Files
Metal AgentandClaude Opus 5 9a29e8e335
CI / rust-headless (push) Successful in 6m38s
CI / rust-ui (push) Failing after 6m20s
Make the gate structural, and correct what it tells an agent
Three gates in this repository were decorative, and each was discovered by
being wrong rather than by failing.

A crate directory in neither members nor exclude is silently not built, which
is how lumbridge-devices shipped 1,127 lines that had never compiled.
scripts/workspace-guard.sh refuses that state, and asserts the gpui source
and version out of Cargo.lock rather than the manifest, because a manifest
states an intent while the lockfile states what would actually be compiled --
and a caret requirement accepts a version nobody reviewed. It needs no
compiler, so it runs first and in the headless job, which unlike the UI job
is not continue-on-error and can therefore actually fail a push.

deny.toml's source policy had never been executed: ci.sh ran `check
licenses` alone, and `check sources` failed immediately on the rev-pinned
buzz-sdk. The permitted Git sources are now named one by one and the check
runs, so a fourth is a decision rather than an accident.

cargo-deny and cargo-nextest being absent was a warning that let a run report
green having skipped the licence gate DISTRIBUTION.md depends on. Under
LUMBRIDGE_CI_STRICT=1 a missing tool now fails; locally it stays a warning so
a contributor is not blocked.

skills/lumbridge-development/SKILL.md told every agent that GPUI and Floem
live in spikes/ and that no framework may be selected until both pass the
hard gates. Decision 0017 settled that a month ago in the opposite direction.
The entry point an agent is meant to read was the least accurate document in
the repository.

Decision 0023 records where the GPUI dependency actually goes. Published gpui
has not been released since 2025-10-22, Zed's main still declares 0.2.2 with
no bump pending, the platform backends moved to crates that inherit
publish = false, gpui's own x11 and wayland features are now empty markers,
and 0.2.2 has no accesskit dependency at all -- so "published now, migrate
later" was never available. The adapter 0017 promised was never written and
the call sites grew from few to 147 against 20 identities, so the adapter is
written first, on 0.2.2, before the dependency moves.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SPYebLiN2w4TqnHUYGdECq
2026-09-01 12:51:25 -07:00

4.2 KiB

AGENTS.md

These rules apply to the entire Lumbridge repository.

Product boundary

Lumbridge is a terminal/workspace runtime and ACP client. It may host, supervise, and observe coding harnesses, but must not silently impersonate them, harvest their private credentials, or claim provider quota data that cannot be verified.

Credential use is narrow and named. A harness's stored credential may be read only to ask that same provider a documented question about the user's own account, and only where the answer cannot be obtained another way. Under that allowance a credential must never be persisted, logged, copied into application state, written to a crash report, or passed as a command-line argument; it must be released as soon as the request it authorises has been made; and any request it authorises must identify Lumbridge as the caller. A refresh token is never used — renewing a credential is the harness's job, not Lumbridge's. Every such use is named in a decision record, and each is switchable off by the user. Reading a credential for anything other than a use recorded that way is out of bounds. See decision 0016.

Research boundary

Upstream repositories are cloned outside this Git repository under the desktop workspace's Research/ directory. They are references, not vendored code.

  • Do not copy upstream implementation code without an explicit license review.
  • Record the source repository and relevant license for any adapted design.
  • AGPL/GPL and unlicensed repositories may be studied for behavior and UX only unless the project deliberately changes its licensing strategy.

Engineering rules

  • Keep provider, harness, protocol, terminal, persistence, and UI boundaries separate. A provider is not a harness and ACP is not a provider API.
  • Preserve a PTY fallback. ACP adds structure but must not be required to open a normal shell or run an arbitrary CLI.
  • Keep secrets out of SQLite, logs, crash reports, command-line arguments, and repository files. Store only opaque secret references in application state.
  • Usage values must include a provenance and confidence classification.
  • Avoid platform behavior hidden behind scattered cfg blocks; isolate it in platform adapters with shared contract tests.
  • No telemetry is enabled by default. Any future telemetry must be documented, opt-in, redacted, and independently disableable.

Dependencies that are not an agent's decision

The native UI framework is settled and its dependency is pinned by a decision record. Do not change the gpui (or gpui_platform) source, revision, or version; do not edit rust-toolchain.toml or the workspace rust-version; do not run cargo update or relax deny.toml to make a build succeed. A headless test asserts the pinned source and revision on every push, so a silent bump fails CI rather than landing.

Changing any of them is a new decision record with a probe run at the new revision, not a commit. If a UI build fails, the failure is the finding: report it rather than routing around it.

Every crate is a workspace member

A crate directory in neither workspace.members nor workspace.exclude inherits no lints, is never built, and its tests never run -- silently. This has happened twice here, and both times the code looked finished and did not compile. Add a crate to members in the same change that creates it, and never move code into an excluded directory to make a build pass. CI fails on a crate that is in neither list.

Verification

Before committing Rust changes, run:

export PATH="$HOME/.cargo/bin:$PATH"   # a non-interactive shell has no cargo
./scripts/ci.sh

cargo is on PATH only via an interactive shell profile, so an agent, an ssh command, or a hook gets cargo: command not found without that export. Never conclude a gate passed from an exit code obtained through a pipe -- a pipeline reports its last command's status, which has hidden a failing gate here already. --headless deliberately excludes apps/lumbridge, so a headless pass is not a product pass.

Keep bacon running during development. Tests must use synthetic fixtures; real agent transcripts, subscription state, private source, and credentials are never test data.