feat(config): render web.backend/search_backend/extract_backend as Config Form selects (#71929)

[PRATHAMESH75]

What & why

Fixes #71929.

The Config page's Form view (/config → Web section) rendered web.backend, web.search_backend, and web.extract_backend as bare text <input>s. Users had to know the exact backend string (e.g. brave-free, not brave) with no hint of valid values and no validation — a silent-failure trap, while sibling fields (terminal.backend, tts.provider, display.skin, …) have long been dropdowns.

Change

Add three _SCHEMA_OVERRIDES entries in hermes_cli/web_server.py turning all three fields into selects, following the existing terminal.backend pattern. The schema feeds every surface that reads /api/config/schema (desktop, CLI, dashboard), so all three get the dropdown.

Option lists are mirrored from the runtime, not hand-copied from the issue:

• web.backend / web.search_backend → the built-in _LEGACY_WEB_BACKENDS set in tools/web_tools.py (firecrawl, searxng, brave-free, ddgs, tavily, exa, parallel, xai). A test asserts the options equal that set so the two never drift.
• web.extract_backend → only the extract-capable subset (firecrawl, tavily, exa, parallel). Search-only providers (searxng, brave-free, ddgs, xai) are rejected at runtime with a "search-only backend" error, so they're excluded here.
• A leading "" keeps the shared-backend / auto-detect fallback selectable.

Plugin-backend safety

Web backends are plugin-extensible (agent/web_search_registry), and the dashboard renders a select as a closed gate (unlike the free-input voice/model fields). A naive static select would silently drop a configured plugin/custom backend and let the next save clobber it. To avoid that regression, _schema_with_dynamic_provider_options() now preserves a configured value that falls outside the built-in list — the same current-value preservation already used for tts/stt/memory providers. The common case (built-in or blank) returns the frozen import-time entry untouched.

Tests

tests/hermes_cli/test_web_server.py:

• the three fields render as selects with a leading "";
• search options equal _LEGACY_WEB_BACKENDS (drift guard);
• extract options are exactly the extract-capable subset and exclude search-only providers;
• a configured plugin/custom backend outside the list stays selectable via the dynamic merge, and the module-level schema is copied not mutated;
• a built-in/blank value produces no overlay (identity with the frozen schema).

uv run pytest tests/hermes_cli/test_web_server.py — all schema tests green.

---

Note: the local footgun preflight flags 10 pre-existing bare write_text() calls elsewhere in test_web_server.py; none are in this change, and the CI Windows footguns check scans production packages only (not tests/), so it passes clean. Those belong to the separate #71014 read_text/write_text campaign, not this PR.

[teknium1]

Thanks for addressing a verified Config Form gap: current main still infers the three web.*_backend fields as strings through hermes_cli/web_server.py:1044-1083.

Problems

• hermes_cli/web_server.py:1245-1248 preserves only an already-configured unknown provider. Web providers are registered and enumerable through agent/web_search_registry.py:78-82, while the Form select renders only supplied options (web/src/components/AutoField.tsx:116-128). A newly installed, unconfigured plugin therefore cannot be selected from this UI.
• tests/hermes_cli/test_web_server.py:5327 and :5339 freeze exact provider sets. AGENTS.md:1333-1380 requires behavior/invariant tests rather than change-detector snapshots.

Suggested changes

• Populate per-capability options from the web registry, filter by supports_search() / supports_extract(), and separately retain stale configured values. Cover an unconfigured custom provider.
• Replace exact-set assertions with capability-based invariants.

Automated hermes-sweeper review.

[teknium1]

This only retains a custom backend after it is already configured. agent.web_search_registry.list_providers() exposes installed providers, but a closed select cannot choose an unconfigured one unless it is added here. Enumerate registry providers per capability and preserve configured-but-undiscoverable names separately.

[teknium1]

Please avoid freezing this current provider enumeration. Test the relation to provider capability flags instead, including a registered custom provider; AGENTS.md prohibits change-detector snapshots for evolving provider lists.

[PRATHAMESH75]

Both points addressed in daca098a6.

1. Enumerate the registry, don't just retain the configured value. You're right — the old helper only preserved a configured backend, so a closed select could never reach an installed-but-unconfigured provider. _web_backend_schema_options now enumerates agent.web_search_registry.list_providers() per capability via a new _registry_web_backend_names(capability):

• web.backend → "any" (supports_search or supports_extract)
• web.search_backend → "search" (supports_search)
• web.extract_backend → "extract" (supports_extract)

Order is base built-ins → newly-discovered registry names → the configured-but-undiscoverable fallback (still preserved so switching away never drops it). The registry read is wrapped fail-safe: any import/registry error returns [] (built-in base stays the floor), and a single provider that raises on introspection is skipped, not fatal.

2. Test capability flags, not a frozen list. Replaced the snapshot approach with relation tests: test_dynamic_merge_enumerates_registry_web_backends_by_capability registers a search-only and an extract-capable fake provider and asserts the search-only reaches web.backend/web.search_backend but is kept out of web.extract_backend, while the extract-capable reaches web.backend/web.extract_backend. test_registry_web_backend_names_filters_by_capability drives the real helper end-to-end against a temporarily-register_provider'd backend (restored in finally). The no-overlay test now stubs enumeration to [] to isolate the "nothing new registered" invariant deterministically.

The two static tests (_web_backend_options_match_runtime_backends, _web_extract_backend_omits_search_only) I left as-is: they guard the import-time base list against tools/web_tools._LEGACY_WEB_BACKENDS drift — a real invariant, not a provider-enumeration snapshot, and unaffected by the dynamic registry merge. Full file green (121 passed), ruff clean.

[GottZ]

This was generated by AI during triage.

Summary

One PR addresses issue #71929. #71958 converts all three reported web backend fields from unrestricted text inputs to selects, adds capability-based discovery of installed registry providers, and preserves configured-but-undiscoverable values, but its tests still include the exact provider-set assertions challenged in review.

Related pull requests

• feat(config): render web.backend/search_backend/extract_backend as Config Form selects (#71929) #71958 best fix — (+289/-0) — best existing fix, review not fully resolved: The diff directly fixes the three free-text fields and now enumerates installed providers by search/extract capability, including unconfigured custom providers. Consistent with the keep_open review on feat(config): render web.backend/search_backend/extract_backend as Config Form selects (#71929) #71958, the PR should remain open because tests/hermes_cli/test_web_server.py still freezes exact built-in provider sets despite the contributor's request for capability-based invariants.

Suggested consolidation

Keep #71958 open with a salvage path: retain the select overrides, capability-filtered registry enumeration, custom-provider discovery, and configured-value fallback, but replace the remaining exact built-in provider-set assertions with behavior/invariant coverage and obtain review clearance. There are no duplicate PRs to close.

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
    I71929(["issue #71929 (open)"])
    P71958["PR #71958 (open)"]
    P71958 -->|best fix| I71929
    class I71929 open
    class P71958 open
    class P71958 best
    class P71958 target
    click I71929 "https://github.com/NousResearch/hermes-agent/issues/71929"
    click P71958 "https://github.com/NousResearch/hermes-agent/pull/71958"

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 (state tag in the node label).

Cross-PR triage: Reviewed 1 pull request and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 16 kB of PR diffs, 5 kB of issue/PR text, 5 kB of discussion (5 comments), 3 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.