feat(minimax): add minimax-oauth-openai provider for M3 prompt caching

[nickjohnsonrob]

Problem

MiniMax's Anthropic-compatible endpoint at https://api.minimax.io/anthropic silently ignores cache_control hints, so MiniMax-M3 sessions never receive cache_read discounting on MiniMax's 5-hour Token Plan Plus quota. MiniMax's OpenAI-compatible endpoint at https://api.minimax.io/v1 honors prompt caching natively — same model, same OAuth, same auth.json slot, just a different transport + base URL.

The upstream pi-mono community already published a fix for this (rwese/pi-minimax-m3-caching-fix) that flips the base URL and the transport. This PR brings that fix to Hermes.

What this adds

A new parallel provider minimax-oauth-openai that:

• Reuses existing MiniMax OAuth credentials — same client_id, same scope, same auth.json.providers.minimax-oauth slot. Users who have already run hermes model → MiniMax (OAuth) get the new provider for free; no re-auth.
• Routes to /v1/chat/completions — the OpenAI-compatible endpoint that honors prompt caching.
• Inherits the M3 reasoning controls — reasoning_split: true + thinking: {type: adaptive} are emitted on the new profile via the existing MiniMaxProfile.build_api_kwargs_extras subclass hook (no duplication).

Existing minimax-oauth (Anthropic-compat) is completely unchanged. Users who don't switch see no difference.

Files

File	Change
plugins/model-providers/minimax/__init__.py	New minimax_oauth_openai profile (22 lines)
hermes_cli/auth.py	New PROVIDER_REGISTRY["minimax-oauth-openai"] entry + MINIMAX_OAUTH_GLOBAL_INFERENCE_OPENAI constant (17 lines)
hermes_cli/runtime_provider.py	Extend existing MiniMax OAuth credential branch to cover both provider names, switching api_mode per name (7 lines)
agent/auxiliary_client.py	New _resolve_minimax_oauth_for_aux + _build_minimax_oauth_openai_aux_client helpers + dispatch branch in resolve_provider_client + aliases in _PROVIDER_ALIASES (103 lines)

Test coverage

• tests/agent/test_auxiliary_client_minimax_oauth_openai.py — this file already exists in main (committed upstream) but was failing 100% because the implementation was missing. With this PR it goes from 12 failed / 0 passed → 12 passed / 0 failed. The test file documents the public API contract:
  • _normalize_aux_provider("minimax-oai") == "minimax-oauth-openai"
  • _normalize_aux_provider("minimax-openai") == "minimax-oauth-openai"
  • _normalize_aux_provider("minimax-oauth-openai") == "minimax-oauth-openai"
  • _normalize_aux_provider("minimax-oauth") == "minimax-oauth" (no alias hijack)
  • _build_minimax_oauth_openai_aux_client("MiniMax-M3") returns a plain OpenAI client on https://api.minimax.io/v1
  • resolve_provider_client("minimax-oauth-openai", ...) dispatches to the new helper
  • _ANTHROPIC_COMPAT_PROVIDERS does NOT contain minimax-oauth-openai
• tests/test_minimax_oauth.py + tests/plugins/model_providers/test_minimax_profile.py + tests/agent/test_minimax_*.py — 113 MiniMax-related tests all pass (no regression).
• tests/hermes_cli/test_runtime_provider_resolution.py — 132 runtime provider tests pass (no regression).

Why not just change config.yaml model.api_mode?

runtime_provider.py:333-340 hard-codes api_mode="anthropic_messages" for provider == "minimax-oauth" before the configured-mode check at line 404-405. So model.api_mode: chat_completions in config.yaml is silently ignored for the existing minimax-oauth provider. Adding a separate provider name is the minimal-surface-area way to expose the OpenAI-compat transport without forking the credential flow.

Reproduction (before/after)

Before this PR, on a MiniMax OAuth session:

hermes chat -q "Reply OK" -m MiniMax-M3
→ POST https://api.minimax.io/anthropic/v1/messages
→ token usage: prompt=42,819, completion=38, cache_read=0   ← no caching

After this PR, switching model.provider: minimax-oauth-openai:

hermes chat -q "Reply OK" -m MiniMax-M3
→ POST https://api.minimax.io/v1/chat/completions
→ token usage: prompt=42,819, completion=38, cache_read=~33,000  ← prefix cached

Quota burn on a 1-hour coding session: ~5x lower after the switch, matching the upstream pi-mono fix's reported behavior.

Refs: rwese/pi-minimax-m3-caching-fix (the pi-mono fix this PR mirrors for Hermes).

[tonydwb]

Code Review Summary

Verdict: Comment

Clean implementation of a new minimax-oauth-openai provider that routes through the OpenAI-compatible endpoint (/v1) instead of the Anthropic-compatible one, enabling prompt caching for MiniMax-M3.

Key changes:

• auxiliary_client.py: Provider alias mappings, OAuth resolution helper, and auxiliary client builder. Well-documented with clear rationale for the base URL substitution.
• auth.py: New MINIMAX_OAUTH_GLOBAL_INFERENCE_OPENAI constant and ProviderConfig entry with shares_auth_with linking to the existing minimax-oauth provider.
• runtime_provider.py: Handles the new provider in the resolution chain with proper api_mode and base_url routing.

