Summary

Area 8 of #2336 (umbrella) / follow-up to #2463. Closes #2468.

• Remove the global vi.mock("./get-reply-run.js", () => ({ runPreparedReply: vi.fn(async () => undefined) })) from src/auto-reply/reply/get-reply.test-mocks.ts — one of the five shields that hid gut(auto-reply): remove dead thinking-level and model-selection plumbing — TypeError on first-turn messages #2334 (TypeError: modelState.resolveDefaultThinkingLevel is not a function) from CI.
• Add src/auto-reply/reply/get-reply.real-chain.test.ts: drives getReplyFromConfig → resolveReplyDirectives → handleInlineActions → runPreparedReply through the real code path on a first-turn webchat message. Only the CLI boundary (runReplyAgent), outbound delivery (routeReply), and session/filesystem boundaries are mocked. No fabricated shapes for the resolveReplyDirectives → runPreparedReply integration boundary.
• Two ancillary mock updates to keep the existing registerGetReplyCommonMocks consumer green:
  • resolveCommandAuthorization stub now returns ownerList: [] and senderIsOwner: false alongside isAuthorizedSender: true (required by buildCommandContext).
  • ./directive-handling.js is partial-mocked via vi.importActual, so real parseInlineDirectives / isDirectiveOnly remain callable.

Why this test has teeth

The #2334-class regression mode was: a caller in runPreparedReply assumed a method existed on a threaded object that resolveReplyDirectives didn't provide. Every mock in the test harness paved over the gap. The new test does not mock resolveReplyDirectives, handleInlineActions, or runPreparedReply — so any future shape drift between those three stages throws at test time. The assertion is runReplyAgent was called with the expected commandBody: proof the chain executed end-to-end.

Scope

Two files:

• src/auto-reply/reply/get-reply.test-mocks.ts — shield block deleted (replaced with doc comment), two mocks adjusted.
• src/auto-reply/reply/get-reply.real-chain.test.ts — new test.

Explicitly out of scope for this PR (per #2468 body)

• The "consider a lint rule for as never / as any in test arrangement code" item — left for a follow-up.
• Other gut(auto-reply): finish thinking-level and model-selection sweep — follow-up to #2335 #2336 areas (3, 5, 6, 7) — directive removal, thinking.ts helpers, session state, directive handling internals, docs, display audit.
• The test-local resolveReplyDirectives mock in reset-hooks-fallback.test.ts — it is intentional for fallback-specific testing, not a global shield.

Test plan

• pnpm check (format + typecheck + lint + fork-integrity gates) → PASS
• get-reply.real-chain.test.ts → 1/1 pass
• get-reply.reset-hooks-fallback.test.ts → 2/2 pass (no regression from common-mock updates)
• get-reply-run.media-only.test.ts → 10/10 pass (no regression from shield removal)
• CI: full pnpm test under project shards (to be verified by the auto-merge gate)
• CI: build + docs jobs

Pre-existing failures in src/auto-reply/reply/session.test.ts and src/auto-reply/reply/agent-runner.media-paths.test.ts were confirmed to reproduce on baseline origin/main — not caused by this PR.

Related

• Parent: gut(auto-reply): finish thinking-level and model-selection sweep — follow-up to #2335 #2336 (umbrella)
• Bug this shield originally hid: gut(auto-reply): remove dead thinking-level and model-selection plumbing — TypeError on first-turn messages #2334
• Prior partial sweep: gut(auto-reply): partial thinking-level + model-selection sweep — Areas 1+2+4 of #2336 #2463

🤖 Generated with Claude Code