diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index d9ac42dd097..eb930c1c4d7 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -877,7 +877,7 @@ jobs:
 
       - name: Run extension shard
         env:
-          OPENCLAW_EXTENSION_BATCH_PARALLEL: 3
+          OPENCLAW_EXTENSION_BATCH_PARALLEL: 2
           OPENCLAW_EXTENSION_BATCH: ${{ matrix.extensions_csv }}
         run: pnpm test:extensions:batch -- "$OPENCLAW_EXTENSION_BATCH"
 
diff --git a/docs/ci.md b/docs/ci.md
index 175f5a6ce38..84fbb21e4b3 100644
--- a/docs/ci.md
+++ b/docs/ci.md
@@ -53,7 +53,7 @@ Local changed-lane logic lives in `scripts/changed-lanes.mjs` and is executed by
 
 On pushes, the `checks` matrix adds the push-only `compat-node22` lane. On pull requests, that lane is skipped and the matrix stays focused on the normal test/channel lanes.
 
-The slowest Node test families are split or balanced so each job stays small: channel contracts split registry and core coverage into six weighted shards total, bundled plugin tests balance across six extension workers, each extension worker runs three config groups to use its 8-vCPU runner without adding machines, auto-reply runs as three balanced workers instead of six tiny workers, and agentic gateway/plugin configs are spread across the existing source-only agentic Node jobs instead of waiting on built artifacts. Broad browser, QA, media, and miscellaneous plugin tests use their dedicated Vitest configs instead of the shared plugin catch-all. The broad agents lane uses the shared Vitest file-parallel scheduler because it is import/scheduling dominated rather than owned by a single slow test file. `runtime-config` runs with the infra core-runtime shard to keep the shared runtime shard from owning the tail. `check-additional` keeps package-boundary compile/canary work together and separates it from runtime topology gateway/architecture work; the boundary guard shard runs its small independent guards concurrently inside one job, and the gateway watch regression uses the minimal `gatewayWatch` build profile instead of rebuilding the full CI artifact sidecar set.
+The slowest Node test families are split or balanced so each job stays small: channel contracts split registry and core coverage into six weighted shards total, bundled plugin tests balance across six extension workers, auto-reply runs as three balanced workers instead of six tiny workers, and agentic gateway/plugin configs are spread across the existing source-only agentic Node jobs instead of waiting on built artifacts. Broad browser, QA, media, and miscellaneous plugin tests use their dedicated Vitest configs instead of the shared plugin catch-all. The broad agents lane uses the shared Vitest file-parallel scheduler because it is import/scheduling dominated rather than owned by a single slow test file. `runtime-config` runs with the infra core-runtime shard to keep the shared runtime shard from owning the tail. `check-additional` keeps package-boundary compile/canary work together and separates it from runtime topology gateway/architecture work; the boundary guard shard runs its small independent guards concurrently inside one job, and the gateway watch regression uses the minimal `gatewayWatch` build profile instead of rebuilding the full CI artifact sidecar set.
 `extension-fast` is PR-only because push runs already execute the full bundled plugin shards. That keeps changed-plugin feedback for reviews without reserving an extra Blacksmith worker on `main` for coverage already present in `checks-node-extensions`.
 
 GitHub may mark superseded jobs as `cancelled` when a newer push lands on the same PR or `main` ref. Treat that as CI noise unless the newest run for the same ref is also failing. Aggregate shard checks use `!cancelled() && always()` so they still report normal shard failures but do not queue after the whole workflow has already been superseded.