fix: probe live Nous inference API for models missing from curated pi…

[Masalale]

What does this PR do?

When you open the /model picker and select Nous Portal, certain
models that the API actually serves were missing from the list.
For example, deepseek/deepseek-v4-flash:free (a free-tier model)
was nowhere to be found even though the API happily serves it.

Why? The model list came from a static catalog — a file shipped
with Hermes and updated periodically. Any model that didn't make it
into that file was invisible.

The fix: When building the model list, we now also probe the
live Nous inference API (GET /v1/models) for extra models.
The static catalog still comes first (so nothing changes for models
already in it), and any new models found via the live API are merged
in via _merge_model_lists(). Nous Hermes foundation models
(e.g. nousresearch/hermes-4) are filtered out — they're served
on the Portal but aren't useful inference-picker options.

Caching: The live result is cached for 2 minutes (_NOUS_LIVE_CACHE_TTL = 120)
so the picker stays fast on repeated opens. Cache-hit path is O(1) —
no redundant iteration. On network failure the cache is preserved
(stale data is better than no data).

Test coverage: Two new tests cover the success path (merge logic
and cache reuse) alongside the existing network-failure fallback tests.

Related Issue

Fixes a gap where free-tier variants like deepseek/deepseek-v4-flash:free
were missing from the /model picker even though they work perfectly
on the API.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✅ Tests

Changes Made

• hermes_cli/models.py:
  • _is_nous_hermes_model() — filter for non-agentic Nous Hermes models
  • _merge_model_lists() — merges additionals into base list, deduping
  • _probe_nous_live_api() — live probe with 120s TTL cache
  • get_curated_nous_model_ids() now delegates to _probe_nous_live_api()
• tests/hermes_cli/test_model_catalog.py:
  • Updated existing tests to mock urlopen (network-free)
  • TestNousLiveProbe — success-path tests for merge + cache

How to Test

1. hermes model → select Nous Portal
2. Free-tier variants like deepseek/deepseek-v4-flash:free should appear
3. Select one and verify inference succeeds
4. Free-tier users: paid models should still be grayed out with upgrade link

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix
• I've run python -m pytest tests/hermes_cli/test_model_catalog.py tests/hermes_cli/test_models.py -q and all tests pass (30 + 76 = 106)
• I've added tests for my changes
• I've tested on my platform: WSL2 (Fedora)

Documentation & Housekeeping

• I've updated relevant documentation — N/A, the fix is transparent to users

Technical notes for reviewers

• _resolve_nous_pricing_credentials() is used to get the API base URL;
  it's safe — returns a default on auth failure (so unauthenticated users
  still get the probe).
• The cache is (list[str], float) stored as a module global with
  time.monotonic() for the timestamp. On miss the fresh result replaces it.
• Network-failure path: returns cached data if available, else curated-only list.
• The Hermes filter (_is_nous_hermes_model) matches hermes-3 and
  hermes-4 patterns — these are Nous Research's own foundation models that
  Hermes itself can't use as an inference provider.

[hclsys]

The feature is genuinely useful — merging live free-tier variants (deepseek/deepseek-v4-flash:free) that the static catalog misses, and the implementation details are careful (normalizes the /v1 double-path, skips Hermes-branded dupes, non-destructive merge, graceful try/except on network failure). But there's a latency regression worth addressing before merge:

get_curated_nous_model_ids() was previously pure/offline (remote manifest → in-repo snapshot fallback). This adds an uncached urlopen(..., timeout=5) live probe that fires on every call. Its callers are interactive paths:

• hermes_cli/auth.py:7546 (model setup)
• hermes_cli/main.py:3082 (model listing/picker)
• hermes_cli/model_switch.py:1169 (the /model switcher)

So opening the model picker now blocks up to 5s on a network round-trip (and the full 5s every time the endpoint is slow/unreachable), where before it was instant. That's a UX regression on a hot path.

Two clean options: (1) cache the live result module-level with a short TTL — there's already an _openrouter_catalog_cache pattern right above (~models.py:1195) you could mirror; or (2) gate the live probe behind a flag / make it opt-in so the default picker path stays offline-fast and only an explicit 'refresh models' action hits the network. Either keeps the new free-tier discovery without making every picker open pay the round-trip. The merge logic itself is sound.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Extends get_curated_nous_model_ids() to probe the live Nous /v1/models endpoint and merge any additional model IDs into the curated list, with corresponding test updates to mock out network calls.

Changes:

• Adds live API probe in get_curated_nous_model_ids to discover models not in the static catalog, filtering out non-agentic Hermes variants.
• Normalizes base URL to avoid duplicate /v1 in path.
• Updates existing tests to patch urllib.request.urlopen so the new network call doesn't reach out during tests.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
hermes_cli/models.py	Merges live /v1/models results into curated Nous model list.
tests/hermes_cli/test_model_catalog.py	Patches urlopen in existing tests to isolate from network.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Masalale]

@hclsys good catch — caching is now in place.

Extracted the live probe into _probe_nous_live_api() with a module-level 120s TTL cache (_nous_live_cache), mirroring the _openrouter_catalog_cache pattern you pointed out at ~models.py:1195.

Repeated picker opens: 2.7s first call → 0.0033s cached (~800x). On network failure the cache is not evicted (stale data is better than no data). On cache hit, we re-merge with the current curated base so a concurrent catalog update does not get masked until TTL expiry.

The commit was force-pushed to the same branch.

[teknium1]

Thanks for addressing stale Nous model discovery and for incorporating the cache feedback.

Problems

• Current main intentionally does not put the full Nous inference /v1/models dump into the interactive picker. hermes_cli/model_switch.py:1872-1883 documents that it is a large alphabetical vendor catalog and instead uses curated models plus Portal free/paid recommendations. This change would reintroduce the behavior that path deliberately avoids.
• The proposed filter only excludes Hermes 3/4 name patterns, while current inference fallback code excludes all Hermes-branded IDs at hermes_cli/auth.py:5202-5205.

Suggested changes

• Please re-scope around a demonstrated failure in the existing Portal recommendation flow, while preserving the curated-plus-recommendations picker contract in hermes_cli/model_switch.py:1872-1908.
• If live inference fallback remains necessary, build on fetch_nous_models() rather than a parallel raw probe.

Automated hermes-sweeper review.