectively).
- .env.example suggests the /anthropic endpoints instead of /v1.
- Explicit base_url or api_mode are preserved when the user sets them.

Tests: 5 new cases covering both providers, explicit overrides, and
prompt-caching flags.

Refs: NousResearch#17332, NousResearch#17333, NousResearch#17425

…base_url

PR NousResearch#17425 (merged) enabled prompt caching for MiniMax models on the
anthropic_messages transport, but users still had to manually configure
both api_mode and base_url to actually benefit from it.

This patch makes the defaults ergonomic:

- AIAgent.__init__ now auto-detects provider=minimax / minimax-cn and
  defaults to api_mode=anthropic_messages + the correct /anthropic base_url
  (global or China endpoint respectively).
- .env.example suggests the /anthropic endpoints instead of /v1.
- Explicit base_url or api_mode are preserved when the user sets them.

Tests: 5 new cases covering both providers, explicit overrides, and
prompt-caching flags.

Refs: NousResearch#17332, NousResearch#17333, NousResearch#17425