Skip to content

AGENTS.md

Latest commit

 

History

History
34 lines (23 loc) · 3.14 KB

AGENTS.md

File metadata and controls

34 lines (23 loc) · 3.14 KB
Raw

AGENTS.md — slmcode

Go, SLM-first coding harness. Deeper: CONTRIBUTING.md, docs/.

Build & test

  • make bootstrap — npm deps + Studio UI build into cmd/slmcode/ui/. go build alone embeds no SPA; the server then serves a placeholder page (pkg/server/placeholder.go) and slmcode studio warns on startup. web/package-lock.json is stale, so npm ci fails and bootstrap falls back to npm install, which regenerates the lock — commit it.
  • make check — the single gate: tidy-check, gofmt, vet, lint, cover (go test ./... + the coverage floor), -race ./pkg/..., web-check (lint + typecheck:test + test + build). CI runs this. The two steps that need the network (module proxy, npm registry) skip with a named reason instead of failing, so it runs offline.
  • Lint baseline is zero: make lint (and make lint-strict, its alias) fail on any finding. Fix it, or add //nolint:<linter> // <why this site is a false positive>.
  • make e2e offline; RUN_E2E=1 make e2e adds live-model tests. Two suites run under plain make test and are the ones to keep green: test/e2e/harness_smoke_test.go drives the harness in-process against a fake OpenAI server, and test/e2e/binary_acceptance_test.go builds the real binary plus test/fakemodel and drives init → doctor → run → task show → diff → apply against a Go and a TypeScript fixture, asserting the bytes on disk.

Layout

cmd/slmcode → harness → orchestrator (phases) → loop (worker/review/correct), over agents, workspace (tools), context+repomap, schema+backends, memory+evolve, server. web/ builds into cmd/slmcode/ui/, go:embed all:uied.

Non-default conventions

  • Normalize → Validate: config/pipeline/block structs default in Normalize(), enforce in Validate(). Call both before persisting.
  • config.Config is the only source of truth (ApplyEnv, ApplyPatch). Precedence: defaults → user file → project file → SLMCODE_* → flags.
  • ANTI-WANDER / HARD SCOPE: agents.AntiWanderCore — workers/correctors touch only focus files and same-package siblings. Tests assert both markers verbatim.
  • Every structured role needs a pkg/schema contract (TestPromptContractsMatchSchemaAndGrammar).
  • Never end a turn on a tool call — emit final JSON after tool use. One tool call per turn.

Gotchas

  • cmd/slmcode/ui/: only .gitkeep is tracked (keeps go:embed all:ui compiling); index.html/assets//vendor/ are gitignored build output. Never commit them, never edit them by hand.
  • Budget context in tokens (pkg/context), never bytes — bytes starved a 32K model to ~3.2K.
  • Prompt assembly stays byte-deterministic, stable prefix first, or KV-cache reuse dies.
  • Every tool result goes through the cap in pkg/workspace; never return unbounded output.
  • .slmcode/ is gitignored runtime state; tools are denied writes into it.

Frontend

react-hooks/exhaustive-deps is an error; run npm run lint && npm test in web/ (make web-deps first). npm run build excludes *.test.ts(x) — tests are typechecked by npm run typecheck:test.