RewardOps

Evidence before effort.

RewardOps is a cross-channel agent that protects developers from wasting hours on closed, stale, vague, or overcrowded software rewards. Send it a public GitHub issue and it verifies the live source, extracts reward evidence, counts visible attempts, applies hard-stop risk gates, and returns an explainable pursue/research/ skip decision.

The same Caspian on_message handler serves every connected channel. The hackathon deployment is designed for email plus Slack; no handler is duplicated and no channel-specific business logic exists.

Why this exists

Reward marketplaces frequently preserve an old card after its underlying GitHub issue has closed. A large dollar amount can also hide dozens of active attempts, an assignee, missing payout mechanics, or a two-line scope. RewardOps checks the source of truth at decision time and refuses to score unreachable sources from stale cached data.

What makes the agent different

• Hard-stop verification: a closed issue always scores SKIP, regardless of the advertised amount.
• Evidence extraction: common $100, 100 USD, 350 USDC, and /bounty 100 formats are supported, with the matching source lines retained.
• Currency-safe scoring: token-only rewards such as MRG, RTC, BTC, or credits are never displayed as USD without independent conversion evidence.
• Competition sensing: visible /attempt, /claim, and work-intent comments reduce the score.
• Explainable decisions: every score includes positive signals and risk flags.
• Persistent watchlist: watch saves evidence to SQLite; digest ranks it.
• One Caspian handler: email and Slack route through exactly the same command and response path, including rich cards and clean text fallbacks.
• Reply-safe email parsing: quoted reply history is removed before a command is evaluated, so email clients cannot accidentally turn one command into many.

Commands

scan https://github.com/<owner>/<repo>/issues/<number>
watch https://github.com/<owner>/<repo>/issues/<number>
digest
help

Quick start

Requires Python 3.11+.

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env

On Windows PowerShell, activate with:

.\.venv\Scripts\Activate.ps1
pip install -e ".[dev]"
Copy-Item .env.example .env

Create a Caspian project and connect two channels:

pip install caspian-cli
caspian init
caspian connect email
caspian connect slack

Authorize the returned Slack link, then start the one-handler agent:

rewardops-agent

For higher GitHub API limits, add GITHUB_TOKEN to .env. Never commit .env.

Local evidence demo

The CLI exercises the exact same service used by Caspian:

rewardops scan https://github.com/BasedHardware/omi/issues/2316

Even if a marketplace still advertises this old $1,000 bounty, RewardOps checks GitHub live and returns SKIP because the issue is closed. That stale-card failure mode is the core demo.

Architecture

Email ─────┐
           ├─ Caspian normalized message ─ one handler ─ evidence verifier
Slack ──────┘                                     │
                                                  ├─ live GitHub API
                                                  ├─ scoring + risk gates
                                                  └─ SQLite watchlist

Verification

ruff check .
ruff format --check .
pytest

Tests cover URL safety, reward extraction, the closed-source hard stop, competition penalties, watchlist persistence, and channel-independent command handling.

The hackathon submission draft and live recording plan are in docs/DEVPOST_SUBMISSION.md and docs/DEMO_SCRIPT.md.

Responsible use

RewardOps only reads public issue metadata. It does not contact maintainers, submit claims, scrape private data, or perform security testing. A high score is evidence for human review, never a guarantee of payment.