LorenEteval_Furious/AGENTS.md

Furious repository guidance

Learn before changing

• Treat the checked-out tree, tests, build configuration, and verified runtime behavior as the immediate source of truth. Existing guidance is a maintained model, not an authority that can make itself true.
• If .codegraph/ exists, use CodeGraph before broad text searches for structural questions; use rg for exact follow-up. Inspect callers, tests, persisted formats, platform branches, and packaging consumers before changing a contract.
• For substantial work, use this loop: understand the intended owner and invariant; form a hypothesis; trace the real call/runtime path; implement at the owning boundary; test real behavior; then re-evaluate the architectural model.
• When guidance says A and code appears to do B, inspect the call path and tests. Decide whether B is intentional evolution, compatibility debt, or a bug; preserve the intended invariant and update either code or the narrowest applicable guidance. Never resolve the contradiction by deleting the scoped guidance file. Do not encode speculation, experiments, or incidental class names as durable rules.

Preserve the guidance hierarchy

• Every existing AGENTS.md and AGENTS.override.md path is an established documentation scope. During ordinary guidance maintenance, do not delete, rename, relocate, merge, consolidate, or change the kind of any existing file. Improve a seemingly redundant scope in place by making its inherited and local rules more precise.
• Add a new AGENTS file only when verified architecture has a durable uncovered scope that no existing file can represent. A new scope never makes an existing one disposable. Keep override files explicit about which inherited assumption they replace and why.
• Before a hierarchy-wide audit, inventory tracked, untracked, hidden, and ignored AGENTS paths, including overrides; record each scope and its nearest ancestor guide. At handoff compare exact path sets and Git status/diff: no original path may disappear or become a rename. Default to exact equality and improve redundant scopes in place.
• Inheritance follows directory ancestry. A sibling guide identifies a contract to consult, not another parent.

Operating model

• Furious is a cross-platform PySide6 desktop client. Entrypoints reach Furious.__main__; a parent process translates child exit/crash results, while the Qt application performs singleton election, composes process-lifetime owners, runs the event loop, and unwinds acquired resources in reverse order.
• Read the packages as cooperating responsibilities rather than a strict stack: Interface and Models define dependency-light contracts/domain values; Repository persists them; Plugins defines backend/extension contracts; bundled Backends and Extensions implement those contracts; Service owns workflows and temporary resources; Controllers owns shared state transitions; Qt, Widget, Window, and Actions present them; Application composes the process; Frozenlib contains compatibility, settings, and host integration boundaries.
• Official proxy backends are Xray, Hysteria 1, Hysteria 2, and External Core. Shared orchestration asks plugin capabilities; backend-specific parsing, runtime preparation, routing, TUN, statistics, and exit interpretation stay behind those capabilities.

Project-wide invariants

• Keep one authority for each state and one owner for each resource. Multiple UI surfaces observe shared controllers/models; they do not copy connection, routing, System Proxy, TUN, subscription, or test state.
• Treat persisted profiles and plugin documents as input. Prepare runtime, routing, probe, and TUN state on explicit copies unless an API deliberately mutates storage. A failed pre-commit stage leaves persistence unchanged; a failed post-commit side effect is reported without pretending the commit rolled back. Identify the unit of commit: cancellation of a batched operation may preserve completed batches rather than roll back the entire command.
• Use stable domain identity, not table rows, proxy indexes, display text, or object position. Async results additionally prove that the target generation/fingerprint is still current before mutation.
• Distinguish profile identity, subscription membership, remote synchronization ownership, and execution snapshots. Moving a profile into a group does not transfer remote ownership; a running core uses its prepared document even when the live profile later changes.
• Startup and other staged workflows own every resource acquired before commit and roll back only that attempt on failure, cancellation, or supersession. Cleanup is bounded where responsiveness requires it, idempotent, and targets exact processes, threads, replies, timers, files, handles, routes, and callbacks—never process names.
• Keep new GUI-thread work bounded through an owned worker or asynchronous Qt boundary; workers publish data back to the owning Qt thread and never mutate widgets or live repositories directly. Existing synchronous compatibility and host-operation paths require explicit responsiveness review: an async entry point alone does not prove non-blocking preparation, cancellation, or shutdown.
• Validate user, network, persisted, and plugin data at boundaries. Keep invariant failures visible, preserve useful diagnostics, and never log credentials, subscription payloads, full share links, environments, or complete core documents.
• Preserve source execution, wheel/sdist installation, and Nuitka/native distributions. Importing dependency-light layers must not construct the application, discover plugins, start runtimes, or create UI.

