[viztor]

Phase P2 of docs/improvement-plan.md. It is stacked on #107 (P1b). Below that come #106 (P1a) and #105.

This is a 1.x minor (D6). Nobody consumes the removed exports, so there is no ! and no BREAKING footer.

The barrel

src/index.ts is the host's contract and nothing more:

export { inject, name } from "./identity.ts";
export { apply } from "./lifecycle.ts";
export { Config, type ModelOverrideConfig, type ModelRouteConfig, type PluginConfig, type ResolvedPluginConfig } from "./config.ts";
export { type EndpointFamily, type KeySourcePolicy } from "./config-values.ts";

test/public-api.test.ts pins this list, so widening it is a reviewed change. It opened as a green pin of the old barrel, then shrank in this PR.

• Added: ModelRouteConfig. PluginConfig.modelRoutes already referenced it, but it could not be named from the barrel.
• Removed (no longer re-exported):
  • Identity: LEGACY_NAME, LEGACY_PKG, OPENCODE_CLI_VERSION, OPENCODE_PATCH_USER_AGENT, OPENCODE_USER_AGENT.
  • Config: CONFIG_DEFAULTS, the ten DEFAULT_* constants, ENDPOINT_FAMILIES, KEY_SOURCE_POLICIES, resolveConfig, ALL_MODELS_MARKER, readBoolean, readString, readStringList, unwrapNode.
  • Session: fallbackSessionId, isOpenCodeMessageId, isOpenCodeSessionId, headerValueFor, newOpenCodeMessageId, openCodeSessionIdFor, OPENCODE_UA, pinnedSessionId, sessionIdForRequest and the four session header constants; also withStore and the ActiveTurnState type.
  • Request path: DUMMY_BASH_TOOL, DUMMY_READ_TOOL, isCoreToolModel, maybeInjectCoreTools, RESPONSES_PATH, hasSessionHeader, isOpenCodeRequest, patchFetch and the CordisContext type.
  • Credentials: clearCapturedApiKeys, extractApiKeyFromHeaders, getCapturedApiKey, isPlaceholderApiKey, recordCapturedApiKey, tierForRequest, discoverGoConfig, effectiveGoKeyRef, resolveGoApiKey, resolveGoBaseURL, resolveOpencodeStoreKey, resolveZenCreditInfo, toGoBaseURL.
  • Usage and cost: GoUsageService, parseGoUsage, parseUsageQuery and the GoUsage, UsageQuery and UsageWindow types; calculateTurnCost, clearSessionUsageStore, describeModel, formatModelRate, formatUsd, getSessionUsage, recordTurnUsage and the ModelCostRate and SessionUsageSnapshot types.
  • Catalog: every models-catalog constant and function, plus the CatalogModelSpec, CatalogProvider and ParsedCatalogs types.
  • Discovery: the models-discovery decorators and helpers.
  • Routes: every responses-routes constant and function, and the responses-provider helpers.
• Deleted outright: resolveRoutedKey and the RoutedKeyDetails type (D4: exported and never called).

scripts/check.ts' consumer probe reads only apply and Config, so it still passes.

Also in this PR

• One door for tests. All 62 test files that imported src now go through test/support/src.ts. It re-exports each name from the module that defines it. A gate in architecture.test.ts rejects a static src import anywhere else. Its exemptions are named: vi.mock targets, re-imports after vi.resetModules(), the architecture and public-API tests, the jsdom client tests, and test/support/. The source-reading gates share a comment blanker that has its own known-answer tests.
• exports ⊆ files. ./src/* is gone from exports. check.ts now fails when an exports target is outside the published files, using the pure helpers in scripts/package-checks.ts, which are tested against fixtures.
• internalRoutes is deleted (D2). The per-protocol routes always mount until P-R1 makes the plugin own the plane routes. A row that still carries internalRoutes validates and the key is ignored; a test feeds such a row and pins that. The bridge-disposer test now targets the "every internal route is already declared" exit.
• @viztor/dsh-opencode is no longer published (D9). The deprecation message already on npm stays as its final word.
  • SCOPED_TARGETS lives in a side-effect-free scripts/scoped-targets.ts.
  • release.yml verifies exactly that list, and check.ts asserts the two agree.
  • P1a's wrapper-forwarder helper and its test are gone with the wrapper.
• Docs. AGENTS.md (barrel scope, the test door), CONTRIBUTING.md, the README's "Changed in 1.x", and three prose mentions of resolveRoutedKey.

Review

Three lenses reviewed the change (correctness and host contracts, tests, repository rules), and two independent skeptics checked each finding. Over three rounds, every upheld finding was fixed: 3 medium and 18 low.

Verification

All run locally:

Node	Step	Result
26	pnpm run check	0 errors; 21 warnings, unchanged from the base
26	unit suite	1234 tests pass, up from 1222
26	shuffled, seeds 1 / 42 / 1234	pass
26	test:mock	51 pass
26	build and scripts/check.ts	pass; this includes the new exports ⊆ files and publish-list checks
26	coverage	96.4% lines, 92.3% branches
24	shuffled (seed 7) and test:mock	pass

The diff contains no local paths.

---

Generated by Claude Code

[viztor]

catalog freshness is red, and the cause is not this PR's diff. The job compares src/catalog-data.ts with the live catalog at models.opencode.ai. Between this branch's base and this run, the vendor added one Go model:

opencode-go: 28 -> 29
  + 1: kimi-k3.1-pro
opencode  : 82 -> 82
shim is STALE — re-run with --write

This PR doesn't touch the catalog or its generator. Any branch that runs the check now fails it the same way, including the base of this stack (#105).

There is no fix yet. The shim has to be regenerated with pnpm run catalog:shim -- --write, which needs network access to models.opencode.ai. The session that built this stack can't reach that host. The generated file must not be hand-edited.

The repository's nightly run (37 4 * * *) regenerates it and opens chore: refresh the generated catalog and limits tables against main. Once that merges, I'll merge main up through this stack. I'm re-running the failed job once to confirm it reproduces identically.

---

Generated by Claude Code