• crates/coral-api: protobuf contract and generated Rust bindings.
• crates/coral-app: local server composition, state, workspaces, source lifecycle, and workspace-scoped catalog discovery behavior.
• crates/coral-cli: terminal adapter.
• crates/coral-client: intentionally thin transport bootstrap plus Arrow IPC decode/render helpers.
• crates/coral-engine: engine-side backend compilation, runtime registration, and query execution.
• crates/coral-mcp: MCP tool core with stdio and Streamable HTTP transport adapters over coral-client.
• crates/coral-spec: declarative source-spec parsing, validation, input discovery, and normalized source-definition models.
• crates/coral-telemetry: cross-crate telemetry helpers that are independent of app bootstrap, query runtime, and adapter surfaces.
• apps/desktop: Electron shell around Reef and the local Coral sidecar.
• apps/docs: Mintlify documentation site.
• apps/reef: React Router/Wax frontend shell. It is validated independently from apps/ui and is not built by Rust crate build scripts.
• apps/ui: embedded Coral app UI built into the CLI release flow.
• plugins/coral: Agent plugin packaging. plugins/coral/skills is the canonical in-repo home for maintained Coral agent skills.

• Run make rust-checks before submitting PRs that include changes to Rust code.
• For Postgres-backed database changes, run make postgres-tests. Keep this as the single entry point for local and CI Postgres coverage; do not duplicate its Cargo test invocations in workflows or contributor instructions. The target uses CORAL_TEST_POSTGRES_URL when supplied. Otherwise it starts a local Docker Postgres and creates a fresh database inside the reusable container. Docker chooses an available localhost port by default; use make postgres-url to print the server URL or LOCAL_POSTGRES_PORT=55432 make postgres-start when you need a stable port. Use make postgres-start when you only need the server, make postgres-stop when finished, and make postgres-clean to remove the reusable container.
• Run make schema-check before submitting PRs that touch generated manifest schemas or the Rust helpers that generate them. Use make schema-generate to refresh generated schema files. The Validate workflow enforces this through its schema-freshness job when schema inputs change.
• UI changes must pass npm run check --prefix apps/ui (oxfmt + oxlint) before submitting.
• Reef changes must pass npm run check --prefix apps/reef, npm run typecheck --prefix apps/reef, npm test --prefix apps/reef, and npm run build --prefix apps/reef, followed by npm run test:server --prefix apps/reef, before submitting. The production server smoke test consumes the build output and runs on every Reef CI job.
• Desktop changes must pass npm run check --prefix apps/desktop and npm test --prefix apps/desktop before submitting.
• Keep Reef Vitest coverage Node-only and focused on atomic deterministic functions and policies or explicit architectural invariants. Do not add broad browser, router, or framework-plumbing tests without a named contract or regression they protect.
• Use Storybook and Chromatic for Reef component visual states.
• Reef styling uses vanilla-extract; do not introduce Tailwind.
• Run make perf-check before submitting PRs that could affect CLI startup, local server bootstrap, source registration, or coral.tables catalog query latency. CI installs the bundled github source with fake credentials and fails when release coral sql "select * from coral.tables" has a hyperfine mean above 750 ms.
• Pull requests do not automatically build macOS Desktop packages. Use manual dispatch for an unsigned packaging preflight when a distribution-sensitive change warrants one. Validation artifacts stay unsigned and must not be reused for a release; Desktop release publishing must rebuild from a clean checkout with signing and notarization.
• The Validate workflow intentionally skips draft pull request runs, starts again on ready_for_review, and still triggers on converted_to_draft so the replacement skipped run cancels any in-progress validation for the PR branch. Keep that draft gate aligned between the initial change detector and final aggregate validate job.
• Keep Release Please branch updates coalesced and cheap to supersede. The release-please workflow may create multiple local regeneration commits, but must push them together through its final push step. Validate intentionally gives release-please--* pull requests a 120-second settle period before checkout and change detection so concurrency cancellation stops intermediate branch states before expensive jobs fan out.
• make rust-checks is the Rust-only local gate and should keep using --all-features; the embedded UI feature is a normal CLI build surface.
• The built UI artifact is produced by repo/CI orchestration (make ui-build or the UI build workflow job), not by crates/coral-cli/build.rs. Local Rust builds may compile without apps/ui/dist, because UI development normally serves assets from Vite while the CLI provides the loopback API server.
• Use make docker-build to compile the current checkout in a native Linux BuildKit stage and package that binary as coral:local. Set DOCKER_IMAGE=coral:test to change the local tag or DOCKER_NO_CACHE=1 to bypass Docker layer caching. Keep the exported-binary layout and runtime platform aligned with the docker-publish workflow; local Docker-exporter loads disable provenance because that exporter cannot load attestation manifests. Local builds must not download a published Coral binary.
• Use make reef-docker-build to build Reef from the current checkout and make reef-docker-smoke to test already-built Coral and Reef images. Use the self-contained make reef-docker-test to build both images and run the topology/configuration matrix. Reef's runtime stage must remain COPY-only and non-root; build and dependency stages run on the build platform. Local builds follow the Docker daemon's architecture, while CI builds and verifies linux/amd64 only. TLS termination belongs to the operator and is not provisioned by the image or its smoke harness.
• Keep adapters thin. If CLI or MCP behavior gets complex, move it inward.
• Keep server topology orchestration private to coral-cli while CLI commands are its only consumers. Do not extract the orchestration into a shared orchestration crate unless it gains a non-CLI consumer; the combined topology is provisional and may be removed rather than promoted.
• Keep transport contract concerns in coral-api, source-spec concerns in coral-spec, app/state concerns in coral-app, and query/runtime concerns in coral-engine.
• Keep app-owned runtime package assembly in coral-app. coral-engine should compile generic runtime components, not interpret DSL v4 authored manifests, materialized fingerprints, semantic IR, or projection catalogs.
• Keep Reef Coral access behind React Router server loaders, actions, or resource routes using apps/reef/app/lib/coral-request.server.ts. Do not expose a generic renderer-to-Coral transport or Desktop sidecar proxy; add an explicit server route when browser-triggered Coral behavior is needed.
• The packaged Reef server resolves its external runtime packages from the Electron app. Keep every apps/reef production dependency represented in apps/desktop production dependencies; the desktop config tests enforce this packaging contract.
• Use CORAL_DESKTOP_APP=1 as Reef's single external desktop build marker. React Router route composition may read it from process.env, while apps/reef/vite.config.ts exposes only its compiled boolean value as import.meta.env.CORAL_DESKTOP_APP. Do not add a parallel VITE_CORAL_DESKTOP_APP marker or expose broader CORAL_* values to browser code.
• For DSL v4 materialization, the user owns when a source is generated or regenerated. Coral materializes at source add, queries only from the installed materialized package, and never silently refreshes descriptors, projections, or persisted artifacts. Treat fingerprints, producer versions, identity metadata, and raw-document hashes as advisory provenance: report mismatches through tracing, but load readable, structurally compatible artifacts. Isolate source-local compatibility failures without hiding operational failures.
