[vite 8.2.0] Fix REFLEX_ENV_MODE leak between AppHarness instances

[masenf]

Type of change

• Bug fix (non-breaking change which fixes an issue)

Description

Fixes a critical bug where AppHarnessProd leaked REFLEX_ENV_MODE=prod into subsequent dev AppHarness instances in the same process. When AppHarnessProd runs export(), it sets REFLEX_ENV_MODE=prod process-wide and never restores it. A dev AppHarness compiling later would then write prerender: true into its dev react-router config, causing the dev server to serve prerendered page HTML. On pages with hydration hazards, this failed hydration on every load and dropped events dispatched during the hydration-recovery window.

Root cause: The environment variable leak was masked by a vite 8.0.16 pin that was added to work around memo component re-rendering failures in the full integration suite. Those failures were actually caused by this REFLEX_ENV_MODE leak, not by vite itself.

Changes

1. reflex/testing.py: Pin REFLEX_ENV_MODE to dev in AppHarness._initialize_app(), mirroring what reflex run does. This ensures dev harnesses always compile in dev mode regardless of prior AppHarnessProd instances.

2. packages/reflex-base/src/reflex_base/constants/installer.py: Bump vite from 8.0.16 to 8.2.0 now that the root cause of the memo regression is fixed.

3. Tests: Added test_app_harness_initialize_resets_leaked_prod_env_mode() to verify that dev AppHarness correctly resets REFLEX_ENV_MODE to dev mode and disables route prerendering, even when the environment variable was previously set to prod.

4. News fragments: Added changelog entries documenting the bugfix and vite bump.

Update (2026-08-07): react-router 8.3.0 merge + docker-test fixes

Merged main (react-router 8.3.0 upgrade, #6854) into this branch so CI exercises the shipping combination, and fixed the two reflex-init-in-docker-test failures that upgrade introduced on main:

5. packages/reflex-base/src/reflex_base/compiler/templates.py: Pin preview.host to 127.0.0.1 in the generated vite config. react-router 8 prerenders by fetching pages from a vite preview server started with the project's config; with no pinned host, the bound socket and the fetched http://localhost:<port>/ URL can resolve to different loopback address families (as in docker containers), failing reflex export with Prerender: Request failed for /: ECONNREFUSED. Same fix TanStack Router shipped for the identical failure (fix: prerendering with docker build TanStack/router#6305). Full diagnosis in this comment.

6. reflex/utils/exec.py: Pre-enable the development export condition (NODE_OPTIONS/BUN_OPTIONS) when starting the dev server. react-router 8's dev CLI otherwise re-executes itself to enable the condition, which bun does not honor on installs without node, tripping the CLI's restart guard and exiting the dev server. Details in this comment.

Test Plan

• New regression test test_app_harness_initialize_resets_leaked_prod_env_mode() verifies the env-mode fix
• New unit tests assert the generated vite config pins the preview host, and that the development condition is merged into NODE_OPTIONS/BUN_OPTIONS
• Existing unit tests pass; memo playwright suites pass 10/10 on the merged react-router 8.3.0 + vite 8.2.0 tree
• Node-free verification (reflex-managed bun 1.3.14, mirroring the docker image): fresh blank app reflex export succeeds, and reflex run --env dev serves 200 with zero restart errors

https://claude.ai/code/session_01AzeRM83L6J367wmwk9Gezk

<img alt="Review in cubic" src="https://www.cubic.dev/buttons/review-in-cubic-dark.svg">

[greptile-apps]

Greptile Summary

The PR prevents production environment mode from leaking into later development harnesses and updates the bundled Vite version. It also pins the prerender preview server to IPv4 loopback and pre-enables React Router’s development export condition for spawned dev servers.

• Resets REFLEX_ENV_MODE during development harness initialization.
• Upgrades Vite from 8.0.16 to 8.2.0.
• Configures the generated Vite preview host as 127.0.0.1.
• Merges the development condition into Node and Bun runtime options.
• Adds regression coverage and changelog fragments.

Confidence Score: 5/5

The PR appears safe to merge because no blocking failure remains.

No blocking failure remains.

Important Files Changed

Filename	Overview
reflex/testing.py	Resets environment mode before harness compilation, while production harnesses still switch explicitly to production during export.
reflex/utils/exec.py	Adds an environment-copy helper that preserves existing runtime options and enables the development export condition for frontend dev processes.
packages/reflex-base/src/reflex_base/compiler/templates.py	Pins the generated Vite preview server to IPv4 loopback for React Router prerendering.
packages/reflex-base/src/reflex_base/constants/installer.py	Updates the generated project’s Vite development dependency to 8.2.0.
tests/units/test_testing.py	Adds regression coverage confirming a development harness clears leaked production mode and disables prerendering.
tests/units/utils/test_exec.py	Covers development-condition insertion, option preservation, deduplication, and non-mutation of the source environment.
tests/units/utils/test_utils.py	Verifies that generated Vite configuration pins the preview host.

