Files
hugocasa c3b2275864 docs(agents): rework agent context, fix dev-env docs, vendor skills (#10667)
* docs(agents): scope agent guidance to where it loads

AGENTS.md loads in every session. Three of its sections only ever applied to
one directory, and docs/autonomous-mode.md was unreferenced by anything in the
repo, so none of its content was in effect.

- Move "Verifying Backend Changes" to backend/CLAUDE.md, "Verifying Frontend
  Changes" and "Banned Patterns" to frontend/CLAUDE.md. They now load when
  working under those directories, which is when they apply.
- Update the two cross-references that pointed at the moved sections (pr and
  svelte-frontend skills).
- Delete docs/autonomous-mode.md. Its "don't stop early" half is already in
  .webmux.yaml's oneshot system prompt, which actually loads; its trigger was
  bypassPermissions, which does not imply an absent user; and it restated
  AGENTS.md and the pr skill with copies that had drifted (hardcoded ports,
  relative screenshot paths). Salvaged the UI traps it uniquely documented
  into frontend/CLAUDE.md and dropped the three stale profile references.

AGENTS.md drops ~3.6k characters with no guidance lost.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* docs(agents): guidance for building a feature — reuse, telemetry, live verification

Three recurring gaps, all cases where a pointer existed but nothing triggered
on it.

Component reuse. The svelte-frontend skill documented three components with
props, which reads as the whole catalog; the barrel exports 23 and common/ has
34 subdirectories against those 23. So "never use raw HTML elements" was an
instruction agents could not follow. Added a mandatory discovery step: read
the barrel, grep the tree, and treat the documented three as examples.

Brand guidelines. frontend/brand-guidelines.md is 34k characters referenced by
bare path, which nothing opens speculatively. Added a table mapping what you
are building to the section that governs it, entered with grep rather than a
full read.

Product telemetry. feature_usage has 14 registered actions across three
features, and an unregistered (feature, kind) pair is dropped by
valid_feature_usage_event with a bare continue — no error, still a 204 — so
frontend-only instrumentation silently records nothing. New
docs/feature-telemetry.md carries the criteria for when to instrument, the
four-step recipe including the allowlist and the InstanceSettings disclosure,
and the privacy rules. Raised in the plan for user-facing work, not as a
separate question, and not at all for bugfixes or refactors.

Also: validation now ends at exercising the change on the running instance,
with standing permission to spin up whatever that takes.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(dev): correct the worktree dev-environment guidance

Several things agents were told to do did not match what the machine does.

- Env discovery pointed at .env / .env.local / backend/.env. In a webmux
  worktree the real values are in $(git rev-parse --git-dir)/webmux/runtime.env
  (BACKEND_PORT, FRONTEND_PORT, DATABASE_URL, CARGO_FEATURES, WM_DB_NAME),
  sourced by every pane and undocumented. Reading it is also not blocked by the
  Read(**/.env) deny rules, which the old instruction walked straight into.
- The database name rule said branch-with-underscores. worktree-common.sh uses
  the worktree directory basename, and Postgres truncates at 63 characters, so
  branch hugo/win-2340-… resolves to windmill_win_2340_…_and_eval with no hugo_
  prefix and the tail chopped. A wrong DATABASE_URL guts the sqlx cache.
- The restart procedure said "tmux pane 1" and sent keys to an undefined
  <pane1>. Pane 1 is the backend under the full profile and the frontend under
  frontendOnly. Replaced with finding the pane by pane_current_command,
  recovering the live feature set from the running process (CARGO_FEATURES in
  runtime.env only records what the pane started with), and restarting in place.
- Added recovery for an orphaned backend holding the port: it reparents to
  systemd when its shell dies, so it survives anything that looks like cleanup.
  Three checks before killing a single pid, because pkill -f windmill takes out
  every sibling worktree.
- Agents spawned their own servers because AGENTS.md opened by telling them to.
  Now it checks for the existing panes first; the spawn commands are scoped to
  a plain checkout.
- New EE worktrees branched from the EE repo's local main, which nothing
  fast-forwards, so they started behind the commit pinned in
  backend/ee-repo-ref.txt — the one CI builds against. They now base on the pin,
  falling back to main only when it is unreadable.
- Enabled webmux autoPull so local main stays current; new worktrees are
  branched from it. Documented what WM_CLONE_DB does, including that it
  terminates every connection to the base windmill database.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* feat(skills): vendor grilling/architecture skills; tighten PR ready and review rounds

