fix(agent): reject stale 32k metadata for MiniMax

[luoxiao6645]

related #24140

Summary

• reject stale 32K context metadata for MiniMax models the same way we already do for Kimi
• invalidate cached 32K MiniMax entries so previously working Telegram and cron setups recover after upgrade
• add regression tests covering MiniMax cache invalidation, OpenRouter fallback rejection, and non-MiniMax safety

Root Cause

get_model_context_length() already had a targeted guard for Kimi models that were wrongly underreported as 32768 by stale third-party metadata. MiniMax models still lacked the same protection.

That left two failure paths for #24140:

• a stale persisted 32768 cache entry for MiniMax-M2.7 or related models would win before any provider fallback
• unknown-provider fallback could still accept OpenRouter metadata reporting MiniMax at 32768

Either path tripped Hermes' existing 64K minimum-context guard and caused Telegram / cron failures even though MiniMax's real context window is 204800.

Testing

• .\venv\Scripts\python -m pytest tests/agent/test_model_metadata.py tests/agent/test_minimax_provider.py

[teknium1]

Thanks for isolating the stale-MiniMax metadata path. The core premise remains present on current main: agent/model_metadata.py:2125-2178 returns a cached MiniMax 32K value, while the MiniMax fallback remains 204,800 at agent/model_metadata.py:281-287.

Problems

• Current main added a separate explicit-OpenRouter path after this PR. At agent/model_metadata.py:2357-2367, a provider="openrouter" MiniMax entry with context_length=32768 is still accepted because the guard is Kimi-only. The submitted diff changes the generic fallback but cannot cover this later-added sibling path unchanged.

Suggested changes

• During salvage, use the shared predicate in the current cache, Nous fallback, explicit OpenRouter, and unknown-provider OpenRouter paths (agent/model_metadata.py:1978-1989, 2125-2145, 2357-2367, 2390-2402).
• Add an explicit provider="openrouter" MiniMax 32K regression case alongside the submitted provider-empty fallback test.

This is an automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Two PRs address the same root cause: stale 32K cache/OpenRouter metadata is accepted for MiniMax models whose actual fallback context is 204,800 tokens. Both diffs add MiniMax detection, unify the existing Kimi/MiniMax stale-underreport predicate, apply it to three resolution paths, and add equivalent regression coverage.

Related pull requests

• #24436 related — (+69/-9) — keep open, but update before merge: the diff invalidates stale MiniMax cache entries and rejects 32K metadata in the Nous and generic unknown-provider OpenRouter paths while preserving unrelated 32K values. As identified by the contributor keep_open review, current main now has an additional explicit provider="openrouter" path that this diff does not guard; salvage should apply the shared predicate there and add the requested regression test.
• #24552 [closed] duplicate — (+73/-8) — duplicate, closed in favor of #24436: its diff implements essentially the same shared MiniMax/Kimi predicate, three guard-site changes, and regression tests, so it remains relevant as the rebased comparison that previously resolved an _safe_ctx conflict. It does not provide a distinct fix to retain separately and likewise predates the explicit-OpenRouter-path requirement identified on #24436.

Duplicates

#24436 and #24552 implement essentially the same stale-32K MiniMax metadata fix; #24552 was closed in favor of #24436.

Suggested consolidation

Merge #24436 after rebasing it onto current main, extending the shared stale-underreport predicate to the explicit provider="openrouter" path, and adding an explicit-provider MiniMax 32K regression test, as required by the contributor keep_open review. Keep #24552 closed as the duplicate superseded by the consolidated #24436 implementation.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    subgraph Dup24436 ["PRs duplicating each other"]
        P24436["PR #24436 (open)"]
        P24552["PR #24552 (closed)"]
    end
    class P24436 open
    class P24552 closed
    class P24436 target
    click P24436 "https://github.com/NousResearch/hermes-agent/pull/24436"
    click P24552 "https://github.com/NousResearch/hermes-agent/pull/24552"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed or no verify verdict yet (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 14 kB of PR diffs, 3 kB of issue/PR text, 3 kB of discussion (4 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[teknium1]

Merged via PR #85506 (merge commit 35720fb) — your commit was cherry-picked onto current main with your authorship preserved in git log. Your fix landed intact through two intervening merges to the same block (the generic pre-catalog guard from #85434 and the non-positive guard from #85507); the shared _model_name_suggests_stale_32k_underreport() predicate now covers Kimi + MiniMax at all three sites exactly as you wrote it. Thanks for covering the Nous-path _safe_ctx too, not just the step-1 cache — that's the half most fixes miss.