Imported from wandb/wsm (.claude/skills/audit-operator-diff/SKILL.md). Install upstream with npx skills add wandb/wsm --skill audit-operator-diff. Copyright stays with the author.

Audit operator diff for wsm impact

Goal: enumerate every change between two operator refs and classify each by its impact on wsm. Output: a structured Markdown report at .claude/audit-report.md with user-approved decisions, ready for /sync-operator to consume. Never edits wsm source files — audit-only.

Precision mandate (read first)

The report is a forensic record, not a summary. Another agent or engineer must be able to reconstruct what changed and to what extent from the report alone, without re-running the diff. Three non-negotiable rules:

1. Total file coverage. Every file in the Step 2 inventory must be accounted for in the report — assigned to a bucket and represented by a finding ID or listed in the coverage ledger. Silent omission is a defect. "≈260 files, mostly internal" is not coverage; the internal files must still be enumerated by logical change-group (see Step 6).
2. Concrete before→after. Every finding cites file:line (or file + symbol) and the actual values that changed — old key → new key, old type → new type, old default → new default, added enum value, removed required entry. Never write "telemetry changed" or "values reworked"; write scrape.kubeStateMetrics added (bool, default true) — deploy/telemetry/values.yaml:17, schema values.schema.json:69.
3. Confirm the negative. When you check a wsm coupling site (a releaseValues key, telemetry.mode, the webhook name, a CR field) and find it unaffected, record that explicitly in a ✅ Verified-safe couplings section with the evidence that proves it safe. An unmentioned coupling reads as "not checked," which is indistinguishable from a miss. Asserting the negative is how the reader knows the audit was exhaustive.

These rules apply to every bucket — including Telemetry and Internal-only, which are the two most often under-reported.

Operator source convention ($OP)

This skill is shared across the team, so it assumes nothing about any particular machine — not that an operator clone exists, and not where it lives. Throughout, $OP denotes the operator repo location resolved in Step 0. Run every operator git command against it (git -C "$OP" …); never hardcode ../operator. Default to the clone-free GitHub path so the audit reproduces identically for any teammate regardless of their working copy.

Sandbox & read actions (read first)

The allowed-tools frontmatter only suppresses permission prompts — it does not lift the command sandbox (network/filesystem isolation). These read-only commands routinely fail under the sandbox and must run unsandboxed (dangerouslyDisableSandbox: true); the failure is the sandbox, not the operation:

Don't treat the first denial as a real failure (a stale clone, a down registry). Retry unsandboxed once, then continue. To remove the friction persistently, 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 (note: the go-proxy hosts are already allowlisted but still break on the sandbox's TLS interception, so go commands — used by /sync-operator, not here — must run unsandboxed regardless).

Step 0: Prerequisites & operator source

1. Resolve the operator source — ask first. The diff needs git history of github.com/wandb/operator at two refs; where that history comes from is a per-machine detail, so decide it explicitly with AskUserQuestion:

   • GitHub upstream (no local clone needed) — the portable default. Make a fresh blobless clone to a temp dir so every branch/tag resolves and the result is independent of anyone's working copy:

     git clone --filter=blob:none https://github.com/wandb/operator "$TMPDIR/operator"

     Then OP="$TMPDIR/operator". (Network + .git writes are sandboxed; if the clone fails on a sandbox denial, retry with dangerouslyDisableSandbox: true. The first time the command runs it may also prompt for permission — that's expected.)
   • Existing local clone — faster and offline, but path and freshness vary per machine, so it can be stale or parked on a feature branch. Resolve the path in priority order — $WANDB_OPERATOR_DIR if set, otherwise ask the user — then set OP to it and verify test -f "$OP/api/v2/weightsandbiases_types.go". Report its branch/HEAD and offer to git -C "$OP" fetch origin --tags --prune so the target refs are current before diffing.

   Record the resolved choice in the report header - **Diff source:** line (e.g. GitHub upstream, cloned <sha> or local clone <path> @ <sha>, fetched <date>).

2. Working tree state — git -C . status --porcelain (wsm) and git -C "$OP" status --porcelain (operator).

   • Dirty $OP (only possible for a local clone): warn the user; the named refs are still authoritative for the diff.
   • Dirty wsm tree: check whether the dirty state is an in-progress sync from a previous run. Heuristic: git diff HEAD -- go.mod | grep -E '^[-+].*github.com/wandb/operator' shows a pin change AND any of cmd/wsm/deploy_v2.go, pkg/operator/operator.go are modified AND/OR .claude/audit-report.md exists from a prior run. In that case, treat the dirty pin as the implicit target (see Step 1.2) and do not ask the user to stash. Otherwise warn that uncommitted wsm edits could confuse the cross-reference in Step 4.

