fix(providers): restore the runtime-agnostic exports the browser barrels dropped

[sroussey]

Stacked on #717 (claude/wonderful-turing-rjtcnx-ai-types), not on main.

#717's 24 manifest edits are correct and are not touched here. What they do is turn an
invisible gap into a hard error: once a customConditions: ["browser"] consumer is
type-checked against the browser declarations, every symbol a browser barrel forgot to
re-export becomes TS2305. Four providers have such omissions. This PR closes them, removes
the one class of nominal browser entry that has no split at all, and adds the guards that
would have caught both.

Why this is stacked on #717 rather than a follow-up

Shipping them apart means one released version in which assertToolChoiceHonored,
DeepSeek_ModelSearch_Stream, registerOpenAiImageValidator, OpenAI_ModelSearch_Stream,
Xai_ModelSearch_Stream and OpenRouter's whole capability/model-search surface are
uncompilable in the browser with no substitute — #717 makes the routing correct, and the
barrels are what the correct routing then points at. In-repo consumers already break:
packages/test/src/test/ai-provider-api/DeepSeek_ToolCalling.test.ts:9-11,
OpenAI_ImageValidation.test.ts:8, and ai-provider/provider-model-search.test.ts:10-16
import these symbols from @workglow/<p>/ai.

M1(b) — the omitted exports, per provider

The evidence these omissions are accidental is stronger than "no node: imports" (also true —
grep 'from "node:"' over all five srctrees returns zero): **every omitted module is already compiled into that provider's browser bundle via the runtime entry.** E.g.providers/openai/src/ai/registerOpenAi.ts:10importsregisterOpenAiImageValidator, and registerOpenAiIS inindex.browser.ts— so it is indist/ai.browser.jstoday and only itsexport *` was missing.

provider	module restored	symbols
deepseek	common/DeepSeek_Client (widened from 2 names to the node barrel's 5)	assertNotTruncatedByReasoning, DEEPSEEK_DEFAULT_REASONING_ALLOWANCE, resolveMaxTokens (plus the 2 already there)
deepseek	common/DeepSeek_ModelSearch	DeepSeek_ModelSearch_Stream
deepseek	common/DeepSeek_ToolCalling (named form)	DeepSeekToolChoiceNotHonoredError, assertToolChoiceHonored, isForcingToolChoice
openai	common/OpenAI_ImageValidation	registerOpenAiImageValidator
openai	common/OpenAI_ModelSearch	OpenAI_ModelSearch_Stream
xai	common/Xai_ModelSearch	Xai_ModelSearch_Stream
openrouter	common/OpenRouter_Capabilities	OPENROUTER_RUN_FN_SPECS, openRouterWorkerRunFnSpecs, deriveCapabilitiesFromMeta, inferOpenRouterCapabilities
openrouter	common/OpenRouter_ModelSearch	OpenRouterRawModel, OPENROUTER_FALLBACK_MODELS, fetchOpenRouterModels, mapOpenRouterModels, OpenRouter_ModelSearch_Stream

DeepSeek_ToolCalling keeps the named form the node barrel uses rather than a export *.
That is what holds DeepSeek_ToolCalling_Stream out of the main-thread barrel on both
platforms — matching index.ts exactly.

_testOnly is deliberately not added anywhere. It is @internal, exists for
@workglow/test alone, and belongs behind a ./test entry the way packages/ai already did
it. After this PR it is the only symbol a browser consumer loses.

Two claims from the original review, refuted — please don't "fix" these

1. ollama has no gap. The "every *_ModelSearch is missing from the browser barrel" claim
   does not hold here: providers/ollama/src/ai/common/Ollama_ModelSearch.ts exists but is
   exported by neither barrel (it is consumed internally by Ollama_JobRunFns /
   Ollama_JobRunFns.browser, which register Ollama_ModelSearch_Stream as a run-fn). The two
   barrels already agree; ollama's only delta is _testOnly. No change made.
   providers/ollama/src/ai-runtime is also untouched — its runtime.ts vs
   runtime.browser.ts split is genuine (import("ollama") vs import("ollama/browser")),
   cactus-class.

2. The nine packages/workglow shims are correctly identical. They contain
   export * from "@workglow/openai/ai" — a bare specifier, re-resolved under the
   consumer's conditions at every hop, so the split really happens one layer down in the
   provider's exports map. The two files being identical is the mechanism. Collapsing the
   browser shim onto export * from "./openai" would resolve the bare specifier once, under
   the shim's own condition, and destroy the split. Left alone, and the new guard has a
   dedicated negative fixture documenting this.

M2(i) — the four genuinely nominal browser entries

providers/llamacpp-server and providers/stable-diffusion-server each had a
src/ai.browser.ts byte-identical to src/ai.ts and a src/ai-runtime.browser.ts
byte-identical to src/ai-runtime.ts, all four naming relative specifiers
(./ai/index, ./ai/runtime). A relative specifier is resolved once, by the importing file's
own path — --target=browser changes the compile target, not the resolver, and a manifest's
browser field applies to bare specifiers only. So both entries already pulled in the same
graph and the declaration split was nominal: two .d.ts files kept equal by hand.