Vendors five skills from https://github.com/mattpocock/skills (MIT, pinned at
84fdeffd12f2ee307994d1eb6feb48173b6e0502). They are one dependency closure:
grill-me is a stub that runs grilling, and improve-codebase-architecture draws
its vocabulary from codebase-design and its CONTEXT.md upkeep from
domain-modeling. .agents/skills/UPSTREAM.md records the license, the pin, and
the four local deltas so a refresh stays a diff:

- flattened the upstream engineering/ and productivity/ split
- rewrote bundled-file links to repo-root paths, since relative links break
  when read through the .claude/skills symlink
- dropped the upstream agents/openai.yaml packaging metadata
- removed every ADR path. This repo has not adopted ADRs, and a skill that
  offers to create them is how the practice arrives by side effect rather than
  by decision.

PR workflow changes, all in the pr skill:

- A round that never starts is usually a conflict with main, not a CI outage.
  Resolve by merging, not rebasing — a rebase rewrites the head SHA that round
  verdicts and the clean-round marker are keyed to. If the merge advances
  backend/ee-repo-ref.txt, the EE worktree has to follow or
  cargo check --features private compiles a tree neither the author nor CI
  intends.
- A clean round no longer means an automatic flip to ready. Wide blast radius
  (*_ee.rs, migrations, OpenAPI or the generated client, auth paths, shared
  worker infrastructure, a new public surface) asks first; self-contained
  changes flip. Unattended, the judgement holds and the action degrades: flip
  the small ones, leave the rest at a clean draft with the reason in the PR
  body.
- Rounds that never converge are usually structural. After three without
  convergence, stop, name the module the findings cluster around, and suggest
  improve-codebase-architecture rather than burning more CI.

AGENTS.local.md (gitignored, with CLAUDE.local.md importing it) holds the
ready/ask calibration, recorded as dated observations rather than a rule.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* docs(dev): state that each worktree gets its own fresh database

The per-worktree section warned which DATABASE_URL to use but never said where
the database comes from: the post-create hook creates and migrates a new one
per worktree, so it starts with none of the main instance's workspaces, scripts
or flows. WM_CLONE_DB was documented only as a comment in .webmux.yaml, which
reads as how things work rather than as a per-project opt-in.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* chore(sqlx): script the cache backup/restore instead of documenting it

The update-sqlx skill spelled out a cp/comm/rm dance around `cargo sqlx
prepare`, which empties backend/.sqlx before regenerating — a failed run leaves
the cache gutted (observed: 2350 -> 142 entries), and a --all-targets run in a
CE checkout fails that way every time. Three problems with documenting it:

- The backup path was the literal /tmp/sqlx_backup, shared by every worktree.
  Two concurrent runs overwrite each other's backup, which is the only thing
  standing between a failed prepare and a gutted cache.
- The restore was a copy-pasted `rm -rf .sqlx && cp -r ... && cp ...` chain.
- Skipping the backup is what turns a routine failure into a lost cache, and a
  convention is easier to skip than a command.

sqlx-cache.sh has backup / newq / restore, keeps state in a per-worktree
directory, and leaves the judgement call where it belongs: `newq` prints each
added entry's query field for review, and only `restore` writes them in.

Also adds the general rule that scratch files belong outside the checkout —
anything written into the tree has to be deleted again, and rm prompts each
time.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* docs(agents): state why a routine cleanup prompts, and where scratch goes

The guard hook already auto-allows a plain rm whose operands are under /tmp or
inside a git checkout in $HOME, so deleting a temp dir or a stale .sqlx entry
costs nothing. What prompts is the command shape: the hook's tokenizer defers on
&&, ;, redirects, quotes and $VAR, so a chained cleanup falls through to the
Bash(rm:*) ask rule.

