feat: bump workflow SDK to v0.15.2, release v0.4.0#2
Merged
Conversation
The plugin was built against workflow v0.9.4 SDK. When invoked under wfctl v0.15.2 (host), the go-plugin handshake failed: the plugin subprocess started but exited without completing the protocol negotiation, leaving the host with "plugin failed to exit gracefully". Bump go.mod to workflow v0.15.2 so the plugin's external/sdk and external/handshake imports match the host's handshake config exactly. Version bumped to 0.4.0 (minor — SDK jumped from v0.9 to v0.15, though the Serve()/ModuleProvider/ServiceInvoker API surface we consume is unchanged). All tests pass (internal + drivers). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Same runner-group issue that blocked the release workflow — self-hosted runners don't pick up jobs for this repo. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
intel352
added a commit
that referenced
this pull request
May 3, 2026
…Copilot finding #2) intFromConfig truncates float64 silently — size_gb: 100.9 (e.g. via JSON or YAML float coercion) was creating a 100 GB volume with no diagnostic. Add intStrictFromConfig that mirrors the float64 fractional-rejection logic already in dropletSSHKeysFromConfig and use it for size_gb in Create, Update, and Diff. Whole-valued float64 (the structpb wire shape) still passes through. intFromConfig is unchanged for callers where rounding is acceptable (node_count, instance_count, probe seconds, etc.). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
intel352
added a commit
that referenced
this pull request
May 3, 2026
…opilot round-2 finding #2) Previously `intStrictFromConfig` returned (defaultVal, false, nil) when the config value was an unsupported type (e.g. `size_gb: "100"` as a string). The caller then emitted "size_gb is required" — a misleading diagnostic that hid the real bug (operator quoted a number). Return present=true with an explicit "expected integer, got <type>" error so the caller surfaces the type problem directly. Tests: string-typed size_gb and bool-typed size_gb both reject with "expected integer, got <type>" before any CreateVolume call is made. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
intel352
added a commit
that referenced
this pull request
May 3, 2026
…+ new infra.volume driver (#55) * feat(drivers): add boolFromConfig and strSliceFromConfig helpers Both helpers are structpb-aware: boolFromConfig falls through to the default for non-bool values, and strSliceFromConfig accepts the typed []string shape Go-native callers emit AND the []any shape that survives a YAML/JSON → structpb round-trip. Empty strings drop silently. Used by the upcoming droplet/volume drivers; placed in util.go alongside the existing strFromConfig/intFromConfig helpers so all drivers can reach them without import cycles. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(droplet): support user_data / vpc_uuid / ssh_keys / volumes / tags / bools Self-hosted services (e.g. Apache-AGE Postgres on a Droplet, since DO Managed Postgres has no AGE extension) need more than the size/image/ region the DropletDriver previously honoured. New optional config keys, all additive and defaulted-empty (no behaviour change for existing callers): - user_data string — cloud-init payload - vpc_uuid string — VPC the Droplet joins - ssh_keys []string|[]int — fingerprints OR numeric IDs; element type detected at runtime (mixed lists OK; structpb floats must be whole numbers) - tags []string - enable_backups bool — godo Backups field - monitoring bool - ipv6 bool - volumes []string of — names resolved to IDs at create volume names time via Storage.ListVolumes(name=, region=). Region-bound: a name not present in the droplet's region returns "droplet volumes: volume %q not found". Outputs gain `private_ip` (godo droplet.PrivateIPv4()) so downstream services in the same VPC can be wired without a second Read. DropletDriver now holds an optional Storage client (set by NewDropletDriver from c.Storage; tests inject via the variadic NewDropletDriverWithClient param). When `volumes` is non-empty but the storage client is nil, Create returns an explicit error rather than silently dropping the attachment. mockDropletClient now captures the godo.DropletCreateRequest so tests can assert on populated struct fields. Adds focused tests for each new key, plus the SSH-keys fractional-float rejection path that prevents silent ID truncation. