Summary

• give @opencode-ai/http-recorder an independent Changesets lifecycle beginning at 0.1.0 and exclude it from OpenCode-wide version synchronization
• restore deliberate, manually dispatched beta publishing with exact-tarball clean-consumer verification, provenance, bootstrap token support, and a documented trusted-publishing transition
• clarify the Effect WebSocket service/layer model, causal frame replay, completion boundary, matching rules, concurrency, and current beta limitations
• replace the public echo test with a finite provider-style conversation and expose socket-specific options so HTTP-only matching cannot be configured accidentally
• align installation and verifier guidance with exact Effect 4.0.0-beta.83 compatibility

Release Flow

Merging this PR does not publish the package. After merge, bun run version:http-recorder creates a focused release PR that consumes the Changeset, advances 0.0.0 to 0.1.0, and generates CHANGELOG.md. Once that release PR reaches dev, an operator manually dispatches the http-recorder release workflow.

The first publication requires the repository NPM_TOKEN secret. packages/http-recorder/RELEASE.md documents switching the workflow to npm trusted publishing through GitHub OIDC afterward.

Verification

• bunx changeset status reports only @opencode-ai/http-recorder at minor
• bun run test in packages/http-recorder (33 passed)
• bun typecheck in packages/http-recorder, packages/llm, packages/core, and packages/opencode
• full Turbo typecheck via the pre-push hook (23/23 tasks)
• bun run verify:package installs, imports, and typechecks the exact tarball in a clean npm consumer
• disposable-worktree bun run version:http-recorder generated only 0.1.0 and CHANGELOG.md; the resulting nine-file tarball passed clean-consumer verification
• Prettier, file-scoped Oxlint, and git diff --check
• bun run release:http-recorder correctly rejects the unversioned 0.0.0 package before registry access