That was recorded only inside a paragraph about screenshot file paths in
frontend/CLAUDE.md, where nobody looking for it would find it. Stated in Core
Principles instead, alongside the rule that scratch belongs outside the tree —
for the reason that actually applies, which is not committing junk rather than
avoiding prompts.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* chore(security): deny agent edits to the permission hooks and project settings

.claude/hooks/guard-rm-outside-tmp.sh and guard-main-branch.sh are the
enforcement points for everything the permission rules are meant to catch, and
nothing stopped an agent editing them. One sed -i disables the guard for every
later command, silently, and the deny list in .claude/settings.json has the same
exposure.

Defence in depth rather than a boundary: an agent with arbitrary bash can still
delete, and this may only close the Edit-tool path if Bash writes are not
covered by Edit deny rules. It costs nothing and removes the cheapest way to
turn the guards off. Changing them now means editing the files by hand, which is
the intent.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix: address review round findings on head 3f47dc1

- backend/ and frontend/ guidance was Claude-only. Codex and Pi read AGENTS.md,
  not CLAUDE.md, so moving "Verifying Backend/Frontend Changes" and the
  $bindable ban out of the root AGENTS.md made them invisible to two of the
  three CLIs this repo supports. Renamed both to AGENTS.md with a one-line
  @AGENTS.md CLAUDE.md beside them, matching what the repo already does at the
  root and in ai_evals/, and retargeted the four references.

- sqlx-cache.sh aborted with exit 2 and no output when .sqlx was empty:
  list_entries ran `ls -1 ./*.json`, and an unmatched glob under
  `set -euo pipefail` killed the script. An empty cache is precisely what a
  failed prepare leaves behind, so it broke in the one case it exists for.
  Replaced with a glob loop; reproduced the failure and verified the fix.

- The oneshot prompt ("never leave the PR sitting in draft") contradicted the
  "Flip, or ask first" rule added in the same PR, which tells unattended runs to
  leave wide-blast-radius changes as clean drafts. The prompt now defers to the
  skill for the flip decision and keeps only "never stop at an unreviewed
  draft".

- Bundled-resource references in the vendored skills were markdown links to
  `.agents/skills/...`, which resolve relative to the file, not the repo root.
  Replaced with inline paths stating they are repo-root relative.

- The PR-ready calibration file was write-only: the skill said to record
  answers there but never to read it. It is now consulted before deciding.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* Revert "chore(security): deny agent edits to the permission hooks and project settings"

This reverts commit 3f47dc1692.

* fix: address round 2 nits

- backend/AGENTS.md told agents to persist CARGO_FEATURES in runtime.env, but
  webmux regenerates that file from metadata and .env.local every time the
  worktree is opened, so the setting is lost on the next reopen. The persistent
  source is .env.local, which scripts/post-create.sh already writes.

- UPSTREAM.md still described the vendoring delta as rewriting bundled-file
  *links* to repo-root paths. 555f063 replaced them with plain paths in prose,
  because a markdown target resolves relative to the file — a repo-root link is
  just as broken as a sibling-relative one through the symlink. Replaying the
  old wording on a refresh would reintroduce the bug UPSTREAM.md exists to
  prevent.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* docs(skills): correct the UPSTREAM.md link-rewrite delta

The delta note still described rewriting bundled-file *links* to repo-root
paths. 555f063 replaced them with plain paths in prose, because a markdown
target resolves relative to the file containing it — a repo-root link is as
broken as a sibling-relative one read through the symlink. Replaying the old
wording on a refresh would reintroduce exactly the bug UPSTREAM.md exists to
prevent.

