Plugin validation

Onboarding preview `0.4.1-dev.1` (2026-09-25)

This development preview selects released CLI 8.0.0-rc.17 for new apps and passes --ui-context prisma-plugin during login. Existing apps retain their toolchains; an older CLI without the option uses standard browser login and an agent explanation to return to chat. No new deployment is required for this instruction change. Composer/cloud 0.21.0, ORM Postgres 8.0.0-rc.11, branding, MCP diagnostics, data-use restrictions, and the two-skill structure are retained.

Directly executed release checks on macOS with Node 24.16.0 and npm 11.13.0:

npm published prisma@8.0.0-rc.17; the publish workflow completed successfully.
A clean installation of the exact CLI/Composer/cloud/ORM set succeeded without --force or --legacy-peer-deps. Composer, cloud-control and ORM target-control imports succeeded. The ORM package exports subpaths, not a root module.
Released auth login --help lists --ui-context with prisma-plugin as its supported value. Inspection of the published bundle confirms context propagation, return-to-ChatGPT success/failure wording and conditional omission of the skills-install section. This inspection is not a completed browser login test.

Executed preview checks:

The existing skill validator, packaging regression (one test), rebuilt-bundle integrity check and whitespace check passed. The installed prisma-preview reports 0.4.1-dev.1; both skills and all seven installed files match the bundle.
A separate seven-file preview ZIP matches the bundle byte-for-byte. SHA-256: c340e05c8951d525f7640ff5c5e5034e7641f917623802a5a7dcf94910da3866. Composer retains SHA-256 67b50e78fbb6cafd00bb99b0e56fe8a49e4a7190219474bf1b9be933f08bbcbb.
auth login, dev, deploy, and project list help commands ran successfully with the released CLI. Existing 0.4.0 submission archive checksums still pass.

Browser acceptance used a new isolated credential file, initially confirmed unauthenticated. Luan completed the standard browser flow with his existing account and selected ChatGPT Plugin Review. The login process exited zero; the subsequent auth whoami --json reported both authenticated: true and verified: true, and a remote project list --json succeeded with three existing projects in that workspace. No personal session was reset and no cloud resource was created or changed.

Luan's screenshot of the actual success page confirms the return-to-ChatGPT wording and absence of the terminal/skills-install section. The page still uses the CLI's previous logo; updating that embedded artwork is a separate upstream follow-up. This was returning-account authorization, not fresh account creation or a new application/deployment test. Failure-page wording was inspected in the published code, not exercised through a new failed browser login.

Instruction review (not executed agent acceptance): existing valid sessions still reuse identity and remote-access checks without login; older CLIs without the option retain standard login without an automatic upgrade. No new test framework or repeated cloud deployment was added.

The submitted 0.4.0 source is preserved in commit 7658737; all seven bundle files matched its saved archive before changes. Its archive and per-skill ZIPs remain unchanged. Luan submitted that version, and the portal was observed in Review on 2026-09-25. Older sections below record preparation-time results and stopping points; they do not describe the current submission status or prove that the new CLI has completed an end-to-end deployment.

Historical data-use restriction revision (2026-09-25)

Version remains 0.4.0. The listing and authored workflow now prohibit processing PHI and PCI-regulated payment card data, including through local files, CLI, database operations, MCP, logs and application verification. The existing diagnostic reference applies the same restriction before log access and preserves secret redaction and the broader-permissions disclosure. No server, authentication, dependency, packaging-behavior or upstream Composer changes were made.

Scenario review of the authored instructions (not executed agent tests or proof of technical enforcement):

Scenario	Reviewed behavior
Ordinary Todo request without restricted-data context	Continues the existing build/deploy journey without an extra questionnaire.
Healthcare prototype explicitly using synthetic data in an isolated environment	Remains supported; no real patient examples are requested.
Request to inspect real patient records	Stops before file/database/tool access, explains the restriction and offers a synthetic environment.
Deployment logs may contain restricted data	Clarifies before retrieval; does not fetch for later redaction or use CLI to bypass the restriction.

