feat(providers): ProviderProfile ABC + plugins/model-providers/ (salvage of #14424 + pluggable migration)

[teknium1]

Provider declarations become one plugin each. Adding a simple api-key provider now requires a single plugins/model-providers/<name>/ directory (2 files: __init__.py + plugin.yaml) and zero edits anywhere else. Bundled plugins can be overridden from $HERMES_HOME/plugins/model-providers/<name>/ without editing the repo.

Salvaged from @kshitijk4poor's PR #14424 onto current main, then migrated to the same plugin pattern used for IRC and Teams gateway adapters.

Commits in this PR

1. feat: provider modules — ProviderProfile ABC, 33 providers, fetch_models, transport single-path — the salvaged feat: provider modules — ProviderProfile ABC, 29 providers, fetch_models, transport single-path #14424 refactor, updated for current main
2. feat(providers): make all 33 providers pluggable under plugins/model-providers/ — moves every profile to a bundled plugin dir, adds user-override path, auto-coerces user-installed plugins to kind: model-provider
3. docs(providers): add model-provider-plugin authoring guide + fix stale refs — new /docs/developer-guide/model-provider-plugin page (authoring guide: layout, fields, hooks, user overrides, testing, pip distribution); stale comment cleanup in main.py/doctor.py/auth.py/models.py; updates to AGENTS.md, user-guide/features/plugins.md (plugin types: 3→4), adding-providers.md + provider-runtime.md cross-references

What changes for plugin authors

Before:

providers/myprovider.py         # one file in repo, required editing the tree

After:

plugins/model-providers/myprovider/
├── __init__.py                 # calls register_provider(ProviderProfile(...))
└── plugin.yaml                 # name, kind: model-provider, version, description

Or for a user override (no repo edit):

~/.hermes/plugins/model-providers/myprovider/
├── __init__.py
└── plugin.yaml

User plugins win over bundled plugins on name collision (last-writer-wins in register_provider()).

Changes (commit 2 of 2)

• plugins/model-providers/ — new category, 28 bundled plugin dirs (covering 33 profiles — minimax, kimi, opencode, gemini each ship two profiles in one plugin)
• providers/__init__.py — _discover_providers() now scans bundled + user plugin dirs first, then falls back to legacy providers/*.py
• providers/ — now only base.py (ABC), __init__.py (registry), README.md
• hermes_cli/plugins.py — adds kind: model-provider to _VALID_PLUGIN_KINDS; general PluginManager skips imports of these (providers/ discovery owns them); heuristic auto-coerces user-installed standalone plugins whose __init__.py calls register_provider with a ProviderProfile to kind: model-provider
• skip_names extended so plugins/model-providers/ isn't double-scanned by the general manager
• 4 new tests in tests/providers/test_plugin_discovery.py — bundled discovery, user override, general-loader isolation

Docs

• plugins/model-providers/README.md — plugin contract + authoring guide
• providers/README.md — rewritten, points at plugin dir
• website/docs/developer-guide/adding-providers.md — fast path now describes plugin layout
• website/docs/developer-guide/provider-runtime.md — references both paths

Changes (commit 1 of 2 — salvage)

Original #14424 changes, same as documented before:

• providers/ — ProviderProfile ABC + registry + 33 profiles
• agent/transports/chat_completions.py — profile path via _build_kwargs_from_profile() for registered providers; legacy flag path retained for lmstudio / tencent-tokenhub
• run_agent.py — profile path integrated; legacy-path variables scoped correctly
• hermes_cli/{auth,config,doctor,main,models,runtime_provider}.py — auto-extend from list_providers()
• agent/{auxiliary_client,model_metadata}.py — auto-wire
• New 3 profiles missing from original PR: alibaba-coding-plan, azure-foundry, minimax-oauth

Salvage notes (from commit 1)

• 965 commits stale; 10 conflict points re-resolved
• agent/copilot_acp_client.py left in place (PR's relocation would have reverted 7 landed fixes)
• lmstudio / tencent-tokenhub stay on legacy flag path — reasoning-effort logic has no clean profile-hook equivalent yet
• Dropped lmstudio aliases from custom profile
• Skipped openrouter / custom from PROVIDER_REGISTRY auto-extension (resolve_provider() special-cases them)
• runtime_provider.py: profile.api_mode is a fallback, not an override (PR's version broke minimax /v1 URL)
• GeminiProfile.build_extra_body translates thinking_config (native + OpenAI-compat nested)

Validation

	Before (main)	After
tests/providers/	—	83 / 83
tests/run_agent/test_provider_parity.py	93 / 93	93 / 93
tests/hermes_cli/test_runtime_provider_resolution.py	109 / 109	109 / 109
tests/hermes_cli/test_plugin* + tests/test_plugin_skills.py	166 / 173	166 / 173
tests/hermes_cli/ (full dir)	3735 / 3744 (9 pre-existing)	3733 / 3744
tests/run_agent/ (full dir)	1224 / 1226 (2 pre-existing)	1224 / 1226
tests/agent/ (full dir)	2482 / 2483 (1 pre-existing)	2482 / 2483

All failing tests are reproduced on origin/main without this PR. Not regressions.

E2E smoke test confirms:

• 33 profiles load from plugins/model-providers/
• Auto-wired into PROVIDER_REGISTRY, CANONICAL_PROVIDERS, OPTIONAL_ENV_VARS, doctor, _URL_TO_PROVIDER
• openrouter / custom correctly absent from PROVIDER_REGISTRY (behavior parity)
• lmstudio / tencent-tokenhub correctly fall through to legacy path
• Gemini thinking_config produces correct native + OpenAI-compat nested shapes
• User override: $HERMES_HOME/plugins/model-providers/gmi/ replaces bundled gmi profile

Closes #14418. Original PR #14424 by @kshitijk4poor — salvage preserves their authorship on commit 1 via rebase-merge.

[alt-glitch]

Supersedes #16326 (which salvaged #14424). This is the most current attempt — 33 providers (vs 29 in #16326), rebased on current main with all conflict points re-resolved. Related tracking issue: #14418.