githubnext · mrjf · Apr 23, 2026 · Apr 23, 2026 · Apr 23, 2026
diff --git a/.autoloop/programs/tsb-perf-evolve/evaluate.sh b/.autoloop/programs/tsb-perf-evolve/evaluate.sh
@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+# Evaluator for the tsb-perf-evolve OpenEvolve program.
+#
+# Both the autoloop agent (Step 6 of the OpenEvolve playbook) and CI (the
+# `benchmark` job in .github/workflows/ci.yml) invoke this script so they
+# produce comparable fitness numbers from identical commands.
+#
+# Output: a single JSON line on stdout with one of these shapes
+#   {"fitness": <number>, "tsb_mean_ms": <number>, "pandas_mean_ms": <number>}
+#   {"fitness": null,     "rejected_reason": "<string>"}
+#
+# Exit code is always 0 — failures are encoded in the JSON so callers can
+# parse the result uniformly. Diagnostics go to stderr.
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"
+
+cd "$REPO_ROOT"
+
+# 1. Validity — existing tests for sortValues must still pass.
+if ! bun test tests/core/series.sortValues.test.ts >/tmp/perf-evolve-tests.log 2>&1; then
+  echo '{"fitness": null, "rejected_reason": "tests failed"}'
+  exit 0
+fi
+
+# 2. Benchmark — tsb side.
+tsb_ms=$(bun run "$SCRIPT_DIR/code/benchmark.ts" \
+  | python3 -c "import json,sys; print(json.load(sys.stdin)['mean_ms'])")
+
+# 3. Benchmark — pandas side. Skip gracefully if pandas isn't available.
+if ! python3 -c 'import pandas' 2>/dev/null; then
+  pip3 install pandas --quiet 2>/dev/null || true
+fi
+pd_ms=$(python3 "$SCRIPT_DIR/code/benchmark.py" \
+  | python3 -c "import json,sys; print(json.load(sys.stdin)['mean_ms'])")
+
+# 4. Fitness = ratio. Lower is better.
+ratio=$(python3 -c "print(${tsb_ms} / ${pd_ms})")
+echo "{\"fitness\": ${ratio}, \"tsb_mean_ms\": ${tsb_ms}, \"pandas_mean_ms\": ${pd_ms}}"
diff --git a/.autoloop/programs/tsb-perf-evolve/program.md b/.autoloop/programs/tsb-perf-evolve/program.md
@@ -55,26 +55,25 @@ Population state lives in the state file on the `memory/autoloop` branch under t
 ## Evaluation
 
 ```bash
-set -euo pipefail
-
-# 1. Validity — existing tests for sortValues must still pass.
-bun test tests/core/series.sortValues.test.ts >/tmp/perf-evolve-tests.log 2>&1 || {
-  echo '{"fitness": null, "rejected_reason": "tests failed"}'
-  exit 0
-}
-
-# 2. Benchmark — tsb side.
-tsb_ms=$(bun run .autoloop/programs/tsb-perf-evolve/code/benchmark.ts | python3 -c "import json,sys; print(json.load(sys.stdin)['mean_ms'])")
-
-# 3. Benchmark — pandas side. Skip gracefully if pandas isn't available.
-if ! python3 -c 'import pandas' 2>/dev/null; then
-  pip3 install pandas --quiet 2>/dev/null || true
-fi
-pd_ms=$(python3 .autoloop/programs/tsb-perf-evolve/code/benchmark.py | python3 -c "import json,sys; print(json.load(sys.stdin)['mean_ms'])")
-
-# 4. Fitness = ratio. Lower is better.
-ratio=$(python3 -c "print(${tsb_ms} / ${pd_ms})")
-echo "{\"fitness\": ${ratio}, \"tsb_mean_ms\": ${tsb_ms}, \"pandas_mean_ms\": ${pd_ms}}"
+bash .autoloop/programs/tsb-perf-evolve/evaluate.sh
+```
+
+The actual evaluator lives in `evaluate.sh` next to this file so the autoloop
+agent (Step 6 of the OpenEvolve playbook) and CI (the `benchmark` job in
+`.github/workflows/ci.yml`) invoke the **exact same** command and produce
+comparable fitness numbers. See that script for details.
+
+It runs the validity tests, then the tsb and pandas benchmarks, and prints a
+single JSON line on stdout:
+
+```json
+{"fitness": <number>, "tsb_mean_ms": <number>, "pandas_mean_ms": <number>}
+```
+
+or, if validity failed:
+
+```json
+{"fitness": null, "rejected_reason": "tests failed"}
 ```
 
 The metric is `fitness` (= `tsb_mean_ms / pandas_mean_ms`). **Lower is better.** A value below `1.0` means tsb is now faster than pandas on this workload.
diff --git a/.autoloop/strategies/openevolve/strategy.md b/.autoloop/strategies/openevolve/strategy.md
@@ -75,6 +75,51 @@ Edit only the files listed in `program.md`'s Target section. The diff style for
 
 Run the evaluation command from `program.md`. Parse the metric.
 
