Skip to content

Fix issues found in a live self-hosted platform end-to-end test - #374

Merged
cpojer merged 49 commits into
mainfrom
fix/platform-e2e-findings
Oct 6, 2026
Merged

Fix issues found in a live self-hosted platform end-to-end test#374
cpojer merged 49 commits into
mainfrom
fix/platform-e2e-findings

Conversation

@Brooooooklyn

@Brooooooklyn Brooooooklyn commented Oct 5, 2026 •
edited
Loading

Copy link
Copy Markdown
Member

Summary

Fixes found during a live end-to-end test of the self-hosted Void Platform on a disposable Cloudflare account: clean up an old installation, install a new platform on a domain, deploy every app kind, run lifecycle commands, and uninstall.

Uninstall

  • New void platform uninstall --purge-resources: deletes every resource the installer created (Workers, Queues, KV, R2, Hyperdrive, dispatch namespace, AI Gateway, routes, DNS records, custom domains, installer Access apps, Email Routing rules that target the platform's gateway), in an order Cloudflare accepts. Each resource is proven against the installation's ownership record right before its DELETE. Changed or foreign resources are kept and listed. Runs are resumable. Zones, adopted resources and external PostgreSQL are never deleted.
  • Plain uninstall now deletes the Workers for Platforms dispatch namespace by name. Before, it was silently left in place.
  • --plan and the final output list every kept resource in manual cleanup order. The docs give an order Cloudflare accepts, with the tail Worker last and Email Routing included.
  • Project creation is refused while an uninstall is in progress, so a purge cannot race a new project.

Install

  • The empty-database check refuses leftover Drizzle migration history before claiming the database.
  • The DNS read in install --plan names the missing token permission. Before, it printed a raw 403.
  • The token instructions say how to scope zone policies; the dashboard links cannot preselect a zone.
  • The tail Worker is no longer published on workers.dev or preview URLs.

Runtime and deploy

  • Legacy multi-statement migrations run statement by statement, so hand-written SQL applies on real D1.
  • Migration warnings are shown after deploy.
  • vinext App Router: the worker entry builds regardless of plugin order, and readiness no longer passes an entry that always throws.
  • Platform ISR stores edge copies as HIT.
  • void project logs shows each runtime log line once.
  • Markdown pages read their frontmatter after a leading <script> block.
  • Smaller fixes: deploy asset count, per-path ISR rules in project status, a log file per concurrent deploy, ISO-8601 token timestamps, readable operator output, a single project delete result line, CLI notes in zero-width terminals.
  • platform status and operator commands report a disabled platform, or an upgrade in progress, clearly.

Playgrounds and docs

  • Playgrounds with hand-written migrations get Drizzle metadata, so deploys pass the drift check. The drift message explains the fix.
  • React Pages keeps its live stream in a scanned directory, and the docs say how to scan other directories.
  • Docs: what disable blocks, the enable redeploy, CI project selection, and committing .void/project.json.

Verification

  • vp run check, vp lint, vp fmt --check: clean.
  • vp test run: 680 files, 12,866 tests passed, 29 skipped.
  • Live, on the branch build:
    • Fresh install on a domain zone.
    • astro, analog, tanstack-start, react-router (original hand-written migration), vinext-app (original plugin order), vinext-pages, react-pages (live SSE), markdown-blog, named-layouts ISR, and the kitchen-sink runtime smoke test (12/12) all pass.
    • Runtime log lines appear once.
    • --purge-resources removed every installer resource in one run, and the account matched its state from before the install.
  • The first live run on main also covered the four Pages adapters, SPA, SSR, SSG, static directories, WebSockets, SvelteKit, Nuxt, rollback, disable/enable, project tokens, secrets and logs.

GitHub OAuth login was not exercised.

Follow-ups (product decisions, not changed here)

  • A platform deploy warns "Missing required secrets … add them to .dev.vars" even when the secrets are set on the project.
  • void project logs rejects --project while sibling commands accept it.
  • Static --dir deploys mark /assets/* immutable even when file names have no hash.
  • No CLI command lists cron triggers.
  • A same-version upgrade or repair is not a no-op, and enable redeploys all Workers.
  • A scoped project token that deploys to another existing project reports "Failed to create project".
  • secret delete of a missing secret reports success.
  • platform status exits 0 for a disabled platform.
  • platform config auth prints text, not JSON, when stdout is not a TTY.
  • The uninstall plan lists adopted DNS records and zones under "Keep for manual cleanup".