• A DSL v4 source declares top-level inputs: and exactly one singular surface:. The source name is its SQL namespace. Do not add surface ids, namespace suffixes, or multiple surfaces to one manifest; represent distinct provider interfaces as distinct source specs instead.
• Keep cross-crate W3C trace-context propagation helpers in coral-telemetry; do not make coral-app, coral-client, coral-engine, or coral-mcp depend on each other just to share telemetry carrier logic.
• Keep shared Arrow IPC decoding and result rendering in coral-client.
• Treat coral-app as an internal composition root even if sibling crates use its bootstrap seam today.
• If a caller needs explicit local server control, prefer coral-client::local over widening the default client surface.
• Keep process environment access owned by the right crate. coral-app owns runtime/bootstrap env reads, coral-cli owns CLI-surface env reads, and other crates should receive explicit values from callers instead of reading ambient process environment directly.
• Keep docs lean and readable. For CLI or MCP changes, update apps/docs/ only when the change affects a public surface or captures important user-facing or contributor-facing knowledge. Do not document every implementation detail. When docs are warranted, choose the best existing location first and make the amount of space match the feature's user-facing weight and visibility.
• Keep stable bundled sources under sources/core/**; put preview DSL v4 source specs under sources/v4/[source]/manifest.yaml with distinct manifest names (defined in the manifest's name field) such as <name>_v4. When a provider has distinct interfaces, use sibling source directories such as github and github_mcp. Do not bundle sources/v4 into the binary; install preview v4 sources with coral source add --file. Do not replace or migrate an existing v3 source merely because a preview v4 spec exists.
• Changes to scripts/install.sh must keep the Validate workflow's install-script matrix in sync with every OS/architecture target that the installer supports.
• Keep general repository automation in xtask; reserve scripts/ for the bash Coral installer and installer-specific support.
• Keep xtask organized by workflow: docs generation lives under xtask/src/docs/, shared source-manifest discovery lives in xtask/src/sources.rs, command-latency checks live in xtask/src/perf.rs, benchmark dispatch lives under xtask/src/benchmarks/, and the isolated benchmark package and fixtures live under xtask/benchmarks/. Skill export lives in xtask/src/skills.rs. Release signing and notarization automation lives in xtask/src/release.rs. The DSL v4 inference report lives in xtask/src/metadata_report.rs.
• Use cargo run --locked -p xtask -- benchmark list-columns to measure the complete MCP list_columns response for the checked-in synthetic wide-table fixture with the o200k_base tokenizer. The benchmark must call the real MCP tool in-process, report without enforcing a token budget, and keep benchmark-only code out of production crates.
• Use cargo run --locked -p xtask -- v4-metadata-report before and after a change to DSL v4 row-path, pagination, or lookup-key inference, and diff the two reports. It imports every non-MCP v4 source under sources/v4 and emits one CSV row per operation, so an unintended reshape in a source nobody was thinking about shows up as a diff hunk. Pass --cache-dir so a before/after pair fetches each descriptor once. It is deliberately not wired into CI: it fetches multi-megabyte vendor descriptors over the network, and vendor descriptors change under us, so a green run proves nothing about the commit that produced it.
• Universal Search relevance benchmarking also lives in the isolated coral-benchmarks package. Keep real catalog inventories, generated questions, collected queries, responses, focused corpora, and replay reports under ignored run directories. Only synthetic, deliberately non-sensitive benchmark fixtures may be checked in. Use frozen-query replay rather than rerunning agents while tuning ranking weights, and do not run agent collection in CI.
• The Electron desktop app version is tied to the CLI release version through release-please. The release workflow builds the macOS desktop app from apps/desktop, uploads its DMG/ZIP/update metadata to the same GitHub Release as the CLI artifacts, and the website should link to the releases/latest/download DMG rather than storing desktop binaries itself.
• make docs-check intentionally skips the aggregate community source catalog. Any PR may leave that generated page stale so unrelated changes do not fail on aggregate community catalog drift; keep docs freshness strict for bundled sources under sources/core/**, apps/docs/docs.json, and the changelog.
• The live docs site deploys from the long-lived docs branch, not main, so the published catalog matches the latest released binary. main still owns docs freshness, but merging to main no longer publishes the site by itself: the release workflow advances docs after release artifacts are published. See apps/docs/AGENTS.md for the full publishing model.
• Keep checked-in generated files marked in .gitattributes with linguist-generated so GitHub collapses them by default in PR diffs.
• Source inputs that carry credentials must be kind: secret, never kind: variable. This includes API keys, bearer tokens, access tokens, passwords, private keys, authorization header values, and admin/read keys, even when the credential is read-only or the source also supports anonymous access.
• When source credential retrieval or auth guidance changes, keep the source spec docs and maintained Coral source-spec skills aligned in the same change. OAuth source-spec behavior needs both reader-facing docs and agent-facing author/review guidance because credential.methods controls setup while auth still controls runtime requests.
• Keep maintained Coral agent skills in plugins/coral/skills. External distribution repos or packages should mirror from that directory rather than becoming a separate source of truth. Use cargo run --locked -p xtask -- export-skills --dest <path> for local export checks and distribution syncs.
• Keep plugins/coral conformant with Agent Plugins 1.0: portable metadata belongs in root plugin.json, skills are discovered from skills/, and MCP servers are declared in root mcp.json. Keep the legacy .codex-plugin, .mcp.json, and .app.json package files for pre-0.147 Codex compatibility. Put current Codex-only metadata under extensions.com.openai, and align shared metadata and MCP invocation across both representations.
• Coral skills must include agents/openai.yaml. Keep interface.display_name in the form Coral or Coral <Title Case Suffix>, keep the top-level SKILL.md heading equal to that display name, and set non-empty short_description and default_prompt values. The default prompt should mention the skill token, such as $coral-create-source-spec.
• When proposing or updating a PR title, use Conventional Commits: type(scope): summary.
• When using a scope, prefer one that matches the primary area changed, usually the crate name minus the coral- prefix, docs, sources/core/<name>, or sources/community/<name>.
• Keep the PR title up to date as the branch evolves. If the change shifts in scope or intent, update the title to match the current final shape of the branch.
• Use ! only for breaking changes, placing it immediately before the colon: type!: summary or type(scope)!: summary. Local WIP commit messages can stay pragmatic unless the user explicitly asks for polished commit history.
• If you add a source using dummy credentials in order to test a change, always configure Coral to store those credentials on the filesystem. Do not store dummy credentials in the OS keychain.

