refactor: simplify provider detection and resolution by emilsvennesson · Pull Request #21 · emilsvennesson/opencode-websearch

emilsvennesson · 2026-04-29T19:36:48Z

Summary

Replaces the multi-strategy provider detection logic and the deeply-layered
resolution helpers with a single canonical-ID lookup and two small functions.
Net −549 lines; bundle size drops from 37.18 KB → 31.91 KB. No behavior
change for any supported provider.

Why

Provider detection layered three independent strategies (npm-package matching,
ID-substring hints, per-model uniformity heuristics) plus five fallback
helpers. OpenCode already exposes canonical, well-known provider IDs
(anthropic, openai, github-copilot, moonshotai, moonshotai-cn) via
provider.id and the chat.message hook's model.providerID. A flat ID map
covers every supported case and makes the rest of the pipeline obvious.

The resolution side had its own bloat: a 6-helper chain for what is a 3-step
priority lookup, a ProviderState wrapper around a single field, a
hasAnyProvider guard that listed every provider type by hand, and three
full provider scans per cold path.

What changed

Detection

src/providers/registry.ts: replaced detectProviderTypeFromNpm,
detectProviderTypeFromProviderID, detectProviderTypeFromModel, hint
arrays, and helpers with a flat PROVIDER_TYPES_BY_ID map and a single
detectProviderType lookup. Added moonshotai-cn.
model.api.npm is no longer consulted anywhere.
Custom-renamed providers (e.g. my-anthropic-proxy) are intentionally not
auto-detected; users must use the canonical provider ID. Custom baseURL
on canonical providers continues to work.

Resolution (`src/index.ts`)

Replaced 6 helpers (resolveSearchProvider → resolveLockedModel/
resolveActiveModel/resolveFallbackModel → resolveModelByPriority +
buildSearchConfig) with findModelByKey + pickModel.
Moved the openai → chatgpt redirect into resolveActiveType, so
resolution no longer needs a special branch.
Replaced hasAnyProvider (23 lines) with
Object.keys(resolutions).length === NO_PROVIDERS.
Inlined hasConfiguredOpenAIBaseURL + shouldAttachChatGPTResolution
(17 lines) as resolutions.openai?.credentials.baseURL?.trim().
Dropped ProviderState wrapper and ResolvedProvider interface.
Replaced inline structural typing of client with PluginInput["client"].

Scanning (`src/config.ts`)

scanProviders now runs once per cold path; loadResolutions threads
the scan state through chatgpt/copilot attachment instead of calling
resolveModelOverrides (deleted) twice more.
4-way repetition (createInitialScanState, buildResolutionMap,
ScanState) collapsed via a SCANNABLE_TYPES array.
Dropped isScannableProviderType guard (detectProviderType returns
ScannableProviderType | null directly).
resolveGeneralModelHint/resolveCopilotModelHint → constants.

Auth (`src/providers/{chatgpt,copilot}/auth.ts`)

Extracted shared readAuthEntry<Entry>(client, directory, key),
PathClient, and path helpers into src/providers/shared/auth.ts.
Each adapter's auth.ts is now ~50–75 lines focused on the entry
shape and buildCredentials. Removes ~170 lines of duplication.

Types (`src/types.ts`)

Dropped providerType field from ProviderResolution (redundant with
the map key).
Dropped SearchArgs interface; adapters now take query: string
directly. Updated all 5 adapter signatures and dispatchSearch.
Moved ScannableProviderType from config.ts to types.ts.
ProviderResolutionMap is now Partial<Record<ProviderType, ...>>.

Lint

Enabled ternaries (no-ternary: off); updated AGENTS.md accordingly.

File-by-file impact

File	Before	After	Δ
`src/index.ts`	416	214	−202
`src/config.ts`	363	264	−99
`src/providers/registry.ts`	112	45	−67
`src/providers/chatgpt/auth.ts`	134	53	−81
`src/providers/copilot/auth.ts`	157	76	−81
`src/types.ts`	104	94	−10
`src/providers/shared/auth.ts`	(new)	82	+82
Bundle (`dist/index.js`)	37.18 KB	31.91 KB	−14%

Behavior preserved

All five supported provider types: anthropic, openai, github-copilot,
moonshotai (now also moonshotai-cn).
Custom baseURL on canonical provider IDs (provider-level options and
per-model api.url), including Anthropic /v1$ stripping.
ChatGPT OAuth detection (only when openai has no configured baseURL).
Copilot enterprise URL via OAuth enterpriseUrl.
websearch: "always" / "auto" flags.
Locked → active → fallback resolution priority.

Behavior intentionally removed

Auto-detection of custom-renamed provider IDs (e.g.
my-anthropic-proxy configured with npm: "@ai-sdk/anthropic"). Users
must use the canonical OpenCode provider ID.

Verification

bun run check  # format:check + lint + typecheck → all pass
bun run build  # ESM bundle 31.91 KB + .d.ts → success

No tests exist yet.

Detect providers by canonical OpenCode ID instead of layering npm-package matching, ID-substring hints, and per-model uniformity heuristics. The provider type now comes from a single flat ID-to-type lookup, with moonshotai-cn added alongside moonshotai. Collapses the 6-helper resolution chain to two functions, moves the ChatGPT-vs-OpenAI routing into detection, drops the redundant ProviderResolution.providerType field and ProviderState wrapper, scans providers once per cold path instead of three times, and extracts the duplicated chatgpt/copilot auth.json reading into a shared helper. Net -549 lines, no behavior change for any supported provider.