🤖 Generated with Claude Code

Brooooooklyn and others added 30 commits October 5, 2026 21:59
@Brooooooklyn @claude
Delete the Workers for Platforms dispatch namespace by name on uninstall
826db02 
Cloudflare addresses dispatch namespaces by name, but uninstall sent the
namespace UUID. Cloudflare answered 404, which uninstall treats as already
gone, so the namespace survived while the manifest recorded it as deleted.

deleteResource now takes the manifest resource and uses its name for
dispatch namespaces and its id for every other kind.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Refuse databases with leftover migration history before claiming them
45e5425 
The installer's empty-database check only looked at tables on the search
path, so a database that still held drizzle.__drizzle_migrations passed as
empty. The leftover history was then read as already-applied migrations
after the claim and after Cloudflare resources were created, and the
install failed late. The check now also looks at the drizzle schema and
refuses the database up front. A resumed install that already holds the
matching claim is unaffected.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
List the resources uninstall leaves behind
af6d776 
`void platform uninstall` now names every resource it keeps (kind, name and
id) in the order of the manual cleanup steps, plus recorded Access
resources. `--plan` shows the same list before any change. When the
installation sends email, both say that Email Routing for the sender domain
is kept. The final message points to the local installation record.

A created AI Gateway that is already gone from Cloudflare is now recorded as
deleted instead of retained.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Draw CLI notes safely in zero-width and tiny terminals
cfccc04 
The shared box renderer threw "Invalid count value: -1" when the
terminal reported fewer than six columns, as a 0x0 pty does, so any
note() call (for example the uninstall plan) crashed the command.

Route every void note() through a small wrapper in output.ts that
treats a non-positive width as unknown (80 columns, like the rest of
the CLI) and draws narrower terminals at the smallest width the box
supports. The vendored renderer stays byte-for-byte upstream.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Document a manual cleanup order that Cloudflare accepts
a92a2b2 
The uninstall guide now deletes the tail Worker last, after the platform
Workers and the dispatch namespace apps that send logs to it, and adds an
Email Routing step before the email gateway Worker is deleted. It also says
where the local uninstall record and recovery file stay, and that both can be
deleted after cleanup.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Add Cloudflare calls needed to tear down a platform
5ce1e49 
The platform Cloudflare client can now delete Worker routes, custom
domains, Queue consumers, dispatch namespace scripts, and Access
applications and service tokens, and list Queue consumers and namespace
scripts. Workers are deleted by script name, and an AI Gateway is
deleted with the runtime credential when one is set, like its other
calls.

Deletes never send force, treat 404 as already gone, accept a success
response without a result, and include Cloudflare error codes in their
errors (for example 100109 for a tail Worker still in use). Emptying an
R2 bucket can run an ownership check before each page of deletes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Add void platform uninstall --purge-all to remove everything Void c…
1794a91 
…reated

`--purge-all` includes `--purge-data` and also deletes the platform
Workers, Queues, the dispatch namespace scripts, the R2 bucket, the AI
Gateway, the platform DNS record, Worker routes and custom domains, and
the Access applications and service tokens the installer created. It
keeps zones, adopted resources, external PostgreSQL, Email Routing
settings, Analytics Engine datasets, and API tokens.

It deletes in an order Cloudflare accepts: traffic and DNS, Access,
Queue consumers, Workers except the tail Worker, Queues, the dispatch
namespace, the tail Worker, Hyperdrive and KV, R2 (emptied first), the
AI Gateway, and routing KV last. Each resource is proven to belong to
the installation right before its DELETE. Routes and custom domains
must still serve the installer Worker they were recorded for. Progress
is checkpointed, so rerunning the command continues after a failure.

Before any change it refuses while projects remain, and checks that the
credential can make every planned deletion: browser login cannot delete
DNS records or Access resources, and missing token grants are named.
`--plan` lists every resource it will delete, in order, and every
resource it keeps, with the token permissions it needs.