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(drivers): add infra.volume driver (DO Block Storage) New ResourceDriver covering CRUD + Diff + HealthCheck for DigitalOcean Block Storage volumes. Backed by godo.StorageService; in-place size growth uses godo.StorageActionsService.Resize. ProviderIDFormat is UUID (DO volume IDs are 36-character UUIDs). Config keys: - name from spec.Name - region defaults to provider region - size_gb required, > 0 - filesystem_type optional ("ext4" / "xfs"); empty = raw block device - description optional - tags optional []string Outputs: id (UUID), name, region, size_gb, filesystem_type. Status is a stable "available" string — godo.Volume exposes no Status field, so a successful Read on a real volume is the strongest health signal we have. Update semantics: - size growth → in-place StorageActions.Resize, then re-read for state - size shrink → explicit error (DO has no shrink API; replace required) - region or filesystem_type change → Diff sets ForceNew so plan classifies as replace, not update Tests cover Create / Read / Update (resize, no-op, shrink-rejected, empty-ProviderID guard) / Delete / Diff (nil current, grow=update-only, shrink=replace, region=replace, filesystem_type=replace, no-changes) / HealthCheck (success, read-error, empty-ProviderID) / ProviderIDFormat / Scale-not-supported. The volume mock pool supports filter-by-name so the same struct exercises both VolumeDriver tests and the droplet volumes-by-name path. providerid_format_test.go adds the volume entry to the manually maintained registry — without it, a future drift in ProviderIDFormat would not be caught. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(provider): register infra.volume driver and capability Wire VolumeDriver into Initialize() alongside the other infra drivers and declare its IaCCapabilityDeclaration with the noScale operation set (volumes have no scale concept; size_gb is dimensional). Add "infra.volume" to the sizing map as a noopSizing entry so ResolveSizing returns "n/a" for every abstract tier — the operator must set size_gb explicitly. Provider Capabilities test gains the matching required entry so removal of the driver registration would be caught. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs: declare infra.volume in plugin.json and CHANGELOG plugin.json: add infra.volume to the iacProvider.resourceTypes array and update the descriptor to mention Block Storage volumes. CHANGELOG: document the new infra.volume driver and the extended Droplet config keys (user_data / vpc_uuid / ssh_keys / volumes / tags / enable_backups / monitoring / ipv6) plus the new private_ip Droplet output. No README exists in this repo, so the CHANGELOG is the canonical surface for these notes. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(droplet): strict validation for volumes config entries (Copilot finding #1) strSliceFromConfig silently drops non-string and empty entries, which is dangerous for volume attachments — a typo or wrong type leaves the Droplet running without an expected disk and emits no diagnostic. Add dropletVolumesFromConfig that returns an explicit error for any non-string or empty entry, mirroring dropletSSHKeysFromConfig's error-message style. Tests cover both shapes: []any{123, "data"} (type mismatch) and []any{""} (empty string). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(volume): reject fractional size_gb instead of silent truncation (Copilot finding #2) intFromConfig truncates float64 silently — size_gb: 100.9 (e.g. via JSON or YAML float coercion) was creating a 100 GB volume with no diagnostic. Add intStrictFromConfig that mirrors the float64 fractional-rejection logic already in dropletSSHKeysFromConfig and use it for size_gb in Create, Update, and Diff. Whole-valued float64 (the structpb wire shape) still passes through. intFromConfig is unchanged for callers where rounding is acceptable (node_count, instance_count, probe seconds, etc.). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(droplet): accept []int and []int64 ssh_keys at top level (Copilot finding #3) dropletSSHKeysFromConfig accepted []any and []string at the top level but bailed on Go-native []int{101} or []int64{101} with "expected list, got []int" — even though numeric IDs are documented as supported and the per-element handler already covers int / int64 / float64. Add type-switches that mirror the per-element validation (non-positive rejection, error message style). Tests cover three shapes: []int{101,102}, []int64{555,666}, and the non-positive rejection path []int{0,101}. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(sizing): cover infra.volume in resolveSizing test cases (Copilot finding #4) infra.volume was added to doSizingMap as noopSizing but the manually- maintained TestResolveSizing cases table was not updated. Without the case, the registration is unverified and a future refactor that drops the entry would not fail this test. Add SizeS + SizeXL coverage so the "n/a" contract is locked in alongside every other noop-sized resource. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(volume): treat description/tags changes as ForceNew (Copilot finding #5) Diff previously tracked only size_gb / region / filesystem_type. Changes to description and tags were silently ignored because godo.StorageActions has no update endpoint for either, and the godo Volume API itself sets both at creation time only — so there is no in-place update path to make available. Resolution: surface drift as ForceNew (matching the way region / filesystem shrinks are handled) so the planner emits a planned replace rather than dropping the change. Document the constraint in the diff comment. Surface description and tags in volumeOutput so Diff has values to compare. Reuse outputsAsStringSlice + equalStringSet (firewall.go) so reordered tags do NOT trigger a needless replace. Tests: tags-add forces replace; reordered tags do NOT; description change forces replace. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(droplet): Diff flags vpc/tags/volumes/backups drift as ForceNew (Copilot finding #6) Extended droplet fields (user_data, vpc_uuid, ssh_keys, volumes, tags, enable_backups, monitoring, ipv6) were wired into Create only; Diff compared just "size", so changing any of these on an existing droplet produced no plan action and the drift was silently dropped. godo explicitly disallows Update on droplets (PUT only resizes), so Diff is the only place to flag these — every detected change must be ForceNew. Surface the comparable fields in dropletOutput (vpc_uuid, enable_backups derived from BackupIDs, tags, volumes) so Diff has values to compare; extend Diff with order-irrelevant set comparison for tags / volumes (reuses outputsAsStringSlice + equalStringSet from firewall.go). user_data, monitoring, ipv6, ssh_keys are NOT drift-checked: godo.Droplet does not surface them on Read, so comparing against current Outputs would produce a perpetually-dirty plan. Documented as a known limitation. Tests: vpc_uuid change forces replace; tags add forces replace; tags reorder does NOT force replace; enable_backups toggle forces replace. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(plugin): include 'Block Storage volumes' in Manifest.Description Manifest.Description must match plugin.json's description (test enforced in plugin_test.go). The merge from main pulled in a longer description that didn't yet mention Block Storage volumes; this aligns both. * fix(ci): wfctl-strict-contracts must copy plugin.contracts.json alongside plugin.json The Validate strict plugin contracts step (added in PR #41) copies only plugin.json to a tempfile when simulating GoReleaser's before-hook version rewrite. But wfctl plugin validate --strict-contracts looks for plugin.contracts.json CO-LOCATED with plugin.json (same directory). The rewrite-validation therefore always fails with "missing_module_contract_descriptor: module type \"iac.provider\" has no strict contract descriptor" — even when plugin.contracts.json IS valid in the repo root. This was masked on main because main has a single configurable contract file and any change rebuilds in-place; surfaced now by PR #55 which introduces a new resource type and exercises the second validation path. Fix: use a temp DIR and copy both plugin.json + plugin.contracts.json into it so the rewritten plugin.json's neighbour is intact. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(droplet): resolve VolumeIDs to names in Outputs (Copilot round-2 finding #1) dropletOutput previously stored godo's raw VolumeIDs (UUIDs) in Outputs["volumes"], while desired config carries volume *names*. After every successful Create/Read, the next Diff would compare e.g. ["vol-uuid-1"] against ["pg-data"] and force-replace the Droplet — catastrophic for stateful workloads (the PG data Droplet would be destroyed and recreated on every deploy, losing all game state). Resolve each VolumeID to its name via Storage.GetVolume in dropletOutput. Cache lookups within a single Read so duplicate VolumeIDs only hit the API once. On resolution failure (e.g. volume deleted out-of-band) fall back to the raw ID and record the unresolved IDs in Outputs["volumes_resolution"] so operators can debug. Update Diff comment to reflect that comparison is now name-vs-name and stable across plans. Tests: - Read with VolumeIDs returns names in Outputs - Diff with matching names returns no change (regression guard) - Diff with desired missing a name returns ForceNew - Resolution failure falls back to ID + records unresolved_ids - Duplicate IDs only call GetVolume once (cache hit) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(util): intStrictFromConfig surfaces type mismatches explicitly (Copilot round-2 finding #2) Previously `intStrictFromConfig` returned (defaultVal, false, nil) when the config value was an unsupported type (e.g. `size_gb: "100"` as a string). The caller then emitted "size_gb is required" — a misleading diagnostic that hid the real bug (operator quoted a number). Return present=true with an explicit "expected integer, got <type>" error so the caller surfaces the type problem directly. Tests: string-typed size_gb and bool-typed size_gb both reject with "expected integer, got <type>" before any CreateVolume call is made. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(droplet): tags clear (non-empty -> empty) surfaces as drift (Copilot round-2 finding #3) The Diff path skipped tag comparison when len(desired.tags)==0, so clearing tags on a Droplet was silently ignored. Operators sometimes need to strip tags to remove a Droplet from a tag-based firewall or backup schedule; "no diff" is dangerously wrong here. Switch the guard to "key present in desired" so empty desired vs non-empty current surfaces as ForceNew, while absent desired still skips (preserves backwards-compat for YAML that predates the field). Tests: clearing tags forces replace; absent tags key does NOT trigger drift; reorder still ignored. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(volume): filesystem_type transitions raw<->ext4 surface as drift (Copilot round-2 finding #4) The empty-side guard skipped raw(empty)->ext4 and ext4->raw transitions, so changing filesystem_type after Create produced no diff. DO Block Storage cannot reformat a volume in place, so any change must surface as ForceNew rather than being silently ignored. Switch to "key present in desired" guard so empty<->non-empty surfaces as drift. Absent desired key still skips (backwards-compat for YAML predating the field). Tests: raw->ext4 and ext4->raw both force replace; absent key does NOT trigger drift. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(volume): description add/clear from empty surfaces as drift (Copilot round-2 finding #5) The empty-side guard skipped both empty->non-empty (adding a description after Create) and non-empty->empty (clearing a description) transitions. DO Block Storage exposes no description-update endpoint, so any change must surface as ForceNew rather than being silently ignored. Switch to "key present in desired" guard. Absent key still skips (backwards-compat). Tests: add-from-empty and clear-to-empty both force replace; absent key does NOT trigger drift. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(volume): tags clear surfaces as drift (Copilot round-2 finding #6) Same pattern as droplet finding #3 and volume description finding #5. Clearing tags (non-empty current -> empty desired) was silently ignored because Diff skipped when len(desired.tags)==0. DO Block Storage has no tag-update endpoint, so any tag change must surface as ForceNew. Switch to "key present in desired" guard. Absent key still skips. Tests: clearing tags forces replace; absent tags key does NOT trigger drift (existing reorder/equal-set behavior preserved). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(droplet): vpc_uuid add-from-empty surfaces as drift (Copilot round-2 finding #7) The curVPC != "" guard skipped vpc_uuid drift when the current state had no vpc_uuid. Pre-release Droplet states (created before vpc_uuid became part of Outputs) would silently ignore an operator adding a vpc_uuid pin to YAML — exactly when explicit drift detection matters most. Switch to "key present in desired" guard. Empty current vs non-empty desired now triggers ForceNew. Absent desired key still skips (backwards-compat). Tests: vpc_uuid add-from-empty forces replace; absent key does NOT trigger drift. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(droplet): explicit limitation block for user_data/ssh_keys/monitoring/ipv6 (Copilot round-2 finding #8) Round-1 documented the gap inline but the silent fallthrough at the end of Diff was easy to miss. Replace with an explicit block that: - Names each undetected field and the godo reason - States the operator workaround (`taint` / delete + re-apply) - References follow-up issue #56 for the resolution path - Updates the function-level Diff doc-comment to point at the inline rationale Issue: #56 "Droplet Diff misses user_data / ssh_keys / monitoring / ipv6 (godo Read limitation)" Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
intel352
added a commit
that referenced
this pull request
May 5, 2026
