taco-plugin

OpenCode plugin that captures raw telemetry for TACO benchmarking.

Captures per-turn, per-tool, and per-step metrics into a SQLite database. TACO CLI reads this database to show cache efficiency, tool I/O detail, retrieval relevance, latency breakdowns, and benchmark comparisons.

Zero configuration — runs entirely on Bun built-ins.

Usage

Register in opencode.json:

{ "plugin": ["taco-plugin"] }

Or from a local path during development:

{ "plugin": ["/path/to/taco-plugin"] }

Then run:

bun install && bun run build

Hooks

The plugin registers 6 OpenCode hooks:

| Hook | Fires When | Writes To | | -------------------------------------- | ------------------------ | --------------------------------------------------------- | | chat.params | Before each LLM request | chat_params, token_estimates | | experimental.chat.system.transform | System prompt assembled | system_prompts | | experimental.chat.messages.transform | Context window assembled | context_snapshots | | tool.execute.before | Tool call starts | tool_calls (partial) | | tool.execute.after | Tool call completes | tool_calls (update), tool_latency_breakdown | | event | SSE stream events | step_metrics, streaming_timing, retrieval_relevance |

Database

Location: ~/.local/share/taco/plugin.db

Uses WAL mode for concurrent reads from TACO CLI while the plugin writes.

Schema (12 tables)

• meta — schema version tracking
• chat_params — LLM request params per turn
• system_prompts — deduplicated system prompt snapshots
• context_snapshots — context window composition per turn
• tool_calls — full tool call data (input/output/timing)
• step_metrics — per-step token data from SSE events
• streaming_timing — streaming latency milestones
• token_estimates — multi-approach token estimates
• retrieval_relevance — how much tool output the model used
• tool_latency_breakdown — per-phase latency for tool calls
• benchmark_tasks — benchmark task definitions
• benchmark_runs — aggregated results per (task, session, strategy)

Tokenizers

Priority-based token counting:

| Model | Package | Accuracy | | ------------------------- | ------------------------------ | ---------- | | OpenAI (GPT-4o, o-series) | gpt-tokenizer | Exact | | Claude | @lenml/tokenizer-claude | Near-exact | | DeepSeek | @lenml/tokenizer-deepseek_v3 | Near-exact | | Everything else | Byte-ratio fallback (~3.8) | Heuristic |

Returns { count, method, warning? } for honest UI.

Commands

bun run build      # Compile TypeScript
bun test         # Run tests
bun run typecheck # Check types

Dependencies

Runtime:

• gpt-tokenizer — exact for OpenAI models
• @lenml/tokenizer-claude — near-exact for Claude
• @lenml/tokenizer-deepseek_v3 — near-exact for DeepSeek

Peer: @opencode-ai/plugin >= 1.14.29

License

MIT