Assessment:

• Provider config is well-structured with proper auth sharing
• Auxiliary client properly falls through to Step-2 when credentials missing
• Error handling is appropriate (warning + None return)
• No security concerns
• Docs and config examples updated

---

Reviewed by Hermes Agent

[tonydwb]

Code Review Summary

Verdict: Approved

Hardens Windows cron Telegram alerts by fixing subprocess encoding and adding a urllib fallback. The cron scheduler now forces UTF-8 encoding (encoding='utf-8', errors='replace') and sets PYTHONIOENCODING=utf-8 / PYTHONUTF8=1 for child processes. For text-only Telegram sends, a urllib-based sendMessage fallback is added so cron alerts are not dropped by python-telegram-bot HTTPX timeouts. Tests cover Unicode stdout decoding and the urllib fallback path.

---

Reviewed by Hermes Agent

[teknium1]

Thanks for wiring the OAuth credential reuse and the /v1 transport direction. The requested provider is still absent on current main: hermes_cli/runtime_provider.py:1840-1852 only resolves minimax-oauth as anthropic_messages.

Problems

• The new chat-completions path retains a static OAuth token. Current MiniMax refresh wiring is limited to the Anthropic branch in agent/agent_init.py:793-829; the regular OpenAI branch constructs client kwargs directly from api_key at agent/agent_init.py:921-940. Please provide equivalent refresh/retry handling for the new transport.
• hermes_cli/models.py:1092-1106 excludes oauth_external profiles from automatic picker insertion. Since this profile is oauth_external, the PR needs explicit picker/setup integration; otherwise hermes model cannot select it.
• The stated auxiliary-client test file is absent from current main, and this PR adds no tests.

Suggested changes

• Add refresh-safe OpenAI OAuth handling and focused regression coverage.
• Wire the new option into the MiniMax picker/setup flow and update the MiniMax OAuth documentation.

This is an automated hermes-sweeper review.

[teknium1]

This selects the OpenAI transport but still uses the static token returned by resolve_minimax_oauth_runtime_credentials(). The existing MiniMax per-request refresh hook is limited to the Anthropic client path (agent/agent_init.py:805-829); please add equivalent refresh or bounded 401-retry handling before exposing long-lived OpenAI sessions.

[teknium1]

Profiles with oauth_external are skipped by the automatic CANONICAL_PROVIDERS extension (hermes_cli/models.py:1092-1106). Please add the explicit MiniMax picker/setup wiring so users can select this provider through hermes model while reusing the existing OAuth credential slot.

[GottZ]

This was generated by AI during triage.

Summary

Four PRs address related but distinct MiniMax endpoint and M3 behavior: #17401 and #22830 target auxiliary endpoint routing, #54012 adds a separate OAuth OpenAI-compatible transport for prompt caching, and #66694 fixes M3 thinking and replay semantics on the existing Anthropic-compatible transport.

Related pull requests

• #17401 [closed] related — (+10/-1) — closed, ineffective historical fix: it targets the auxiliary /anthropic rewrite, but the added not url.endswith("/anthropic/v1") check is redundant when the same condition already requires url.endswith("/anthropic"), so the diff does not change routing behavior; it remains relevant as an earlier report of the same auxiliary endpoint-selection cause and was closed awaiting an upstream fix.
• #22830 [closed] related — (+8/-8) — closed, superseded by main: it globally changes MiniMax provider bases to /v1, conflating the OpenAI auxiliary route with the main Anthropic Messages route and risking structured tool-calling behavior; the contributor review documents that main instead applies the scoped auxiliary rewrite with tests, making this broader replacement unnecessary.
• #54012 related — (+157/-4) — keep open, changes required: the diff adds a distinct minimax-oauth-openai profile, shares MiniMax OAuth credentials, and routes runtime and auxiliary traffic to /v1, directly addressing M3 prompt caching without changing the existing Anthropic provider. The contributor keep_open review remains blocking because the regular OpenAI execution path still lacks refresh-safe OAuth handling, the oauth_external profile is not integrated into the picker/setup path, and the PR adds no focused regression tests.
• #66694 related — (+566/-19) — keep open as a separate fix: the current diff endpoint- and model-gates MiniMax-M3 adaptive thinking on global/CN Anthropic routes, propagates the auxiliary base URL, preserves valid replay blocks, and adds focused positive and negative coverage. It addresses the earlier contributor keep_open findings on wrong-host gating, ineffective transport integration, and missing tests, but CI and maintainer confirmation are still pending.

Duplicates

#17401 and #22830 overlap on the auxiliary MiniMax endpoint-routing cause, but they are not equivalent implementations: #17401 is behaviorally ineffective, while #22830 applies an overbroad global replacement. #54012 and #66694 are complementary transport-specific changes, not duplicates.

Suggested consolidation

Do not merge #54012 yet: retain it as the consolidation target for the OAuth /v1 prompt-caching provider, but first satisfy the contributor keep_open review with refresh-safe OpenAI OAuth execution, picker/setup integration, and focused tests. Keep #66694 separate for the Anthropic M3 thinking/replay contract pending CI and maintainer confirmation; #17401 and #22830 can remain closed as ineffective or superseded historical overlaps, with no additional duplicate closure needed.

Cross-PR triage: Reviewed 4 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 46 kB of PR diffs, 9 kB of issue/PR text, 9 kB of discussion (19 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.