Substance looks well thought out: cursor-based session walk, cache with proper invalidation on every write path, nonces to dodge stale results, reset on session change. Tests cover dupes, workspace scoping, and cross-session walks. Closes #2051 properly.

Three things before merge:

1. Scope creep — please split. Two changes here are unrelated to prompt history persistence:

   • desktop/src/i18n/ja.ts — sendFeedback / skipFeedback / cancelFeedbackLabel / refineFeedbackLabel backfill. This is exactly the scope of #2069, which is open in parallel; they will conflict.
   • tests/web-tools.test.ts — a new vi.mock("../src/config.js", …) block. I can't see how that's related to this PR.

   Please drop both from this PR and rebase; #2069 will land separately and the web-tools change should have its own PR with a motivating description.

2. Public API surface — is this intentional? You're adding promptHistoryStep, PromptHistoryCursor, and PromptHistoryEntry to src/index.ts (and bumping the public-api allowlist). That's a stability commitment to downstream consumers, but the function reads like an internal desktop helper. Unless there's a planned external consumer, I'd prefer keeping it un-exported and having desktop.ts import directly from ./memory/session.js.

3. 5s cache TTL — minor. All mutating helpers correctly call invalidatePromptHistoryCache(). Add a one-line comment near the cache constant calling out that invalidation is the primary correctness mechanism and 5s is just a backstop for external mtime changes. (Optional, but helps the next reader.)

CI is green and the core change is good. Fix 1, decide on 2, and we're ready.

Re-reviewed — both blockers from my original review are resolved:

1. Scope creep dropped: the desktop/src/i18n/ja.ts backfill (that's #2069's territory) and the unrelated tests/web-tools.test.ts vi.mock are both gone. The diff is now focused on prompt-history persistence (session.ts/config.ts) + its composer wiring.
2. Public API kept internal: src/index.ts is untouched — promptHistoryStep/PromptHistoryCursor/PromptHistoryEntry are no longer exported, so desktop.ts consumes them directly from ./memory/session.js without committing to a downstream-stability surface, as preferred.
   The core (cursor-based session walk, cache with per-write invalidation, nonces, reset-on-session-change) was already solid. CI green. Approving — merging if it's conflict-free.