Step 1: Resolve refs

1. Read committed pin from wsm's go.mod (always git show HEAD:go.mod, never the working-tree file — the working tree may already be bumped by a prior incomplete sync):

   git show HEAD:go.mod | grep 'github.com/wandb/operator '

   Parse the version string. Decode pseudo-versions (v1.21.3-0.20260424225346-b25b977174fd) to commit hashes (here b25b977174fd); stable tags (v1.2.3) map directly. This is the audit's "from" ref.

2. Pick the target ref. Two paths:

   • In-progress sync detected in Step 0 — read the bumped pin from the working-tree go.mod and default the target to that SHA. Confirm with the user in one line ("Auditing from committed pin <old> to in-progress bump <new>?") rather than asking open-ended.
   • Clean tree, no in-progress sync — ask the user. Typical answers: a branch HEAD (v2, main), a tag (v1.22.0), or a specific commit hash. List candidate refs (works for both source types — remote-tracking refs and tags in the temp clone):

     git -C "$OP" for-each-ref --sort=-creatordate \
       --format='%(refname:short)  %(creatordate:short)' refs/remotes refs/tags | head -15

3. Resolve both refs to commit SHAs and confirm they're reachable: git -C "$OP" rev-parse <ref> for each.

Step 2: File-level diff inventory

git -C "$OP" diff <current>..<target> --name-status | sort
git -C "$OP" diff <current>..<target> --stat | tail -1   # total file/insertion/deletion count for the ledger

Assign every file to exactly one bucket. Build a coverage ledger as you go (bucket → file count → which files / finding IDs) — this is reproduced verbatim in the report (Step 7) and is the proof of total coverage. Do not let any file fall through; if a path doesn't obviously fit a bucket, default it to Operator internals and note it. Renames (R###) count as one file and must record both old and new path.

Canonical classifier (test-first, then path-prefix — avoids double-counting). Tests live in many dirs, so classify *_test.go first, then by prefix; this guarantees mutually-exclusive buckets that sum to the total on the first pass:

git -C "$OP" diff <from>..<to> --name-status | awk -F'\t' '{print $NF}' > /tmp/op_paths.txt
python3 - /tmp/op_paths.txt <<'PY'
import sys, re
def bucket(p):
    if p.endswith("_test.go"): return "tests"
    if p.startswith("api/v1/"): return "api_v1"
    if re.match(r"api/v2/.*\.go$", p): return "go_api"
    if p.startswith("config/crd/bases/"): return "crd"
    if p.startswith("config/rbac/"): return "rbac"
    if p.startswith("deploy/operator/"): return "op_chart"
    if p.startswith("deploy/telemetry/"): return "telemetry"
    if p.startswith("deploy/"): return "deploy_other"
    if p.startswith("hack/"): return "hack"
    if p.startswith("docs/"): return "docs"
    return "internals"   # internal/, pkg/, cmd/, config/* non-crd/rbac
b = {}
for line in open(sys.argv[1]):
    p = line.strip()
    if p: b.setdefault(bucket(p), []).append(p)
tot = 0
for k in sorted(b, key=lambda k: -len(b[k])):
    print(f"{k}: {len(b[k])}"); tot += len(b[k])
print("TOTAL:", tot)   # must equal `git diff --name-only | wc -l`
PY

For renames, --name-status emits the new path in $NF (one file). Internals (internal/, pkg/vendored, cmd/, config/ non-crd/rbac) still get logical change-group enumeration per the Precision mandate.

Categorize each changed file into one of:

Step 3: Run the targeted checks per bucket

The Precision mandate applies to every bucket equally — telemetry was only the example that exposed a miss; it is not special. For each bucket: diff every file in it, and enumerate every change with concrete before→after, not a representative sample. The per-bucket notes below say what kind of change matters most in each, but they never license skimming. Run git -C "$OP" diff <from>..<to> -- <bucket-paths> for the whole bucket and account for each hunk — an unmentioned change is a coverage gap, the same way an unlisted file is. The Telemetry subsection shows the depth expected; replicate that depth here for Go API, CRD, chart, RBAC, and internals.