Each .browser.ts now re-exports its node peer. Cycle safety was checked: under
moduleResolution: "bundler" a file (./ai.ts) beats a directory (./ai/index.ts), bun
agrees, ai.ts re-exports ./ai/index and never ./ai.browser, and nothing in the toolchain
(examples/web/vite.config.ts, bunfig.toml, bun's --target=browser) substitutes
X.browser.ts for X.ts on a relative specifier. Each file keeps a one-line comment saying a
future .browser.ts resolver plugin would break this.

The browser condition is kept for both packages — --target=browser still produces a
genuinely different bundle, so the entry earns its keep. Confirmed by build: all four bundles
are byte-identical before and after (see table below).

Bundle-size cost — measured, and the plan's prediction was wrong in an interesting way

bun run build:packages at the base commit vs. after, dist/ai.browser.js:

provider	before	after	Δ raw	gzip before	gzip after	Δ gzip
deepseek	7,842	13,028	+5,186	2,680	4,155	+1,475
openai	10,300	12,941	+2,641	3,400	4,054	+654
xai	6,319	8,427	+2,108	2,256	2,803	+547
openrouter	8,786	11,358	+2,572	2,694	3,333	+639
llamacpp-server	28,270	28,270	0	5,855	5,839	−16
stable-diffusion-server	15,368	15,368	0	3,560	3,536	−24

Corrections to what the plan predicted:

• The SDK was already there. grep -o 'import("[^"]*")' finds import("openai") in the
  baseline ai.browser.js of all four providers, reached through registerDeepSeek /
  registerOpenAi / registerXai / registerOpenRouter. getClient is not pulled in by
  this change — it was already in every browser bundle.
• No code-splitting happens, and none is needed. bun emits no lazy chunk (ls dist/*.js
  is unchanged: ai.js, ai.browser.js, ai-runtime.js, ai-runtime.browser.js and nothing
  else). Under --packages=external the await import("openai") stays a literal dynamic
  import of a bare specifier, which the downstream bundler defers — so the SDK never
  contributes eager weight either way.
• The whole delta is the restored modules' own source, and it is eager. Isolated by
  rebuilding openai with only OpenAI_ImageValidation added: 10,300 → 10,331 (+31 bytes,
  effectively the plan's "literal zero"); adding OpenAI_ModelSearch on top accounts for the
  remaining +2,610.
• OpenRouter is not free. The plan predicted "adds nothing (uses global fetch, no SDK)".
  It adds +2,572 raw / +639 gzip — its own capability-inference and model-mapping code,
  including the OPENROUTER_FALLBACK_MODELS literal. Correct, but not zero.

Worst case is deepseek at +1.5 KB gzipped, for symbols a browser consumer currently cannot
import at all.

M2(ii) + M1(b)'s regression net

• duplicateBrowserEntryViolations (ExportTypesPairing.test.ts, beside
  browserSplitViolations): reports a src/*.browser.ts identical to its src/<stem>.ts
  peer whose specifiers are all relative. The relative/bare distinction is the entire rule
  and is documented in full in the doc comment, including why the packages/workglow shims
  are correctly identical. No violation survives after M2(i), so both branches carry fixtures:
  a positive pair (export * from "./ai/index" on both → one violation), a negative pair
  (export * from "@workglow/openai/ai" on both → zero violations, commented with the
  refutation), a differing-graph pair, and one proving a comment cannot silence the rule.

• packages/test/src/test/util/ExportBarrelParity.test.ts (new): parses each provider's
  src/ai/index.ts + src/ai/index.browser.ts into a surface set and asserts
  node \ browser equals a pinned INTENTIONAL_NODE_ONLY fixture (each of the five providers
  → ["_testOnly"]), with a comment explaining why _testOnly is node-only and pointing at
  the ./test migration. Includes a staleness check (a pin that no longer describes a real
  difference fails), mirroring ALLOWED_MISMATCHES. Source-only — no build, works under
  use-source. The parser is regex-based, so an unclassifiable statement is reported, not
  skipped, matching how buildEntryViolations reports an underivable dist stem.

• LOW, findBrowserBlock: browserSplitViolations did a top-level ?.browser lookup
  while its sibling nodeImportTarget recurses, so a nested
  { import: { browser: {…}, default: … } } was invisible — every rule keyed on the branch
  behaved as if no browser condition existed. Replaced with a walker; fixture added for the
  no-implementation case reached that way.

• SKIPPED, as planned: hardening buildEntryViolations' regex against glob/variable-driven
  entry lists. Fixing it properly means executing the build, and it is already fenced by
  GLOB_BUILT_PACKAGES.

