Files
pikasTech-HWLAB/docs/dev-gate-preflight.md
T
Lyon 7b928bd863 fix: classify DEV DB and provider blockers
fix: classify DEV DB and provider blockers
2026-05-22 23:38:11 +08:00

148 lines
6.9 KiB
Markdown

# HWLAB DEV Gate Preflight
This preflight is the read-only gate before a real DEV deploy or DEV smoke on
D601. It decides whether the current `origin/main` can enter the real
`hwlab-dev` runtime path, or whether it is blocked by missing contract,
artifact, cluster, registry, edge, or safety evidence.
It does not authorize PROD, restart runtime services, read secrets or tokens,
substitute UniDesk services for HWLAB runtime, or run heavyweight/browser e2e.
## Command
Run from the repository root:
```sh
node --check scripts/dev-gate-preflight.mjs
node --check scripts/src/dev-gate-preflight.mjs
node --check scripts/src/registry-capabilities.mjs
node --check scripts/deploy-desired-state-plan.mjs
node --check scripts/src/deploy-desired-state-plan.mjs
node --check scripts/refresh-artifact-catalog.mjs
node scripts/deploy-desired-state-plan.mjs --check
node scripts/dev-gate-preflight.mjs
```
The command writes `reports/dev-gate/dev-preflight-report.json` and prints a
short JSON summary. A `blocked` conclusion means the preflight ran correctly but
found blockers; the command exits zero by default so the report can be committed
as evidence. Use `--fail-on-blocked` only in CI jobs that should fail on an open
gate.
Useful options:
```sh
node scripts/dev-gate-preflight.mjs --target-ref origin/main
node scripts/dev-gate-preflight.mjs --report reports/dev-gate/dev-preflight-report.json
node scripts/dev-gate-preflight.mjs --timeout-ms 5000
node scripts/dev-gate-preflight.mjs --no-write
```
## Read-Only Scope
The preflight checks:
- `deploy/deploy.json` and `deploy/artifact-catalog.dev.json` are internally
consistent and DEV-only.
- `scripts/deploy-desired-state-plan.mjs --check` can be run as a read-only
source support check for deploy commit/image/tag/workload/env mirror
convergence. It does not contact a registry or prove DEV runtime state.
- The deploy manifest, artifact catalog, catalog service commits, and image tags
are covered by the selected `origin/main` target. If `origin/main` is a
report-only evidence commit whose changes do not touch artifact build inputs,
the preflight may accept the artifact source recorded in
`reports/dev-gate/dev-artifacts.json`.
- The artifact catalog clearly separates source commit, artifact commit,
`ciPublished`, `registryVerified`, and per-service digest state.
- The artifact catalog has real publish and registry digest evidence, or stays
blocked with `not_published` digests.
- `reports/dev-gate/dev-artifacts.json`, when present, proves all frozen DEV
service artifacts were published for the selected target or for an artifact
source commit that still covers the target.
- Registry capability evidence is split into `process-http-access`,
`docker-daemon-push-access`, and `k3s-pull-access` instead of one ambiguous
"registry reachable" result.
- `deploy/k8s/base` and `deploy/k8s/dev` parse and remain scoped to
`hwlab-dev`.
- `hwlab-cloud-api` declares the DEV DB env contract:
`HWLAB_CLOUD_DB_URL` from Secret reference
`hwlab-cloud-api-dev-db/database-url` and
`HWLAB_CLOUD_DB_SSL_MODE=require`. The non-secret DB DNS contract is
`cloud-api-db.hwlab-dev.svc.cluster.local:5432` via selectorless
`Service/cloud-api-db` in `hwlab-dev`.
- The local cloud-api health payload reports DB env presence/missing status with
redacted values only. This is a readiness gate, not live DB evidence.
- `deploy/frp` and `deploy/master-edge` describe the D601-to-master DEV route
on public frontend `:16666` and API/edge/live `:16667`. Legacy public
`:6666/:6667` observations are historical/deprecated only and must not be
used as current green evidence or current active blockers.
- The runner can perform read-only `kubectl` probes against `hwlab-dev`.
- `reports/dev-gate/dev-edge-health.json`, when present, records read-only
public edge, frps, tunnel-health, and k3s observability evidence.
- `http://74.48.78.17:16667/health/live` responds.
- Catalog image manifests are visible without reading credentials. A failed
manifest read is tracked as a #66 registry reachability dimension and must not
be conflated with missing publish digests.
- Deploy images do not point at UniDesk, provider-gateway, or microservice-proxy
substitutes.
## Registry Capability Dimensions
`registryCapabilities` appears in both the preflight report and the artifact
publish report. The dimensions are:
- `process-http-access`: runner-process HTTP access to `/v2/`. A failed
`127.0.0.1` request can happen while Docker daemon push access still works,
so this dimension is diagnostic and may be `degraded` without blocking
publish.
- `docker-daemon-push-access`: read-only Docker daemon evidence for the
local/internal registry target. This is the publish-path capability.
- `k3s-pull-access`: read-only `kubectl` evidence for whether the DEV cluster
can inspect image pull state. This is the deploy-path capability.
## Verdict Rules
`ready` requires all checks to pass and no open blocker.
`degraded` means a non-publish-path capability is observable but incomplete,
such as runner-process loopback HTTP refusing `127.0.0.1:5000/v2/` while Docker
daemon registry visibility is still present. `blocked` is expected until the
real DEV runtime path can be proven. Each blocker includes a `type`, `scope`,
`summary`, and `nextTask` so the next task is the smallest repair needed before
rerunning the preflight.
The DB gate has two common blocked scopes:
- `cloud-api-db-env-contract`: manifest/Kubernetes placeholders are incomplete.
- `cloud-api-db-health-gate`: runtime env presence is missing. The next task is
to configure the DEV Secret/env names, not to print or commit any secret
value.
For catalog commit mismatch, the preflight emits an
`artifactIdentity.refreshCommands.blocked` command. Use it when the source
commit changed but publish is still blocked:
```sh
node scripts/deploy-desired-state-plan.mjs --target-ref origin/main --pretty
node scripts/refresh-artifact-catalog.mjs --target-ref origin/main --blocked
```
Only use the published refresh path after `reports/dev-gate/dev-artifacts.json`
proves every frozen service has a registry digest for that same source commit:
```sh
node scripts/deploy-desired-state-plan.mjs --target-ref origin/main --pretty
node scripts/refresh-artifact-catalog.mjs --target-ref origin/main --publish-report reports/dev-gate/dev-artifacts.json
```
Committed DEV artifact reports can prove required service digests only for the
source commit they name. If the selected `origin/main` target drifts onto files
that change artifact build inputs, the preflight can still reopen
`artifact-source-commit` or `dev-artifact-publish` even though the base-image
blocker stays closed. Remaining registry manifest-read failures are #66
reachability evidence, not a reason to reopen the base-image blocker.
This is M3 virtual hardware trusted-loop support only. It does not claim M3
DEV-LIVE unless a real `DO1 -> patch-panel -> DI1` chain is observed through
the current DEV endpoint.