opencode-cache-hit

OpenCode TUI sidebar plugin for prompt cache hit rate, token usage, and cost—with first-class sub-agent (child session) rollup. Standalone by default (main + sub-agents in one panel). Optional coexistence with opencode-visual-cache.

Languages: English (this file) · 简体中文 · Documentation

Why this plugin

opencode-visual-cache already covers main-session cache visualization (token distribution, savings, slash-driven settings). This project exists because that scope does not fit several real workflows:

1. Sub-agent visibility — OpenCode spawns child sessions for Task / explore agents; you need rolled-up cache, tokens, and cost per sub-session, not only the main thread.
2. One panel for the whole session — Main session Hit/tokens/cost and a collapsible Agents section for sub-agent rollup.
3. Analysis off the TUI — Optional timeline JSONL (per assistant turn) for charts, jq, and billing post-mortems without scraping platform logs.
4. Shared TUI building blocks — src/tui-panel/ extracted so other sidebar plugins can reuse the same layout language as visual-cache.

Roadmap items (sidebar Timeline section, metric windows, nested sub-agents) are described in docs/en/timeline.md and docs/en/design.md.

Acknowledgments

This plugin is not part of opencode-visual-cache. Its sidebar layout, panel components (src/tui-panel/), and coexistence patterns are heavily inspired by opencode-visual-cache. visual-cache focuses on main-session context / token distribution; cache-hit focuses on per-turn metrics and sub-agent totals.

The cache TTL feature (elapsed time display with color-coded status) is inspired by opencode-cache-timer by nero-sensei. The original plugin provides a standalone sidebar countdown for prompt cache expiration; this plugin integrates the concept directly into the cache-hit panel.

The tool-part TTFT fallback (capturing tool.pending as first-response time when no text/reasoning streaming part exists) is inspired by oc-tps by Tarquinen, which is the first OpenCode plugin to properly handle finish=tool-calls for TTFT measurement.

Features

• Cache hit rate: session-level total (DB aggregate when available) + per-turn rate with trend (↑ / ↓ / -) on the main block
• Token breakdown: cache read / write / miss / output (aligned rows with visual-cache)
• Cost: per-message model rates with multi-currency config (USD, CNY, EUR, GBP, JPY); read savings, write premium, and net cache value; dynamic pricing for time-of-day tiers (DeepSeek peak/off-peak) and context tiers (context_over_200k, e.g. GPT-5.6)
• Sub-agents: Agents section rolls up child sessions only (scope labeled in UI); each row shows model name + session ID suffix with vendor-tinted label (cost in muted gray)
• Main + Agents: main block always shown; Agents section when sub-agents exist (foldable)
• Model lineages: separate provider/model buckets; model switches show switch or warming instead of a cross-model trend
• Cache TTL: active and recent model lineages have independent timers
• Compaction filtering: summary: true and agent: "compaction" messages do not change interactive metrics
• Collapsible sections: Detail / Model / Models (and Agents); theme-adaptive hit bar colors
• i18n: display.lang — en / zh / auto via config (no slash commands yet)
• Timeline (optional): daily JSONL per assistant turn for jq / scripts

Comparison with opencode-visual-cache

Standalone use is the default (main + sub-agents in one panel). Layout patterns were inspired by visual-cache; that package is not required.

	visual-cache	opencode-cache-hit
Main session context / token distribution estimate	Yes	No — use visual-cache
Per-role token breakdown (system / tools / …)	Yes	No
Cache savings estimate	Yes	Yes (from provider pricing)
Model per-million pricing from provider	Yes	Yes (from SDK provider config)
Slash commands (/cache-lang, /cache-currency, …)	Yes	Config file only
Fold state in api.kv	Yes	In-session (not persisted)
Loaded skills panel	Yes	No
Sub-agent session rollup	No	Yes
Combined hit (main + subs)	No	Yes when sub-agents exist
Per-call JSONL export	No	Optional timeline

Quick start

Option A: OpenCode command palette (recommended)

Ctrl+P → type install plugin → press Tab to switch scope to global (default is local) → type opencode-cache-hit@latest → press Enter.

Global plugins install to ~/.cache/opencode/packages/opencode-cache-hit@latest/. Create config at ~/.config/opencode/cache-hit.json:

Option B: Manual

Create or edit ~/.config/opencode/tui.json / tui.jsonc:

{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": ["opencode-cache-hit@latest"]
}

Local development: use an absolute path to the checkout, such as "/path/to/opencode-cache-hit", instead of the npm name. OpenCode loads the package root and its index.tsx entry.

