OpenCode CLI Notes

Log Sources

• OpenCode session usage is recorded under ${OPENCODE_DATA_DIR:-~/.local/share/opencode}/storage/message/ (the CLI resolves OPENCODE_DATA_DIR and falls back to ~/.local/share/opencode).
• Each message is stored as an individual JSON file (not JSONL like Claude or Codex).
• Message structure includes tokens.input, tokens.output, tokens.cache.read, and tokens.cache.write.

Token Fields

• input: total input tokens sent to the model.
• output: output tokens (completion text).
• cache.read: cached portion of the input (prompt-caching).
• cache.write: cache creation tokens.
• Pre-calculated cost field may be present in OpenCode messages.

Cost Calculation

• OpenCode messages may include pre-calculated cost field in USD.
• When cost is not present, costs should be calculated using model pricing data.
• Token mapping:
  • inputTokens ← tokens.input
  • outputTokens ← tokens.output
  • cacheReadInputTokens ← tokens.cache.read
  • cacheCreationInputTokens ← tokens.cache.write

CLI Usage

• Treat OpenCode as a sibling to apps/ccusage and apps/codex.
• Reuse shared packages (@ccusage/terminal, @ccusage/internal) wherever possible.
• OpenCode is packaged as a bundled CLI. Keep every runtime dependency in devDependencies.
• Entry point uses Gunshi framework.
• Data discovery relies on OPENCODE_DATA_DIR environment variable.
• Default path: ~/.local/share/opencode.

Testing Notes

• Tests rely on fs-fixture with using to ensure cleanup.
• All vitest blocks live alongside implementation files via if (import.meta.vitest != null).
• Vitest globals are enabled - use describe, it, expect directly without imports.