+The in-sandbox evaluation is a *cheap pre-filter only* — the agent sandbox often cannot install `bun`, run `python3 -c 'import pandas'`, or otherwise reproduce realistic conditions (the `releaseassets.githubusercontent.com` firewall block is the common culprit). A null/missing metric here is **not** grounds for rejecting the candidate; that decision is deferred to Step 6.5.
+
+### Step 6.5. Wait for CI
+
+Before recording the candidate in the population (Step 7) or posting *any* iteration comment on the program issue / PR, wait for CI on the pushed commit. CI is the authoritative source of both correctness (Test & Lint / Build / Validate Python Examples) and fitness (the `OpenEvolve benchmark` check, which runs `bash .autoloop/programs/{program-name}/evaluate.sh` on a real runner with `bun` + `python3` + `pandas` installed).
+
+This step extends — and ties into — the generic `Step 5a → 5b → 5c` flow described in the autoloop workflow. OpenEvolve's only added requirement is that you must reach Step 5c (or the budget-exhausted handler) **before** writing the iteration comment, never after a speculative push.
+
+```bash
+# Resolve the PR — prefer the pre-step lookup, fall back to gh.
+PR=$(jq -r '.existing_pr // empty' /tmp/gh-aw/autoloop.json 2>/dev/null || true)
+if [ -z "$PR" ]; then
+  PR=$(gh pr list --head autoloop/{program-name} --json number -q '.[0].number')
+fi
+
+# Block until every required check terminates (or the wall-clock cap fires).
+gh pr checks "$PR" --watch --interval 30 --fail-fast || true
+
+# Determine an aggregate status. Same awk classifier as Step 5a in the
+# generic autoloop playbook — keep them in sync.
+status=$(gh pr checks "$PR" --json conclusion,state \
+           -q '.[] | (.conclusion // .state // "")' \
+         | awk '
+             BEGIN { r = "success" }
+             /^(FAILURE|CANCELLED|TIMED_OUT|ACTION_REQUIRED|STARTUP_FAILURE|STALE)$/ { r = "failure" }
+             /^(PENDING|QUEUED|IN_PROGRESS|WAITING|REQUESTED)$/ { if (r == "success") r = "pending" }
+             END { print r }')
+
+# Read the fitness from the OpenEvolve benchmark check-run (created by the
+# `benchmark` job in .github/workflows/ci.yml). Title format: `fitness=<num>`
+# or `fitness=null`. SHA = the HEAD of the PR after the latest push/fix.
+SHA=$(gh pr view "$PR" --json headRefOid -q '.headRefOid')
+fitness=$(gh api "repos/${GITHUB_REPOSITORY}/commits/${SHA}/check-runs" \
+           --jq '.check_runs[] | select(.name == "OpenEvolve benchmark") | .output.title' \
+         | sed -n 's/^fitness=//p' | head -n1)
+```
+
+Branch on `$status`:
+
+- **`success`** → record the candidate in the population with `fitness: <number>` from the check-run (or `fitness: null` only if the `OpenEvolve benchmark` check explicitly reported it that way — e.g., correctness held but the benchmark itself errored). Proceed to Step 7. The iteration comment is `✅ Accepted` with the real numeric fitness.
+- **`failure`** → enter the fix-retry loop from the generic autoloop Step 5b (up to 5 attempts, no-progress guard, 60-min wall-clock cap). Do **not** post an "accepted" comment. On a successful fix, loop back through the `gh pr checks --watch` block above on the new HEAD. On exhausted budget, mark the candidate `status: error` in the population with `fitness: null` and `pause_reason: "ci-fix-exhausted: <signature>"`, and post a `❌ Rejected` (or `⚠️ Error`) iteration comment that links to the failing run.
+- **`pending`** (the wall-clock cap fired before CI concluded) → don't post a speculative `⏳ Pending CI` comment. Record the candidate in the population with `fitness: null` and `status: pending-ci`, and leave a single reconciliation-pending comment on the PR/issue that the next iteration's Step 6.5 is allowed to overwrite when it reads the now-concluded status for this same SHA.
+
+In all three branches, the iteration comment posted to the program issue and PR must reflect *terminal* state — never `⏳ Pending CI` as a permanent label. Comments live forever; the pending placeholder is what produced the bug this step exists to fix.
+
 ### Step 7. Update the population
 
 Regardless of whether the iteration is accepted or rejected at the branch level, the candidate has been tried and should be recorded in the population — the population is a memory of what's been explored, not just what's been kept.
@@ -88,7 +133,8 @@ Append a new entry to the `## 🧬 Population` subsection in the state file usin
 
 Continue with the normal autoloop Step 5 (Accept or Reject → commit / discard, update state file's Machine State, Iteration History, Lessons Learned, etc.) as defined in the workflow. The only additional requirements from OpenEvolve are:
 
-- The Iteration History entry must include `operator`, `parent_id(s)`, `island`, and `fitness` fields (in addition to the normal status/change/metric/notes).
+- The Iteration History entry must include `operator`, `parent_id(s)`, `island`, and `fitness` fields (in addition to the normal status/change/metric/notes). The `fitness` value comes from the `OpenEvolve benchmark` check-run resolved in Step 6.5 — never from the in-sandbox Step 6 estimate.
+- The iteration comment posted to the program issue and PR must use the terminal status from Step 6.5 (`✅ Accepted` / `❌ Rejected` / `⚠️ Error` / `⏸ Pending-CI` only when the wall-clock cap genuinely fired). Never post `⏳ Pending CI` as a final state — that placeholder is what Step 6.5 exists to eliminate.
 - Lessons Learned additions should be phrased as *transferable heuristics* about the problem space, not as reports of what this iteration did. (E.g. "Hex layouts dominate grid layouts above n=20" — not "Iteration 17 tried a hex layout.")
 