The same paragraph requires stopping further access and avoiding reproduction in responses or artifacts after unexpected exposure. It covers application/browser verification as well as MCP, so switching tools does not remove the restriction.

Executed checks:

The existing skill validator, packaging regression (one comprehensive test), rebuilt-bundle integrity check and whitespace check passed. Initial packaging attempts could not reach npm from the sandbox; rerunning with registry access and Node 24.16.0 passed without changing dependencies.
Both installed skills and all seven bundle files match the rebuilt source byte-for-byte. Composer 0.21.0 retains SHA-256 67b50e78fbb6cafd00bb99b0e56fe8a49e4a7190219474bf1b9be933f08bbcbb.
Complete and individual workflow ZIP contents were compared with source bytes. Checksums were refreshed; the prior skills-only fallback and Composer skill ZIP remain unchanged. No cloud deployment or new recording was performed.

The public description was appended and verified after reloading the existing portal draft. The revised workflow skill replaced the previous upload and passed its new scan; Composer remains passed and exactly two skills are listed. All other Info fields, icon asset identities, video URL and release notes were unchanged. Publisher declarations remain unchecked and the draft has not been submitted. Existing reviewer-access notes explicitly disclose missing dedicated credentials; clearing form validation does not satisfy that requirement or establish an OpenAI exception. Publisher declarations and submission remain Luan's decision.

Current candidate: MCP draft preparation (2026-09-24)

The user approved resuming the existing With MCP draft after the isolated server fix shipped. Version remains 0.4.0; the two-skill structure, dependency pins, upstream Composer reference, and CLI build/deploy path remain unchanged. Optional diagnostic instructions are restored in the authored workflow and its existing reference. No new upstream or packaging-behavior changes are part of this work. The stopping point is Luan's review of the saved draft, not submission.

Direct production checks after PR prisma/pdp-control-plane#5458 merged:

The challenge returned HTTP 200, text/plain, and the exact draft token; the OpenAI portal now shows Domain verified.
A fresh OAuth-authorized scan in Plugins returned 31 tools. All three missing openWorldHint values are present; all 93 annotation explanations are saved.
OAuth resource discovery returned HTTP 200; unauthenticated MCP returned 401; an unknown well-known path returned 404.
A fresh read-only draft audit confirmed the five positive and three negative cases and listing copy match prior work. Icons, skills, reviewer credentials, and demo URL were still absent. The directory version is Draft. There is no visible editor audit trail, so this does not prove nobody else accessed it.

The authenticated scan verifies discovery, not execution of diagnostic tools. Application/log smoke results and final packaging/upload outcomes follow below. Dedicated reviewer access, reviewer-account tests, the hosted recording, and publisher review remain pending. No submission or publication is authorized by this preparation task.

Draft preparation checks (2026-09-24–25)

