The backend OpenAICompatibilityAPIKey struct only has api-key and
proxy-url (internal/config/config.go:570-577). PUT /openai-compatibility
unmarshals into this struct (config_lists.go:439) and discards any
unknown field, so per-entry headers submitted from the UI never persist
and never reach the runtime request path. The GET response wrapper
also has no headers field (config_auth_index.go:31-34).
Drop the per-entry headers input from BaseProviderForm and the
ApiKeyEntry/ApiKeyEntryInput types. Connectivity test and model
discovery stop reading entry-level headers — they continue to apply
provider-level headers, which is the supported contract.
Backend's POST /auth-files and DELETE /auth-files single-item paths
return only {status:"ok"}, with no uploaded/deleted/files
(auth_files.go:680 and :794). Multi-item paths return the full payload.
85c8b34 simplified the normalizer assuming the full payload was always
present, which caused single-file uploads and single-item deletes to be
read as "0 succeeded" — the upload page skipped its success toast and
list refresh, and batch delete reported "(0)" with no row removal.
Re-introduce a narrow fallback: when failed is empty and the count
field is omitted, derive uploaded/deleted and files from requestedNames.
The fallback only kicks in for the documented single-item shape, not
for the partial/failure paths.
All bool-tagged config fields on the backend are Go bools and serialize
as JSON true/false (e.g. Debug, RequestLog, WebsocketAuth, Disabled,
Websockets, ForceModelPrefix). normalizeBoolean is only called against
those response paths, so the number / string / Boolean(value) fallbacks
never fire.
Backend batch upload/delete handlers always return a complete payload:
status (string), uploaded/deleted (number), files (string array), and
failed (only on partial). See auth_files.go:702-711. The fallback chains
that re-derived uploaded/deleted counts and file names from the
requestedNames list were unreachable.
Likewise, runtime_only is always a real bool from Go (auth_files.go:403),
and entry.modified is never emitted, so isRuntimeOnlyEntry collapses
to a strict equality check and readDateField drops the modified alias.
The backend api-call response schema (api_tools.go:54-56) fixes the
field names to status_code / header / body. camelCase statusCode and
plural headers are never emitted.
Backend list endpoints (config_lists.go: GetGeminiKeys / GetClaudeKeys /
GetCodexKeys / GetOpenAICompat / GetVertexCompatKeys) always wrap the
result as {"<kebab-section>": [...]}, never as a raw array, never under
"items" or "data". And /config emits the same kebab keys, so the camelCase
aliases (geminiApiKey, openAICompatibility, ...) are unreachable.
Remove RAW_SECTION_ALIASES and inline the lookup; drop the raw-array
and items/data fallbacks from extractArrayPayload.
The backend emits provider config with kebab-case JSON tags only
(internal/config/config.go: ClaudeKey/CodexKey/GeminiKey/OpenAICompatibility,
internal/api/handlers/management/config_auth_index.go for auth-index).
The camelCase / snake_case entries in PROVIDER_KEY_FIELDS,
OPENAI_PROVIDER_FIELDS, MODEL_ALIAS_FIELDS, API_KEY_ENTRY_FIELDS,
CLOAK_FIELDS, RESPONSE_ONLY_FIELDS were dead — same for the identity
helpers and the apiKeyEntries fallback in mergeOpenAIProviderPayload.
The backend always serializes config fields with kebab-case JSON tags
(internal/config/config.go). The camelCase and snake_case fallbacks in
normalizeApiKeyEntry / normalizeProviderKeyConfig / normalizeGeminiKeyConfig /
normalizeOpenAIProvider / normalizeAmpcode* / normalizeConfigResponse are
dead paths. Read kebab-case only.
The legacy openai-compatibility `api-keys` (flat string array) flat-string
branch is also gone — backend startup migration is disabled
(internal/config/config.go:657).
Prepending new API key entries shifted all existing indices, causing useConnectivityTest index-based status tracking to desync.
Fix: keep array operations as append (stable indices) and reverse the rendering order so new entries appear at the top visually. Use realIdx for status lookups, React keys, and all updateField operations.
Add eye toggle button for all API key input fields so users can verify their keys before saving.
