opencode-agent-pulse

Agent effectiveness analytics for OpenCode.

What it does

Every other OpenCode analytics tool counts tokens and dollars. This one answers the questions that actually matter:

• Is my agent working? → Health score (0-100) per session
• Which tools keep failing? → Tool success rates and latency, grouped by agent version
• Did my prompt change help? → Agent config change tracking with before/after metrics

Prerequisites

• Bun — required runtime (the plugin and CLI use bun:sqlite)

Install

bun add @arcevico/opencode-agent-pulse

Add to your opencode.json:

{
  "plugin": ["@arcevico/opencode-agent-pulse"]
}

Restart OpenCode. Done.

Data is stored locally at ~/.local/share/opencode/agent-pulse.db.

Usage

bunx @arcevico/opencode-agent-pulse

Example output:

=== Agents (last 7 days) ===
┌───┬────────┬────────┬──────────┬──────────┬────────────┬─────────┐
│   │ agent  │ health │ sessions │ avg cost │ avg tokens │ error % │
├───┼────────┼────────┼──────────┼──────────┼────────────┼─────────┤
│ 0 │ coder  │ 82     │ 15       │ $0.1200  │ 3800       │ 5%      │
│ 1 │ fast   │ 71     │ 8        │ $0.0400  │ 1200       │ 12%     │
└───┴────────┴────────┴──────────┴──────────┴────────────┴─────────┘

=== Evolution: coder ===
┌───┬──────────┬───────────────────┬──────────┬────────────┬──────────────────┬────────────────┬─────────┐
│   │ version  │ changed           │ sessions │ health     │ avg cost         │ avg tokens     │ error % │
├───┼──────────┼───────────────────┼──────────┼────────────┼──────────────────┼────────────────┼─────────┤
│ 0 │ a3f2...  │ 4/10/2026, 14:00  │ 5        │ 65         │ $0.1500          │ 4500           │ 12%     │
│ 1 │ b7c1...  │ 4/12/2026, 09:30  │ 8        │ 82 (+17.0) │ $0.1000 (-0.05)  │ 3200 (-1300)   │ 5%      │
└───┴──────────┴───────────────────┴──────────┴────────────┴──────────────────┴────────────────┴─────────┘

=== Tools by Agent Version (last 7 days) ===
┌───┬────────┬──────────┬────────────┬───────┬────────┬─────────┬────────┐
│   │ agent  │ version  │ tool       │ calls │ errors │ error % │ avg ms │
├───┼────────┼──────────┼────────────┼───────┼────────┼─────────┼────────┤
│ 0 │ coder  │ b7c1...  │ file_write │ 42    │ 1      │ 2.4%    │ 120    │
│ 1 │ coder  │ b7c1...  │ bash       │ 31    │ 3      │ 9.7%    │ 850    │
│ 2 │ coder  │ a3f2...  │ file_write │ 28    │ 4      │ 14.3%   │ 145    │
└───┴────────┴──────────┴────────────┴───────┴────────┴─────────┴────────┘

How it works

The plugin hooks into OpenCode's event system and tracks every session automatically.

When you modify an agent config file (.opencode/agents/*.md), the plugin detects the change in real-time via the file watcher — no restart needed. It hashes the new content, creates a new version, and all subsequent sessions are tagged with that version so you can compare performance across config changes.

Health score

Each session gets a 0-100 score based on:

| Signal | Weight | Logic | |---|---|---| | No errors | 30 | Full if clean, partial if 1 error | | No reverts | 25 | Did the user undo the agent's work? | | Low retries | 15 | Degrades per retry | | Tool success | 15 | Ratio of successful tool calls | | Reasonable steps | 15 | Sweet spot is 2-15 steps |

Data stored

Every session writes structured data to SQLite:

| Table | What's in it | |---|---| | sessions | Health score, cost, tokens, tool stats, errors, reverts, config version | | tool_executions | Per-tool success/failure, duration | | model_calls | Per-LLM-call tokens, cost, latency, errors | | step_metrics | Per-step token breakdown | | agent_changes | Every config change with timestamp, hash, and content snapshot |

License

MIT