Ownership checks now list dispatch namespaces only when one is being
checked, so installations without Workers for Platforms make no such
call. The kept-resource list after uninstall now puts the tail Worker
after the dispatch namespace, matching the manual cleanup steps.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Prove each --purge-all deletion against the right owner
7331f9b 
- Look up and delete Access resources in the account that holds them, so a
  service token in a separate identity account is deleted instead of being
  recorded as gone. The credential check names that account too.
- Delete the AI Gateway only when Cloudflare reports it was created next to
  the installer's provision intent. A gateway created later under the same
  id, or one without a reported creation time, is kept and listed.
- Read the project list again before deleting the application Workers in
  the dispatch namespace, and stop if a project was created during the run.
- Save the R2 bucket's generation when its ownership marker is removed, so
  a rerun can finish buckets recorded before creation fingerprints existed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Keep changed resources, detach Email Routing, and finish stopped unin…
878cbf6 
…stalls with --purge-all

- A route, custom domain, DNS record, or Access resource that no longer
  matches what the installation recorded is kept instead of stopping the
  run. It is listed with the reason in the final kept note, and a rerun
  does not look at it again.
- When the platform sends email, --purge-all deletes the Email Routing
  rules in the sender zone that send mail to this installation's email
  gateway Worker, and resets such a catch-all rule to its disabled drop
  default, after proving the Worker still carries the installation
  marker and before deleting it. Rules for other Workers, Email Routing
  itself, and the sender domain's MX, SPF, and DKIM records are kept and
  listed. The credential check names Zone → Email Routing Rules → Edit,
  and only when email is enabled.
- If a plain or --purge-data uninstall stopped before it finished,
  --purge-all now deletes the resources that run recorded as retained,
  including routing KV, with the same proof as ready ones. A finished
  uninstall still cannot be reopened.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Tell installers to switch token zone policies to Specified Domains
c6d5c64 
Cloudflare's account token links cannot preselect a zone, so the
dashboard opens zone permissions under All Domains. The management and
runtime token prompts now name the exact change (All Domains to
Specified Domains and the zone to select), keep All Domains when Void
creates the zone, and the installation docs say the same.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Read markdown page frontmatter after a leading script block
08e5226 
Page title, description and layout in .md files are now found when the
page opens with a <script> block, matching how @void/md reads the same
frontmatter.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Keep the React Pages live stream in a scanned directory
c368981 
The playground defined its live stream in lib/, which binding inference
does not scan, so builds emitted no VOID_LIVE Durable Object and
/api/live/publish failed. Move it to src/live.ts as the live guide shows,
document that the module importing void/live must be in a scanned
directory, and list void/live in the resource inference table.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Prove reopened resources before any change, and list kept Access reso…
9103b46 
…urces with other kept items

- When --purge-all finishes an uninstall that stopped before it finished,
  the resources that run recorded as retained are now proven before the
  first change, like ready ones. A Worker that no longer carries this
  installation's marker stops the run before anything is deleted, and
  the retained resources stay recorded as retained.
- An Access resource that an earlier --purge-all kept because it changed
  is listed under the kept resources with its reason, not under "Keep
  Access resources you connected".

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Add Drizzle metadata to playgrounds with hand-written migrations
faf9dfd 
analog, react-pages, tanstack-start, vinext-app, vinext-pages, vue-pages
and pg-test shipped a hand-written initial SQL migration without a Drizzle
journal or snapshot, so every deploy stopped at the schema drift check.
Their schemas also disagreed with the SQL (missing unique email, missing
created_at, serial author_id).

Align each db/schema.ts with its SQL and add a journal plus snapshot under
the existing migration tag, keeping the SQL files unchanged, so the drift
check passes and later `void db generate` runs diff from this baseline.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Explain how to add Drizzle metadata when deploy finds hand-written mi…
b9ea2ce 
…grations

