| name | sync-operator |
| description | Sync wsm to be e2e compatible with a target github.com/wandb/operator ref — always invokes /audit-operator-diff first, then applies the user-approved fix list, updates docs, and smoke-tests on Kind. Triggers on phrases like "sync wsm with operator", "bump operator dep", "update for new operator release", "match operator changes". v2 surface only by default; v1 only on explicit user request. Requires a local clone of the operator repo and OrbStack running. |
| allowed-tools | ["Bash(git -C *operator *)","Bash(git log:*)","Bash(git diff:*)","Bash(git show:*)","Bash(git status:*)","Bash(git branch:*)","Bash(git rev-parse:*)","Bash(git describe:*)","Bash(grep:*)","Bash(rg:*)","Bash(find:*)","Bash(ls:*)","Bash(cat:*)","Bash(head:*)","Bash(tail:*)","Bash(wc:*)","Bash(helm show chart:*)","Bash(helm pull oci://*)","Bash(gh pr view:*)","Bash(gh pr list:*)","Bash(gh repo view:*)","Bash(gh api -X GET *)","WebFetch(domain:github.com)","WebFetch(domain:api.github.com)","WebFetch(domain:raw.githubusercontent.com)","Bash(go get github.com/wandb/operator*)","Bash(go mod tidy:*)","Bash(go mod verify:*)","Bash(go build:*)","Bash(go vet:*)","Bash(make fmt)","Bash(make lint)","Bash(make build)","Bash(golangci-lint run:*)","Bash(gofmt:*)","Bash(./wsm cluster create:*)","Bash(./wsm cluster cleanup:*)","Bash(./wsm cluster destroy:*)","Bash(./wsm cluster list:*)","Bash(./wsm deploy-v2:*)","Bash(helm --kube-context kind-* uninstall:*)","Bash(kubectl --context kind-* get:*)","Bash(kubectl --context kind-* describe:*)","Bash(kubectl --context kind-* logs:*)","Bash(kubectl --context kind-* apply --dry-run:*)","Bash(docker info:*)","Bash(docker ps:*)","Bash(orbctl status:*)","Bash(lsof -nP -iTCP:*)"] |
Sync wsm with current operator
Goal: bring wsm into compile + runtime compatibility with a current operator release, with the smallest possible diff. Do not introduce new packages, abstractions, build-time flags, or refactors. Defer all "what changed" analysis to /audit-operator-diff; this skill only applies the audit's approved fixes.
Sandbox & read actions (read first)
allowed-tools only suppresses permission prompts — it does not lift the command sandbox. Every command below is read-only yet fails under the sandbox and must run unsandboxed (dangerouslyDisableSandbox: true); the failure is the sandbox, not the operation:
| Command | Failure under sandbox |
|---|
docker info | docker unix socket blocked — a denial here is NOT evidence the daemon is down; retry unsandboxed before telling the user to start it |
git -C "$OP" fetch … | .git/FETCH_HEAD write denied (Operation not permitted) |
go get … / go mod tidy / go mod verify | proxy.golang.org TLS interception (OSStatus -26276) — the host is already allowlisted; allowlisting does not fix it, so go runs unsandboxed |
go build / go vet / golangci-lint run | module cache / proxy under sandbox |
helm show chart oci://us-docker.pkg.dev/… | host not allowlisted (tls: failed to verify certificate) |
Retry unsandboxed once on any of these; don't misread the first denial as a real failure. To persist the friction-removers, offer the update-config skill to add us-docker.pkg.dev to the network allowlist and $OP/.git to the filesystem-write allowlist in .claude/settings.local.json. (The go/docker cases can't be allowlisted away — they need unsandboxed execution.)
Step 0: Prerequisites — block on these
- Local operator clone. Resolve the path the same way
/audit-operator-diff does — $WANDB_OPERATOR_DIR if set, else ask the user (do not hardcode ../operator; see feedback_shared_skill_portability in user memory). Verify the resolved $OP/go.mod contains module github.com/wandb/operator. If absent, ask for the path or abort. The audit (Step 1) does the authoritative resolution and may also use the portable GitHub-clone path; this is just the blocking prerequisite check.
- Container runtime running. Prefer
docker info (works for OrbStack — reports Name: orbstack — and Docker Desktop). orbctl status is OrbStack-specific and not always on PATH. Run it unsandboxed (the socket is sandbox-blocked); a sandboxed failure is not a down daemon. If genuinely down, ask the user to start it.
- Kind ports free. OrbStack binds
:8080 and :8443 for its proxy, colliding with wsm's default Kind ports. Check lsof -nP -iTCP:8080 -sTCP:LISTEN and :8443; if bound, use higher ports later (--http-port 18080 --https-port 18443).
