refactor(test): extract shared SessionPrompt layer to test/lib/prompt-harness (F2)

[tesdal]

Summary

Audit finding F2 from 2026-04-22 diamond audit: prompt.test.ts and subagent-hang-regression.test.ts both composed the same ~85-line Effect Layer graph with identical MCP/LSP/Summary stubs. Regression test explicitly called this out with a "Copied verbatim" comment. Task 2 extracts the shared graph into a single helper.

Changes

• New: packages/opencode/test/lib/prompt-harness.ts exporting makePromptLayer(). Contains the three inline stubs (summary/mcp/lsp) plus module-scope status/runState/infra constants and the full SessionPrompt.layer composition.
• packages/opencode/test/session/prompt.test.ts: deleted inline stubs + constants + makeHttp() function; now imports makePromptLayer from the new harness.
• packages/opencode/test/session/subagent-hang-regression.test.ts: same deletions, "Copied verbatim" comment removed, now imports makePromptLayer.

Net −120 LOC (147 added in harness, 267 removed across both test files).

Verification

• bun typecheck — clean
• bun test test/session/subagent-hang-regression.test.ts — 2/2 pass, verified 4/4 clean runs on this branch + 1 clean run on base for sanity
• bun test test/session/prompt.test.ts — 42 pass / 2 fail. Both failures are in the known-flaky PT0 pattern (cancel tests spawning sleep 30 shells with 3s it.live timeout). Failure count is within / below the documented 3–11 range. Refactor cannot introduce new failures — it's pure Layer-composition reuse, no test-body changes. See anomalyco/opencode#24060.

Diamond review status

• Spec compliance (codex-5.3): 9/9 ✅. Initially flagged a regression-test hook-timeout blocker; investigation (controller ran test 4 more times on branch + once on base) confirmed it was a non-reproducing runner artifact unrelated to the refactor. Verdict overridden to APPROVE.
• Code quality (Opus): APPROVE WITH NITS. Zero semantic drift (byte-equivalent Layer wiring), clean import trimming, correct file location in test/lib/. One actionable nit (duplicate import from \"../../src/tool\") — fixed and amended into the commit. Other nits (naming, module-scope constant inlining) are judgement calls, left as specified.

Context

Part of 14-task audit remediation plan (docs/superpowers/plans/2026-04-23-audit-remediation.md). Base phase-ab-base is 2309cc89d (F1 merge on local/integration-v2). Draft PR opened for Copilot review only — merge target is local integration branch, not upstream.

[Copilot]

Pull request overview

This PR addresses audit finding F2 by extracting the shared Effect/Layer composition used by SessionPrompt loop tests into a single reusable test harness, eliminating duplicated stub/service wiring across multiple test suites.

Changes:

• Added makePromptLayer() in a new shared test harness (test/lib/prompt-harness.ts) containing the common SessionPrompt Layer graph and MCP/LSP/Summary stubs.
• Refactored prompt.test.ts to remove the inline Layer graph and import makePromptLayer() instead.
• Refactored subagent-hang-regression.test.ts similarly, removing the prior “Copied verbatim” duplication and using the shared harness.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
packages/opencode/test/lib/prompt-harness.ts	New shared helper exporting makePromptLayer() with consolidated stubs and SessionPrompt Layer composition.
packages/opencode/test/session/prompt.test.ts	Removes duplicated Layer/stub setup; now uses makePromptLayer() from the harness.
packages/opencode/test/session/subagent-hang-regression.test.ts	Removes duplicated Layer/stub setup; now uses makePromptLayer() from the harness.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[tesdal]

Copilot review: no comments. All three reviewers (codex-5.3 spec 9/9 after reproducibility override, Opus quality APPROVE WITH NITS fixed, Copilot clean) agree. Merging --no-ff into local/integration-v2. PR closed as review-only.