fix(browse): worker-free JSON in record editors to end Monaco worker OOM (#1370/#1499)

[dawsontoth]

Summary

RUM flagged a recurrence of the closed #1370 / #1499 Monaco-worker-OOM class through the browse record editors (EditTableRowModal, AddTableRowModal), which hardcoded language="json". Monaco's json language runs a worker and clones the model's full text (and every change delta) to it over postMessage; a large record — a row with a big blob/array field, or a bulk-insert array pasted into "Add row" — overflows the structured clone and crashes the worker with DataCloneError: ... out of memory., which Monaco reports as an unhandled error and then retries in a tight loop, flooding the session.

This revision replaces the earlier size-based json→plaintext switch with a worker-free JSON language, after review (@kriszyp) correctly pointed out the switch can't close the paste path (details below).

Approach

• New Monaco language json-worker-free (workerFreeJsonLanguage.ts): registers Monaco's main-thread JSON tokenizer plus the standard JSON language configuration (brackets, auto-closing pairs), and none of the worker-backed providers (validation, folding, colors, completion, hover, formatting). Registered in setup.ts alongside the built-in json contribution, which is left untouched for the config/roles editors that rely on its schema validation.
• Both record editors switch to json-worker-free. There is no language worker to overflow, at any size; syntax highlighting and bracket/auto-close are unchanged.
• Worker validation markers go away with the worker, so the onValidate signal is replaced by recordEditorJson.ts: a cheap main-thread parse for content under MAX_WORKER_MODEL_CHARS keeps the "Save disabled while JSON is invalid" affordance for normal-size records without parsing a huge buffer on every keystroke, and a caught JSON.parse at submit time toasts on malformed input (for both Add and Edit) instead of throwing an uncaught SyntaxError.
• EditTableRowModal resets its draft/validity when a different record is opened (that modal instance is reused across rows).

Why not the size-based switch

@monaco-editor/react applies a changed language prop in a later React effect, and Monaco syncs the model to the worker on a debounce; once synced, the worker's own onDidChangeContent listener posts the oversized delta on the next change. So starting from a small sample and pasting a >512 KB array hands the payload to the worker before any json→plaintext switch can run — the switch also doesn't stop an already-running model sync. A worker-free language sidesteps the race entirely rather than trying to win it.

Where to focus review

• Tradeoff — live validation: worker-free means no inline red-squiggle validation. Mitigated by the main-thread Save gating (normal records) + submit-time parse/toast (all records). This is a deliberate behavior change for the browse record editors; if the team wants richer inline validation back, a main-thread JSON diagnostics provider would be a follow-up.
• createTokenizationSupport import (monacoJsonTokenization.d.ts + workerFreeJsonLanguage.ts): reuses Monaco's own JSON tokenizer from an internal ESM path (same style as the existing deep imports in setup.ts/languageServices.ts) for pixel-identical highlighting. It ships without types, so it's declared locally. A Monaco version bump could move it — tsc -b / vite build would catch that.
• Out of scope: Monaco's base editor worker still syncs models for link detection (a '*' provider), exactly as it does for every plaintext editor the app already treats as the safe fallback. This PR closes the json-language-worker path the RUM data implicates; the base-worker exposure is pre-existing and shared, not addressed here.

Testing

• workerFreeJsonLanguage.test.ts: asserts the language registers tokenization + config and no worker-backed provider.
• recordEditorJson.test.ts: validity/parse helpers (valid, malformed, empty, oversized-deferred).
• AddTableRowModal.test.tsx (new) + EditTableRowModal.test.tsx (updated): assert both editors use json-worker-free, a valid record saves the parsed value, and an oversized malformed paste is caught at submit time (toast, no save) — the large-paste integration regression asked for in review.
• Full suite green (1786 passed / 11 skipped), tsc -b clean, oxlint + dprint clean, vite build clean.

Not browser-verified: reproducing the crash needs a live instance with a >512 KB record; verified at the mechanism/unit level and via a production build (the tokenizer import bundles into the editor.api2 chunk).

🤖 Generated by an LLM (Claude Opus 4.8) via Claude Code.

[gemini-code-assist]

Code Review

This pull request introduces a helper function, chooseRecordEditorLanguage, to dynamically switch the Monaco editor's language from 'json' to 'plaintext' when the payload size exceeds a safe threshold. This prevents the JSON worker from crashing on oversized records in both the AddTableRowModal and EditTableRowModal. The review feedback recommends defaulting empty or nullish content to 'json' instead of 'plaintext' to preserve helpful editor features (such as auto-closing brackets and syntax highlighting) when a user begins typing in an empty editor, along with updating the corresponding unit tests.

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	52.28%	5618 / 10745
🔵	Statements	52.8%	6007 / 11375
🔵	Functions	44.64%	1385 / 3102
🔵	Branches	45.19%	3765 / 8330

File Coverage

File	Stmts	Branches	Functions	Lines	Uncovered Lines
Changed Files
src/features/instance/databases/modals/AddTableRowModal.tsx	70.45%	47.82%	55.55%	70.45%	59, 80-98, 113, 123-124, 173, 202, 208, 215
src/features/instance/databases/modals/EditTableRowModal.tsx	82.85%	77.5%	71.42%	82.85%	62, 91-92, 165-167
src/features/instance/databases/modals/recordEditorJson.ts	100%	100%	100%	100%
src/lib/monaco/workerFreeJsonLanguage.ts	100%	100%	100%	100%

Generated in workflow #1573 for commit 7647364 by the Vitest Coverage Report Action

[dawsontoth]

Good call — applied in 26f5fc3. Defaulted empty/nullish content to json so the record editor keeps auto-close brackets, indentation, and highlighting from the first keystroke; the OOM guard now only trips once content actually exceeds MAX_WORKER_MODEL_CHARS. Updated the doc comment and the unit test accordingly.

[kriszyp]

The suggestions might help protect against OOM?