refactor(prisma): tighten the $transaction adapter cast

[joryirving]

Summary

• asPrFixQueueClient / asAgentWorkClient used a double as any on $transaction ((fn) => client.$transaction(fn as any) as any), silencing all typechecking on the callback + its return. Replace with a single explicit cast of the callback to (tx: Prisma.TransactionClient) => Promise<T> and drop the result as any, so the transaction's return type T is checked again.

Why

The only genuine type gap is the tx type (the adapter interfaces type it as the narrowed client; Prisma types it as Prisma.TransactionClient) — structurally compatible for the model delegates these callbacks use. The narrow, documented cast preserves that bridge without blinding the compiler to callback return types.

Verification

• tsc → 0; eslint → clean; vitest pr-fix-queue + agent-work → 50 passing.

Closes #510 · part of #498

[its-saffron]

AI Automated Review

Full PR review.

Analysis engine: MiniMax-M2.7@https://litellm.jory.dev/v1 (anthropic) — escalated (fast_low_confidence)

Recommendation: Approve

This is a clean, targeted type-safety improvement that directly addresses issue PR 510. The change is minimal (10 additions, 3 deletions in a single file), all CI checks pass, and the implementation is well-reasoned.

Change-by-Change Findings

src/lib/prisma.ts

Import addition (+2): Added Prisma to the import from @prisma/client to access Prisma.TransactionClient for the explicit cast.

asPrFixQueueClient refactor:

• Before: (fn) => client.$transaction(fn as any) as any — double as any silenced all typechecking on callback parameter and return type.
• After: <T>(fn: (tx: PrFixQueueClient) => Promise<T>): Promise<T> => client.$transaction(fn as (tx: Prisma.TransactionClient) => Promise<T>) — single narrow cast only on the tx parameter type (PrFixQueueClient → Prisma.TransactionClient). The generic return type T is now preserved and typechecked.

asAgentWorkClient refactor: Same pattern applied to the agent-work adapter.

Comments: A 6-line explanatory comment documents why the cast is safe (structural compatibility for model delegates) and why it's preferable to as any.

Standards Compliance

• Code Standards - "Error handling": Uses error instanceof Error pattern elsewhere in the file; this PR is type-only, no runtime error paths introduced.
• No commit of secrets: No secrets or env files touched.
• Prisma conventions: Follows existing src/lib/prisma.ts patterns; Prisma namespace is already used in this codebase (via generated client).
• No agent-specific names in generic docs: N/A — internal refactor.

Linked Issue Fit

Issue PR 510 specified:

• Problem: Double as any on $transaction (lines 102, 110) silences all typechecking — addressed
• Fix: "type the adapter's $transaction wrapper properly (or narrow the cast)" — done: single narrow cast preserves return type T
• Effort: M — proportionate to the scope (single file, minimal changes)

This PR is part of PR 498 (dispatch audit), consistent with a systematic type-safety audit.

Evidence Provider Findings

No evidence providers configured; skipped.

Tool Harness Findings

• read_file confirmed the final state of src/lib/prisma.ts matches the diff.
• git_log confirmed commit d97acf0 is the PR head.
• git_grep on asPrFixQueueClient found 16 usage sites across routes, tests, and library code — all consuming the narrowed client; none call $transaction directly, confirming the narrow cast is sufficient.
• git_grep on asAgentWorkClient found 7 usage sites — same conclusion.

CI Verification

All five CI checks concluded success on commit d97acf0:

• Docker Build ✓
• Build ✓
• Tests ✓
• Typecheck ✓
• Lint ✓

The PR body claims tsc → 0, eslint → clean, vitest (pr-fix-queue + agent-work) → 50 passing, consistent with CI passing.

Unknowns / Needs Verification

None. The diff is self-contained, CI green, and the change is straightforward to reason about: structural compatibility between PrFixQueueClient/AgentWorkClient (narrow client subsets) and Prisma.TransactionClient (Prisma's own tx client) is well-established — these are just subsets of the same model delegates. The single explicit cast is more expressive and safer than as any.