 ## Feature dimensions

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -11,6 +11,7 @@ on:
 
 permissions:
   contents: read
+  checks: write
 
 jobs:
   test:
@@ -76,3 +77,111 @@ jobs:
 
       - name: Validate Python playground examples
         run: python scripts/validate-python-examples.py playground/
+
+  benchmark:
+    # Run the OpenEvolve benchmark for autoloop *-evolve PRs so the autoloop
+    # agent can read a real fitness number from CI (see .autoloop/strategies/
+    # openevolve/strategy.md, Step 6.5). The sandbox the agent runs in cannot
+    # install bun reliably and so cannot measure fitness itself.
+    name: OpenEvolve benchmark
+    if: |
+      (github.event_name == 'pull_request' && startsWith(github.head_ref, 'autoloop/') && contains(github.head_ref, '-evolve'))
+      || (github.event_name == 'push' && startsWith(github.ref_name, 'autoloop/') && contains(github.ref_name, '-evolve'))
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      checks: write
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: latest
+
+      - name: Install dependencies
+        run: bun install
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install Python dependencies
+        run: pip install pandas numpy
+
+      - name: Resolve program directory
+        id: program
+        run: |
+          # Resolve the program directory from the branch name:
+          #   autoloop/<program-name>  →  .autoloop/programs/<program-name>/
+          BRANCH="${GITHUB_HEAD_REF:-${GITHUB_REF_NAME}}"
+          PROGRAM="${BRANCH#autoloop/}"
+          PROGRAM_DIR=".autoloop/programs/${PROGRAM}"
+          echo "program=${PROGRAM}" >> "$GITHUB_OUTPUT"
+          echo "program_dir=${PROGRAM_DIR}" >> "$GITHUB_OUTPUT"
+          if [ -x "${PROGRAM_DIR}/evaluate.sh" ]; then
+            echo "has_evaluator=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "No evaluate.sh for program '${PROGRAM}' — skipping benchmark." >&2
+            echo "has_evaluator=false" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Run OpenEvolve benchmark
+        id: bench
+        if: steps.program.outputs.has_evaluator == 'true'
+        run: |
+          PROGRAM_DIR="${{ steps.program.outputs.program_dir }}"
+          # evaluate.sh is contracted to always exit 0 and encode failures in
+          # the JSON, but we tolerate non-zero exits anyway and fall back to a
+          # null fitness so the check-run still gets created.
+          set +e
+          bash "${PROGRAM_DIR}/evaluate.sh" >/tmp/bench-result.json 2>/tmp/bench-stderr
+          rc=$?
+          set -e
+          if [ ! -s /tmp/bench-result.json ]; then
+            echo "{\"fitness\": null, \"rejected_reason\": \"evaluator produced no output (exit ${rc})\"}" \
+              > /tmp/bench-result.json
+          fi
+          cat /tmp/bench-result.json
+          fitness=$(jq -r '.fitness // "null"' /tmp/bench-result.json)
+          echo "fitness=${fitness}" >> "$GITHUB_OUTPUT"
+          # Compact JSON for the check-run output below.
+          echo "result_json=$(jq -c . /tmp/bench-result.json)" >> "$GITHUB_OUTPUT"
+
+      - name: Upload benchmark result
+        if: steps.program.outputs.has_evaluator == 'true'
+        uses: actions/upload-artifact@v4
+        with:
+          name: benchmark-result
+          path: /tmp/bench-result.json
+
+      - name: Attach fitness as check-run
+        if: steps.program.outputs.has_evaluator == 'true'
+        uses: actions/github-script@v7
+        env:
+          FITNESS: ${{ steps.bench.outputs.fitness }}
+          RESULT_JSON: ${{ steps.bench.outputs.result_json }}
+        with:
+          script: |
+            const fitness = process.env.FITNESS;
+            let result;
+            try {
+              result = JSON.parse(process.env.RESULT_JSON);
+            } catch {
+              result = { raw: process.env.RESULT_JSON };
+            }
+            const sha = context.payload.pull_request
+              ? context.payload.pull_request.head.sha
+              : context.sha;
+            await github.rest.checks.create({
+              ...context.repo,
+              name: "OpenEvolve benchmark",
+              head_sha: sha,
+              status: "completed",
+              conclusion: fitness === "null" ? "neutral" : "success",
+              output: {
+                title: `fitness=${fitness}`,
+                summary: "```json\n" + JSON.stringify(result, null, 2) + "\n```",
+              },
+            });