_{Reviews (5): Last reviewed commit: "Rename news fragments to PR 6857 and ded..." | Re-trigger Greptile}

[codspeed-hq]

Merging this PR will not alter performance

✅ 26 untouched benchmarks
⏩ 8 skipped benchmarks¹

---

_{Comparing claude/vite-8-2-0-test-memo-regression-h6shlg (327cecd) with main (4363732)}

Footnotes

1. 8 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[cubic-dev-ai]

No issues found across 5 files

_{Re-trigger cubic}

[masenf]

The reflex-install-and-init failures on this PR are not caused by this diff — and the root cause is now fixed here.

What broke: reflex-init-in-docker-test fails on every run whose checkout contains the react-router 8.3.0 upgrade (4116ff2 / #6854), starting with its own push run on main at 16:51Z — identical Prerender: Request failed for /: ECONNREFUSED at vite.js:887. PR runs check out the merge with current main, so this PR inherited it on both attempts, as did claude/auth-overview-redirect-rsj733 (19:53Z, 20:06Z). Re-runs of workflow runs created before the upgrade merged still pass on their pre-upgrade merge commits (e.g. fix/splat-catchall at 18:08Z) — that's why the failure looked flaky. Note #6854's head ran no Actions workflows at all (only cubic + Greptile), so this path was first exercised post-merge.

Mechanism: react-router 8 prerenders through the Vite Environment API path: startPreviewServer boots a vite preview server with the project's vite config (inline options set only port: 0) and fetches each path from resolvedUrls.local[0], i.e. http://localhost:<port>/. In the docker container localhost maps to both 127.0.0.1 and ::1, and the bound socket vs. the prerender fetch can resolve to different address families under bun → connection refused. Dev machines with IPv4-only localhost never hit it, which is why #6854's local verification (and my sandbox) passed. react-router 7 was immune because its build doesn't use the HTTP prerender path (the passing logs have no "Using Vite Environment API"). TanStack Router hit the identical bug in their preview-server prerender — builds fine locally, ECONNREFUSED in docker — and fixed it by pinning the preview host: TanStack/router#6275 → TanStack/router#6305.

Fix (d5cc49a): pin preview: { host: "127.0.0.1" } in the generated vite config. react-router's inline preview options don't set host, so the config-file value survives the merge; the server binds the literal IPv4 loopback and resolvedUrls.local[0] becomes http://127.0.0.1:<port>/ — no hostname resolution left in the loop. Unit test added asserting the pin.

Verification on this branch merged with main (react-router 8.3.0 + vite 8.2.0): fresh blank app reflex export with node and bun stripped from PATH (reflex downloads its pinned bun 1.3.14, mirroring the docker image) succeeds — the preview server binds 127.0.0.1 and serves 200; the memo playwright suites pass 10/10 on the merged combination. My sandbox has no docker daemon and no IPv6 stack, so the in-container repro is validated by this PR's own CI run — watching it now.

Also merged main into this branch (417889f) so CI exercises the real shipping combination. If you'd rather land the prerender fix on main independently of this PR, d5cc49a cherry-picks cleanly.

---

Generated by Claude Code

[masenf]

Second react-router 8.3.0 fallout found and fixed (84adcbf). The reflex-install-and-init run on d5cc49a got past the prerender failure — confirming the preview.host fix in the actual docker environment — and then crashed in the dev phase: react-router 8's dev CLI requires the development export condition (probed via its #development-condition-enabled import map) and re-executes itself with NODE_OPTIONS=--conditions=development to enable it. bun does not apply that option when it executes the CLI — and with no node in the container, bun is what runs it — so the restarted process probes false again and trips the CLI's restart guard: restartWithMergedOptions() was called, but the process has already been restarted. The dev server exited, and the workflow hung waiting on the frontend port until cancelled (that's the wedged 55-minute run — I cancelled it after it blew past the 30-minute job timeout). Environments with node installed never see this because the CLI's #!/usr/bin/env node shebang wins, which is why the 24-job integration matrix stayed green while the docker test died.

Fix: run_frontend now appends --conditions=development to NODE_OPTIONS and BUN_OPTIONS (merged with any existing value, idempotent) before spawning the dev server. Under bun, the CLI's restarted process picks BUN_OPTIONS back up from the environment and comes up with the condition enabled; under node, the condition is active up front and react-router's self-restart is skipped entirely. Prod is untouched — run_frontend_prod serves compiled static files with no react-router CLI, and vite's bundle resolution conditions are mode-driven.

Verified node-free (reflex-managed bun 1.3.14, mirroring the docker image): plain bun run dev reproduces the CI crash exactly; with the fix, reflex run --env dev on a fresh blank app serves 200 with zero restart errors. Unit tests added.

Heads-up for main: after cherry-picking the prerender fix, main will also need this one — the docker workflow runs reflex run --env dev right after the export. Both fixes are self-contained commits (d5cc49a, 84adcbf).

---

Generated by Claude Code

[cubic-dev-ai]

All reported issues were addressed across 4 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>