Cache Control v1.0.0

Cache Control v1.0.0

See actual prompt-cache reuse in OMP and maintain eligible context prefixes during idle pauses—without adding hidden replies to chat history.

Install

omp plugin install github:ubranch/omp-cache-control

Fully exit and restart OMP. /reload-plugins does not reload TypeScript extensions. For visibility, select the custom Status Line preset in Appearance settings and add status to Left Segments.

After a real cacheable request, inspect maintenance and its estimated cost:

/cache warm

Paid automatic warming is enabled by default. To keep visibility without paid upkeep for the current process:

/cache warm off

Commands are process-local; restarting restores defaults. Uninstalling the plugin is the persistent opt-out. Cancelled calls may still be billable; unpriced calls are not free, and cache-storage charges are separate from estimated inference cost.

What ships

• Actual cache-read percentage and estimated/unknown expiry, plus /cache and /cache warm [on|off|now] controls and a separate warm-cost ledger.
• A stable context prefix captured from a successful foreground payload, refreshed only while the native TUI is idle through isolated, bounded maintenance. Explicit Google cached resources use verified lease renewal.
• Model, branch, and session invalidation; foreground priority and cancellation; shutdown cleanup; miss-aware estimates and stop-on-failure behavior. Warm replies stay outside session history and the foreground response chain.

Eligibility is guarded, not universal. Unsafe payloads are refused, Antigravity Gemini is deliberately unavailable, and provider retention is never guaranteed. Maintenance is not a background daemon and does not run in headless mode.

Publication hardening

• Request-local privacy guards intercept HTTP failures, supported SSE in-band/parse errors, refusals and failed terminal states, plus Bedrock binary errors/exceptions before native SDK diagnostics. The public Smithy codec validates binary framing/CRC and preserves valid bytes; it is a normal production dependency, not a bundled OMP SDK. Foreground diagnostics are unchanged. Successful SSE cleanup closes only the response transport, preserving warm completion/accounting.
• Hidden inference dispatch is capped at one actual send, even when the native transport would retry; authentication and metadata requests are not counted as inference.
• Explicit-cache identity comes from the final successful settled payload, including later in-place cachedContent additions, changes, and removals. Whole-body replacement still requires extension ordering.

Evidence and limits

The historical Codex workflow ran for 80.0475 real minutes, with 3 automatic cache-hit refreshes, unchanged branch/assistant-message IDs, and successful resumed tool work. It used OMP CLI 18.5.1 with locally installed SDK 18.6.0—not a new public-release soak. Claude passed a bounded live 1-token warm-up, not an equivalent elapsed-time soak.

Strict TypeScript checking and all 189 permanent tests passed (1,448 expectations) with published OMP SDK 18.6.1 and Bun 1.4.2. Controlled loopback regressions consume native SSE/binary responses with genuine active diagnostic-writer positive controls: Completions HTTP 400/413, Google false completion, and Bedrock unknown binary exceptions leaked private body/echo data before protection and pass afterward; valid cached binary output remains usable. CLI nested-response/failed-finish cases, single-send retry enforcement, and settled explicit-cache identity are also covered. This is not new native-TUI/live-provider proof, provider-retention evidence, or compatibility with every model. No CI or post-publication pass is claimed. See verification evidence and boundaries; Actions reports current public CI status.

A separate actual native-TUI smoke on SDK 18.6.1/Bun 1.4.2 passed against a local Google loopback with synthetic usage and an isolated unauthenticated home. Exactly one foreground and two 1-token hidden requests preserved the prefix: successful [DONE] cancelled only the response transport before shutdown, an in-band 400 stopped warming, and another warm now while off sent nothing. Session/history and foreground event counts stayed unchanged; the private canary was absent from both scanned native SDK logs. This is controlled native-TUI evidence, not paid-provider retention or billing proof.

Source, the complete regression suite, CI configuration, and the MIT license are included. Read the lifecycle guide and safety/compatibility guide for scheduling, safeguards, costs, and refusal conditions. Distribution uses OMP's native GitHub plugin installer; no npm publication is claimed.