Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
281 lines (223 loc) · 12.8 KB

CONTRIBUTING.md

File metadata and controls

281 lines (223 loc) · 12.8 KB
Raw

Contributing To LoopX

You want to help improve LoopX? Great, and thank you. Contributions come in many shapes, and not all of them are code:

  • filing clear bug reports with reproduction steps;
  • triaging and reproducing issues;
  • improving documentation, examples, and smoke tests;
  • answering questions in issues or discussions;
  • reviewing pull requests and helping contributors navigate the review flow;
  • following the Code of Conduct;
  • implementing a public task or fixing a bug.

Every one of these helps. The rest of this guide covers finding work, keeping public/private boundaries intact, validating changes, and getting a pull request merged.

The best code contributions are small, reviewable, and tied to a public task or a clear bug.

Code Of Conduct

Everyone participating in LoopX community spaces, including maintainers and contributors, is expected to follow the Contributor Covenant. Reports of unacceptable behavior can be sent to huangrt01@163.com; maintainers review every report promptly and keep the reporter's identity confidential to the extent possible.

Find Work

Start with the current technical directions to understand the active programs and their maturity, then use docs/development/contributor-tasks.md to find public work that is useful, claimable, and safe to discuss in the repository.

If you do not see a matching task:

  1. open a GitHub issue with the contributor task template;
  2. explain the problem, proposed scope, touched files, and validation command;
  3. wait for maintainer feedback before starting large or behavior-changing work.

Small docs typo fixes and obviously safe cleanups can go straight to a pull request.

Public And Private Boundaries

LoopX coordinates local agent state, so some files are runtime data and must stay out of public contributions:

  • do not commit .loopx/, .codex/goals/, or live ACTIVE_GOAL_STATE.md files;
  • do not publish private benchmark traces, verifier output, raw agent sessions, credentials, internal document links, or local machine paths;
  • do not run or duplicate maintainer-owned benchmark cases unless a maintainer has split out a public issue for that work.

Safe contribution surfaces include docs, examples, smoke tests, CLI diagnostics, schema docs, dashboard UI code, and sanitized fixtures.

Run the public/private scan before sending docs or examples:

loopx check \
  --scan-path README.md \
  --scan-path CONTRIBUTING.md \
  --scan-path docs/development/contributor-tasks.md \
  --scan-path docs/ \
  --scan-path examples/

Local Development

Use the developer guide as the stable entry point. Before changing scheduler, quota, todo/gate, onboarding, agent-facing output, or release behavior, read the bilingual testing and quality guide. Before adding or consolidating a public smoke, use the bilingual good smoke guide to define its durable invariant, independent oracle, cadence, and public-safe fixture boundary.

For source development, run commands from the repository or dedicated worktree root with uv. It manages a compatible Python and installs the current checkout in the project environment, keeping checks separate from a globally installed LoopX release. See the local validation commands for environment, lockfile, and CI boundaries.

uv sync --extra test
uv run --extra test python -m ruff check tests loopx/canary loopx/control_plane loopx/domain_packs loopx/presentation
uv run --extra test python -m mypy
uv run --extra test python examples/control_plane/cli-output-budget-regression-smoke.py
uv run --extra test python -m pytest -q
uv run --extra test loopx canary premerge --from-git-diff
uv run --extra test loopx check --scan-path loopx/ --scan-path tests/ --scan-path examples/ --scan-path docs/
git diff --check

Choose focused smokes and broader canaries by change risk; do not run every public smoke or a live model call for every patch. The quality guide explains the CI, local/manual, and release-only boundaries.

License And DCO Sign-Off

LoopX's unified open source core is licensed under the Apache License 2.0. Unless you explicitly state otherwise before submission, contributions accepted into this repository are submitted under Apache-2.0 without additional terms or conditions. LoopX does not require a copyright assignment or contributor license agreement.

Every pull-request commit must certify the Developer Certificate of Origin 1.1 by including a sign-off trailer:

Signed-off-by: Your Name <your.email@example.com>

Create that trailer with Git's -s option:

git commit -s -m "feat: describe the change"

The name and email must identify the person making the certification and must be information you are permitted to publish in the permanent Git history. If a commit is missing the trailer, amend it with git commit --amend -s or use an interactive rebase to sign the affected commits, then update the pull-request branch. The DCO pull-request check rejects unsigned commits.

Releases through v0.4.7 remain under their original MIT terms. See the licensing and v0.4.8 transition policy for the historical notice, patent-grant boundary, and open-core scope.

Experimental Features

Use loopx/experiments/<experiment-id>/ for an opt-in prototype that does not yet have a stable caller contract or participate in LoopX's default lifecycle. Keep its tests, examples, and scripts under matching experiments/ paths so the prototype can be evaluated or removed as one unit. Core modules must not import experimental packages. See the experiment placement and promotion policy.