Changes:
- BaseProviderForm: wrap single API key field (Gemini/Codex/Claude/Vertex) with passwordField + toggle button
- BaseProviderForm: wrap per-entry API key fields (OpenAI-compatible) with passwordField + toggle button
- BaseProviderForm: extract applyRawApiKey helper to populate apiKey from resource.raw in edit mode
- BaseProviderForm: sync initialFormSignature with applyRawApiKey to prevent false isDirty
- sharedForm.module.scss: add .passwordField, .passwordInput, .passwordToggle styles
- i18n: add showApiKey/hideApiKey keys to en, zh-CN, zh-TW, ru locale files
- accessibility: add aria-label and title attributes on both toggle buttons
Previously confirmDiscardIfDirty only fired for Sheet-driven close paths
(Cancel button, backdrop, escape). Clicking a different provider in the
left rail bypassed the prompt and silently dropped the form changes.
ProviderSheet now exposes the existing guard via an imperative handle,
which the workbench page calls before swapping the active brand.
- useModelDiscovery: openaiCompatibility branch now forwards
resolvedAuthIndex into modelsApi.fetchModelsViaApiCall, mirroring the
gemini/codex/claude branches. Configs that rely on a backend-resolved
auth index without a local plaintext key can now list models.
- useConnectivityTest: runOpenAIKey accepts an authIndex fallback when
entry.apiKey is empty, sets Authorization: Bearer $TOKEN$ in that
case, and forwards authIndex on the /api-call request so the gateway
can substitute the upstream token.
- useModelDiscovery: reset stale results when baseUrl / apiKey /
apiKeyEntries / headers / authIndex change, so reopening the discovery
panel after editing fields re-fetches instead of showing the previous
endpoint's models.
The new workbench only wired refreshRecentRequests into the manual refresh
button; the hook itself just set an interval (240s). As a result the
status bar showed empty/zero until the first poll. Load on mount when
enabled (cache hit avoids a network round-trip).
The actual login page lives at src/pages/LoginPage.{tsx,module.scss};
this stale src/pages/Login/Login.module.scss has had no importers since
it was committed in 450964f.
- Remove the legacy ai_providers.* block (~268 keys per locale) — only
nav.ai_providers under a different namespace is still referenced
- Remove providersPage.{disconnected,error,actions.back,
actions.openModelCatalog,ampcode.behaviorSection,detail.empty,
status.notAvailable,modelCatalog.{loading,openAction,summaryCount,
summaryError,summaryTitle}} which were never wired up after the
ProvidersWorkbench rewrite
These helpers were only consumed by the old AiProviders edit pages and
became orphaned after the providers refactor. modelInputListUtils.ts was
only used by ModelInputList, so it goes too.
- Drop empty barrel src/components/providers/index.ts (no external importers)
- Drop src/components/providers/types.ts (only referenced in a stale comment)
- Trim src/components/providers/utils.ts to the 10 exports actually used
externally; remaining helpers (ampcode mappers, key builders, base URL
normalizers, redundant recent-stats wrappers) were dead after the
ProvidersWorkbench rewrite
fix(provider): update button icon styles in ProviderHeaderCard
style(provider): enhance layout and typography in ProviderHeaderCard and ProviderResourcePanel
fix(provider): update status styles in ProviderResourceTable and sharedForm
The detail metadata section was rendering authIndex with the
status.notAvailable label (zh: "不可用"), which reads like the
provider doesn't support the feature at all. Every provider supports
authIndex; it's just absent until you link an auth file. Reuse the
status.notSet label so the wording matches baseUrl / proxyUrl and
makes it clear that the entry simply doesn't have a value yet.
The form's apiKey field is intentionally empty in edit mode (the
placeholder reads "leave blank to keep unchanged"), so model discovery
and the Claude connectivity test were sending requests with no
credentials and getting 401 for Codex / Claude / Gemini providers.
- BaseProviderForm derives a fallbackAuthIndex from resource.raw.authIndex
alongside the existing fallbackApiKey and passes both into the
discovery and connectivity hooks
- modelsApi.fetch{V1,Claude,Gemini,}ModelsViaApiCall accept an optional
authIndex and forward it on apiCallApi.request so the backend can
inject the OAuth/auth-file token for providers that store the key
externally
- useModelDiscovery threads the authIndex through to every brand
- useConnectivityTest threads authIndex through the Claude path and no
longer rejects the request as "API key required" when an authIndex is
available
Replace the free-form test-model input on the OpenAI-compatible and
Claude edit sheets with a Select whose options are derived from the
already-added models. An Auto entry (annotated with the first model
name) maps to an empty form.testModel so connectivity tests keep
falling back to the first configured model when the user hasn't made
an explicit pick. If an existing config carried a test model that
isn't in form.models, surface it as a custom option so the value is
preserved instead of silently dropped.
