npx skills add ...
npx skills add isala404/forge --skill forge-idiomatic-engineer
Forge-focused engineering workflow for Rust applications with generated frontend bindings. Activate this skill for repositories containing a `forge.toml` file, Forge macros, or code generated by the Forge CLI.
npx skills add isala404/forge --skill forge-idiomatic-engineer
Full-stack Rust framework. Single binary, PostgreSQL-backed. Axum + Tokio + SQLx. Macros generate runtime wiring and frontend bindings; each handler must be registered in src/main.rs (macros alone do not wire it in).
These cause hours of wasted debugging if missed. Internalize before writing code.
SQLX_OFFLINE=true is mandatory for any cargo check / cargo build you run by hand. CI sets it globally. Without it, sqlx tries to validate every sqlx::query!() against your live DATABASE_URL — including queries inside published forge-runtime crate files you cannot edit — and you get a wall of "column does not exist" errors in third-party code. The simplest fix is eval "$(forge env)" in your shell rc; otherwise export SQLX_OFFLINE=true by hand.forge check auto-prepares the offline cache. It detects when src/ is newer than .sqlx/ and runs cargo sqlx prepare --workspace before the rest of the pipeline, so you don't need to think about prepare ordering. (For raw cargo check, you still need to run forge migrate prepare after editing any sqlx::query!().) Pass --no-prepare in CI where the cache should already be correct.forge migrate prepare hard-fails if cargo-sqlx is missing. Install with cargo install sqlx-cli --no-default-features --features postgres and re-run.forge commands walk up to find forge.toml. Run them from any subdirectory; the resolved root is printed at start. No need to cd first.forge doctor first. It checks rustc, cargo-sqlx, SQLX_OFFLINE, DATABASE_URL reachability, Docker, frontend tooling, forge.toml syntax, .sqlx/ freshness, and that the latest migration file isn't empty in one shot.cargo sqlx prepare is a passing compile. Don't run forge check purely to "confirm" what prepare just proved — prepare invokes a full cargo check internally. Run forge check only when you have new edits since the last prepare, or to exercise the rest of the validation suite (registration, schema, clippy).Run bash docs/skills/forge-idiomatic-engineer/scripts/orient.sh first. The script walks up to find forge.toml, then prints a structured dump: project name + auth mode + frontend, environment readiness (SQLX_OFFLINE, DATABASE_URL, cargo-sqlx, Docker), .sqlx/ cache freshness, the contents of src/main.rs / src/functions/mod.rs / src/schema/mod.rs, every registered handler grouped by kind, the latest migration, reactivity-enabled tables, and concrete NEXT action hints. One invocation replaces five separate reads.
Always read references/pitfalls.md and references/resilience.md once per session — the script doesn't print them.
Fallback when the script is unavailable (older checkout, sandboxed env): read forge.toml, Cargo.toml, src/main.rs, src/functions/mod.rs, and the most recent file under migrations/.
These files change only when you write to them — re-reading mid-session is almost always wasted context.
After auto-compaction, trust the summary's file inventory. Don't re-read main.rs / mod.rs to confirm something the summary already documented — only re-read if you're about to write to them and need exact current content. If a compaction attaches a file as an <attachment>, treat it as already in context; don't issue a fresh Read.
When you do need to read a file, read it fully in one call. Don't issue overlapping ranges (offset 200 then offset 1) — combine into a single read at offset 1 with a wide limit. Forge handler files are rarely larger than 600 lines.
| Concept | Macro | Struct suffix | Registration |
|---|---|---|---|
| Read-only query | #[forge::query] | Query | .register_query::<FnNameQuery>() |
| Data mutation | #[forge::mutation] | Mutation | .register_mutation::<FnNameMutation>() |
| Background job | #[forge::job] | Job | .register_job::<FnNameJob>() |
| Scheduled task | #[forge::cron] | Cron | .register_cron::<FnNameCron>() |
| Durable workflow | #[forge::workflow] | Workflow | .register_workflow::<FnNameWorkflow>() |
| Long-running process | #[forge::daemon] | Daemon | .register_daemon::<FnNameDaemon>() |
| External HTTP event | #[forge::webhook] | Webhook | .register_webhook::<FnNameWebhook>() |
| AI agent tool | #[forge::mcp_tool] | McpTool | .register_mcp_tool::<FnNameMcpTool>() |
Or use .auto_register() to pick up all handlers via inventory.