- Working tree state.
git status --porcelain. Three cases:
- Clean — proceed normally.
- In-progress sync from a previous run — do not ask to stash. Recognize this state when
git diff HEAD -- go.mod shows an operator pin bump AND cmd/wsm/deploy_v2.go / pkg/operator/operator.go are modified AND/OR .claude/audit-report.md already exists. Treat the in-progress edits as the baseline and continue: /audit-operator-diff will pick up the in-progress pin as the target (see its Step 0/1), and Step 2 below applies the remaining fix list on top of what's already there.
- Unrelated dirty edits — ask the user whether to stash / continue / abort, as before.
Step 1: Audit the diff
Invoke /audit-operator-diff. Hand off control; the audit will:
- Ask the user for the target operator ref.
- Enumerate all changes between the current pin (read from
go.mod) and the target.
- Cross-reference each change with wsm's coupling sites.
- Classify findings into 🔍 Pre-flight, 🚫 Breaks, ⚠️ Mismatches, ➕ Missing capabilities, ℹ️ Internal-only.
- Probe OCI chart freshness vs the operator-repo Chart.yaml (run early when chart deps / RBAC changed).
- Run the systematic
SecretKeySelector zero-value-trap sweep and emit the mandatory full-surface capability ledger (every CR field → exposed / new / gap), so operator↔wsm parity is reconciled in the report — not left for the user to push back on.
- Present the report inline and via
.claude/audit-report.md.
- Collect the user's batched decisions (
all / none / abort / comma-separated IDs like B1=C, M1, M2, C1).
- Update the report's
## Decisions and ## Recommended action sections.
When the audit hands back control, read .claude/audit-report.md. The Decisions section is the source of truth for what to apply in Steps 2–3.
Audit verdicts that change the flow
- Recommended action:
ABORT-AND-ROLLBACK — typically when a pre-flight is fatal (chart-Go drift) or a Break has no acceptable resolution. Default behavior: do NOT proceed to Step 2. Roll back if a previous run already bumped go.mod: go get github.com/wandb/operator@<previous-pin> && go mod tidy. Tell the user the report path and stop.
- Recommended action:
WAIT-FOR-UPSTREAM — same default: do not proceed. Surface the upstream dependencies (e.g. "operator team needs to re-push chart at tag X") and stop.
- Recommended action:
INFORMATIONAL-ONLY — no fixes selected. Stop.
- Recommended action:
PROCEED-TO-SYNC — continue to Step 2.
Override path on ABORT / WAIT-FOR-UPSTREAM
The user can legitimately want to stage the wsm-side bump even when the audit recommends abort — so that when the upstream gap closes (chart republish, missing field added, etc.) no wsm work blocks the release. Recognize this when the user explicitly says "proceed anyway", "override P1", "stage the wsm side", or picks an AskUserQuestion option that says so. In that case:
- Apply the approved M+C fix list normally (Step 2).
- Skip Step 4c smoke test — it will fail at the known upstream gap, producing no new signal. Run Step 4b server-side dry-run if useful; document the known-broken runtime gap in the PR description instead.
- Do NOT roll back
go.mod. The whole point of the override is to land the wsm-side bump.
- Tag every fix whose runtime verification was blocked as
UNVERIFIED (matches the audit report's per-Break Verification: field). A B-fix like an OIDC strip that depends on the blocked 4b/4c is staged on static reasoning only — the static build proves it compiles, not that the CRD/operator accept it. List the UNVERIFIED fixes explicitly in the PR description so a reviewer knows what still needs a real-cluster pass once the upstream gap closes.
- The PR description should call out the upstream dependency and what unblocks the runtime side (e.g. "blocked by operator chart republish at tag X — wsm side ready"), plus the
UNVERIFIED fix list and the audit's full-surface capability ledger (so deferred gaps are visible, not silently dropped).
Step 2: Apply the approved fix list
Iterate the Decisions block in .claude/audit-report.md. For each approved finding:
Bump operator dep (always — implied by any non-INFORMATIONAL run)
go get github.com/wandb/operator@<target-ref>
go mod tidy
Baseline build before any edits (mandatory). Immediately after go get/go mod tidy, run go build -tags=containers_image_openpgp ./cmd/wsm ./pkg/operator before touching a line of wsm source. This isolates dep-bump breaks (operator renamed/removed a type wsm references → compile error now) from fix-induced breaks (errors you introduce later). A clean baseline build also positively confirms the audit's "all v2.* types survive" verified-safe claim. If the baseline build fails, the failing symbols are exactly the mechanical-rename Breaks — fix those first.