5/5 findings addressed: #1 (validate_plan.go:52) — byName index must exclude delete-action resources. Previously a vpc_ref pointing to a VPC scheduled for deletion in the same plan would silently 'resolve' as if live; now delete-targets are excluded from the index so cross-resource references to them surface as Severity=Error dangling references. New regression test: TestDOProvider_ValidatePlan_DatabaseVPCRefToDeleteTargetIsDangling. #2 (validate_plan.go:91) — appendDatabaseDiagnostics docstring incorrectly said 'Warning when missing'; the implementation has always emitted PlanDiagnosticError, and the conformance scenario + TDD tests both REQUIRE Error severity. Doc rewritten to match the implemented contract. #3 (validate_plan.go:225) — zonesInGroup docstring promised sorted output but returned the raw underlying slice unsorted. Now copies + sorts (lexicographic) so diagnostic messages are deterministic and the returned slice is owned by the caller (safe to mutate). nil for unknown groups; the caller's strings.Join still works. #4 (provider_conformance_test.go:70) — comment described a 'stub-then-real swap' with two Initialize calls; the implementation has always made one call with the right token chosen up-front. Doc rewritten to match the actual single-call flow. #5 (codemod-report.yml:73) — fork PRs run with a read-only GITHUB_TOKEN per GitHub's pull_request workflow security model, so issues:create-comment + issues:update-comment fail 403 and would block CI. Gate the comment step on github.event.pull_request.head.repo.fork == false. The artifact upload step still runs unconditionally so the report remains reachable from the Actions tab. Verified locally: go test ./... -count=1 -race PASS, go test -tags=conformance ./internal/ -run TestConformance PASS, go vet clean, codemod-report.yml YAML valid. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
intel352
added a commit
that referenced
this pull request
May 5, 2026
…lyPlan) (#61) * chore(deps): bump workflow to e2c582b (W-7 conformance + W-8 codemod) Pseudo-version v0.20.6-0.20260505011403-e2c582bece90, the workflow main HEAD that includes: - W-7: iac/conformance scenario suite (12 scenarios) and DO smoke gate - W-8: cmd/iac-codemod 4-mode AST tool Required for TP1-TP5 of PR P-DO (IaC conformance plan §P-DO). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(plugin): codemod-report workflow uploads dry-run output as artifact + sticky PR comment summary PR P-DO TP1: per pull_request, runs the iac-codemod refactor-apply mode in -dry-run against the plugin source, uploads the full Markdown report as a 90-day retention GitHub Actions artifact, and posts/updates a sticky PR comment with the top-30 lines of the report so drive-by reviewers see the key findings without downloading the artifact. Supply-chain note: actions/github-script SHA-pinned per workflow security policy (Renovate tracks upstream releases via .github/renovate.json). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * refactor(provider): collapse Apply to wfctlhelpers.ApplyPlan PR P-DO TP2: replace the in-Apply per-action switch (create/update/replace/ delete + upsert recovery + nil-out diagnostic) with a single dispatch to wfctlhelpers.ApplyPlan. The helper handles: - ErrResourceAlreadyExists upsert recovery via interfaces.UpsertSupporter (DO drivers AppPlatform, VPC, Firewall, Database already implement SupportsUpsert() bool, so they satisfy the canonical interface without code change — the local upsertSupporter declaration is now removed). - JIT ${MODULE.id} / ${VAR} substitution (W-5). - Replace decomposition + ReplaceIDMap propagation (W-3a/W-3b). - Input-drift postcondition (W-3a). - Per-action context cancellation between iterations. The DO-plugin-specific deferred-update flush (DatabaseDriver type=app trusted_sources referencing apps created later in the plan; regression gated by provider_deferred_test.go and CHANGELOG entry for staging-deploy- blockers Blocker 2) is preserved by wrapping ApplyPlan with the second- pass loop that calls FlushDeferredUpdates on any deferredUpdater driver. The wrapper deviates from the codemod's canonical 'return wfctlhelpers.ApplyPlan(ctx, p, plan)' single-statement shape; the deviation is documented and marked with // wfctl:skip-iac-codemod so AssertApplyDelegatesToHelper recognizes the intentional shape. When wfctlhelpers grows a deferred-update lifecycle hook, the wrapper can collapse and the marker can drop. The provider_apply_test.go DeleteAction_MissingCurrent regression test was a v1 pre-flight defense (synthesize an error when action.Current is nil before calling Delete). Under v2 dispatch the contract is 'driver is the authority on what an empty ProviderID means' (per wfctlhelpers/ apply.go::doUpdate's analogous comment). The test was rewritten to lock the new contract: dispatch IS made with an empty-ProviderID ref; real drivers like FirewallDriver surface the diagnostic via their typed validation. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(provider): ValidatePlan for App Platform region-VPC constraints PR P-DO TP3: implement interfaces.ProviderValidator (W-4) on DOProvider to surface DO region constraints at `wfctl infra align` time before any cloud API call. Per the W-4 design, R-A10 invokes ValidatePlan via type-assertion; providers that do not implement it continue to work unchanged. The first pass covers three constraint families: 1. App Platform infra.container_service requires a region GROUP slug (nyc, ams, fra, sfo, sgp, syd, tor, blr, lon). Zone slugs (nyc1, sfo3 …) are rejected with PlanDiagnosticError. This is the "copy-pasted nyc3 from a Droplet config" defense. 