Existing packaging regression passed, and the authored skill passed the skill-creator validator. The validator needed PyYAML in an isolated temporary environment; no dependency was added to this repository.
The focused preview was installed as prisma@prisma-preview, version 0.4.0. The bundle and installed files were compared byte-for-byte. The original broad installation was preserved. The Composer reference SHA-256 remains 67b50e78fbb6cafd00bb99b0e56fe8a49e4a7190219474bf1b9be933f08bbcbb.
CLI auth whoami, remote project listing, and service inspection succeeded for the existing Plugins Todo target. The service had a running live version. No deployment, restart, or cloud resource mutation was performed.
MCP Inspector 2.8.0 completed a separately authorized OAuth connection to Plugins, using isolated temporary storage. All five diagnostic reads succeeded: workspace identity, project-filtered apps, branch-filtered builds, app deployments, and deployment logs. Workspace, project, app, live-version and endpoint identifiers matched the CLI target. There were no build-history records; logs returned a clean end at cursor 0. CLI log inspection for the same version also succeeded with no retained lines. Empty history/logs are not represented as a build failure or proof of application health.
The authenticated server advertised 31 tools with explicit annotation booleans. Inspector reported zero schema portability errors and 29 warnings across 13 tools; the portal's successful scan remains the submission scan result. No upstream schema changes were made.
An allowlisted 9-file source fixture was prepared outside this repository. It excludes credentials, local state, reports, dependencies, and machine paths. Reviewer-account testing and the hosted recording remain pending.
Luan created ChatGPT Plugin Review and approved one isolated Todo deployment there. Its source copy has a distinct project name and passed clean dependency installation, typechecking, and build under Node 24.16.0. The first installation context resolved Node 23.11.0; it was repeated with the selected Node executable pinned in PATH. Luan completed its isolated CLI OAuth login. Authenticated identity and a remote empty project list confirmed the requested workspace. The approved fixture then deployed successfully with the previously working Bun invocation. CLI inspection confirmed a running live version; browser checks verified creation, completion, persistence after reload, and deletion. One sample Todo remains for review. Local restart persistence uses the earlier baseline; it was not repeated. The missing Git history warning was reported separately from application success. No commit was created to suppress it. This is owner-authorized setup, not a dedicated reviewer's sign-in test.
Chrome upload permissions were enabled by Luan. Both approved PNG icons were uploaded and saved. The full plugin ZIP left no skills in the With MCP draft; individual skill ZIPs with root SKILL.md were accepted. Both skills are now listed and scanning. Skill source and supporting-reference bytes are unchanged; no packager behavior changed and the complete release/fallback ZIPs remain.
The saved deployment test expectation was corrected to distinguish deployment records from live runtime status. The final portal page currently flags missing icons, private test credentials, and publisher declarations before the icon uploads. Release notes are present; availability remains the existing Allow all selection for Luan's review. Skill safety scans remain pending and the hosted recording is still absent. No declarations were accepted and nothing was submitted.

Static scenario review of the small instruction change: ordinary build/deploy does not require MCP; unavailable/denied/mismatched access preserves the selected target and app progress; partial deployment retains state and does not trigger MCP writes or automatic cleanup. Explicit Vercel hosting, general PostgreSQL education, and AWS diagnostics from supplied logs do not select this workflow. These are instruction reviews, not executed fresh-agent or reviewer-account tests. Inspection of the server schema also confirms deployment listing returns records and version IDs, not runtime status; the reference now makes that limit explicit. Existing local restart and live CRUD evidence remains the baseline.

Historical skills-only candidate (2026-09-24, superseded)

The user ruled out all changes outside the plugin repository. The optional MCP integration was removed from the candidate and listing; Composer/CLI remains the working build-and-deploy path. MCP metadata, domain verification, and a new CLI release are not dependencies of this package. The prior MCP scan/draft evidence below is historical, not the current publication route. The existing unpublished draft was preserved.

The official Skills only route is documented but absent from the verified Prisma organization's Create plugin menu. Draft menus expose no type-conversion action. OpenAI must provide the supported submission route; the cause of the missing option has not been established. No upstream changes or public submission occurred.

Initial submission candidate (`0.4.0`, 2026-09-24)

The initial release uses the verified prisma@8.0.0-rc.15 package set and standard prisma auth login --json. Plugin-specific browser completion wording is deferred; the agent explains that the user should return to chat and continues to require login completion, authenticated identity, and remote workspace access. The listing now states the desktop-local requirement and macOS validation scope.

The CLI feature and release PRs were merged, but publication failed. A fresh npm lookup on 2026-09-24 still returns E404 for prisma@8.0.0-rc.16; no unpublished version or unsupported context option is used. Installed rc.15 login help confirms the normal browser login command and JSON output are supported.

Direct checks: the existing packaging regression test, authored skill validator, and rebuilt bundle integrity check passed. A stray macOS .DS_Store in the bundle root caused the first integrity check to fail; removing that metadata file restored the expected file set. The two skills, supporting reference, approved branding, license, and provenance are included. The imported Composer 0.21.0 reference is unchanged. No personal sessions, installed plugins, or cloud resources were changed.

