sentry-rice 0.1.0

RICE-prioritise your Sentry issues, with AI scoring that traces each issue into your codebase (via Claude Code).

sentry-rice

RICE-prioritise your Sentry issues — with AI scoring that traces each issue into your codebase. Every issue gets one number, so the stuff actually worth fixing floats to the top. Browse it in a Sentry-styled web UI; override anything you disagree with; resolve straight through to Sentry.

RICE  =  (Reach × Impact × Confidence) / Effort

• Reach and the final RICE are computed deterministically — recent event volume on a log curve, relative to the busiest issue in the same environment (so a spicy staging issue ranks alongside a spicy prod one), decayed by recency. Never AI-assigned.
• Impact, Confidence, Effort come from an AI agent that reads the error and traces it into your actual code before judging. Impact is fixed per category; confidence/effort come from what the trace finds.

Everything opinionated — categories, thresholds, Sentry projects, the scoring rubric — lives in a single config.yaml. Nothing is baked to one org.

---

Requirements

• Python 3.11+
• A Sentry auth token (SENTRY_AUTH_TOKEN, or [auth] token in ~/.sentryclirc). Read scope is enough for syncing; resolving issues from the UI needs write scope.
• Claude Code — only for the AI scoring step. The deterministic engine, sync, web UI, overrides and resolve all work without it; AI scoring is the Claude-Code-native layer (see AI scoring).

How you consume it

This is a library you install + a config.yaml you own — not a repo you fork. Keep your own small repo with just your config, rubric, and .claude/ scoring commands; install the engine as a dependency and pull updates with a version bump instead of a fork-merge. (Fork only if you want to change the engine itself.)

Since it isn't published to a package index, install from a local clone:

git clone <this repo> ~/Server/side/sentry-rice

# in YOUR project repo:
python -m venv .venv && source .venv/bin/activate
pip install -e ~/Server/side/sentry-rice      # editable: engine updates as you pull

Once you host it on a private remote, pip install "git+ssh://…/sentry-rice.git" works too.

Quickstart

# 1. Configure
cp ~/Server/side/sentry-rice/config.example.yaml config.yaml
cp ~/Server/side/sentry-rice/rubric.example.md   rubric.md
$EDITOR config.yaml      # set sentry.org, projects, categories, codebase_path

# 2. Pull issues from Sentry (creates the DB, imports, scores reach/RICE)
export SENTRY_AUTH_TOKEN=...        # or rely on ~/.sentryclirc
sentry-rice --config config.yaml sync

# 3. Browse
sentry-rice --config config.yaml serve      # http://127.0.0.1:5001

At this point issues are imported and have a deterministic reach, but no AI judgment yet. To score them, set up the AI layer.

Tip: set SENTRY_RICE_CONFIG=config.yaml once and drop the --config flag.

AI scoring

Scoring is done by Claude Code sub-agents — one per issue — that read your rubric, trace the issue into project.codebase_path, decide impact/confidence/effort, and upsert the result. Reach is never sent by the agent; it's computed.

# Render the scoring commands into your repo's .claude/ (paths baked from config)
sentry-rice --config config.yaml init-claude .

This writes:

• .claude/commands/reimport.md — sync the last N days + score the new issues
• .claude/commands/reclassify.md — re-score everything against the current rubric
• .claude/commands/score-issue.md — score a single issue
• .claude/workflows/score-issues.js — the per-issue fan-out the commands launch

Then, inside Claude Code in your repo, run /reimport. It syncs, dumps the unscored issues, fans out one agent per issue, and recomputes. Edit rubric.md to change how everything is scored, then /reclassify.

No Claude Code? You can still use the tool: reach/RICE, the UI, overrides and resolve all work. Score impact_category/confidence/effort yourself by piping JSON to sentry-rice upsert, or via the override form in the UI.

Configuration

A single YAML file. See config.example.yaml for the full annotated schema. The shape:

Section	What it sets
project	name, codebase_path (where agents trace), rubric_file
sentry	org, region_url, projects (id → app), environments, prod_environments
thresholds	sync_days, per-env min_events floors, reference_floor, recency_decay, rice_bands
categories	name → { score 0–10, icon (lucide), color } — impact is fixed per category
ui.fix_prompt_template	the "copy fix prompt" body (optional; {sentry_id} etc. interpolated)
db.path	SQLite path (default <config dir>/db/rice.db; or RICE_DB_PATH)

CLI

sentry-rice initdb                 create / migrate the database
sentry-rice sync [--days N]        pull Sentry, import new, prune stale, recompute
sentry-rice serve [--port 5001]    run the web UI
sentry-rice dump [--all] [PATH]    write issues to JSON for the scoring fan-out
sentry-rice recompute              re-derive reach + RICE for all issues
sentry-rice upsert [PATH]          upsert one scored issue (JSON; stdin if no PATH)
sentry-rice init-claude [DEST]     render the .claude scoring commands into DEST

A worked example

examples/acme/ is a complete configuration for a fictional B2B SaaS, "Acme Cloud" (all values made up) — 10 categories, three projects (backend/web/worker), prod+staging+dev, and a matching rubric.md. Copy it as a starting point.

Development

pip install -e ".[dev]"
pytest          # 41 tests: config, db, scoring, store, sentry (mocked), web, cli

Disclaimer

This is an independent, unofficial project. It is not affiliated with, endorsed by, or sponsored by Functional Software, Inc. or the Sentry project. "Sentry" is a trademark of its respective owner and is used here only to describe what this tool interoperates with.

License

MIT — see LICENSE.