Copy cache-hit.config.example.json → ~/.config/opencode/cache-hit.json (recommended) or next to the plugin root. Restart OpenCode after changing plugin code or config.

Install	After update
Local ./plugins/...	Full restart
npm @latest	Restart; if UI is stale, remove ~/.cache/opencode/packages/opencode-cache-hit@latest

Load errors: ~/.local/share/opencode/log/ (search cache-hit or failed to load tui plugin).

Configuration

Configuration files accept JSONC (line/block comments and trailing commas). cache-hit.config.example.json ships as strict JSON so it stays tool-friendly (jq, editors).

Cost display (USD → CNY example)

{
  "currency": "CNY",
  "costUnit": "USD",
  "rate": 6.77
}

Field	Meaning
costUnit	Currency of msg.cost (usually USD)
currency	Sidebar display currency
rate	Multiply costUnit → currency

rate is a manual snapshot — the 6.77 in the example is the USD→CNY rate at the time it was written and does not auto-update. Omitting rate falls back to the same built-in default, so it only needs to be set when you display a non-USD currency and want a current rate (e.g. from Xe, Wise, or OANDA). Use "currency": "USD", "costUnit": "USD" when no conversion is needed.

Supported display currencies in config: USD, CNY, EUR, GBP, JPY (see cache-hit.config.example.json). Runtime slash switching like visual-cache’s /cache-currency is not implemented yet.

Display (display)

"display": {
  "lang": "en",
  "panelBorder": true,
  "showSpeed": true,
  "speedUnit": "tpot"
}

Field	Default	Meaning
lang	"en"	en / zh / auto
panelBorder	true	Border/padding
mainHitLabel	(i18n)	Optional override for the Hit row label
showSpeed	true	Show/hide speed section
speedUnit	"tpot"	"tpot" (ms/tok) or "tps" (tok/s)

Agents totals sum child sessions only, not the main session (see agentsScopeHint). Main session metrics stay in the block above; collapse Agents to save space. Per-child rows use the same model slug as the main Model line (truncated when the sidebar is narrow); see docs/en/design.md § Sub-agent row display.

Model lineages and metric history

The main Hit, Total Hit, and cost values use the DB-level session aggregate when available (session.get, complete regardless of message-source caps), keeping the active lineage as the model identifier; the cache TTL follows the active lineage. Read savings, write premium, and net cache value sum eligible messages with each message's provider/model rates. The foldable Models section shows per-lineage buckets separately.

The plugin excludes assistant messages with summary: true or agent: "compaction" from interactive metrics. Timeline JSONL keeps these rows when timeline.logSummaryMessages is enabled and marks them with skippedForMetrics: true.

The plugin requests main-session history directly with a limit of 10,000 messages. When that request fails and the mirror has filled its 100-message cap, or the returned history reaches the 10,000 limit, the panel shows a * history truncated hint. Streaming speed and TTFT use this live mirror so they can update during a call. If a message has no matching rate, the cost row keeps OpenCode's reported blended cost instead of showing a partial recomputation.

Timeline logs (timeline, default off)

Per assistant turn → JSONL (tokens, cache, cost, TTFT, per-tool toolDurations, and skippedForMetrics). docs/en/timeline.md · 中文.

"timeline": {
  "enabled": true,
  "dir": "",
  "rotateMaxBytes": 16777216,
  "retainRotated": 5,
  "maxAgeDays": 30,
  "maxLogFiles": 20,
  "toolSummary": { "allTools": true, "bash": false }
}

Field	Default	Meaning
enabled	false	Master switch
toolSummary	{ allTools: true, bash: false }	Controls the privacy-sensitive toolDurations[].summary field (secure-by-default: bash off). true/false for all tools, or { allTools, bash?, … } per-tool. Durations (tool, durationMs) are always recorded. See docs/en/timeline.md
dir	""	logs/timeline-YYYY-MM-DD.jsonl under plugin root
rotateMaxBytes	0	Same-day size roll to .jsonl.1
retainRotated	5	Backups kept per day
maxLogFiles	0	Cap file count; deletes earliest logs first

LOG=~/.config/opencode/plugins/opencode-cache-hit/logs/timeline-$(date +%Y-%m-%d).jsonl
tail -f "$LOG"
# time fields are ISO 8601 strings with local timezone (e.g. "2024-05-30T08:00:00.000+08:00")
jq -r 'select(.rootSessionId=="YOUR_ROOT") | [.created,.scope,.hitPercent,.cost]|@tsv' "$LOG"

Retention details: Rotation and retention. Charts: scripts/README.md.