2. Zone-bound resources (infra.vpc, infra.droplet, infra.volume) MUST use a zone slug. Bare group slugs are rejected with PlanDiagnosticError. Inverse of (1). 3. Cross-resource: an App Platform with vpc_ref pointing to a VPC in the same plan must have a region group whose zones include the referenced VPC's region. This locks the recurring 'App Platform in nyc cannot reach VPC in sfo3' production bug class (root-cause issue D from the conformance design). Cross-resource resolution looks at desired spec first, falls back to action.Current's Outputs["region"] for unchanged-VPC scenarios. vpc_ref pointing to a name not in the plan emits a Severity=Warning so non-strict align tolerates external VPCs while --strict escalates. ValidatePlan is read-only and makes no remote calls per the W-4 contract. Compile-time interface assertion lives at the bottom of validate_plan.go. 11 TDD tests in validate_plan_test.go cover: nil/empty plan, group-slug accepted, zone-slug rejected for AP, zone-slug accepted for VPC/Droplet/Volume, group-slug rejected for VPC, mismatch error (flagship), happy-path match, unknown-vpc_ref Warning, current-state fallback, delete-action skipped, compile-time assertion. Future extensions (deferred follow-ups): database/cache zone slugs, load balancer zone matching against attached droplets, registry regional restrictions. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(plugin): opt into computePlanVersion: v2 PR P-DO TP4: declare iacProvider.computePlanVersion=v2 at the TOP LEVEL of plugin.json so wfctl's runtime dispatcher (cmd/wfctl/deploy_providers.go ::iacPluginManifest) routes Apply through wfctlhelpers.ApplyPlan instead of the legacy in-provider switch. Schema note: the SDK manifest schema (plugin/sdk/manifest_schema.json) expects iacProvider.computePlanVersion at the top level of plugin.json. The existing capabilities.iacProvider sub-block (name, resourceTypes, configSchema) is a DIFFERENT consumer (plugin discovery + capability declaration); the two structures coexist in the same file. The runtime loader unmarshals both into one struct (one Capabilities.IaCProvider.Name field plus one top-level IaCProvider.ComputePlanVersion field) so a single plugin.json read serves both code paths. Validated via three checkers: - go run ./cmd/wfctl plugin validate --file plugin.json --strict-contracts (OK) - JSON-schema validation against plugin/sdk/manifest_schema.json (OK) - sdk.ParseManifest decode confirms EffectiveComputePlanVersion()==v2 Backward compat: wfctl < v0.21.0 ignores the new field; the legacy v1 dispatch (provider.Apply switch, now wrapping wfctlhelpers.ApplyPlan) continues to work for all existing callers. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(provider): add conformance test + extend ValidatePlan for database vpc_ref + bump to v0.10.0 PR P-DO TP5: ship the conformance test entry point, extend ValidatePlan to satisfy the conformance suite's cross-resource constraint scenario, and bump plugin.json + CHANGELOG to v0.10.0. provider_conformance_test.go (new, behind 'conformance' build tag): - Invokes iac/conformance.Run against a freshly-constructed and initialized DOProvider. Initialize is always called with a stub token so the driver registry is populated for the non-cloud scenarios that probe ResourceDriver lookups (structpb-roundtrip, cross-module resolution, etc.) — these exercise read-only or pure-data paths that don't hit DO's API. - LiveCloud (CONFORMANCE_LIVE_CLOUD=1 + DIGITALOCEAN_ACCESS_TOKEN) swaps the stub for the real token before driver instantiation. - SmokeOnly = testing.Short() limits to Smoke=true scenarios for the per-PR smoke gate's narrow contract. ValidatePlan extension (validate_plan.go): - Added appendDatabaseDiagnostics: infra.database vpc_ref pointing to a name not in the plan emits a Severity=Error diagnostic. Closes the conformance Scenario_CrossResourceConstraintRejection assertion that 'at least one Severity=Error diagnostic' fires for a dangling cross-resource reference. The assertion was failing before this change (database vpc_ref was previously unhandled by ValidatePlan). - Two new TDD tests: dangling-vpc_ref → Error (mirrors the conformance contract in-tree); in-plan vpc_ref → no diagnostic (happy path). plugin.json: - version 0.9.0 → 0.10.0 - download URL paths v0.9.0 → v0.10.0 (per-OS/arch) CHANGELOG.md: - New v0.10.0 section catalogues TP1-TP5 changes (ValidatePlan, computePlanVersion: v2 opt-in, Apply collapse, conformance test, codemod-report workflow), the Apply-delete v2-contract change, and the workflow dep bump. - Migrates the previously-Unreleased infra.vpc id-output fix into the v0.10.0 Fixed section (it ships in this release). Test results: - go test -tags=conformance ./internal/ -run TestConformance: 6/6 non-cloud scenarios PASS (CrossResourceConstraintRejection, DiffSurvivesGRPCRoundTrip, InfraOutputCrossModuleResolution, PlanStaleDiagnostic, ProtectedReplaceWithOverride, ProtectedReplaceWithoutOverride). Upsert-on-already-exists + grpc-roundtrip skip when their opt-in driver types are absent (DO does not expose infra.compute). - go test ./... -count=1 -race: ALL packages PASS. - go run ./cmd/wfctl plugin validate --strict-contracts: OK. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: Copilot review round 1 — 5 substantive findings 5/5 findings addressed: #1 (validate_plan.go:52) — byName index must exclude delete-action resources. Previously a vpc_ref pointing to a VPC scheduled for deletion in the same plan would silently 'resolve' as if live; now delete-targets are excluded from the index so cross-resource references to them surface as Severity=Error dangling references. New regression test: TestDOProvider_ValidatePlan_DatabaseVPCRefToDeleteTargetIsDangling. #2 (validate_plan.go:91) — appendDatabaseDiagnostics docstring incorrectly said 'Warning when missing'; the implementation has always emitted PlanDiagnosticError, and the conformance scenario + TDD tests both REQUIRE Error severity. Doc rewritten to match the implemented contract. #3 (validate_plan.go:225) — zonesInGroup docstring promised sorted output but returned the raw underlying slice unsorted. Now copies + sorts (lexicographic) so diagnostic messages are deterministic and the returned slice is owned by the caller (safe to mutate). nil for unknown groups; the caller's strings.Join still works. #4 (provider_conformance_test.go:70) — comment described a 'stub-then-real swap' with two Initialize calls; the implementation has always made one call with the right token chosen up-front. Doc rewritten to match the actual single-call flow. #5 (codemod-report.yml:73) — fork PRs run with a read-only GITHUB_TOKEN per GitHub's pull_request workflow security model, so issues:create-comment + issues:update-comment fail 403 and would block CI. Gate the comment step on github.event.pull_request.head.repo.fork == false. The artifact upload step still runs unconditionally so the report remains reachable from the Actions tab. Verified locally: go test ./... -count=1 -race PASS, go test -tags=conformance ./internal/ -run TestConformance PASS, go vet clean, codemod-report.yml YAML valid. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: Copilot review round 2 — 2 vpc_ref type-mismatch findings 2/2 findings addressed (both legitimate type-safety improvements): #6 (validate_plan.go:123) — appendDatabaseDiagnostics validated that vpc_ref name resolved in the plan but did NOT validate that the target type is actually infra.vpc. A vpc_ref pointing to a Droplet or App Platform resource silently passed prior validation. Now surfaces a typed Severity=Error diagnostic when target.spec.Type != "infra.vpc". New regression test: TestDOProvider_ValidatePlan_DatabaseVPCRefToNonVPCType. #7 (validate_plan.go:189) — same bug in appendAppPlatformDiagnostics: vpc_ref pointing to a non-VPC resource would silently bypass the region-match check (target.spec.Config["region"] would be a region GROUP for an App Platform target, not a zone for a VPC) and the operator would never see a clear diagnostic. Same fix: typed Error when target.spec.Type != "infra.vpc". New regression test: TestDOProvider_ValidatePlan_AppPlatformVPCRefToNonVPCType. Both diagnostics carry the offending type in the message body so the operator immediately knows whether they typo'd a name (resolves to nothing) vs. typo'd a TYPE in a name (resolves to wrong resource). Verified locally: go test ./... -count=1 -race PASS, go test -tags=conformance ./internal/ -run TestConformance PASS, go vet clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: Copilot review round 3 — 4 deeper findings (vpc_ref UUID/template + classifyRegion + delete-test comment) 4/4 findings addressed (deeper architectural correctness): #8/#9 (validate_plan.go:137,205) — vpc_ref's accepted DO API shapes are: (a) a VPC UUID literal, OR (b) a wfctl JIT template like ${vpc.id} that resolves to a UUID at apply time via wfctlhelpers.ApplyPlan's jitsubst.ResolveSpec. The prior validator unconditionally treated vpc_ref as an in-plan resource Name and would have rejected production configs that use the canonical UUID shape (godo.AppVpcSpec{ID: vpcID} consumes it directly per internal/drivers/app_platform_buildspec.go:674). New looksLikeResourceName() helper detects UUID literals (RFC-4122 pattern) and JIT templates (${...