pub async fn handlers only — private functions fail codegen.snake_case fn names → macro generates PascalCase + type suffix. Do not include the type in the fn name (heartbeat, not heartbeat_daemon, or you get HeartbeatDaemonDaemon).#[forge::model] must be the first attribute on a struct.Memorize so you don't reach into the forge-core source mid-implementation:
ctx.db() → ForgeDb (a sqlx::Executor). Pass directly to query macros: sqlx::query_as!(...).fetch_one(ctx.db()).await?.ctx.conn().await? → ForgeConn<'_> (transactional, mutations only). Pass &mut conn to query macros.ctx.user_id() → Result<Uuid> on QueryContext and MutationContext. Returns ForgeError::Unauthorized if no principal. There is no ctx.auth() method on MutationContext.ctx.db_conn() → DbConn<'_> for shared helpers that must work in both queries and mutations. DbConn has an inverted convention (call .fetch_* on the DbConn, passing the query) — see references/patterns.md.let mut conn = ctx.conn().await?). Don't import ForgeConn / ForgeDb / DbConn and write explicit types unless a helper signature requires it.references/pitfalls.md and references/resilience.md. Detect frontend via frontend/package.json (Svelte) or frontend/Cargo.toml (Dioxus).forge new <kind> <name> instead of writing the file by hand. It scaffolds the right macro defaults, appends pub mod <name>; to src/functions/mod.rs, and inserts mod functions; in src/main.rs if missing. Kinds: query, mutation, job, cron, workflow, daemon, webhook, mcp_tool, model, enum.forge check. It auto-prepares the .sqlx/ cache when sources are newer, so you don't need a separate forge migrate prepare step in the common case.forge generate after backend changes settle. Never edit generated files.forge test for Playwright E2E.forge check clean, forge test green, write a brief change summary.When the user says "fix it" / "can you fix it" after you've diagnosed a problem, fix it — don't ask for re-confirmation. Only pause to confirm for destructive actions (data deletion, schema drops, force pushes).
patterns.md). Social logins must link via the user_identities table.ctx.env_require() / ctx.env_or(), never std::env::var().ctx.http() for RPC, ctx.raw_http() when you need bytes_stream() or custom redirect policy.sqlx::query!() / query_as!() bang-macros only. Run forge migrate prepare after schema or query changes — see Compile-Loop Hard Rules.transactional = false on a mutation that dispatches.src/utils/ the moment two handlers need it.| Task | Reference |
|---|---|
| Any new handler (mandatory read) | references/resilience.md |
| Macros, context, errors, configuration, CLI | references/api.md |
| Backend patterns: jobs, workflows, auth, webhooks, compile loop | references/patterns.md |
| Copy-paste recipes (user fetch, plan gating, email, S3, payments, AI) | references/recipes.md |
| Frontend principles (reactivity, subscriptions, errors, uploads) | references/frontend.md |
| SvelteKit specifics (runes, stores, auth helper) | references/frontend/svelte.md |
| Dioxus specifics (hooks, signals, auth keying) | references/frontend/dioxus.md |
| Writing tests (backend builders + Playwright scenarios) | references/testing.md |
| Debugging build or runtime errors | references/pitfalls.md |
resilience.md.dead_code, unwrap_used, panic, indexing_slicing, unsafe_code.ForgeError variants (never unwrap/expect/panic!).user_id / owner_id. ctx.user_id() for the principal. Opt out with #[query(unscoped)] only for shared/admin data.dispatch_job and start_workflow require a transactional mutation. Since transactions are on by default, just don't set transactional = false on mutations that dispatch.forge check enforced): never INSERT/UPDATE/DELETE on forge_* tables. Use dispatch_job, start_workflow, record_signal.-- @down. Optional -- @up marker is stripped. Enable reactivity with SELECT forge_enable_reactivity('table_name');. No IF NOT EXISTS.fetch_optional().await?.ok_or_else(|| ForgeError::NotFound(format!(...))).On completion, report:
forge check result.Every handler must survive revoked auth, deleted entities, concurrent modification, and network drops.