A meta change modifies how contributors or agents should work in this repo, not only runtime behavior. Examples include repo layout, crate ownership, source directory conventions, docs generation behavior, CLI/MCP surface rules, PR title/scope guidance, verification commands, and agent-facing review or source-authoring instructions.

For meta changes:

• Update the nearest relevant AGENTS.md in the same change.
• Update apps/docs/, generated docs, or docs tooling only when the changed behavior is user-facing or docs-authoring-facing, and use the smallest useful edit in the best existing location.
• Preserve provenance: keep observed repo facts, project direction, local preferences, and generated context separate instead of merging them into one untraceable rule.
• Treat repeated human steering as a defect in the operating loop. Identify the failure class, update durable context or tooling when that can prevent recurrence, and verify the new rule before resuming unrelated work.
• Include explicit validation showing the guidance matches the implemented behavior.
• Mention in the PR description what agent or contributor behavior changed.

For a CLI, the user interface is the API.

A change is breaking if it can break existing:

• commands people run manually
• scripts and CI jobs
• documented workflows
• integrations that parse output

Treat these as stable contract surfaces:

• command/subcommand names
• flags and positional arguments
• exit codes
• structured output (for example JSON)
• config file keys, format, and location
• environment variables and precedence rules

If any of those change incompatibly, it is a breaking change.