refactor(ci): the timing-history read becomes the action #69 added

.github/workflows/sharded-tests.yml

@@ -281,52 +281,22 @@ jobs:
         with:
           submodules: ${{ inputs.submodules }}
 
-      # INLINE, and only until the next PR. #68 moved this into a file so that
-      # `read-history.test.sh` could put the SAME code in the states a consumer is in — the right
-      # idea, at a path that cannot work: a `run:` in a REUSABLE WORKFLOW executes in the CALLER's
-      # checkout, so `run: .github/scripts/read-history.sh` found the file in this repository's own
-      # CI and exited 127 on every consumer (ParaLinearAlgebra.jl run 30959798081: five jobs red,
-      # eight shards never started, six minutes after #68 merged).
+      # AN ACTION, not a `run:`, and the reason is where each one runs. #68 moved this into a
+      # file so `read-history.test.sh` could put the SAME code in the states a consumer is in —
+      # the no-history state cannot be reached from a run of this repository, which is how it
+      # shipped broken (#54) and surfaced on the first adopter. That part was right.
       #
-      # The fix is `actions/read-history`, added in this PR — an action's own files are fetched
-      # with it, so `${{ github.action_path }}` resolves for every caller. It cannot be USED here
-      # yet: `uses: …@main` resolves against main, where the action does not exist until this
-      # merges. So this step is the pre-#68 block again, which unbreaks every consumer on merge,
-      # and the switch to `uses:` is the immediate follow-up. The duplication is deliberate, is
-      # one PR long, and is the reason the follow-up is not optional.
-      - name: Read the timing history, once, for every shard
-        run: |
-          set -uo pipefail
-          # `ls-remote --exit-code` distinguishes the cases: 0 = branch exists, 2 = no matching
-          # ref, anything else = error. Measured.
-          #
-          # Captured through an `if`, and that is not style. A `run:` step's shell is
-          # `bash -e {0}`, so errexit is ON from the command line and `set -uo pipefail` does not
-          # take it off — it only ADDS options. A bare `out=$(...); rc=$?` therefore kills the
-          # step the moment ls-remote exits non-zero, BEFORE `rc` is read: a repository with no
-          # ci-timings branch yet died with exit 2 and no output, which is the one case this
-          # code was written to handle. An assignment inside an `if` condition is exempt from
-          # errexit however the shell was invoked.
-          if out=$(git ls-remote --exit-code --heads origin ci-timings 2>&1); then rc=0; else rc=$?; fi
-          if [ "$rc" -eq 0 ]; then
-            if ! git fetch --depth=1 origin ci-timings; then
-              echo "::error::The ci-timings branch exists but could not be fetched. Failing here, before any shard starts, rather than letting the shards disagree about the split."
-              exit 1
-            fi
-            if git show FETCH_HEAD:timings.tsv > timings.tsv 2>/dev/null; then
-              echo "$(wc -l < timings.tsv) timing rows."
-            else
-              echo "ci-timings carries no timings.tsv yet — round-robin."
-              : > timings.tsv
-            fi
-          elif [ "$rc" -eq 2 ]; then
-            echo "No ci-timings branch yet — round-robin until the first push to the default branch."
-            : > timings.tsv
-          else
-            echo "::error::Could not tell whether the ci-timings branch exists (git ls-remote exit $rc). Failing rather than guessing: guessing wrong puts the shards on different assignments, which double-runs some units and skips others while every shard stays green."
-            echo "$out"
-            exit 1
-          fi
+      # What could not work is a `run:`: a `run:` step in a REUSABLE WORKFLOW executes in the
+      # CALLER's checkout, so `run: .github/scripts/read-history.sh` found the file here and
+      # exited 127 on every consumer (#69: ParaLinearAlgebra.jl run 30959798081, five jobs red,
+      # eight shards never started). An action's own files are fetched WITH it, so
+      # `${{ github.action_path }}` resolves for every caller and nothing lands in the consumer's
+      # workspace — which a second `actions/checkout` of this repository would not manage, since
+      # its `src/*.jl` would then be inside the consumer's coverage sweep.
+      #
+      # Same split as `actions/upload-coverage`: the workflow owns what needs JOBS, an action owns
+      # what has to run from this repository's own tree.
+      - uses: QAtlasHub/TestShards.jl/actions/read-history@main
 
       - uses: actions/upload-artifact@v7
         with:

Project.toml

@@ -1,6 +1,6 @@
 name = "TestShards"
 uuid = "acceef1d-f5e0-4fe4-a546-818dc56ce7b2"
-version = "0.3.32"
+version = "0.3.33"
 authors = ["sota shimozono <shimozono-sota631@g.ecc.u-tokyo.ac.jp>"]
 
 [deps]