The older root 0.3.0 installation is sourced from the local repository; its visibility in Codex is not evidence of public directory publication. The user subsequently obtained access to the existing verified Prisma business organization and created its dedicated Prisma Plugin project. An unpublished With MCP draft exists there. The portal offers separate MCP and Skills sections, so the server is registered in the portal and the two-skill bundle is uploaded separately. This supersedes the earlier unverified-organization/skills-only investigation.

Application/deployment evidence remains the previously completed macOS runs below; no fresh-account signup or additional deployment was performed for this candidate. The original fully verified novice-onboarding acceptance remains follow-up work. The following sections are historical records; their dev.4 release hold does not apply to this rc.15-based initial submission candidate.

Optional MCP diagnostics (2026-09-24)

The workflow now routes explicit diagnostics or failure investigations to five read-only operations on the existing Prisma MCP server. It preserves the CLI target, requires matching workspace/application IDs, keeps MCP and CLI sessions separate, and falls back to supported CLI inspection when MCP is unavailable. The server's broader write tools and workspace:admin/offline_access scopes are not narrowed by these instructions. No server, packaging behavior, or API changed.

Direct local checks passed: packaging regression, skill-format validation, bundle integrity, and diff whitespace. The upstream Composer reference SHA-256 remains 67b50e78fbb6cafd00bb99b0e56fe8a49e4a7190219474bf1b9be933f08bbcbb. Approved artwork was rendered to 512px and 96px square PNGs for the portal.

Static scenario review confirms that ordinary build/deploy requests do not require MCP, missing/denied/mismatched MCP access preserves app progress, and partial deployment recovery does not claim success or invoke MCP writes. These are instruction reviews, not fresh-agent executions or authenticated MCP smoke tests. Workflow, recovery, and the portal's five positive/three non-trigger cases are in submission materials; unexecuted cases are not represented as passed. Existing macOS deployment evidence below remains the application baseline.

Authenticated submission scan (2026-09-24)

The user approved ChatGPT scanner OAuth access to the isolated Plugins workspace. The scan succeeded and returned 31 tools, including all five intended diagnostics. The draft contains explanations for all 90 supplied annotation values. It also reports missing openWorldHint on create_prisma_postgres_database, create_prisma_postgres_connection_string, and create_prisma_postgres_recovery. Those explicit booleans must be supplied by the MCP server and rescanned; portal text cannot repair them. The three source definitions were located in pdp-control-plane/services/mcp-server/mcp/tools/; a three-line annotation patch was prepared outside this repository and passed git apply --check against that checkout. It was not applied or deployed. No MCP resource operation was executed during the scan, so the app/state/log smoke test remains pending.

The domain challenge URL returned HTTP 404. The exact token and maintainer handoff are kept with local submission artifacts, outside the repository. OAuth discovery also reported enterprise domain restrictions unavailable; this was a warning, not an observed blocker requiring an OIDC implementation in this revision.

The existing draft now contains listing copy, verified Prisma business identity, starter prompts, release notes, and the required review-case drafts. Icon/ZIP uploads were blocked by Chrome extension file access; native computer-use fallback was also unavailable. Reviewer credentials, fixture checks, demo recording, domain verification, missing server annotations, and publisher declarations remain incomplete. No submission, legal acceptance, or public publication occurred.

Local PR review (2026-09-24)

Review of PR #7 found and resolved two issues before merging the preview source:

The default repository marketplace pointed at an incomplete, release-gated bundle. It now remains byte-for-byte identical to main's existing root-plugin route. The opt-in prisma-preview marketplace lives under plugins/; its documented installation step is conditional on the onboarding release gate.
The ignore-rule regression assertion skipped tracked files. It now uses git check-ignore --no-index; an isolated tracked-file fixture reproduced the original false negative and confirmed the corrected command detects it.

