OpenCode Cache View

@mtayfur/opencode-cache-view adds a compact sidebar to the OpenCode TUI for cache efficiency, estimated context usage, generation speed, and loaded skills. The latest cache-hit ratio and trend also remain visible in the prompt status row.

Cache View
Hit [████████████████] 98.7% ↑0.8%
Session Hit                96.0%
Read                    8.0M tok
Miss                    326K tok
▼ Speed ─────────────────────
TTFT                       ~1.4 s
TPS                       47 tok/s
Trend                    ▂▅▄▇▆█
▼ Estimated Tokens ──────────
Prompt                  186K tok
Tool Call                59K tok
Tool Result              80K tok
Agent Reasoning          14K tok
Agent Output             21K tok
Total                   326K tok
▼ Loaded Skills (2) ─────────
playwright-mcp-ops       2.4K tok
customize-opencode       3.1K tok

Metrics

• Cache: Hit is cache.read / (input + cache.read + cache.write) for the latest regular LLM step. Session Hit is the aggregate OpenCode session ratio, including compaction calls. Miss combines uncached input and cache writes.
• Estimated Tokens: Uses one character-based estimator for the visible system, user, tool, reasoning, and output content from the latest completed compaction summary, its preserved tail, and subsequent messages loaded in the TUI. Provider system prompts, tool schemas, and media tokens are not included.
• Speed: During streaming, TPS shows a ~-prefixed estimate from visible text and reasoning deltas over the latest five seconds. Hidden reasoning with no deltas uses the previous completed turn or remains … tok/s when none exists. Completed tool-call steps and tool execution time are excluded. The rate remains ~-prefixed while the turn is incomplete; the final non-tool-call assistant step replaces it with exact (output + reasoning) / duration throughput. TTFT retains the estimated delay to the first visible reasoning or text block, and Trend shows the latest eight completed-turn TPS values.
• Loaded Skills: Shows active, non-compacted skill outputs and their estimated token counts. When a skill is loaded more than once, the latest load is shown.

Rows and the cache-hit bar adapt to the available sidebar width. The arrow next to the hit rate compares the latest two LLM steps. Sections can be expanded or collapsed with the mouse.

The prompt status indicator uses the composable session_prompt_right slot, so it does not replace the prompt or conflict with plugins that customize prompt behavior. It shows the same completed-step TPS as the sidebar, followed by a single separator and the cache ratio; only the cache trend is color-coded.

Requirements

• OpenCode 1.18.16 or a newer compatible 1.x release
• Bun 1.3.14 or newer for local development

Installation

Add the package to ~/.config/opencode/tui.json:

{
  "plugin": ["@mtayfur/opencode-cache-view"]
}

OpenCode installs package plugins automatically. Restart the TUI after changing the configuration.

Local checkout

bash ./install.sh

The installer resolves dependencies, builds dist/index.js, and adds its local file:// URL to the TUI plugin configuration. Remove the local entry with:

bash ./install.sh --uninstall

Restart OpenCode after either operation.

Architecture

src/
├── index.tsx             Plugin entrypoint
├── cache-status.tsx      Prompt status-row cache ratio
├── cache-view.tsx        Solid/OpenTUI view and event subscriptions
├── format.ts             Display formatting
├── token-estimator.ts    Cached character-based token estimation
└── metrics/
    ├── cache.ts          Cache usage and hit ratios
    ├── context.ts        Active context, token, and skill calculations
    ├── helpers.ts        Shared message and part helpers
    ├── read.ts           Metric orchestration
    ├── speed.ts          Completed-step generation speed
    └── types.ts          Metric data contracts

Metric calculations are isolated from the TUI rendering layer. read.ts snapshots OpenCode state and coordinates the dedicated metric modules, while cache-view.tsx owns reactive updates and presentation.

Development

From the repository root:

bun install --frozen-lockfile
bun run --filter @mtayfur/opencode-cache-view typecheck
bun run --filter @mtayfur/opencode-cache-view build
npm pack ./packages/cache-view --dry-run