Host Loops And LoopX Turn

Treat LoopX Turn and a long-running host loop as separate layers:

  • loopx turn run-once is one atomic governed transaction. It may decide, invoke one bounded host segment, validate independently, write back, spend once, and project the latest scheduler phase.
  • A Turn Loop Controller is an outer runtime owner. It decides when to wake, invokes run-once, consumes the typed result, applies the shared scheduler_hint, and either waits, routes a user action, repairs, replans, continues, or stops.
  • A host adapter translates one typed request and result. It owns the opaque host session and tools, but it does not own LoopX state, quota, completion, validation, scheduler policy, or replan policy.

Do not add a sleep loop, cron implementation, recurring daemon, operator notification path, or multi-Turn replan loop inside run-once. Do not copy Codex App heartbeat prompt rules into a second scheduler. Reuse the existing interaction, scheduler, autonomous-replan, todo, and TurnEnvelope contracts; only the runtime-specific act of applying a wakeup belongs in a scheduler adapter.

A replan_required result is not permission to invoke the same todo again. A controller must first record a bounded todo or vision delta, obtain a fresh TurnEnvelope, and preserve the causal (goal_id, agent_id, todo_id) frontier. An opaque resumable host session is recovery metadata, not authority to bypass that decision.

Stage host-loop contributions in reviewable slices:

  1. characterize current Codex App and Turn behavior with independently derived fixtures;
  2. add a pure next-disposition decision table with no host or state effects;
  3. add one scheduler-owner adapter with a fake clock and fake host;
  4. add runtime-specific wakeup, notification, or presentation only after the shared transition contract is stable.

For every controller or host-loop change, prove:

  • scheduler owner, host surface, and execution mode are explicit and valid;
  • wait, user action, monitor-only, and cadence-only paths make no model call and spend no quota;
  • material progress requires independent postcondition validation before durable writeback and spend;
  • replay and interrupted-phase recovery are idempotent;
  • repair and replan remain distinct, and replan produces a fresh frontier before another Turn; and
  • fixtures contain no raw prompts, transcripts, credentials, private state, or host-local paths.

See the LoopX Turn protocol and the Contributor Task Board for the staged controller plan.

Governance And Attribution

Repository roles and decision authority are defined in Governance. Creator and contributor attribution is recorded in docs/project/authors.md, while path-scoped maintenance and preferred review assignments are recorded in the same governance document. The public Git history records individual contributions. Contribution does not automatically grant merge or release authority, and an agent or automation identity is not a human maintainer.

When naming or packaging a fork, integration, or hosted service, follow the project's name and marks guidance.

For dashboard changes:

cd apps/presentation/dashboard
npm install
npm run build
npm run smoke:demo-readiness

Claiming A Task

  • Comment on the issue before starting non-trivial work.
  • If a maintainer marks it claimed or assigns it to you, keep the scope close to the issue.
  • If you get stuck, comment with the blocker and what you already tried.
  • If you need to change the scope, ask first.
  • If there is no update for 14 days, maintainers may release the task so someone else can pick it up.

Pull Request Checklist

Before opening a pull request:

  • link the issue or task ID when one exists;
  • state the requested outcome, current gap and observable before/after result;
  • distinguish completion of the scoped task from a justified increment; for an increment, name the remaining gap, next owner/dependency and why the boundary is independently testable and reversible;
  • link decisive validation to that outcome, including relevant user-entrypoint readback and failure/recovery cases;
  • keep unrelated formatting or refactors out of the PR;
  • include docs or tests when changing user-visible behavior;
  • confirm that no private/local runtime state was committed.

Use the overall roadmap for cross-cutting work, without inventing roadmap ids for ordinary fixes. Existing issues and canonical Todos own execution; update them instead of duplicating follow-up work. A completed task needs no invented successor. Prerequisites, research, docs and maintenance can be useful delivered outcomes. A schema, message, mock or passing suite alone does not complete a promised user journey.

Maintainers may request consolidation when a useful outcome was unnecessarily split, or a smaller PR when unrelated concerns were mixed. Review evaluates the verified goal delta and evidence, not minimum size, model identity or PR count.

Validation disclosure

Use the PR template to report facts, not a self-assessed quality grade. The enum values are author declarations, not an automated proof or merge gate. finished means execution ended, not that all checks passed or that coverage is sufficient. Record the tested commit; after changes, rerun affected checks or identify the stale evidence and remaining gap.

Use one row per relevant check, including failures, skips (not_run), blocked checks and work still running. Static checks do not prove runtime behavior; unit/mocked tests do not prove the public entrypoint or a real backend. State the behavior checked and why the set covers the changed paths. Refactors need the real-path and parity evidence required by the testing guide. Small documentation changes can report a link/render/static check and explain why runtime testing is not applicable; do not launch unrelated suites to fill rows.

Keep evidence public-safe at the point of entry, including hidden HTML comments and attachments. Report repository-relative test commands, public fixture names, aggregate outcomes, backend product/version and isolation mode, or public CI links. Do not copy commands containing private arguments, raw logs, snapshots, prompts, private screenshots, infrastructure addresses, connection strings, local paths or credentials. authorized_private_read_only discloses only a data category: it neither grants permission to access live state nor requires publishing its content, identifiers or fingerprints. When evidence cannot be shared, describe the tested behavior and verification limitation; public reproducibility remains an explicit gap until a safe reproducer or authorized review is available. Never upload private evidence to make a checkbox green.