Cache TTL (cacheTTL, default on)

Shows how long the active provider/model prompt cache has been alive. Each model lineage has an independent timer. Color changes when exceeding TTL:

• Green: elapsed < TTL
• Yellow: TTL ≤ elapsed < 2×TTL
• Red: elapsed ≥ 2×TTL

"cacheTTL": {
  "enabled": true,
  "providers": {
    "anthropic": "5m",
    "openai": "5m",
    "deepseek": "2h",
    "google": "1h"
  }
}

Field	Default	Meaning
enabled	true	Master switch
providers	{}	TTL per provider (or provider:model). Human-readable: 30s, 5m, 1.5h

Built-in defaults (used when provider not in config):

Provider	Default TTL	Source
anthropic	5 min	Anthropic docs
openai	5 min	OpenAI docs
deepseek	2 hours	DeepSeek docs
google	1 hour	Google docs
xai	5 min	xAI docs
minimax	5 min	MiniMax docs
xiaomi	5 min	Implicit caching
qwen	5 min	Implicit caching
moonshot	5 min	Implicit caching

Default TTL: 5 minutes for all providers not listed above. Color changes based on elapsed time vs TTL: green (< TTL), yellow (TTL-2x TTL), red (≥ 2x TTL).

Dynamic pricing (dynamicPricing, default on)

Model rates are normally static per-million USD from OpenCode's provider registry. Some models price by time of day (DeepSeek V4: peak Mon–Fri 09:00-12:00 / 14:00-18:00 Beijing time, off-peak half price — weekends are off-peak) or by context size (context_over_200k, e.g. GPT-5.6: rates roughly double above 200k tokens).

Without config, the plugin already:

• Reads the model's context tier from state.provider (runtime tiers / experimentalOver200K, normalized internally) and shows the right tier based on total context (input + cache read) vs the threshold.
• Applies a built-in DeepSeek off-peak 0.5× multiplier when the schedule below matches (weekdays only for peak, weekends fall to the off-peak fallback).

"dynamicPricing": {
  "enabled": true,
  "timezone": "Asia/Shanghai",
  "schedule": [
    { "level": "peak", "windows": [
      { "start": "09:00", "end": "12:00", "days": [1, 2, 3, 4, 5] },
      { "start": "14:00", "end": "18:00", "days": [1, 2, 3, 4, 5] }
    ] },
    { "level": "offpeak", "windows": [] }
  ],
  "contextThreshold": 200000,
  "providers": {
    "deepseek": {
      "models": {
        "deepseek/deepseek-v4-flash": {
          "multipliers": { "peak": 1, "offpeak": 0.5 }
        }
      }
    }
  }
}

Field	Default	Meaning
enabled	true	Master switch. Disable to restore fully static pricing
timezone	Asia/Shanghai	IANA zone used for schedule matching (DeepSeek prices are Beijing-time based)
schedule	DeepSeek peak/off-peak	{level, windows:[{start,end,days?}]} list; HH:MM, cross-midnight windows allowed. days is an optional ISO weekday list (1=Monday … 7=Sunday); omitted or [] = every day. A level with empty windows is the catch-all fallback: it applies whenever no windowed level matches (e.g. weekends with the default DeepSeek schedule)
contextThreshold	200000	Token threshold for the context tier; per-model contextThreshold wins over the runtime tier size from state.provider
providers	{}	Per providerID → modelID rules

Per-model rules support two forms (explicit config wins over the built-in DeepSeek default):

• multipliers: factor applied to the static rates per schedule level (e.g. {"offpeak": 0.5})
• levels: absolute rates per level, e.g. {"peak": {"input": 0.44, "output": 0.88, "cacheRead": 0.01}, "offpeak": {"input": 0.22, ...}}. Cache rates may be written as flat cacheRead/cacheWrite (or cache_read/cache_write) or nested cache: {"read": …, "write": …} (both are accepted; flat wins if both present). Default unit is USD per 1M (same as state.provider). To write prices in another currency, set "currency": "CNY" and either make it match the display cost.currency (converted via cost.rate) or provide the per-rule "rate" (USD → that currency, e.g. "rate": 1.08 for EUR). If the currency cannot be converted (no rate, currency ≠ display currency), a warning is logged to stderr and the values are treated as USD. multipliers are ratios and have no currency.
• contextThreshold: per-model override of the global threshold (wins over the runtime tier size from state.provider).