The packaging regression test, skill-format validator, bundle integrity check, marketplace target/asset/skill checks, and diff whitespace check passed. Composer's imported reference is unchanged. No installed plugin, personal credentials, or cloud resources were changed. This repository currently has no GitHub CI checks; these are locally executed results, not hosted CI or fresh-user acceptance.

Desktop onboarding revision (`0.4.0-dev.4`)

The authored workflow now requires desktop-local execution, agent-managed setup, one managed browser login, a real emitted authorization link when opening fails, and successful login/identity/remote-access checks before confirming connection. It preserves the workspace selected during consent and keeps local work moving while signup is pending. Recovery retains app progress and never requests codes, tokens, or callback URLs. Web/cloud execution is explicitly deferred.

Upstream implementation: Prisma CLI PR #281, commit b50b587, based on main 6a34270. The optional context is validated by the normal argument parser and propagated only to browser completion rendering. OAuth, scopes, consent, callback validation, credential persistence, and terminal output are unchanged. No source changes were made to Composer or the authentication service. PR #281 was marked ready for review, approved by CodeRabbit with no actionable findings, and merged as 12c9663 on 2026-09-24 after all checks passed, including the repository's end-to-end job and Ubuntu/Windows package tests. No Prisma bot review appeared; the user explicitly authorized proceeding with CodeRabbit's approval. Its advisory docstring-coverage warning was assessed without adding boilerplate to existing callbacks and test helpers. These CI results do not establish a fresh desktop signup or Windows/Linux plugin journey. No review requirement was bypassed. The coordinated release remains a separate human gate; the plugin must still wait for an exact released and verified CLI version.

The version-only release PR #282 prepares 8.0.0-rc.16 from merged main and is deliberately left unmerged with auto-merge disabled. Its checks passed for build, types, lint, 80 versioning/script tests, skill packaging, CLI tests (1,019 passed; 2 skipped), wrapper tests (3 passed), and clean tarball installs. Release conformance reports zero failures and six allowed findings under main's existing engine 0.4.0/0.6.0 transition exceptions. The PR also records a local macOS engine prompt-test timeout that reproduces before the bump at b50b587; this requires coordinated review rather than being reported as a fully green release. No published CLI pin or installed plugin was changed during this release preparation.

Directly executed checks on the available macOS desktop, Node 24.16.0 and pnpm 11.6.0 (not Windows/Linux acceptance):

Check	Observed result
CLI static checks	`pnpm typecheck` and `pnpm lint` passed
CLI package regression suite	1,019 passed, 2 skipped
Focused auth suite	85 passed, including context propagation, invalid/missing values before login, unchanged credential/session shape, success/failure wording, OAuth denial, workspace escaping, and abort handling
Default browser page	Compared rendered output with main for known, missing, and escaped workspace names: byte-for-byte identical
Browser-opening failure	A simulated opener failure in a persistent TTY attempt completed through the same listener; the exact verification URL was emitted once and token exchange ran once, without a pasted callback
End-to-end runner	After building the `prisma` wrapper: 7 local checks passed, 48 cloud lifecycle tests skipped without isolated test credentials; this is not a real OAuth or cloud acceptance result
Built CLI help	Unified `prisma auth login --help` exposes `--ui-context` with `prisma-plugin` as its supported value
Skill/package checks	Authored skill format, rebuild, offline integrity check, and existing packaging regression test passed; both skills and the single supporting reference are in the six-file bundle
Preservation	Imported Composer `0.21.0` SHA-256 remains `67b50e78fbb6cafd00bb99b0e56fe8a49e4a7190219474bf1b9be933f08bbcbb`; branding and starter prompts unchanged
Installed copy	Both v3 skills and all installed files still match their provenance; deliberately not refreshed with the release-gated v4 draft

The first auth test run could not bind local callback servers in the sandbox; rerunning with local-server permission passed. The first end-to-end runner attempt required the wrapper build; after building it, local checks passed. Neither is recorded as an authentication/product defect. No personal Prisma session was changed and no live deployment or cloud resource creation was performed.