Resolution of 🚫 Breaks
Use the resolution the user selected for each Break finding (e.g. B1=C). Per-finding edits live in the audit report. Typical patterns:
- CR field renamed/restructured: mechanical rename in
cmd/wsm/deploy_v2.go:47-96 (template) and :897-1021 (processWandbCR). Driven by the compiler — go build ./cmd/wsm reports every stale construction.
- Zero-value serialization trap (OIDC-style): minimal Unstructured strip in
pkg/operator/operator.go:ApplyCR if the user picked the wsm-side workaround. Document the upstream dependency that allows the workaround to be removed.
Resolution of ⚠️ Mismatches
Apply the fix recipe from the audit report. Common edit sites:
- Flag descriptions / defaults:
cmd/wsm/deploy_v2.go:115-134 (DeployV2Cmd().PersistentFlags()).
- Flag value validation: in
processWandbCR (cmd/wsm/deploy_v2.go:897-1021), operatorDeployCmd().RunE, or wandbCreateCmd().RunE.
- Chart values wsm sets:
pkg/operator/operator.go:DeployOperator releaseValues map.
Resolution of ➕ Missing capabilities
When approved (add):
- Register the new flag in
DeployV2Cmd() (~line 115).
- Read it in
wandbCreateCmd().RunE (~line 250) and operatorDeployCmd().RunE (~line 331).
- Plumb through
processWandbCR signature (~line 897), performDeploy signature (~line 429), and operator.DeployOperator signature in pkg/operator/operator.go (~line 535).
- Set the corresponding
wandbCR.Spec.<...> field or releaseValues["..."] key.
Pattern: new corev1.SecretKeySelector field on the CR
When the operator adds a new SecretKeySelector-typed field (or a struct of them, like OidcSpec), expose one flag per leaf field taking <secret-name>:<key> form. Do not collapse into a single --<feature>-secret <name> flag that assumes conventional keys — pre-existing secrets in real clusters rarely conform to wsm's chosen key names, and forcing rename is operational friction. See feedback_secret_key_selector_flags.md in user memory for the why.
- Flag naming: after the leaf field (
--oidc-client-id, --smtp-username), not the parent struct (--oidc-secret).
- Value format:
<secret-name>:<key>, parsed via strings.Cut(value, ":"). Empty value = leave the SecretKeySelector zero — the existing ApplyCR strip handles serialization.
- No separate
--<feature>-file flag — --cr-file is already the escape hatch for users with complex CR shapes.
- Canonical example:
cmd/wsm/deploy_v2.go processWandbCR parseSecretRef helper + oidcRefs slice.
Drive cleanup with the compiler
After each round of edits: go build -tags=containers_image_openpgp ./cmd/wsm (the tag avoids the pkg-config/gpgme system dep that make build doesn't gate). Iterate until clean. Operator-side renames cascade — don't declare done after the first clean build mid-edit.
Step 3: Update user-facing docs
For every flag whose name, default, or description changed (per Decisions), update:
docs/reference/commands.md — flag table for the affected command.docs/reference/cr-fields.md — if a CR field reference changed.
Diff each affected docs section against the new flag set before editing.
Cross-check enums/defaults on adjacent flags too, not just the changed ones. While editing a flag table, verify the documented enum/default of neighboring rows still matches the code — pre-existing doc bugs cluster in the same table. (This run found --observability-mode documented as off|on when the real enum is off|full|forward — an unrelated stale doc fixed opportunistically.) Fixing an adjacent doc error you're already looking at is in scope; chasing doc bugs in unrelated files is not.
Step 4: Verify
4a. Static checks (mandatory)
make fmt is safe. make lint runs go vet without the containers_image_openpgp tag and will fail locally with pkg-config: executable file not found unless libgpgme-dev is installed (CI installs it). To verify locally:
make fmt
go vet -tags=containers_image_openpgp ./...
GOGC=off golangci-lint run --build-tags=containers_image_openpgp --timeout=5m
go build -tags=containers_image_openpgp ./cmd/wsm
Pre-existing lint findings unrelated to the operator bump (e.g. unchecked resp.Body.Close(), deprecated certmanagermetav1.ObjectReference) are out of scope — note them in the PR description, don't fix.