The preceding commit's message claimed this fix; the edit had failed on a
stale anchor and only the backend/AGENTS.md half landed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* docs(dev): describe what a fresh worktree database actually contains

Exercising a real worktree creation showed the previous wording ("none of your
workspaces, scripts or flows") reads as an empty database. It is a bootstrap
instance: the admins workspace, the admin@windmill.dev superadmin, the license
key copied from the base database, and the migration seeds — observed as
u/admin/hub_sync and the default app theme resource.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-13 11:12:20 +00:00

4.3 KiB

Frontend (Svelte 5)

  • Coding patterns: MUST use the svelte-frontend skill when writing Svelte code
  • Validation: docs/validation.mdnpm run check:fast (2s) for iteration, npm run check (50s) for final PR
  • UI components: use Windmill's design-system components — never raw HTML elements. Start from the barrel src/lib/components/common/index.ts and grep src/lib/components/; the component you need almost certainly exists
  • Brand/design: frontend/brand-guidelines.md — read the relevant section before building UI, not after; the svelte-frontend skill maps which section covers what
  • Backend API: routes in ../backend/windmill-api/openapi.yaml, generated types in src/lib/gen/
  • Regenerate client: npm run generate-backend-client after backend API changes

Key Frontend Patterns

Prefer Composable State Over Two-Way Binding

// Use resource() from runed for async data
import { resource } from 'runed'
let items = resource(() => args, (args) => SomeService.list(args))
// items.loading, items.current

// Use composables for shared reactive state
function useLoader(argsGetter: () => Args) {
  let items = $state([])
  let loading = $state(false)
  $effect(() => { /* react to argsGetter() */ })
  return { get loading() { return loading }, get items() { return items } }
}

Two-way binding is fine for simple form inputs. Avoid it for component-to-component state.

Verifying Frontend Changes

After modifying frontend code, drive the running dev server with the Playwright MCP to verify the change in a real browser — don't claim a UI change works without exercising it.

Two MCP servers are registered in .mcp.json:

  • playwright — headless Chromium, default for devboxes (no display required)
  • playwright-headed — windowed Chromium, when a display is available

One-time setup: run npx playwright install chromium to download the browser binary (Playwright won't fetch it automatically on first use).

Typical flow:

  1. Ensure backend (cargo run) and frontend (REMOTE=http://localhost:8000 npm run dev) are running
  2. mcp__playwright__browser_navigate to the relevant page (login at admin@windmill.dev / changeme)
  3. mcp__playwright__browser_snapshot to inspect the accessibility tree (preferred over screenshots for reading the DOM)
  4. mcp__playwright__browser_click / browser_fill_form / browser_type to interact
  5. mcp__playwright__browser_take_screenshot for visual confirmation
  6. mcp__playwright__browser_console_messages / browser_network_requests to surface errors

Write screenshots to an absolute path under /tmp (the MCP servers already do; standalone Playwright scripts must be told): moving a PNG out of the checkout afterwards needs a mv the permission hooks always prompt on. Same reason to run rm/mv/cp as one plain command per Bash call: those hooks defer on &&, ;, redirects, quotes and $VAR.

Attach the screenshots to the PR. For any change under frontend/, embed screenshots of the affected UI in the PR body — the pr skill requires this and carries the upload recipe.

If you cannot exercise a UI change (no dev server, etc.), say so explicitly rather than claiming success.

Traps while driving the UI

  • critical_alerts 404s are expected on CE builds (EE-only endpoint) — ignore them.
  • VSCode worker 404s are dev-mode artifacts — ignore them.
  • <Toggle> hides its checkbox (sr-only). Click the <label> wrapper, not the checkbox.

Banned Patterns

$bindable(default_value) on optional props

Using $bindable(default_value) on props that can be undefined is banned. This pattern causes subtle bugs because the default value masks the undefined state.

Bad:

let { my_prop = $bindable(default_value) }: { my_prop?: string } = $props()

Correct alternatives:

  1. Use $derived with nullish coalescing — handle the potential undefined at the usage site:

    let { my_prop = $bindable() }: { my_prop?: string } = $props()
    let effective_value = $derived(my_prop ?? default_value)
    
  2. Create a useMyPropState() helper — encapsulate the undefined-handling logic in a reusable function and call it higher in the component tree, so the child component always receives a defined value.