opencode-perf-stats

Performance analytics for OpenCode sessions — TPS, TTFT, tokens, cost, and comparisons, with an interactive web UI.

Key Features

• Interactive web UI — discover, aggregate, trend, and compare in the browser
• CLI available — analyze sessions straight from the terminal
• Self-contained HTML reports — shareable Chart.js charts, no server needed
• Model-aware aggregation — per-model p50/p95 breakdowns
• Low-confidence filtering — noisy short messages auto-excluded from stats

Demo

Table of Contents

• Key Features
• Demo
• Install
• Quick Start
• Web UI
• Modes
• HTML Reports
• How Metrics Are Calculated
• CLI Reference
• Database Location
• Roadmap
• Development
• Contributing
• License

Install

Requires Python 3.12+. Developed and tested on Python 3.12.

# Recommended (includes the web UI)
pip install "opencode-perf-stats[web]"

Or with pipx (recommended for CLI tools):

pipx install "opencode-perf-stats[web]"

Or with uv — as a persistent tool install:

uv tool install "opencode-perf-stats[web]"

Or run without installing:

uvx opencode-perf-stats --help

⚠️ uvx runs the CLI in an isolated environment without the [web] extra, so opencode-perf-stats serve will fail with an ImportError for Flask. To use the web UI, install with pip or pipx as shown above.

CLI-only install (no web UI):

pip install opencode-perf-stats

Quick Start

# Launch the interactive web UI (recommended)
opencode-perf-stats serve

# …or analyze from the terminal:
opencode-perf-stats                  # most recent session
opencode-perf-stats --days 7         # aggregate, last 7 days
opencode-perf-stats --list           # list recent sessions

Web UI

An interactive browser UI covers every mode with modern, intuitive UX: session discovery, single-session reports, aggregate views, per-period time analysis, and side-by-side comparison. Requires the optional web extra:

pip install -e ".[web]"      # or: pip install "opencode-perf-stats[web]"
opencode-perf-stats serve    # opens http://127.0.0.1:5000/ in your browser

The UI reuses the same dark theme and Chart.js charts as the standalone HTML reports. Features:

• Discover (/) — filter by days/model, click a row to open a session, checkbox-select 2–4 sessions and use the sticky "Compare selected" basket.
• Aggregate (/aggregate) — TPS/TTFT/tokens/cost across filtered sessions with per-model charts and top-sessions tables.
• Trends (/trends) — time-series analysis: bucket metrics by day, week, month, or year (shared period selector), with stacked per-model charts for TPS, TTFT, tokens, cost, sessions, and messages. Each chart is independently toggleable. Defaults to the last 30 days.
• Compare (/compare) — compare sessions or models side-by-side with grouped bar charts and a comparison table.
• Single session (/session/<id>) — full report with TPS/TTFT/token charts and a per-message detail table; final_only toggle.

opencode-perf-stats serve --port 5001      # custom port
opencode-perf-stats serve --no-browser     # don't auto-open a browser
opencode-perf-stats serve --db /path/to/opencode.db

⚠️ Binding to a non-loopback --host is insecure (no auth/CSRF). Keep it on 127.0.0.1 for local use.

Modes

Single-Session (default)

Reports TPS, TTFT, reasoning TTFT, session duration, and token/cost breakdown for one session.

opencode-perf-stats                         # most recently updated session
opencode-perf-stats ses_abc123              # specific session
opencode-perf-stats --final-only            # only finish='stop' messages

Aggregate

Aggregates TPS/TTFT/tokens across all matching sessions, with per-model breakdown.

opencode-perf-stats --days 7                # last 7 days
opencode-perf-stats --days 30 --model mimo  # filtered by model
opencode-perf-stats --days 7 --json         # JSON output

Discovery

List recent sessions with IDs, titles, and metadata.

opencode-perf-stats --list                  # recent sessions table
opencode-perf-stats --list --days 7 --json  # filtered, JSON

Comparison (experimental)

Compare sessions or models side by side.