4b. Server-side CR dry-run (cheap, catches schema mismatches the typed compile misses)
Before launching the full smoke test, dry-run the CR shape against the cluster's installed CRD. This catches OIDC-style zero-value serialization traps and other CRD-vs-Go mismatches without spinning up the operator pod.
kubectl --context kind-wsm-sync-test apply --server-side --dry-run=server -f <(cat <<'EOF'
apiVersion: apps.wandb.com/v2
kind: WeightsAndBiases
metadata: {name: wandb-probe, namespace: wandb}
spec:
size: small
retentionPolicy: {onDelete: detach}
wandb: {hostname: "http://localhost:8080", version: "<defaultWandbVersion>", features: {}, internalServiceAuth: {enabled: false}}
mysql: {managedMysql: {telemetry: {enabled: false}}}
redis: {managedRedis: {telemetry: {enabled: false}}}
kafka: {managedKafka: {telemetry: {enabled: false}}}
objectStore: {managedObjectStore: {telemetry: {enabled: false}}}
clickhouse: {managedClickhouse: {telemetry: {enabled: false}}}
EOF
)
A serverside-applied (server dry run) response means the CRD accepts wsm's shape. Failures here = additional findings missed by the audit; loop back to Step 2.
4c. Kind smoke test (catches runtime gaps Step 4b can't predict)
./wsm cluster create --cluster-name wsm-sync-test --workers 0 --http-port 18080 --https-port 18443
./wsm deploy-v2 operator --context kind-wsm-sync-test --include-cr
Do not pass --wait — the flag is declared but commented out in cmd/wsm/deploy_v2.go (TODO note), so it errors unknown flag: --wait. Poll manually:
kubectl --context kind-wsm-sync-test -n wandb get weightsandbiases wandb \
-o jsonpath='{.status.ready}' --watch
Watch for true. If the install times out on a pre-install hook (context deadline exceeded on a cold cluster — chart has many sub-chart hooks), that's an environmental timeout, not a regression. Retry with cached images:
helm --kube-context kind-wsm-sync-test -n wandb-operators uninstall wandb-operator
./wsm deploy-v2 operator --context kind-wsm-sync-test --include-cr
4d. Diagnostic decision tree on failure
| Failure | Likely cause |
|---|
unknown flag: --wait | The skill's command, not wsm. Drop --wait. |
failed pre-install: ... context deadline exceeded | Helm timeout vs slow image pulls. Retry once cached. |
Operator pod never Ready, WaitForOperator hangs | Webhook config rename (audit's Category E) — confirm wandb-operator-mutating-webhook-configuration name in pkg/operator/operator.go:649. |
ApplyCR rejected with missing required field | Audit missed an added required field; loop back to Step 1. |
ApplyCR rejected with unknown field or expected X, got Y | Audit missed a CRD shape change (often zero-value serialization). Loop back. |
failed calling webhook ... connection refused | Webhook not ready; should be caught by WaitForOperator. Check pod logs in wandb-operators namespace. |
Operator pod runs but CR status.ready never flips | Reconcile failure. kubectl -n wandb-operators logs -l app.kubernetes.io/name=wandb-operator — often chart-Go RBAC drift (the audit should have caught this as a pre-flight). |
4e. Cleanup
./wsm cluster cleanup --context kind-wsm-sync-test
./wsm cluster destroy --cluster-name wsm-sync-test
v1 path (only when user explicitly asks)
v1 (wsm deploy, wsm list, wsm download, wsm console) is in maintenance mode. If a v1 change is requested, defer to the "v1 update playbook" table in CLAUDE.md. Edit sites: cmd/wsm/list.go, cmd/wsm/download.go, pkg/crd/weightsandbiases.go, pkg/spec/spec.go, pkg/helm/consts.go, pkg/deployer/deployer.go. Verify with ./wsm list (no cluster needed). The audit skill does not currently cover v1.
What this skill does NOT do
- No analysis or classification of operator changes — that's
/audit-operator-diff's job.
- No new packages, build-time flags, or abstractions.
- No refactor of the typed CR template.
- No automatic PR creation — leave the working tree clean for review.
- No consolidation of helm v3/v4 imports.
- No edits to v1 surface unless explicitly requested.
Failure modes
- Audit recommends ABORT: stop here. Restore go.mod if needed. Surface the upstream blockers.
- Build cascades: rerun after each fix round; operator changes routinely surface more compile errors after the first.
go mod tidy adds unexpected indirect deps: define "unexpected" so you know when to stop. A new top-level / direct module, or a major-version jump of any module → stop, git diff go.mod go.sum, and ask the user. Patch/minor bumps of already-present transitive deps (e.g. k8s.io/*, go-openapi/* moving a patch level — common, since the operator bump drags its own dep graph) → expected; note the count in the PR description and proceed without asking. When unsure, show the go.mod (not go.sum) diff — it's the signal; go.sum churn is noise.- Smoke test fails with a category not in 4d: treat as new audit finding; loop back to Step 1 with the new evidence.
- Docs already in sync: skip Step 3 edits, report "docs already match".