Go API

• Diff the whole file: git -C "$OP" diff <from>..<to> -- api/v2/*.go.

• For every modified type X struct: enumerate each field change (added / removed / type change / JSON-tag change) with concrete before→after (e.g. OidcSpec.ClientId: string → corev1.SecretKeySelector; *Connection.URL json tag URL → url,omitempty). Do not stop at the first interesting type.

• Systematic zero-value / SecretKeySelector trap sweep (mandatory — do not check just the one flagged field). The OIDC trap is a pattern, not a one-off: any non-pointer struct field declared <Field> <StructType> json:",omitempty"`` is a candidate, because Go's omitempty does NOT drop a zero-value struct (only nil pointers, empty slices/maps/strings — confirmed: the k8s DefaultUnstructuredConverter.isZero returns false for reflect.Struct). A zero corev1.SecretKeySelector serializes as {"key":""} (its Key field has no omitempty), so a by-value wrapper emits a non-empty object the operator may reject or mis-resolve. Run the sweep every time:

  1. Enumerate every struct in the target API that contains a SecretKeySelector (or any nested no-omitempty field):

     git -C "$OP" show <to>:api/v2/weightsandbiases_types.go | grep -nE 'corev1\.SecretKeySelector|^type .* struct'

  2. For each such struct, trace how wsm's CR reaches it and classify — record the verdict for all three classes:
     • By-value omitempty on a path wsm builds (e.g. WandbAppSpec.OIDC OidcSpec) → TRAP. wsm emits {"key":""} whether or not it sets the field. Candidate 🚫 B (strip in ApplyCR, or per-leaf flags — see sync skill).
     • Reached only via a *pointer field (e.g. MySQLSpec.ExternalMysql *MysqlConnection) → nil-safe, omitempty drops it on wsm's managed-only path. Record in ✅ Verified-safe with the pointer evidence (NOT "operator-generated" — the real reason is pointer-gating).
     • .status / operator-output only → not sent on apply.
  3. Also sweep the managed *Config / ManagedInfraSpec / Telemetry structs wsm builds by value for any nested no-omitempty field, not just SecretKeySelector.
  • Cross-reference each candidate against wsm's cmd/wsm/deploy_v2.go:wandbCR template and processWandbCR. A SecretKeySelector-bearing struct that is neither a B finding nor a ✅ Verified-safe line is a coverage gap — every one must be accounted for.

CRD schema

• Diff the full file: git -C "$OP" diff <from>..<to> -- config/crd/bases/*.yaml. Account for every hunk.
• Enumerate every added/removed required: entry (each can reject or relax a CR that previously validated — note which direction).
• Enumerate every +kubebuilder:validation:Enum=... / enum: change (tightening or widening) with the old and new value sets.
• Enumerate every type: change (e.g. string → object) and new property block, with the path.
• For each, state whether wsm's CR is affected, and put the unaffected ones in ✅ Verified-safe couplings (e.g. "removed required: -URL relaxes the schema; wsm never sent URL → safe").

Chart structure

• Diff dependencies: in Chart.yaml and Chart.lock. List every dep added / removed / renamed / version-bumped with old→new (name + version + repository). If a dep was renamed (e.g. mysql-operator 2.1.10 → moco 0.24.0), check the chart version: field too — if the chart version string was NOT bumped, the OCI tag is mutable and runtime is unpredictable (P-tier; see Step 5).
• Early-gate trigger: the moment this bucket shows a dep rename/add/remove or config/rbac/role.yaml (RBAC) changed, run Step 5 (OCI freshness) immediately — before the deep per-bucket dive. If the published OCI chart is frozen, most downstream findings become "staged-only" and the user should learn that up front, not at the end. The OCI drift check is the single most decision-determining probe in the audit; do not leave it for last when these signals are present.
• Diff all top-level keys in values.yaml (added/removed/renamed, each with old→new) and values.schema.json. Cross-reference with pkg/operator/operator.go:DeployOperator releaseValues (currently sets wandb.install, wandb-operator.image.pullPolicy, telemetry.mode). Flag any key wsm sets that no longer exists (🚫 B); record the keys wsm sets that still exist in ✅ Verified-safe.
• Operator image tag, not just chart version. The deployed binary is wandb-operator.image.{repository,tag} in deploy/operator/values.yaml, separate from the chart version:. Record both at <to> (git -C "$OP" show <to>:deploy/operator/values.yaml | grep -A4 '^wandb-operator:'). If the image tag is unchanged while the binary (Go API / RBAC / reconcilers) changed, the published image is frozen too — reinforces the P-tier chart-Go drift. A chart-version bump with a stale image tag (or vice-versa) is itself a drift signal.