Still required before activating this revision:

Release the merged CLI change through the coordinated maintainer process. Select an exact published version containing the flag, clean-install it alongside the unchanged Composer/cloud 0.21.0 and ORM peer 8.0.0-rc.11, and verify it. Then replace the reference's old CLI pin/install command and remove its release gate; do not use a guessed version, local package patch, or floating dependency.
With isolated credential storage and a human completing signup/consent, test an ordinary prompt from a fresh desktop account and clean local setup. Verify pending signup never triggers provisioning, cancelled/expired login can retry without duplicate processes, a returning session is reused, explicit workspace choices are respected, and an empty authorized workspace succeeds. The unit simulations above do not replace this acceptance run.
Verify actual local restart persistence and live service version, reachable URL, database-backed actions, and browser behavior with that released CLI.
Record the exact version/results here, clear the README's pending notice, rebuild with a fresh dev.4 cache suffix, refresh the installed copy, and verify both installed skills and their references. A fresh task must pick up the copy.

Listing and branding follow-up (2026-09-24)

Applied the approved platform description, 26-character listing subtitle, and new/existing-app starter prompts. Display name and publisher remain Prisma. Both listing accents use the website's coral #F34A60; contrast checks passed against white and #212121. Preserved the supplied logo in the source asset and centred its unchanged artwork on a 264 × 264 canvas.

The rebuilt bundle passed integrity checks and the existing packaging regression test. Removed a Finder .DS_Store file from the bundle after integrity validation identified it. All skill files remained byte-for-byte unchanged, including the Composer 0.21.0 reference. Codex installation succeeded; all six installed bundle files and provenance matched the source. No cloud deployment was run.

Generic-prompt follow-up (2026-09-23)

This revision introduced “Build a simple Todo app and deploy it” as the starter prompt and README example; the listing prompts were broadened on 2026-09-24. The authored workflow's discovery metadata and introduction explicitly interpret Prisma selection as the Composer/Compute default, retaining local and live checks. An independent offline routing evaluation selected this workflow for the generic prompt with Prisma selected, preserved local-only scope and existing Next.js/pnpm conventions, and did not route explicit Vercel or unrelated SQL requests to it. An unselected installed plugin was not treated as the user's provider choice. Skill-format and bundle-integrity checks passed; installed files matched the bundle and the upstream Composer reference remained unchanged. This evaluated routing decisions, not another live deployment or the host's plugin-selection UI.

Previous revision checks (`0.4.0-dev.3`)

The skill-format validator, bundle rebuild and integrity check, and existing packaging regression test passed. All six installed bundle files and provenance matched the checkout; fresh native Codex discovery found both enabled skills at 0.4.0-dev.3. The Composer reference matched the original 0.21.0 archive byte-for-byte. Packaging code, its tests, marketplace configuration, README, and ignore rules were unchanged by this revision.

An independent agent evaluated 13 supplied scenarios (including failure variants) using only the authored workflow and its reference. It executed no Prisma commands, network calls, or mutations. The resulting decisions covered:

Supplied evidence	Observed decision
Explicit target; a different workspace is active	Select the chosen session, then verify the effective workspace; no repeated target question
Authoritative single/multiple memberships	Select the sole workspace or ask among multiple; combine missing region selection
One stored session with unknown membership; membership lookup fails	Ask once, offer known sessions as suggestions, and preserve the distinction between a discovery failure and an empty result
New project; existing region; conflicting requested region	Present all supplied supported regions for the new project, preserve the existing region, or clarify the conflict before provisioning
Pending target/login; local-only request	Continue local work; neither case permits cloud provisioning
Changed command runtime; DNS-only failure	Restore the verified executable pair in the actual context; handle network access separately from dependency resolution
Exact gzip/JSON failure; generic container, quota, authentication, or startup failures	Apply Bun only to the matching signature, preserve state/reports, and diagnose the other failures independently
Bun recovery succeeds with an existing build-and-deploy script	Retain the working invocation while preserving the npm build step, target configuration, and external credentials
Running verified app without Console history	Report live success and unavailable history separately; create no Git commit

Review clarified selection-before-access-validation ordering, existing-region precedence, and a distinct retry-report filename. A focused follow-up found those ambiguities resolved. These are offline behavioral evaluations, not new live deployments or tests of actual account-wide workspace discovery.

