Use spare tokens for good.
Turn unused agent/model capacity into careful open source maintenance. OSS Supporter is an operating playbook for finding small, useful OSS work, validating it, doing it well, and leaving maintainers with less burden than before.
It is model-agnostic, agent-agnostic, and OS-agnostic: use any assistant, CLI, editor, or manual workflow that can follow the gates.
Latest public receipt: June 7, 2026 checkpoint.
| Tokens used | Tracked public items | Open PRs tracked | Merged PRs visible on GitHub | Accepted/adopted outcomes |
|---|---|---|---|---|
| 1,523,011,192 | 126 | 93 | 52 | 60+ |
Recent accepted work includes focused fixes in OpenTelemetry JS, OpenTelemetry .NET, containerd/nerdctl, DataDog Saluki, Microsoft Physical AI Toolchain, Microsoft VS Code DocumentDB, Elastic CLI, JFrog Credentials Provider, NVIDIA-NeMo Curator, NVIDIA NemoClaw, Mellanox Network Operator Docs, Loxia OnBuzz, Open QA, monday vibe, Shopify Roast, and shadcn-vue.
Give this repo link to your agent, then paste:
You are helping me use OSS Supporter.
Read this repo as the public playbook. Keep private workbench files, local
clones, raw logs, credentials, account state, and personal notes outside this
repo.
Start by asking me where I want to begin:
1. Set up my private OSS workbench
2. Find new OSS projects to help
3. Follow up existing PRs or issues
4. Prepare a focused PR or comment
5. Publish an impact update
6. Just explain the workflow
Do not scan, clone, patch, post, or open a PR until I choose a starting point.
Do not take public action until I approve the target and message.
See START.md for first-run setup, modes, and next steps. Agents can use AGENT_START.md as the shortest entrypoint.
Pick one mode:
| Mode | Use when | First result |
|---|---|---|
| Learn | You are new to the workflow | A short explanation and recommended next step |
| Set up | You want a private local workbench | A safe folder layout and starter ledgers |
| Scout | You want new OSS leads | Candidate issues after reject checks only |
| Follow up | You already have PRs or comments open | A compact action list |
| Prepare | You approved one target | A focused comment, repro, docs fix, test, or PR plan |
| Impact | You want a public receipt | Sanitized aggregate impact draft |
Default first run: ask where to start. No surprise scans, no public action.
These helpers automate the boring checks while leaving judgment with the operator:
| Step | Helper | Result |
|---|---|---|
| Score a lead | tools/lead-score |
proceed / watch / skip signal |
| Preflight a repo | tools/repo-capability |
likely checks and local blockers |
| Check generated surfaces | tools/source-truth |
template/source hints |
| Draft PR text | tools/pr-body-builder |
body-file safe Markdown |
| Check submitted work | tools/status |
compact follow-up actions |
| Count impact | tools/token-meter |
aggregate token totals |
| Guard public export | tools/public-boundary |
private-data leak check |
| Find cleanup wins | tools/cleanup-doctor |
dry-run heavy-folder list |
See docs/workflow-upgrades.md.
Use these defaults even when the agent is autonomous:
| Boundary | Default |
|---|---|
| Public action | Human approves the target and message |
| Private data | Keep raw logs, credentials, local notes, and claims outside this repo |
| Scouting | Reject stale, duplicate, crowded, or unclear targets before cloning |
| Patches | Prefer focused bug fixes, tests, repros, CI fixes, SDK/API correctness, and small code changes |
| Disclosure | Mention agent assistance plainly in PRs/comments |
| Impact | Publish aggregate totals and links, not raw sessions |
Most of the repo is plain Markdown. Tools are optional:
| Need | Requirement |
|---|---|
| Read the playbook | Any browser, editor, or agent |
| Run Python helpers | Python 3.10+ |
| Check GitHub PR/issue status | GitHub CLI authenticated with gh auth login |
| Use PowerShell status helpers | PowerShell 7+ preferred; Windows PowerShell usually works |
| Parse Codex impact locally | Access to your own local Codex session logs |
| Publish the repo | Run tools/public-boundary before pushing |
Many developers have unused model capacity: idle subscription windows, spare API budget, local agent time, or small gaps between larger work. This project turns that capacity into practical OSS help:
- bug reproduction
- issue triage
- docs fixes
- focused tests
- small CI repairs
- low-risk patches
- respectful review follow-up
The point is not contribution graph farming. The point is useful maintenance. If the action mostly helps you and not the maintainer, skip it.