Chart templates

• Diff every modified template: for f in $(git -C "$OP" diff <from>..<to> --name-only -- deploy/operator/templates/); do echo "=== $f ==="; git -C "$OP" diff <from>..<to> -- "$f"; done.
• Enumerate every Helm validation rule change — eq $.. "...", case, fail, required blocks — with old→new accepted values; cross-reference enum values with wsm's flag-accepted values.
• Always inspect by name and report what changed (even if "no semantic change"):
  • deploy/operator/templates/telemetry-validation.yaml — telemetry mode enum
  • deploy/operator/templates/*-role.yaml — RBAC scope changes (renames like *-minio-role → *-seaweedfs-role tie to the P-tier chart-Go break)
  • deploy/operator/templates/webhooks.yaml — the MutatingWebhookConfiguration name WaitForOperator depends on
  • deploy/operator/templates/hooks/* — crd-installer / helm-hook changes (interact with wsm's WaitStrategy="hookOnly")

RBAC

• Diff config/rbac/role.yaml fully and enumerate every apiGroup/resource/verb added or removed with old→new (e.g. mysql.oracle.com removed, moco.cybozu.com/mysqlclusters added; minio.min.io removed, seaweed.seaweedfs.com/seaweeds added).
• Any resource-group change is a chart-Go mismatch indicator: the operator binary needs the new RBAC, but the chart at the current OCI tag may still grant the old. Cross-check against Step 5 — if role.yaml changed but the chart version: didn't bump, this is a fatal P-tier finding regardless of any wsm code fix.

Telemetry subchart

This is the worked example of the per-bucket depth every bucket above must match — not a uniquely-detailed bucket. wsm sets telemetry.mode in releaseValues, so it is also directly wsm-coupled. Run all of:

# operator chart's top-level telemetry block (the mode enum wsm depends on)
git -C "$OP" diff <from>..<to> -- deploy/operator/values.yaml | grep -A30 'telemetry:'
git -C "$OP" show <to>:deploy/operator/values.yaml | grep -A20 '^telemetry:'   # confirm mode comment block: off/full/forward
# telemetry subchart — diff EVERY file, not just values
git -C "$OP" diff <from>..<to> -- deploy/telemetry/values.yaml
git -C "$OP" diff <from>..<to> -- deploy/telemetry/values.schema.json
git -C "$OP" diff <from>..<to> --stat -- deploy/telemetry/templates/ deploy/telemetry/dashboards/
for f in $(git -C "$OP" diff <from>..<to> --name-only -- deploy/telemetry/templates/); do echo "=== $f ==="; git -C "$OP" diff <from>..<to> -- "$f"; done

For each result, record concretely:

• Every added/removed/renamed key in values.yaml + values.schema.json with its type and default (e.g. scrape.kubeStateMetrics: true added). New optional keys are candidate ➕ C findings; removed keys wsm sets are 🚫 B.
• The telemetry.mode enum (off / full / forward). Cross-reference with wsm's releaseValues["telemetry"]["mode"] value and any wsm flag that feeds it. If the enum is unchanged, say so in ✅ Verified-safe couplings; if a mode was added/removed/renamed, it's an M or B.
• Template reworks — scrape-target selector/namespace changes, new VMServiceScrape/VMPodScrape objects (e.g. kube-state-metrics, seaweedfs replacing minio), honorLabels changes. These are usually I-tier (internal scrape wiring), but enumerate each as a one-liner; a scrape renamed minio→seaweedfs ties back to the P-tier chart-Go break.
• Dashboard JSON add/modify — I-tier, but list each file by name and whether added or modified.

Step 4: Cross-reference findings with wsm

For each operator change identified in Step 3, find the wsm coupling site:

Reconcile defaultWandbVersion. CLAUDE.md calls cmd/wsm/deploy_v2.go:34 defaultWandbVersion "hand-bumped per release" but neither audit nor sync verifies it. Compare wsm's defaultWandbVersion against the chart's wandb.version default at <to> (git -C "$OP" show <to>:deploy/operator/values.yaml | grep -A3 '^wandb:'). If they diverge, flag it (⚠️ M) with both values — wsm may be defaulting to a server manifest version the matched operator/chart doesn't expect. If they agree (or wsm intentionally pins newer), record it in ✅ Verified-safe with both values.