Baseline checks (`0.4.0-dev.2`)

Check	Evidence
Clean installation	In an empty temporary project, npm installed Composer/cloud `0.21.0`, ORM Postgres `8.0.0-rc.11`, and `prisma` `8.0.0-rc.15` without `--force` or `--legacy-peer-deps`; Node `24.16.0`, npm `11.13.0`
Required peer	`npm ls --depth=0` showed all four exact versions; imports of `@prisma/composer`, `@prisma/composer-prisma-cloud/control`, and `@prisma/orm-postgres/control` succeeded
CLI surface	Installed dev, deploy, project-list, and service-show help loaded; deploy supports `--stage` and `--report`
Existing authentication	`auth whoami --json` identified a stored session; `project list --json` successfully verified remote access without new credentials or browser login
Packaging	`node --test scripts/package-plugin.test.mjs` passed: repeat builds preserve authored files, additional authored references, and identical provenance; changed/missing files and unexpected skills fail verification
Skill format	The skill-creator validator accepted the authored workflow
Installed plugin	Native Codex plugin inspection and fresh skill discovery found exactly two enabled Prisma skills at `0.4.0-dev.2`, no MCP servers; installed content matched the source bundle
Upstream preservation	Installed Composer SKILL.md matched the original pinned npm archive byte-for-byte

The clean install's npm audit reported 14 upstream dependency advisories (10 moderate, 4 high). This check did not assess exploitability or change the pinned dependency tree. Installation success is not a security or production-readiness assessment.

Baseline offline workflow scenarios (`0.4.0-dev.2`)

An independent fresh agent read the installed skills and reference, then evaluated these supplied scenarios without executing Prisma commands or making cloud calls. All five produced the intended actions:

Scenario	Observed behavior
Existing Next.js/pnpm app with newer Composer and ORM migrations	Preserves framework, manager, versions, and schema strategy; consults matching installed documentation
Stored session and remote access work; token variables unset	Reuses the session without requesting service tokens or login
Local-only request with no session	Continues local verification without cloud authentication or provisioning
Quota refusal after project/database/service creation; no live version	Reports partial provisioning and known IDs, preserves state, stops retries, avoids inventing quota limits
Failed update while an earlier version still serves	Reports the failed update and still-live previous version separately; does not claim outage or successful rollout

These are behavioral checks against supplied evidence, not live failure injection.

Application acceptance baseline (`0.4.0-dev.2`)

Use a fresh task and this ordinary prompt with the installed Prisma plugin:

Build a simple Todo app with Prisma Composer and Postgres persistence. Run it locally, then deploy it to Prisma Compute and verify it works.

Pass criteria:

Locally: typecheck/build pass; add/list/complete/delete work in the API and UI; data survives a service restart without resetting the database.
Remotely: the resolved project/stage has a live version and reachable URL; a database-backed user action and browser interaction succeed.
If deployment fails, record the actual state and blocker; do not mark the live test passed because local tests or resource creation succeeded.

Local result: passed

An independent fresh agent received only the installed plugin and the local portion of the ordinary request above. It built a Todo app in an empty temporary directory using Composer, a Node HTTP service, and raw Postgres with pg. It did not read the plugin source repository or parent conversation. It consulted public Composer documentation and installed package types where needed. No cloud credentials or resources were needed for local development.

Typecheck and build passed. API create/list/complete/delete passed, and a blank title was rejected with HTTP 400.
Chrome UI create/complete/delete passed. After an actual service-process restart, the same row ID, title, and completed state remained in Postgres and appeared on browser reload.
The local service was stopped after verification; its database and shared emulators were preserved. The in-app browser refused localhost, so browser verification used Chrome.