When db/migrations has SQL files but no Drizzle journal or snapshot, the
drift error told users to run `void db generate`, which cannot work: it
emits a migration that creates the existing tables again and fails the
fresh-database check. The error now names the missing metadata and the
steps to add it, and keeps the generate advice for Drizzle-managed
projects.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Give the 101-project Zero Trust disable test room under a loaded machine
210342a 
The test drives two full project batches and passes in about 1.5s alone,
but took 10.4-14.9s when the platform API suite ran alongside other test
runs, so its 10s limit timed it out. Under that load it timed out in 8 of
9 runs before and 0 of 9 after (slowest 15.5s). This matches the 30s limit
the 103-project domain activation test already uses.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Keep the runtime token on All Domains when Void creates the zone
445bf27 
The runtime token prompt told installers to switch to Specified Domains
even when the install plan creates the zone, which cannot be selected
before it exists. It now keeps All Domains in that case so Cache Purge
reaches the new zone, matching the management token prompt and the
credentials guide.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Say that a custom live stream directory must keep the default scan list
34d7387 
Setting inference.scanDirs replaces the default directories, so the live
guide now tells users to list the defaults they still use alongside lib/.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Read the queue consumer Worker from script_name during --purge-all
89dd0ca 
Cloudflare names a Worker queue consumer in script_name. Teardown only read
script and service, so it skipped the installation's own consumers and then
failed to delete their Workers and Queues. Read script_name first, keep the
older names Wrangler still reads, and refuse a Worker consumer that has no
name instead of skipping it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Refuse new projects once purge-all passes its final project check
1ea4335 
Project creation now takes the shared configuration lock that
`void platform uninstall` holds across its last project check. A create
request admitted before traffic was disabled either commits before that
check, so --purge-all refuses before changing anything, or waits and is
refused with 409 once the uninstall is recorded. Before, such a project
could commit right after the check and was only noticed after traffic,
Access, Workers and Queues were already deleted.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Warn against running purge-all while the platform's Workers or routes…
f47ef16 
… are changing

--purge-all proves each resource right before deleting it, but Cloudflare
has no conditional delete, so a change made in between is not detected.
Say so in the uninstall guide and next to the proof helper.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Tell project creation the platform is being uninstalled
e54ba51 
Creating a project while `void platform uninstall` is in progress now
answers 409 with "The platform is being uninstalled, so new projects
cannot be created." instead of the Zero Trust and deployments message
shared by other refused requests.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Write each user log line once when the dispatch Worker also tails it
d16e31d 
A tail batch from the dispatch Worker's consumer holds the dispatcher item
and the user Worker items it invoked. Each user Worker's own tail consumer
already delivers those items, so the tail Worker now writes only the
dispatcher items from such batches. Batches without a dispatcher item
(including crons, migrations, alarms and queues) are written as before.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Store platform ISR edge copies as cache hits
66cf307 
On an ISR miss the dispatch Worker wrote its HTML and JSON edge copies
with `x-isr-cache: MISS`, so every later request served from the edge
still reported MISS. The edge copies now say HIT, matching the direct
runtime; only the response that triggered the render reports MISS.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Run legacy multi-statement migrations statement by statement
c144304 
D1 exec() treats each line as its own statement, so a multi-statement
migration without Drizzle breakpoints failed on any multi-line statement
such as CREATE TABLE. Legacy migrations now split on top-level semicolons
(ignoring strings, quoted identifiers and comments) and run the statements
plus the tracking insert in one D1 batch. SQL that may hold a BEGIN ... END
body, such as CREATE TRIGGER, still falls back to exec().

The legacy-plan warning now tells users to add `--> statement-breakpoint`
between statements, and the playground migrations use breakpoints.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Show migration warnings after deploy and describe the migration check…
b68a979 
… accurately

The CLI dropped migration_warning events, so warnings from the migration
runner never reached the user. They now print with log.warn, which stays
visible above the spinner.

The migration_start count covers every bundled migration, not only the
unapplied ones, so the spinner now reads "Checking N bundled migration(s)..."
instead of "Running N migration(s)...".

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Build the vinext worker entry regardless of plugin order
7bd7095 
With `[voidPlugin(), vinext()]`, vinext's RSC entry replaced the Cloudflare
worker entry for the rsc environment, so the build shipped vinext's bare
handler instead of Void's wrapper and the worker could not serve requests.
Void now restores the Cloudflare worker entry after all plugin configs are
applied.