Confirm the negative. Walk this whole table every run, even when nothing changed. For each coupling site that you checked and found unaffected, add a line to the report's ✅ Verified-safe couplings section stating what you verified and the evidence (e.g. "wsm releaseValues keys wandb.*, wandb-operator.image.pullPolicy, telemetry.mode all still present in <to>:deploy/operator/values.yaml — no break"; "all 22 v2.* type names wsm references survive at <to> — grep -rohE 'v2\.[A-Z][A-Za-z0-9]+' cmd/ pkg/ cross-checked against the api/v2 diff"). A coupling that is neither a finding nor a verified-safe line is a coverage gap.

Step 4.5: Parity validation gate (operator ↔ wsm completeness — mandatory)

The audit must prove it reconciled the operator surface against wsm on its own — the reader should never have to push back with "are you sure you got everything." Two reconciliations, both required, both written into the report:

1. Diff-range completeness (the coverage gate, Step 6) — every changed file is bucketed and every hunk is a finding or verified-safe. This proves nothing was missed in the diff.

2. Full-surface capability ledger (new, mandatory) — enumerate the entire WeightsAndBiasesSpec tree at <to> (not just the diff range) and mark each field's wsm exposure. This proves wsm parity is understood, and surfaces pre-existing gaps proactively instead of reactively.

   # Dump the full spec tree to enumerate every field
   git -C "$OP" show <to>:api/v2/weightsandbiases_types.go | sed -n '/type WeightsAndBiasesSpec struct/,/^}/p'
   # …then recurse into WandbAppSpec, NetworkingSpec, each *Spec / Managed*Spec / *Config

   For every leaf field, assign one status:

   • ✅ exposed — wsm sets it (flag or template). Cite the flag / template line.
   • ➕ new-this-diff — added in <from>..<to>; this is a C finding (in scope).
   • ⬜ gap → --cr-file — operator supports it, wsm has no flag, pre-existing. Out of fix scope, but must still be listed so the user sees the full picture without asking.

   Write this as a ## Full-surface capability ledger section in the report (table: CR field → status → wsm site / note). The C-bucket fix list stays diff-range-scoped; this ledger is the visibility mechanism, not a fix list. If the ledger is large, group by spec (Wandb / Networking / each infra) — but enumerate every field; "mostly covered" is not a ledger.

This gate is the structural answer to "did you reconcile everything": the SKS sweep (Step 3) covers secret-bearing correctness, the coverage gate covers the diff, and this ledger covers full-surface capability. All three are mandatory.

Step 5: Check chart vs Go-API drift (OCI freshness)

Run this as soon as Step 3's Chart-structure or RBAC bucket shows a dep rename/add/remove or role.yaml change (see the early-gate trigger there) — not only at this position. Its verdict reorders the whole audit.

helm show chart oci://us-docker.pkg.dev/wandb-production/public/wandb/charts/operator \
  --version <wsm-default-chart-version>

Sandbox note: us-docker.pkg.dev is typically outside the default network allowlist. Expect the first attempt to fail with tls: failed to verify certificate or similar; retry with dangerouslyDisableSandbox: true. Don't treat the first failure as a real OCI outage.

Compare its dependencies: against $OP/deploy/operator/Chart.yaml at <target> ref (git -C "$OP" show <target>:deploy/operator/Chart.yaml). If they differ AND the chart version: string didn't bump → fatal pre-flight condition. Note the OCI manifest digest in the report for traceability.

Also: if config/rbac/role.yaml changed but the chart version: didn't, surface this prominently — runtime will fail regardless of any wsm code fix.

Step 5.5: Check for removable wsm-side workarounds

When prior syncs added wsm-side strips/workarounds (e.g. pkg/operator/operator.go:ApplyCR removes zero-value OIDC or .status), the audit should look for whether the upstream condition that justified each workaround has resolved:

• OIDC strip is removable when either:
  • WandbAppSpec.OIDC in the new operator types is *OidcSpec (pointer), so Go's omitempty correctly drops the nil; the strip's == v2.OidcSpec{} comparison won't even compile.
  • The OCI chart at the deployed tag has been re-published with the new SecretKeySelector-typed OIDC CRD (so wsm sending zero-value {"key":""} validates).
• Status strip is permanent best practice; flag only if a reviewer wants to revisit.