} / $(...)), and the in-plan- name lookup is skipped for both. Only plain-name vpc_ref values trigger the dangling-reference + non-VPC-type checks. Diagnostic messages updated to call out the 'plain resource name' branch explicitly so an operator who hit the diagnostic understands it does not apply to UUID/template forms. Two new TDD tests cover the UUID-literal + JIT-template paths for both database and App Platform. #10 (validate_plan.go:307) — classifyRegion emitted 'a zone slug in group ""' for legacy zones (nyc2, ams2) that intentionally map to an empty group. Now special-cases the empty-group case to emit 'a zone slug not in any App Platform region group'. New TDD test exercises the path via an App Platform action with region=nyc2. #11 (provider_apply_test.go:94) — the test comment claimed real drivers reject empty-ProviderID deletes via typed validation; FirewallDriver.Delete actually resolves by name when ProviderID is empty. Comment rewritten to reflect that v2 dispatch contract is 'driver knows what an empty ProviderID means for its resource shape', not 'all drivers reject empty ProviderID'. Net new test count: 3 (UUID-deferred, JIT-template-deferred, classify-empty-group). All existing tests still pass. Verified locally: go test ./... -count=1 -race PASS, go test -tags=conformance ./internal/ -run TestConformance PASS, go vet clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: Copilot review round 4 — 3 polish findings (UUID case + t.Context + workflow perms) 3/3 findings addressed: #12 (validate_plan.go:16) — uuidPattern was lowercase-only; UUIDs are case-insensitive in practice (operator clipboards, templating engines, mixed-case API responses). Added (?i) flag so upper-case and mixed-case VPC UUIDs also classify as deferred-to-apply, not as plain resource names that would trigger false dangling-reference diagnostics. New test: TestDOProvider_ValidatePlan_VPCRefAsUpperCaseUUIDIsDeferred. #13 (provider_conformance_test.go:86) — switched p.Initialize(context.Background(), ...) to p.Initialize(t.Context(), ...) so live-cloud Initialize is interrupted promptly when the test is canceled or hits its deadline. Removed the now-unused "context" import. #14 (codemod-report.yml:12) — dropped the unused pull-requests:write permission. The workflow only creates/updates an issue comment (PR comments are issues at the GitHub API layer) so the surviving issues:write is sufficient. Inline doc-comment captures the reasoning so future maintainers don't restore the broader grant. Aligns with ci.yml's contents:read-only baseline. Verified locally: go test ./... -count=1 -race PASS, go test -tags=conformance ./internal/ -run TestConformance PASS, go vet clean, codemod-report.yml YAML valid. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: Copilot review round 5 — 2 findings (Initialize-ctx comment + unknown-region forward-compat) 2/2 findings addressed: #15 (provider_conformance_test.go:83) — the prior comment claimed t.Context() makes Initialize cancelable, but DOProvider.Initialize today constructs its godo client with its own context.Background() and ignores the passed-in ctx. Comment rewritten to reflect that the change is forward-prep (so any future rev of Initialize that honors ctx picks up the test-scoped cancellation/deadline) rather than an immediate behavior fix. Tracked as a follow-up to thread ctx through the godo client construction. #16 (validate_plan.go:334) — the hardcoded region/zone allowlists would Severity=Error any brand-new DO region the plugin hasn't caught up to, blocking apply until the plugin is bumped. Severity is now two-bucket: - Documented misconfig (group-where-zone-required, zone-where- group-required) → Error (the original anti-pattern stays loud). - Unknown slug (neither known group nor known zone, e.g. a hypothetical 'atl' or 'atl1') → Warning so non-strict align lets operators on bleeding-edge DO regions proceed; --strict still escalates for cautious operators. New regression test: TestDOProvider_ValidatePlan_UnknownRegionSlugWarnsNotErrors covers both VPC zone and AP group unknown-slug paths. Net new test count: 1. All existing tests still PASS (the documented misconfig branches use known specific slugs that hit the Error path). Verified locally: go test ./... -count=1 -race PASS, go test -tags=conformance ./internal/ -run TestConformance PASS, go vet clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes the go-plugin handshake failure observed when v0.3.1 was loaded by wfctl v0.15.2 (BMW deploy #134's latest run failed with
plugin failed to exit gracefully).Root cause: the plugin's go.mod pinned workflow v0.9.4, compiling the plugin against an outdated
plugin/external/sdkandplugin/external/handshake. Under the v0.15.2 host those types and protocol details had drifted enough that the subprocess exited before the handshake completed.Bump to workflow v0.15.2. Build + all tests pass (
internal,internal/drivers). Version → 0.4.0 (minor — SDK jumped, API surface we use is stable).🤖 Generated with Claude Code