Rates shown in the sidebar switch automatically at schedule boundaries (no polling). The peak/off-peak badge on the rate row appears only when the model actually prices that level (the level exists in its explicit levels/multipliers, or the built-in DeepSeek default applies); plain static models and unpriced levels (e.g. a peak-only rule at an off-peak moment) show no badge. Session cost shown is recomputed per message from its provider/model, request time, and context tier when dynamic rules apply; otherwise OpenCode's own msg.cost is used. If a message has no matching rate, the cost row keeps OpenCode's reported blended cost instead of showing a partial recomputation. Read savings, write premium, and net cache value use the same per-message rates. Sub-agent rows use their session creation time (session.list) for time-of-day pricing.

Note

Migration (weekday-aware schedules). Schedules now support an optional days field (ISO weekday, 1=Monday … 7=Sunday; omitted = every day) and a catch-all fallback level (a level with empty windows). DeepSeek's official peak is Monday–Friday 09:00-12:00 / 14:00-18:00 Beijing time; weekends are off-peak. The built-in default schedule is weekday-aware, and new configs written from the examples above are too. Legacy configs without days keep the old behavior — weekends are still billed as peak. To pick up the fix, add "days": [1,2,3,4,5] to your peak windows (or use the new default schedule with an offpeak fallback). If you resolve a DeepSeek model while your configured schedule has windowed levels but no days, the plugin logs a one-time hint to stderr.

Timeline dashboard (docs/en/timeline.md) also recomputes costs offline: it reads ~/.config/opencode/opencode.json (JSONC-aware) for provider rates and injects dynCost per record (shown with ≈ and in charts/totals when it differs from the original).

Refresh official DeepSeek rates: bun scripts/fetch-deepseek-pricing.ts prints a ready-to-paste dynamicPricing snippet — a weekday-aware schedule (peak Mon–Fri, off-peak catch-all fallback) plus providers rates (CNY by default with "currency": "CNY", --usd --rate 6.77 for USD).

Updating

Important

OpenCode pins @latest to the first-resolved version and never re-fetches it (opencode#6774, #25293, #30631). A restart alone will not pick up a newer npm release. If you are running an old cached build, you may hit crashes already fixed upstream (e.g. #3 — the sidebar vanishing with undefined is not an object (evaluating 'config.providers')). OpenCode wraps each plugin slot in a per-slot <ErrorBoundary> (via @opentui/solid, present since v1.17.0). A SolidJS component render error is caught and logged to stderr — the broken slot unmounts silently (no crash screen, easy to miss) while the rest of the TUI survives. Note: lower-level opentui/yoga renderer errors (e.g. layout-phase faults) can still bypass this boundary and crash the whole TUI.

To force an update, delete the cached package, then reinstall and restart:

rm -rf ~/.cache/opencode/packages/opencode-cache-hit@latest

Then reinstall via Ctrl+P → install plugin, and restart OpenCode.

To avoid the pinning issue entirely, install a pinned version instead of @latest:

{ "plugin": ["opencode-cache-hit@0.7.3"] }

Compatibility

Model-agnostic: any OpenCode provider that exposes assistant tokens / cost on messages (DeepSeek, Claude, GPT, etc.). Data comes from the OpenCode session API, same as visual-cache.

Requires OpenCode with TUI plugin slots (@opencode-ai/plugin ≥ 1.14). Works alongside visual-cache; no extra dependencies at runtime beyond peers in package.json.

Documentation

Audience	English	中文
Users	This README	README.zh-CN.md
Maintainers	docs/en/design.md	docs/zh-CN/design.md
Timeline / JSONL	docs/en/timeline.md	docs/zh-CN/timeline.md
TUI panel reuse	src/tui-panel/README.md	src/tui-panel/README.zh-CN.md
Contributing / npm	CONTRIBUTING.md	—
Coding agents	AGENTS.md	—
Index	docs/README.md

Project layout

index.tsx
cache-hit.config.example.json
src/
  plugin.tsx              # sidebar_content slot
  sidebar-host.tsx        # messages, child sync, timeline
  widget.tsx
  stats.ts / timeline/ / tui-panel/
tests/

Development

bun test

See CONTRIBUTING.md for setup, PR notes, and npm publishing. Architecture: docs/en/design.md.

License

MIT