The deploy wrapper's readiness probe now answers 503 with ready:false when
the wrapped entry has no fetch handler, so such a build is not promoted.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Name the DNS permission when the install plan cannot read DNS records
ba4b877 
When Cloudflare answers 401 or 403 while the install plan reads DNS records
for the application wildcard or the custom API hostname, report which zone
was denied and that the management token needs Zone → DNS → Edit, instead
of surfacing the raw "Authentication error". Other errors pass through.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Keep the platform tail worker off workers.dev and preview URLs
b163a16 
The tail worker only receives trace events from the dispatch worker, so it
no longer publishes a workers.dev subdomain or preview URLs when platform
endpoints are public. Upgrades turn off an existing tail subdomain because
the live subdomain no longer matches the desired trigger state.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Report the deployed asset count without the internal cache-buster
02f1b5a 
Worker deploys add a per-deploy cache-buster file to the asset manifest
sent to Cloudflare. The asset_upload event and the final `assets` count
included it, so a deploy of N files reported N + 1. Report the count of
the user's assets instead, and keep the internal count for gating
Cloudflare asset work.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Brooooooklyn and others added 17 commits October 6, 2026 03:53
@Brooooooklyn @claude
Show per-path ISR rules in the project status ISR column
b066962 
`void project status` showed `-` in the isr column for deployments that
only configure per-path revalidation. The column now shows the global
revalidate interval when set, otherwise the number of per-path rules
(for example `1 rule`), and `-` only when ISR is not configured.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Give each concurrent deploy its own log file
b6a29f9 
Deploy log names only carried a millisecond timestamp and were opened in
append mode, so two deploys started at the same moment on one machine
wrote interleaved records into one file. Add the process id and a short
random suffix to the name and open it exclusively.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Report a disabled platform clearly in platform status and operator co…
4134a65 
…mmands

`void platform status` now prints that the platform is disabled and the
`void platform enable` command when discovery returns the traffic gate's
503 response, instead of exiting with a raw discovery error. Operator
requests that fail without a message now include the API error, so a
disabled platform reads "Platform disabled. Run `void platform enable`."

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Describe the platform disable plan in plain terms
49e1a32 
The disable plan now says it blocks all platform traffic and the
platform API, and that routes, custom domains, Workers, and data stay
installed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Document platform disable scope, enable redeploy, SPA asset 404s, and…
fa26db4 
… CI project selection

Disable also blocks the platform API and CLI, leaving only the
administrator status page and sign-in. Enable redeploys and verifies the
platform Workers before restoring traffic. On a Void platform, SPAs return
404 for missing files under the hashed assets folder. CI deploys must
choose a project with VOID_PROJECT, --project, or a committed
.void/project.json.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Return ISO-8601 timestamps from project token endpoints
782516b 
Token create, list, and renew responses passed Postgres timestamp strings
through unchanged (for example "2026-10-06 04:01:25.899+08"). They now
return UTC ISO-8601 strings; a missing revocation time stays null.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Show readable labels in operator output and platform auth config
0856cf2 
Operator human output now uses sentence-case field labels that keep
known acronyms (API version, GitHub OAuth, Callback URL, Installation ID).
`void platform config auth` commands print the same readable text by
default and keep JSON for --json.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Delete the platform API and proxy Workers that bind to each other dur…
c805240 
…ing --purge-all

The API Worker checks the proxy through a service binding and the proxy
calls the API through another, so Cloudflare refused whichever went first
and --purge-all stopped at the proxy Worker. When Cloudflare refuses a
platform Worker, purge-all now reads what depends on it the way
`wrangler delete` does and forces the DELETE only when every dependent is a
Worker this installation created. A Worker outside the installation still
stops the run, and the message now names it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Read AI Gateway creation times as UTC so --purge-all deletes the inst…
8808165 
…aller's gateway

Cloudflare reports an AI Gateway's created_at without a zone, for example
"2026-10-05 18:29:14". JavaScript read that as local time, so outside UTC
the gateway the installer created seemed hours away from its creation
record, and --purge-all kept it as not created during installation.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Print one result line when deleting a project
dc00c83 
`void project delete` printed "Project deleted." from its spinner and then
"Deleted <slug>". The spinner now clears, so only "Deleted <slug>" remains.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Prove a Worker's dependents before --purge-all forces its delete
d2a37db 
--purge-all forced a platform Worker's DELETE when each Worker that bound
to it had the name of a Worker the installation created. A Worker that
took one of those names during the run, or after the installation's own
Worker was recorded as deleted, passed that check, so the forced DELETE
left it with a broken binding. Each dependent must now be a Worker this
run still deletes and must carry this installation's ownership marker;
otherwise the run stops before forcing anything.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Tell an upgrade or repair apart from a disabled platform
b8fda92 
An upgrade's staged window and a Sandbox drain close the API traffic gate
with the same 503 as `void platform disable`, so `void platform status`
reported the platform as disabled, suggested `void platform enable` on
machines without the saved operation, and exited 0. Operator commands
suggested enabling it too.