# Compare up to 4 sessions
opencode-perf-stats compare sessions ses_a ses_b ses_c ses_d

# Compare models
opencode-perf-stats compare models mimo gpt-4 claude

# JSON output
opencode-perf-stats compare sessions ses_a ses_b --json

HTML Reports

Generate self-contained HTML reports with interactive Chart.js charts:

# Save a single-session report to a file
opencode-perf-stats --html report.html

# Aggregate report filtered by days
opencode-perf-stats --days 7 --html report.html

# Pipe HTML to stdout (bare --html and --html - are equivalent)
opencode-perf-stats --html > report.html
opencode-perf-stats --html - | head -n 20

Charts include:

• TPS per message — bar chart with low-confidence markers
• TTFT per message — time to first token visualization
• Token breakdown — doughnut chart (input/output/reasoning/cache)
• Message summary — final vs tool-calls vs incomplete
• Per-model comparison — grouped bars (aggregate mode)
• Top sessions — horizontal bar chart (aggregate mode)

How Metrics Are Calculated

All metrics are derived from the timing and token data OpenCode records in its SQLite database. This tool reads that data — it does not time requests itself.

TPS (Tokens Per Second)

• Measured per assistant message.
• Formula: output tokens / generation duration.
• Generation duration = time between when the message was created and when it completed (both recorded by OpenCode).
• Only output tokens are counted — input, reasoning, and cache tokens are excluded.

TTFT (Time To First Token)

• Measured per assistant message, in milliseconds.
• The time from when the message was created to when the first streaming part (text or reasoning) started.
• Considers all part types, so messages that only contain tool calls still get a TTFT via their reasoning part.

Cost

• Read directly from OpenCode's database — this tool does not compute pricing.
• Per-message cost and per-session cost are summed for aggregates and per-model breakdowns.

Low-confidence filtering

• Messages with very few output tokens (< 20) or very short durations (< 1 s) produce noisy TPS values and are flagged as low-confidence.
• In aggregate views, low-confidence messages are excluded from TPS stats.
• When fewer than 20 samples are available, p95 values are flagged as unreliable.

Aggregation

• TPS and TTFT are reported as mean, median (p50), p95, min, and max.
• p95 uses linear-interpolation percentile (the NumPy-default method).
• Per-model breakdowns group messages by provider/model identity.

CLI Reference

opencode-perf-stats [session_id] [options]

Positional:
  session_id              Session ID (default: most recently updated)

Options:
  --version               Show version
  --final-only            Only include finish='stop' messages
  --json                  Output as JSON
  --html [FILE]           Generate interactive HTML report;
                          writes to FILE, or stdout when FILE is
                          omitted or '-'
  --db PATH               Path to opencode.db
  --days N                Filter to last N days (triggers aggregate mode)
  --model SUBSTRING       Filter by model ID (triggers aggregate mode)
  --list                  List recent sessions and exit

Subcommands:
  compare                 Compare sessions or models
  serve                   Launch the web UI (requires the [web] extra)

Database Location

The tool reads from OpenCode's SQLite database:

$XDG_DATA_HOME/opencode/opencode.db
# or
~/.local/share/opencode/opencode.db

Override with --db /path/to/opencode.db.

Roadmap

These features are planned for upcoming releases:

• Web UI color customization — let users switch the dark-theme palette (chart colors, accent, background) via config or environment variables.
• Budget tracking — define per-model prices and monthly/weekly budgets. The tool will warn when a session or period exceeds the configured threshold and show remaining budget in aggregate views.
• More analytics splits — group metrics by message type (stop vs tool-calls) and by individual tool name, so you can see which tools or message categories are the most expensive or the slowest.

Development

Python 3.12 is pinned in .python-version; dependencies are locked in uv.lock.

git clone https://github.com/mtsanaissi/opencode-perf-stats.git
cd opencode-perf-stats
uv sync --extra web     # creates .venv from uv.lock (includes the dev group)

# Run tests
uv run pytest tests/

# Run the tool
uv run opencode-perf-stats --help

Contributing

Contributions are welcome! Here's how to get started:

1. Fork the repository and create a feature branch.
2. Install development dependencies: uv sync --extra web
3. Make your changes and add tests if applicable.
4. Run the test suite: uv run pytest tests/
5. Open a pull request against main.

For bugs and feature requests, please open an issue.

License

MIT