- A public, reusable method for agent-assisted OSS support.
- A set of validation gates before public action.
- Templates for scoped slices, comments, PRs, and status ledgers.
- Portable skill packs for scout, PR, and review workflows.
- A privacy-safe way to publish aggregate impact.
- A durable memory system for lessons learned from maintainer feedback.
- Not a bounty-hunting system.
- Not a bot for mass comments.
- Not a replacement for human judgment.
- Not a way to hide AI assistance.
- Not a public dump of private agent logs or working directories.
- Scout: find likely-useful issues or PR follow-ups.
- Reject fast: archived, stale, duplicate, crowded, unclear, or high-risk.
- Claim locally: avoid duplicate action across operators or agents.
- Slice: create a small work record with target, scope, validation, and risks.
- Verify: reproduce, patch, and test at the right depth.
- Disclose: be plain about agent assistance when publishing work.
- Follow up: read reviews, fix small valid feedback, and record lessons.
- Clean up: move merged/closed/dropped work out of active state.
This repo is designed to be public. Keep private workbench data elsewhere:
- raw chats and prompts
- credentials and account state
- CLA/legal status details beyond public facts
- private queue notes
- local clone directories
- unredacted session logs
- private identity details
Publish methods, templates, sanitized examples, and aggregate impact.
See docs/publication-boundary.md.
assets/ public README and docs visuals
docs/ reusable playbooks and policies
examples/ sanitized public case studies
impact/ aggregate public impact snapshots
skills/ portable Markdown agent workflows
templates/ copyable issue, slice, PR, and status templates
tools/ optional local helpers; never require a specific agent
The public repo includes a local token usage scanner:
python tools/token-meter/token_meter.py --include-cwd oss-supporterIt reads Codex JSONL logs and emits aggregate-only token totals. It does not publish prompts, responses, raw paths, or session IDs.
The repo also includes GitHub CLI helpers for checking submitted PRs/comments:
pwsh ./tools/status/oss-status.ps1 -LeadClaimsPath work/lead-claims.mdSee tools/status.
Score a lead before cloning:
python tools/lead-score/lead_score.py owner/repo#123Check a repo before promising tests:
python tools/repo-capability/repo_capability.py path/to/repoCheck generated/docs source-truth risk:
python tools/source-truth/source_truth_check.py path/to/repo --touched docs/example.mdUse the meter to create a private .local.* snapshot, review it, then publish a
sanitized public receipt under impact/.
Latest public receipt: June 7, 2026 checkpoint.
Public receipts should show the useful part:
- tokens spent on OSS support
- projects helped
- PRs, comments, reproductions, and review fixes
- merged or maintainer-accepted outcomes
They should not show private logs, local paths, raw conversations, account data, or hidden queues.
Read these in order:
- START.md
- docs/operating-model.md
- docs/validation-gates.md
- docs/public-voice.md
- docs/review-followup.md
- docs/maintainer-lessons.md
- docs/workflow-upgrades.md
- docs/impact-ledger.md
- docs/adapters.md
Then copy templates/work-slice.md for the first small support attempt.
Optional agent workflows live in skills/. For a fake private workbench shape, see examples/minimal-workbench. For a merged-work lifecycle example, see VS Code DocumentDB. For a meatier code review loop, see DataDog Saluki.
Before publishing copied workbench material, run the public-boundary checker:
python tools/public-boundary/check_public_boundary.pyMIT. Take it, adapt it, remix it, and use it to help maintainers.