The existing-app and absent-session cases above remain offline scenario checks; this test does not claim a separate existing-app migration or real login exercise.

Cloud result: passed

The same app was deployed in a user-selected empty workspace, region us-east-1, stage demo, reusing the existing authenticated session. The first Node-based attempt failed before creating a project; the confirmed runtime recovery is recorded below. Retrying the same target with Bun succeeded.

The structured deployment report recorded success with database and Compute service IDs. Service inspection independently showed a running, live version and a public URL.
The live HTML returned HTTP 200. API create/list/complete/delete passed, with updated state retained by a subsequent read and invalid input rejected.
Chrome UI create/complete/delete passed, with completion retained on page reload. Only disposable smoke-test rows were removed.

The smoke project and database remain available for manual review. The app is an anonymous shared Todo demo without accounts or per-user ownership. This is a release smoke test, never part of packaging or an automatic cloud deployment on every change. Persistence across service restart was verified locally; no cloud restart was performed. Quota failure recovery was evaluated offline, not induced against the workspace.

Second run: agent-reported evidence

A user-supplied report, reviewed on 2026-09-23, describes another successful Todo deployment in the selected workspace, Frankfurt (eu-central-1), stage demo. It reports local CRUD and persistence across an actual service restart, a running live version, deployed database/API/browser checks, and cleanup of its test rows. It also reports successful recovery from the same npm resolver and gzip/JSON failures, plus a live app without a Console history entry because Git metadata was absent. These outcomes were not rerun or independently inspected for dev.3.

The report's cookie-ownership and cross-origin checks apply to that app's design; they do not impose an authentication model on plugin-generated applications. No starter or generic smoke-test helper was added based on this report.

Upstream findings kept outside the workflow

Composer 0.21.0's bundled reference has stale CLI packaging language. The core package's manifest has no binary and points to @prisma/composer-cli, while the skill says the core carries the CLI. The plugin toolchain reference corrects the installation path without editing the imported skill.
Authentication guidance needs CLI scope. Unified prisma@8.0.0-rc.15 successfully reused a stored session. The standalone/control-API token path remains separate. A direct invocation of only the upstream concepts skill can still miss this distinction; use the build-and-deploy workflow for the journey.
The missing ORM peer was not reproduced with normal installation. Cloud 0.21.0 declares @prisma/orm-postgres@8.0.0-rc.11 as a required peer. The clean install and cloud-control import passed with it present. The earlier failure after bypassing peer resolution does not establish an undeclared-dependency bug.
The npm resolution failure reproduced in the fresh Todo test. Node 23.11.0/npm 10.9.2 failed with Cannot read properties of null (reading 'edgesOut'). A process-scoped Node 24.16.0/npm 11.13.0 invocation completed installation without bypass flags. The toolchain reference records this conditional recovery; it does not prescribe a global runtime change.
The compressed-response failure reproduced during the cloud smoke test. With the pinned package set and Node 24.16.0, prisma deploy module.ts --stage demo --report ... failed with DEPLOY.CONTAINER_FAILED, a Management API container-resolution JSON parse error, and leading gzip bytes. The report had no resource nodes; a remote project listing confirmed the workspace was still empty. The failed report and target were preserved. Project-local Bun 1.4.2 with bun run --bun prisma deploy ... successfully deployed the same target. Both parent and child runtime probes reported Bun 1.4.2. This is a verified conditional recovery, not an established root cause or a default runtime change.
Local CLI exit did not stop the service in the tested invocation. With unified CLI 8.0.0-rc.15, Node 24.16.0, and a Bun 1.1.18 service, Ctrl-C exited npm exec -- prisma dev module.ts while its listener still served. Inspection verified that the process belonged to this app. A supervised process restart proved persistence; the installed local target's app-scoped stop operation then stopped it without deleting data or shared emulators. The workflow requires observing the actual service restart. It does not prescribe the internal emulator endpoint as a general public command.

No upstream repository or platform changes are part of this revision.

FilesExpand file tree

validation.md

Latest commit

History