[XieX]

initClient() reads LD_BASE_URI / LD_STREAM_URI for the base SDK, but FDv2SkillStore always defaulted to the production hosts. A process pointed at catfood, staging or a private instance by its environment resolved flags there, then sent skill delivery to production with a non-production key. That shows up as a fatal 401 in store.failed, not as a configuration error.

With neither URI option given, the constructor now reads both variables in their place, once:

• A lone LD_BASE_URI moves both endpoints, under the store's existing rule for a lone baseUri. This differs on purpose from initClient(), where flag streaming stays on its default host.
• An option given turns the environment off for both endpoints, so an explicit host is never mixed with one from the environment.
• A blank value counts as unset.
• A refused, non-https value is reported under the variable's name.
• SKILLS_DEFAULT_BASE_URI / SKILLS_DEFAULT_STREAM_URI keep the production values as the fallback.

The client README and agents.md are updated.

Spec: TESTING.md §3.25 and A.12, in launchdarkly/ai-sdks-monorepo#59. Python twin: launchdarkly/python-ai-sdk#207, merged.

Tests: new cases in the endpoints block of skills-fdv2.test.ts. That block now clears both variables around each test, so a developer shell can't move the default-host assertions. The 371 FDv2 tests pass. etc/ai-server-experimental.api.md is regenerated, since the new doc lines on FDv2SkillStoreOptions move a forgotten-export reference.

Found while dogfooding AgentControl (Queenbee) on staging.

🤖 Generated with Claude Code

---

Note

Overview
Aligns FDv2SkillStore skill delivery hosts with initClient() env configuration. When neither baseUri nor streamUri is passed, the constructor now reads LD_BASE_URI and LD_STREAM_URI once at construction (blank values treated as unset), using the same lone-base-URI rule as explicit options: only LD_BASE_URI drives both poll and stream. Any explicit URI option disables env lookup entirely. Invalid http:// values fail requireHttpsUri with the env var name in the error.

Docs in README and agents.md describe the behavior and note that lone LD_BASE_URI differs from initClient(), where streaming can stay on the default host. Tests add an endpoints suite for env resolution and isolate LD_* vars so local shells do not skew default-host assertions.

^{Reviewed by Cursor Bugbot for commit 5ddfedc. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jeffdupont]

Merged main in and regenerated packages/client/etc/ai-server-experimental.api.md (c586bbe). The new doc lines on FDv2SkillStoreOptions move the Requester forgotten-export reference from skills-fdv2.d.ts:454 to :460. Locally, typecheck/api:check pass and client tests pass (1837 passed). This needs a human reviewer; the Python twin, python-ai-sdk#207, is already approved.

[jeffdupont]

Approving. This matches python-ai-sdk#207 and TESTING.md §3.25 on every bullet: options win as a set, a lone LD_BASE_URI moves both endpoints, blank or whitespace counts as unset, the https error names the variable, the vars are read once, and the published defaults are unchanged. There is a test for each one. The API report is current (CI "Check API reports" passes at c586bbe). Heads-up for the merge order: #159 touches the same Requester line in etc/ai-server-experimental.api.md and moves it to ai-server.api.md:934. Whichever merges second needs yarn workspace @launchdarkly/ai-server api:update after the rebase.