Copilot

Pull request overview

This PR refactors the web-search plugin’s provider detection and model resolution pipeline to rely on canonical OpenCode provider IDs and a simplified scan/resolve flow, while also de-duplicating shared auth.json reading logic.

Changes:

Replaced multi-strategy provider detection with a canonical provider-id → adapter-type map (detectProviderType) and consolidated scan state (scanProviders).
Simplified model selection/resolution logic to a small set of functions (resolveActiveType, findModelByKey, pickModel) and reduced repeated provider scanning.
Extracted shared OpenCode auth.json reading into src/providers/shared/auth.ts and updated provider adapters to accept query: string directly.

Reviewed changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/index.ts	Simplifies resolution loading, active-provider routing (incl. OpenAI→ChatGPT), and model selection.
src/config.ts	Collapses provider scanning and resolution map building; removes npm-based detection inputs.
src/providers/registry.ts	Introduces canonical provider-id lookup map and exports scan/type priority constants.
src/types.ts	Refactors shared types (adds `ScannableProviderType`, removes `SearchArgs`, simplifies resolution map typing).
src/providers/shared/auth.ts	New shared helper to locate and read OpenCode `auth.json` entries.
src/providers/chatgpt/auth.ts	Uses shared auth reader; trims and simplifies OAuth credential extraction.
src/providers/copilot/auth.ts	Uses shared auth reader; refactors enterprise URL normalization and credential extraction.
src/providers/index.ts	Updates adapter interface to accept `query: string` and streamlines dispatch functions.
src/providers/anthropic/index.ts	Updates adapter signature to `executeSearch(config, query)`.
src/providers/openai/index.ts	Updates adapter signature to `executeSearch(config, query)`.
src/providers/moonshot/index.ts	Updates adapter signature to `executeSearch(config, query)`.
src/providers/chatgpt/index.ts	Updates adapter signature to `executeSearch(config, query)`.
src/providers/copilot/index.ts	Updates adapter signature to `executeSearch(config, query)`.
AGENTS.md	Updates style guidance to allow ternaries in simple cases.
.oxlintrc.json	Disables `no-ternary` to match updated style guidance.

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

+  const copilot = await resolveCopilotCredentials(client, directory);
+  if (copilot) {
    resolutions.copilot = {
-      credentials: copilotCredentials,
-      fallbackModel: modelOverrides.fallbackModel,
-      lockedModel: modelOverrides.lockedModel,
-      providerType: "copilot",
+      credentials: copilot,
+      fallbackModel: scan.copilot.fallbackModel,
+      lockedModel: scan.copilot.lockedModel,
    };


+ * Provider types that are scannable from OpenCode provider config,
+ * in the order they should be preferred when multiple providers offer
+ * a `lockedModel` or `fallbackModel`.


+const PROVIDER_TYPES_BY_ID: Record<string, ScannableProviderType> = {
+  anthropic: "anthropic",
+  "github-copilot": "copilot",
+  moonshotai: "moonshot",
+  "moonshotai-cn": "moonshot",
+  openai: "openai",
+};


Switch from a type-keyed resolution map to a list of resolutions keyed by OpenCode provider ID. This lets users have multiple providers of the same type (e.g. `openai-prod` and `openai-staging`) with their own credentials and baseURLs, all auto-detected via npm-package matching on `@ai-sdk/openai`, `@ai-sdk/anthropic`, or `@ai-sdk/github-copilot`. Detection now falls back to npm matching when the provider ID is not canonical. Active model routing matches by provider ID, so each renamed provider serves traffic on its own credentials. ChatGPT OAuth shadowing is scoped to the canonical `openai` provider with no baseURL; renamed openai-typed providers keep their explicit credentials.

…feedback Allow scanProvider to emit credential-less scan results when a provider exposes websearch flags but no apiKey, so an OAuth-only github-copilot config that pins a model via `"websearch": "always"` is no longer silently dropped. The OAuth attachment phase fills in credentials with a unified baseURL guard that also preserves any explicit Copilot proxy configuration. Use the canonical `moonshotai` provider ID in the formatNoProviderError example block.

Copilot AI review requested due to automatic review settings April 29, 2026 19:36

Copilot started reviewing on behalf of emilsvennesson April 29, 2026 19:37 View session

Copilot AI reviewed Apr 29, 2026

View reviewed changes

emilsvennesson added 2 commits April 29, 2026 22:06

emilsvennesson merged commit fb85188 into main Apr 29, 2026
1 check passed

emilsvennesson deleted the refactor/simplify-provider-detection-and-resolution branch April 29, 2026 20:29

Navigation Menu

refactor: simplify provider detection and resolution - #21

refactor: simplify provider detection and resolution#21
emilsvennesson merged 3 commits into
mainfrom
refactor/simplify-provider-detection-and-resolution

emilsvennesson commented Apr 29, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

emilsvennesson commented Apr 29, 2026

Summary

Why

What changed

Detection

Resolution (src/index.ts)

Scanning (src/config.ts)

Auth (src/providers/{chatgpt,copilot}/auth.ts)

Types (src/types.ts)

Lint

File-by-file impact

Behavior preserved

Behavior intentionally removed

Verification

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Resolution (`src/index.ts`)

Scanning (`src/config.ts`)

Auth (`src/providers/{chatgpt,copilot}/auth.ts`)

Types (`src/types.ts`)