- BaseProviderForm and AmpcodeForm now snapshot the initial form
signature on mount and report dirty/clean state to the parent via a
new onDirtyChange callback
- ProviderSheet tracks the form's dirty flag and wraps the cancel
button plus the new Sheet.confirmClose hook with a confirmation
prompt; the prompt is skipped during submit and in detail mode
- Sheet gains an optional confirmClose hook so that Escape, overlay
click and the close button all route through the same async check
before starting the close animation
- Add i18n keys (en/zh-CN/zh-TW/ru) for the unsaved-changes prompt
When the OpenAI-compatible endpoint rejects the first authenticated
/models request (some upstreams expose the route unauthenticated, or
the configured key only covers chat completions), retry once with no
key and no custom headers before surfacing the original error so the
discovery panel can populate.
ModelDiscoveryPanel now hands the full ModelInfo objects to onApply
instead of just the model names, and BaseProviderForm copies each
incoming alias into the new form.models row so suggestions from
endpoints that return alias metadata are no longer silently dropped.
- Wire useProviderRecentRequests into ProvidersWorkbenchPage and pass
usageByProvider down to the resource table; the header refresh now
refreshes recent requests in parallel with the provider snapshot
- ProviderResourceTable status column hosts the existing status badge
plus an inline ProviderStatusBar driven by the recent-requests usage;
column widths rebalanced so the bar fits on a single row
- New OpenAIBrandToolbar gives the OpenAI-compatible brand a sort
control (name / priority / recent-success with direction) and a
multi-select model filter dropdown; ProviderResourcePanel renders it
via the new openaiControls prop
- ProvidersWorkbenchPage computes the available model union, sorts /
filters resources when the OpenAI brand is active, and resets the
selected models when switching brands
- Add providerStatusBar.module.scss for the in-row status bar styling
and i18n keys (en/zh-CN/zh-TW/ru) for the new toolbar
- New useModelDiscovery hook fans out to modelsApi.fetchGemini /
fetchV1 / fetchClaude / fetchModels based on brand and surfaces
loading, error and the normalized model list
- New ModelDiscoveryPanel renders an inline search, select-all and
per-model checkboxes with an Already added marker for entries that
already exist in form.models
- BaseProviderForm: show a Fetch from endpoint button at the top of
the models section for the supported brands, with apply merging
the picked names into form.models while skipping duplicates and
replacing the empty placeholder row
- Add i18n keys (en/zh-CN/zh-TW/ru) for the discovery toolbar, list
states and apply count
Track per-entry signatures (apiKey + proxyUrl + headersText) so editing
one OpenAI key entry no longer clears the success/error state of the
other entries. The global signature (baseUrl/headers/models/testModel)
keeps clearing every entry as before.
- Introduce useConnectivityTest hook that drives per-key and bulk tests
via apiCallApi against the chat-completions / messages endpoints, with
signature-based status reset
- OpenAI form: each API key entry header now shows a status icon plus a
per-key Test button, and the entries section gains a Test all button
- Claude form: enable the Test model field and add a Test button with
inline status; falls back to the persisted apiKey when the edit form
field is left blank
- Add ConnectivityStatusIcon, related styles (spin keyframes, buttons,
error banner), and i18n keys (en/zh-CN/zh-TW/ru) for the test labels
and validation messages
- Wire useProviderWorkbench.toggleDisabled through ProviderResourcePanel
to ProviderResourceTable as an inline ToggleSwitch in the actions cell
- Hide the toggle for the Ampcode singleton brand
- Add toast on success/failure and i18n keys (en/zh-CN/zh-TW/ru) for the
enable/disable labels and a toggleFailed error message
- Deleted AiProvidersVertexEditPage and its associated routes.
- Consolidated AI provider routes under ProvidersWorkbenchPage.
- Removed Claude and OpenAI draft stores as they are no longer in use.
- Updated theme styles to include new workbench tokens for UI consistency.
LTbinglingfeng
gwdgithubnom
Luis PaterandGitHub
Dylan