The Console’s policy editor (Console › Policies › a policy › Editor) is a plain text field with a
diagnostics layer behind it. It behaves like a developer tool, not a textarea: line numbers, the
server’s lint at the exact line and column, quick fixes where the lint names the replacement, and
a dry run and tests against the unsaved buffer. Source: web/src/controllers/policyeditor.ts,
web/src/lib/{editormath,lintloop,policyfix,linediff}.ts (S26).
- One editable surface. The text is a native
<textarea>, so the browser’s undo, IME input,
spellcheck-off, selection and screen readers all work as usual.
- A mirror behind it with identical metrics (Geist Mono 13/20, tab size 2,
white-space: pre,
the same padding) draws the diagnostic underlines (errors in the mismatch colour, warnings in the
unverifiable colour; underlines, never shaded blocks) and the highlighted line. Scrolling is
synced on each animation frame; only lines with a diagnostic are rebuilt.
- A gutter with line numbers and a marker per line with a problem (✕ error, ○ warning), whose
tooltip carries the message.
- Columns count characters the way the server does (Unicode code points: a tab is one, an emoji
is one), so
Line 9, col 5 in the list is the place the underline marks.
- No dependency: about 500 lines of TypeScript. A CodeMirror-class library would need an ADR.
| Key |
What it does |
| Tab |
inserts two spaces, or indents every selected line |
| Shift Tab |
removes up to two leading spaces from each selected line |
| Enter |
a new line with the current indent (after - id: it lines up with the item’s keys) |
| ⌘ S / Ctrl S |
Save as vN (admins; when the buffer differs from the version you opened) |
| Esc, then Tab |
leaves the editor (Tab is the editor’s own key until you press Esc; the field’s description says so) |
| ⌘ Z / Ctrl Z |
the browser’s own undo, including undoing an applied fix |
- 400 ms after you stop typing, the buffer goes to
POST /v1/policies/lint. A newer change aborts
the call before it, so there is never more than one in flight.
- Results appear three ways: the gutter marker, the underline, and the list under the editor:
Line 40, col 5 · Unknown key "desicion" in rule "block-destructive-sql". Did you mean "decision"?
with Apply fix and Docs. Clicking a row reveals its line.
- Quick fixes: S10’s diagnostics carry a
hint, not a structured fix. When the hint is exactly
Did you mean "x"? (today the unknown_key and cel_no_field codes), the Console replaces the
diagnostic’s range with x. After a fix it lints again at once and only says
✓ No problems · 16 rules · 1 budget when the server does.
- Limits: a policy is at most 256 KiB (UTF-8); a bigger buffer shows
Policies are limited to 256 KiB. Split this one. and nothing is sent. Lint, dry run and tests
share S12’s 60 calls a minute per person; a 429 shows Lint paused for 20 seconds (rate limit).
and lint resumes on its own.
Save as v5 posts the buffer as a new version. Versions are immutable and hashed; saving content
whose normalized form is unchanged (a comment, whitespace, key order) answers
No changes: this is already v4.- Leaving with unsaved changes (a link in the Console, or closing the tab) asks first.
- A viewer gets the same editor, read-only:
Viewers can read policies. Editing needs the Admin role. Dry run and tests still work against the buffer.
- The textarea has a label (the policy and version) and a description that includes the Esc, then
Tab rule and the diagnostics list.
- Diagnostics are text with a word (
Error, Warning) and a glyph; colour never carries meaning
alone. The status line under the editor is a polite live region with the count of problems; it
never announces on every keystroke.
- The focus ring is drawn around the whole editor.
| Budget |
Measured (S26) |
| Typing in a 2,000-line policy: < 16 ms per keystroke |
0.32 ms per keystroke with the mirror render (happy-dom); the mirror updates once per frame |
| Lint round trip < 150 ms locally, never more than one in flight |
one in flight by construction (AbortController) |
| Diff of two 2,000-line versions < 50 ms |
1.3 ms for a 5 % edit, 0.9 ms for a moved 200-line block, 48 ms for a full shuffle |
- No syntax colouring or completion.
- No collaborative editing: two people saving produce two versions; the last is the newest.
- No fix where the lint doesn’t name the replacement.