plugin-sdk-impact

Imported from openclaw/openclaw (.agents/skills/plugin-sdk-impact/SKILL.md). Install upstream with npx skills add openclaw/openclaw --skill plugin-sdk-impact. Copyright stays with the author.

Plugin SDK impact

Combine current ClawHub inventories with a source scan of every bundled OpenClaw plugin, including plugins absent from ClawHub, to inform compatibility decisions. Static usage identifies dependencies; it does not prove execution or reliance on a particular behavior. Keep the approval and deprecation rules in the SDK boundary guide authoritative.

Investigate

1. Identify every affected public surface from source, including alternate SDK exports and runtime paths. Verify spelling against the definition and known callers; an incorrect key can return a convincing zero. Trace grouped facades, SDK helpers, and registration readers as well: a flat alias can have zero direct plugin calls while implementing a supported grouped method.

2. Scan every bundled plugin in a clean OpenClaw checkout using Plugin Inspector's static source analysis. Reuse an installed Inspector 0.4.3 or newer; otherwise install it in a separate temporary directory, without lifecycle scripts:

   inspector_dir=$(mktemp -d)
   npm install --prefix "$inspector_dir" --ignore-scripts @openclaw/plugin-inspector@0.4.3
   node {baseDir}/scripts/scan-bundled.mjs \
     --inspector "$inspector_dir/node_modules/@openclaw/plugin-inspector" \
     --out /tmp/openclaw-bundled-sdk-usage.json

   The scanner defaults to this skill's checkout; --root <checkout> selects another. It discovers extensions/*/openclaw.plugin.json, never executes plugin code, and records the commit, input fingerprint, Inspector version, every plugin, and incomplete analyses. A failed or timed-out scan remains a gap. Plugins without declared entrypoints also remain coverage gaps, because Inspector may miss implicit entrypoints. An interrupted corpus scan cannot be used as a finished inventory. Reuse the report for multiple queries while its source remains unchanged. A changed commit, plugin roster, or uncommitted source invalidates it; rescan instead of editing the receipt.

3. Query each canonical key with both sources:

   node {baseDir}/scripts/query.mjs 'api-member:api.registerTool' \
     --bundled /tmp/openclaw-bundled-sdk-usage.json

   For “updated within the last three months”, add --months 3. This means the latest eligible public release was uploaded within three calendar months of now, measured in UTC (month ends are clamped). Listing edits and rescans do not count as updates. The stored release date is its creation/upload time; public approval can happen later. --since 2026-07-16T12:00:00Z accepts an explicit cutoff. Report the exact inclusive cutoff from releaseWindow, not a guessed date. Public consumers and coverage use the same window on every page. Bundled source has no comparable upload date: report it separately, unfiltered, even for a recent-release question. The combined count is the union of that public release window and current bundled source, not a count of recent releases.

   The helper follows all usage and coverage pages, including empty intermediate pages. Counts and coverage include every result; items, gaps, and warnings contain bounded examples with omitted counts in output.omitted. Mixed references count in both runtime and type totals. For an investigation needing every consumer or gap, save --full output to a file and read it in sections:

   node {baseDir}/scripts/query.mjs 'api-member:api.registerTool' \
     --bundled /tmp/openclaw-bundled-sdk-usage.json --full > sdk-usage.json

   --help documents local endpoint and traversal limits.

4. Inspect matched evidence against the proposed change. A method call may not use the behavior being changed. Type-only consumers still matter for type/export removal. Query related members separately; a module import does not prove use of every export. Repeated references in a plugin count as one consumer.

5. For a usage-count question, report the key, query time, exact release window (if requested), ClawHub and bundled counts, their combined count, bundled commit, and both coverage gaps. Top-level distinctPlugins and coverage remain ClawHub-only; bundled describes source results; combined merges exact package identities once. Some bundled plugins have no package name: when they could overlap public matches, combined.distinctPlugins is null and distinctPluginBounds reports the possible range. observedPluginIdentities counts source identities before that uncertain overlap. Report the range, not its upper bound as an exact count. Different names remain distinct; never guess aliases from display names. Missing bundled evidence makes combined impact unavailable. For a proposed SDK change, also report matched evidence, indirect callers, and your compatibility recommendation. Finish change investigations with the maintainer's decision: preserve behavior, deprecate with migration, or proceed under existing policy. Never infer permission from the count.

Canonical keys

Keys describe SDK identities, not local aliases. api.runtime members normalize to runtime.*. The inventory includes supported and deprecated references; nonempty deprecationCodes selects entries associated with existing warnings. Per-plugin inspection is available at https://clawhub.ai/api/v1/sdk-usage/plugin?package=<encoded-name>; follow its nextCursor until null to obtain the full inventory.

Interpret the result

Coverage is tied to current release/artifact and analysis versions. Eligible ClawHub reports have inputKind: "artifact"; bundled reports use "source". A source-checkout benchmark cannot establish what a published package contains. Source-only type references still matter for SDK type and export removal; preserve the source of each match. An old scan timestamp alone does not make an immutable artifact stale. Evidence is sampled; each scan reports its evidence/unknown limits. Pages are live reads, not one snapshot: rerun if releases or visibility change during the investigation. The helper flags changes it observes but cannot establish a transaction-wide snapshot.

Do not dismiss a security issue just because no consumers were observed. Describe how usage evidence affects compatibility risk; security and SDK owners still decide the fix and any exception to the normal migration window.