Surface removable workarounds as ⚠️ Mismatches with a "remove cleanup" recipe.

Step 6: Classify findings into 5 buckets

Assign each finding a stable ID (P<n>, B<n>, M<n>, C<n>, I<n>):

1. 🔍 Pre-flight (P) — independent gates. If any P is fatal, the audit's recommended action is ABORT regardless of other findings.
2. 🚫 Breaks (B) — wsm fails to compile or deploy without resolution. Per finding: resolution options (e.g. operator-side fix vs wsm workaround vs rollback).
3. ⚠️ Mismatches (M) — existing wsm drift caught by the audit (flag descriptions wrong, missing enum validation, etc.).
4. ➕ Missing capabilities (C) — operator added something new in this diff range that wsm doesn't expose. Scope: only changes within <current>..<target>. Pre-existing gaps are out of the fix scope here — but they are NOT invisible: they must all appear in the mandatory ## Full-surface capability ledger (Step 4.5) so the user sees full parity without asking.
5. ℹ️ Internal-only (I) — operator-internal changes, listed for context, no action. Still enumerated by logical change-group per the Precision mandate — never collapsed to "internal refactor."

Coverage gate (run before writing the report)

Reconcile the finding set against the Step 2 inventory:

# total files changed
git -C "$OP" diff <from>..<to> --name-only | wc -l

Sum the file counts across your coverage-ledger buckets and confirm it equals the total. If they don't match, you dropped files — find them and assign them before proceeding. Every changed file must be either (a) cited by a P/B/M/C/I finding, or (b) present in the coverage-ledger row for a bucket (e.g. "Hack/docs/tests — 41 files, skipped"). The ledger's "skipped" rows are legitimate coverage; an unlisted file is not. Do not write the report until the gate balances.

Step 7: Write the audit report to .claude/audit-report.md

Template:

# Operator sync audit

- **Generated:** <ISO timestamp>
- **From:** <current-pin-ref> (<short SHA>)
- **To:** <target-ref> (<short SHA>)
- **Diff source:** <GitHub upstream, cloned <sha> | local clone <path> @ <sha>, fetched <date>>
- **OCI chart probed:** <chart URL>@<version> (digest <sha256>)
- **Diff size:** <N files changed, +X/-Y lines> (`git diff --stat` total)

## Change coverage ledger

Every changed file is accounted for here (the Coverage gate must balance: Σ counts == total files changed).

| Bucket | Files | Finding IDs / disposition |
|---|---|---|
| Go API (`api/v2`) | <n> | B1, C1, I8 |
| CRD schema | <n> | (covered by B1/relaxations) |
| Operator chart structure | <n> | P1 |
| Operator chart templates | <n> | P1, I-roles |
| RBAC | <n> | P1 |
| Telemetry | <n> | C-ksm, I-scrapes |
| Operator internals (`internal/`, `pkg/vendored/`, `cmd/`) | <n> | I1–I6 |
| Hack / docs / tests | <n> | skipped (informational) |
| v1 conversion (`api/v1`) | <n> | skipped (wsm always v2) |
| **Total** | **<N>** | == `git diff --name-only \| wc -l` |

## ✅ Verified-safe couplings

Coupling sites checked and confirmed unaffected (proves the negative — see Precision mandate):

- <coupling> — <evidence, e.g. "wsm `releaseValues` keys all present in `<to>:deploy/operator/values.yaml`">
- <coupling> — <evidence>
- <SecretKeySelector sweep result — every SKS-bearing struct classified trap / pointer-safe / status, per Step 3>

## Full-surface capability ledger

Scope note: the ➕ Missing-capabilities fix list below is **diff-range only**. This ledger is the full-surface view (Step 4.5) — every `WeightsAndBiasesSpec` field at `<to>`, so pre-existing gaps are visible without the user asking.

| CR field | Status | wsm site / note |
|---|---|---|
| `spec.size` | ✅ exposed | `--size` |
| `spec.wandb.oidc.*` | ⬜ gap → `--cr-file` | no flag; pre-existing |
| `spec.<field>` | ✅ exposed / ➕ new-this-diff / ⬜ gap | <flag or note> |
| … (enumerate every leaf field) | | |

## 🔍 Pre-flight conditions

[P1] **Chart-Go drift** — <description>. Severity: <fatal | warning>. Resolution: <ABORT | WAIT-FOR-UPSTREAM | acknowledge>.