Verification

$ bun scripts/test.ts util vitest
 Test Files  55 passed (55)
      Tests  810 passed | 10 skipped (820)
   Duration  177.72s
exit=0

$ bun scripts/test.ts provider-api provider vitest
 Test Files  3 failed | 87 passed | 3 skipped (93)
      Tests  10 failed | 1021 passed | 61 skipped (1093)

All 10 failures are live OpenAI integration tests
(OpenAI_Generic.integration.test.ts, OpenAI_ImageGeneration.integration.test.ts,
OpenAI_UsageAccounting.integration.test.ts) failing on billing, unrelated to this change:

Error: OpenAI Responses 429: {"message":"You have no credits remaining. ...",
  "type":"insufficient_quota","code":"credit_balance_exhausted"}

The three consumer tests this PR unblocks pass:

$ bunx vitest run --project test src/test/ai-provider-api/DeepSeek_ToolCalling.test.ts \
    src/test/ai-provider-api/OpenAI_ImageValidation.test.ts
 Test Files  2 passed (2)      Tests  11 passed (11)

$ bunx vitest run --project test src/test/ai-provider/provider-model-search.test.ts
 Test Files  1 passed (1)      Tests  10 passed (10)

Lint (the repo has no root lint script; each package carries one) — clean, no output, on
every changed package:

$ for p in providers/{deepseek,openai,xai,openrouter,llamacpp-server,stable-diffusion-server} packages/test; do (cd $p && bun run lint); done
$ eslint . --ext ts,tsx --report-unused-disable-directives --max-warnings 0   # x7, all silent

Build:

$ bun run build:packages
 Tasks:    81 successful, 81 total
  Time:    49.415s
exit=0

The guards actually guard

Deleting the one added Xai_ModelSearch line:

AssertionError: expected [ Array(1) ] to deeply equal []
+ [
+   "providers/xai: providers/xai/src/ai/index.ts exports * from \"./common/Xai_ModelSearch\"
+    but providers/xai/src/ai/index.browser.ts does not",
+ ]

Restoring providers/llamacpp-server/src/ai.browser.ts to its duplicated form:

FAIL  workspace exports maps > re-exports rather than duplicates a browser entry that has no split
+ [
+   "providers/llamacpp-server/src/ai.browser.ts is identical to providers/llamacpp-server/src/ai.ts
+    and names only relative specifiers (\"./ai/index\"), so both entries resolve the same module
+    graph and the declaration split is nominal — re-export the node entry from it instead of
+    duplicating it",
+ ]

Both restored afterwards.

tsgo probe — #717's own methodology, reproduced

A throwaway file (not committed) importing the two symbols, compiled twice under
moduleResolution: "bundler", once with customConditions: ["browser"] and once without:

import { assertToolChoiceHonored, DeepSeek_ModelSearch_Stream } from "@workglow/deepseek/ai";
void assertToolChoiceHonored;
void DeepSeek_ModelSearch_Stream;

Before M1(b) (deepseek's browser barrel reverted, rebuilt):

=== BEFORE M1(b): NODE ===
exit=0
=== BEFORE M1(b): BROWSER ===
probe.ts(1,10): error TS2305: Module '"@workglow/deepseek/ai"' has no exported member 'assertToolChoiceHonored'.
probe.ts(1,35): error TS2724: '"@workglow/deepseek/ai"' has no exported member named 'DeepSeek_ModelSearch_Stream'. Did you mean 'DeepSeekModelSchema'?
exit=1

After M1(b):

=== NODE (no customConditions) ===
exit=0
=== BROWSER (customConditions: ["browser"]) ===
exit=0

No tsgo-spawning vitest test was added — it needs a built dist, is slow, and has no
precedent in the repo.

Release metadata

The repo uses bunset, not changesets (there is no .changeset/), and CHANGELOG.md files
are generated from conventional-commit subjects — none were hand-edited. Once M1(b) lands, the
only symbols leaving the browser type surface are _testOnly (internal) and ollama's genuine
platform split, so a patch bump is honest and the normal bunset --patch --all path
applies. The three commit subjects are the changelog entries:

• fix(providers): restore the runtime-agnostic exports the browser barrels dropped
• refactor(providers): re-export the node entry from the browser entry where no split exists
• test(exports): flag a duplicate browser entry and pin browser barrel parity

Suggested description fix for #717

Not edited by this PR — for the author to apply:

• It says 25 branches; the actual count in the diff is 24.
• It says "10 in workglow including ./worker"; there are 9 packages/workglow
  shims (anthropic, deepseek, google-gemini, hf-inference, hf-transformers,
  ollama, openai, openrouter, xai). ./worker was already correct on main and is
  untouched by fix(providers): point each export condition's types at its own declaration #717.
• It says "Three assertions"; ExportTypesPairing.test.ts as merged carries 8 tests
  in its workspace exports maps suite (finds branches, pairs types, source entries, types
  ordering, browser split, condition ordering, build entries, glob exemption) plus the
  fixture suites.
• Worth adding: once this parity fix lands, a customConditions: ["browser"] consumer loses
  only _testOnly relative to the node surface.

---

Generated by Claude Code