The API now adds `reason: "maintenance"` to that 503 while an upgrade or
repair has traffic paused. `void platform status` then exits with an error
that says an upgrade or repair is in progress, as the docs describe, and
operator commands say to retry when it finishes instead of suggesting
`void platform enable`. A platform without the field is still reported as
disabled.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Warn only about migrations that have not been applied yet
b255cd6 
Deploys now print migration warnings, and the runner built them from
every bundled migration. An app with an older migration that lacks
statement breakpoints saw the advice to add them on every deploy, though
that migration never runs again. The runner now drops the warning once
the migration is recorded as applied.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Explain how to commit .void/project.json in an app that ignores .void
7dad32a 
The CI guide suggests committing .void/project.json, but new apps list
.void in .gitignore, so git refuses to add it. The guide now gives the
ignore rules that keep the rest of .void out and let the link in.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Correct which invocations skip the dispatch Worker in the tail Worker…
61a1637 
… note

Queue deliveries reach user Workers through the dispatch Worker, so they
are not among the invocations whose only log path is the user Worker's
own tail consumer.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Load the platform discovery module under Node's TypeScript type strip…
b03ab84 
…ping

The maintenance flag on PlatformDisabledError was a constructor parameter
property, which Node's strip-only TypeScript mode rejects, so scripts that
import the CLI source directly failed to load. It is now a plain field.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Collapse duplicate runtime log rows when reading project logs
31cf75f 
A request routed through the dispatch Worker reaches the tail Worker
twice: through the dispatch Worker's tail consumer and through the user
Worker's own consumer. On a live install Cloudflare delivered each copy
in its own tail invocation, never in one batch with the dispatcher item,
so the tail Worker could not drop the second copy and every request
appeared twice in `void project logs` and the deployment log views.

The tail Worker again writes every item it receives. The project log
query and the operator deployment log query now group by every stored
column, so each invocation is returned once, and paging still counts
unique rows.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Oct 5, 2026

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Oct 5, 2026 •
edited
Loading

Copy link
Copy Markdown

🚀 Deploying Preview to Cloudflare 🚀

Preview URL: https://fix-platform-e2e-findings-void-cloud.voidzero-docs.workers.dev (commit 277617d)

This URL reflects your latest Preview deployment

Preview Deployments by commit

Status Deployment URL Commit Updated (UTC) See this deployment's details
  • Build: Success ✅
  • Deployment: Success ✅

View logs ↗		 https://b869a3a2-void-cloud.voidzero-docs.workers.dev 277617d 2026-10-06T03:53:28.711Z Visit the dashboard ↗
  • Build: Success ✅
  • Deployment: Success ✅

View logs ↗		 https://c8ad427f-void-cloud.voidzero-docs.workers.dev edd5b17 2026-10-06T03:25:43.719Z Visit the dashboard ↗
  • Build: Success ✅
  • Deployment: Success ✅

View logs ↗		 https://86463218-void-cloud.voidzero-docs.workers.dev 31cf75f 2026-10-05T21:43:42.460Z Visit the dashboard ↗

@cpojer

cpojer commented Oct 6, 2026

Copy link
Copy Markdown
Member

Nice. fwiw I left resource removal intentionally out initially (and kinda forgot to come back to it). I'm wondering if --purge-all would be better as --delete-resources or --delete-all-resources. Descriptive names for destructive commands you'll only run once or twice are ok to be long.

Brooooooklyn and others added 2 commits October 6, 2026 11:23
@Brooooooklyn @claude
Rename platform uninstall --purge-all to --delete-resources
edd5b17 
The flag does not delete everything: zones, adopted resources, and
external PostgreSQL stay. The new name says what it does and matches
--delete-zone.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn @claude
Rename platform uninstall --delete-resources to --purge-resources
277617d 
Both destructive uninstall flags now use the same verb: --purge-data
and --purge-resources.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@Brooooooklyn

Brooooooklyn commented Oct 6, 2026

Copy link
Copy Markdown
Member Author

Renamed it to --purge-resources, so it matches the existing --purge-data.
It still keeps zones, reused resources, and external Postgres, so I left "all" out of the name.

@Brooooooklyn
Brooooooklyn requested a review from cpojer October 6, 2026 03:59
@cpojer
cpojer merged commit 71c8403 into main Oct 6, 2026
21 checks passed
@cpojer
cpojer deleted the fix/platform-e2e-findings branch October 6, 2026 05:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@cpojer cpojer cpojer approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@Brooooooklyn @cpojer