Working in the tree

• Preserve unrelated and unstaged user changes. Do not revive deleted experiments from history or broaden a task to nearby technical debt.
• Before Python work, prefer an existing root .venv*/venv* interpreter. Do not create or mutate an environment without need. Format only touched Python files with python -m black <files>, then python -m black --check <files>; pyproject.toml preserves string quotes. Syntax/import checks supplement behavior tests. Documentation-only work does not require unrelated formatting or generated-file refreshes.
• Preserve GPL headers, from __future__ placement, import grouping, and established naming. Search consumers before changing public exports, plugin APIs, persisted keys/schemas, IDs, aliases, migrations, package data, or semantic exit codes.
• Generated and curated artifacts have separate sources of truth: never hand-edit Furious/Frozenlib/AppResources.py; update Resources.qrc/resource inputs and regenerate it. Follow Furious/Externals/AGENTS.md for the translation catalog and Furious/Data/AGENTS.md for bundled assets.
• Dependency, Python/Qt floor, entrypoint, package-data, version, or artifact changes may span pyproject.toml, setup.py, requirements.txt, Deploy.py, and the release workflow. Review every applicable surface rather than assuming one declaration is canonical. Networked Deploy.py --download and destructive build cleanup run only when explicitly in scope.

Verification

• Run the narrowest relevant test first, then the affected tier documented in tests/README.md. Tests use isolated settings and mocked host/network boundaries; they never mutate a real proxy, TUN, routing table, startup registration, desktop, or unrelated process.
• Match evidence to the contract: round trips/migrations for models and repositories; exact transitions/signal counts for controllers; stale/cancel/rollback/cleanup paths for services; partial startup and resource reaping for runtimes; mocked OS branches for host helpers; import/discovery and packaged checks for compiler-sensitive changes. Test a resource that refuses cleanup as well as one that exits normally. A terminal flag, cleared reference, or elapsed timeout is not evidence of native resource release; distinguish that observation from the intended guarantee.
• Report source inspection, executed tests, mocked platform evidence, and packaged validation separately. A passing source suite does not prove native distributions or every declared Python/Qt floor. Release import checks do not replace behavioral tests; record untested targets and compatibility gaps explicitly.
• Use real Qt semantics when focus, selection, keyboard modifiers, proxy mapping, event delivery, queued callbacks, geometry, or QObject destruction matters. Prefer semantic state and destroyed/resource counts; use targeted rendering assertions when pixels are the defect, without relying on whole-window snapshots or arbitrary sleeps.
• Before handoff, review for duplicate authorities, persisted-data mutation during preparation, stale async write-back, swallowed diagnostics, unowned resources, unbounded external-input caches, plugin-specific branches in shared code, and source-only assumptions at packaging boundaries.

Maintaining this guidance

• AGENTS files are maintained architectural memory, not immutable truth. Update durable rules only when supported by current architecture, tests, verified runtime behavior, explicit design, or an intentional refactor completed in the same change.
• Put a rule at the narrowest scope where it helps future decisions; let child guides specialize rather than repeat parents. For each local rule, identify the decision it protects and the implementation or test that could disprove it. Remove obsolete content inside files, distinguish preferred architecture from compatibility paths, and preserve every established file path while doing so. A guidance audit must not turn an observed defect into a required design.
• After significant architectural work, re-read the applicable hierarchy as a fresh agent: can it identify the owner, invariant, failure boundary, and relevant tests without relying on conversation history? Challenge rules likely to become stale, circular references, and wording that freezes incidental structure.
• Distinguish observed behavior from design requirements and name tests/consumers that can challenge a local rule. A known limitation is not a desired invariant. During guidance-only work, report code defects separately instead of changing production code to satisfy the prose.