[continue for each P...]

## 🚫 Breaks
[B1] **<title>**
- Operator-side: `<file>:<line>` (commit `<sha>`) — **before→after:** `<old>` → `<new>`
- wsm-side: `<file>:<line>`
- Why: <one-line root cause>
- Resolution options: A) <option> · B) <option> · C) <option>
- Default if unselected: <option>
- **Verification:** `<how this fix is confirmed>` — tag **`UNVERIFIED`** if runtime proof (Step 4b dry-run / 4c smoke) was blocked (e.g. by a P-tier gate). An `UNVERIFIED` fix is staged on faith; say so explicitly here and in the PR description.

[continue...]

## ⚠️ Mismatches
[M1] **<title>**
- wsm-side: `<file>:<line>` (current state) — **before→after:** `<old>` → `<new>`
- Operator-side expectation: `<file>:<line>` or schema rule
- Fix recipe: <1-2 lines>

[continue...]

## ➕ Missing capabilities (new in this diff range)
[C1] **<title>**
- Operator added: `<file>:<line>` (commit `<sha>`) — **what:** `<concrete field/key/type + default>`
- wsm would need: <flag / CR field / chart value plumbing>
- Edit recipe if approved: <1-2 lines>

[continue...]

## ℹ️ Internal-only

One entry per logical change-group, each with a concrete description and a representative path — never a bare "internal refactor".

- I1: <change-group> — <what changed, concretely> (`<representative file>`, <N files>)
- I2: ...

## Decisions

(Filled in by user via batched approval. Format: comma-separated IDs with optional resolution choice. Examples:
- `B1=C, M1, M2, C1` — abort on B1 via option C, fix M1+M2, add C1
- `all-mismatches, C1` — fix every M, add C1, leave B/P unresolved
- `none` — informational audit only, no fixes
- `abort` — pre-flight fatal, stop here)

Decisions: <FILLED IN>

## Recommended action

<Based on Pre-flight + Breaks + user decisions: PROCEED-TO-SYNC | ABORT-AND-ROLLBACK | WAIT-FOR-UPSTREAM | INFORMATIONAL-ONLY>

Step 8: Present batched approval to user

Print the full report inline (or its summary if long), then collect batched decisions. Prefer AskUserQuestion when the decision space collapses to 2–4 distinct paths — typical when there's one P-tier fatal whose resolution determines everything else (e.g. "abort", "override and apply M+C fix list", "override and add capabilities"). The structured UI gets clearer responses than free-text. Fall back to free-text only when the report has 5+ independent findings the user might want to mix-and-match.

Accept formats (whether via AskUserQuestion options or free text):

• all — fix every M, add every C, resolve every B with its default
• none — informational audit, no fix list
• abort — pre-flight fatal, stop
• Comma-separated IDs (B1=C, M1, M2, C1) — explicit per-item, with =<choice> for B items that have options

Update the ## Decisions section in .claude/audit-report.md with the user's response, then set the ## Recommended action based on the combined state.

Step 9: Hand off

Print the report path: .claude/audit-report.md. Caller (user or /sync-operator) reads from here.

If invoked standalone (not via /sync-operator), exit. If invoked from /sync-operator, control returns to that skill's next step.

What this skill does NOT do

• No edits to wsm source files. Audit-only.
• No bumping of go.mod. That's /sync-operator's job.
• No verification (lint, smoke test). Also /sync-operator's job.
• No fix for the v1 surface. v1 is frozen — see CLAUDE.md's v1 update playbook.
• No fixing of pre-existing gaps wsm doesn't expose — the C fix list stays diff-range-scoped (<current>..<target>). It DOES, however, list every pre-existing gap in the mandatory ## Full-surface capability ledger (Step 4.5) for visibility; listing ≠ fixing.

Failure modes

• Operator clone missing or wrong module: Step 0 catches; ask user for path.
• Refs unreachable: git rev-parse will error; surface and ask user to fetch.
• OCI chart probe times out / 401: continue audit without the OCI drift check, flag P-tier "OCI freshness unverified" with severity warning.
• No changes at all between refs: acceptable — report with all empty buckets, recommended action INFORMATIONAL-ONLY.
• .claude/audit-report.md already exists: overwrite without prompt; previous report is single-run scratch.