# GitNexus skill evolution: runs the offline propose → benchmark → gate loop
# (eval/workflow_bench/evolve.py) on a schedule and, when the deterministic
# promotion gate passes, opens a human-reviewed PR with the promoted skill
# overlay. The gate is evidence FOR a PR, never a bypass of one — nothing
# merges without review.
#
# Activation and operations checklist.
# [x] Configure at least one model secret on the `gitnexus-evolution`
#     Environment: GITNEXUS_BENCH_ANTHROPIC_API_KEY (Anthropic API key — not
#     the Claude Code OAuth token; legacy GITNEXUS_BENCH_AUTH_TOKEN is still
#     accepted) and/or GITNEXUS_BENCH_OPENAI_API_KEY. Sessions bill real usage.
#     OpenAI keys are not native to Claude Code; the loop starts a loopback
#     LiteLLM proxy and keeps the OpenAI key off the sandboxed agent. With only
#     the OpenAI secret, or with provider=openai, dispatch-time Claude model
#     defaults are gpt-5.6-sol with xhigh reasoning effort.
# [x] Configure the RELEASE_APP_ID and RELEASE_APP_PRIVATE_KEY secrets (the
#     App that opens the promotion PR). The Mint-App-Token step hard-fails
#     without them once a promotion is detected. Verify the App installation
#     is scoped to this repo with only Contents: RW + Pull requests: RW.
# [x] Create the protected Environment `gitnexus-evolution` with a
#     deployment-branch rule restricting it to `main`, and ideally scope the
#     four secrets above to that Environment. workflow_dispatch runs this
#     workflow (and eval/workflow_bench/evolve.py) from the *dispatched ref*,
#     so this server-side rule — not a code-side guard the branch could edit
#     away — is what stops a non-main branch from running with the secrets.
# [x] Register a self-hosted runner labeled `gitnexus-evolution` (a dedicated
#     EC2 box works well). GitHub-hosted runners hard-cap job execution at 6
#     hours, non-configurable — too short once a benchmark session actually
#     invokes Skill/MCP tools for real. Self-hosted runners cap at 5 days
#     instead. This job only ever runs on schedule/workflow_dispatch, never
#     on fork-PR content, so the usual public-repo self-hosted-runner risk
#     doesn't apply — still keep the box dedicated to this workflow, with
#     outbound-only network access, and prefer on-demand over Spot (a Spot
#     reclaim mid-run loses the same way a 6-hour timeout does). Instance,
#     security group, and IAM setup are documented privately, not in this
#     repo — publishing the exact topology of a real, live AWS account
#     isn't safe to do in a public repo even without literal secrets.
#     Accepted tradeoff: the box is stopped between runs (an EventBridge
#     schedule starts it ~15min before the Saturday cron and stops it 24h
#     later) but is not destroyed/recreated per run, so it isn't fully
#     ephemeral — a compromise between the review-flagged ideal (re-image
#     between runs, bounding how long the injected model API key could
#     matter if the box were ever compromised some other way) and the added
#     complexity of per-job ephemeral provisioning for a job that runs at
#     most weekly. Revisit if run frequency increases or the threat model
#     changes; stopping already bounds the exposure window to the job's own
#     runtime on 1 day out of 7.
# [ ] Install and verify the runner survival policy below before enabling
#     scheduled runs. A run
#     spans ~15h and apt-daily-upgrade.timer fires daily (~06:34), so every
#     scheduled run crosses it. On 2026-08-02 unattended-upgrades upgraded
#     openssl at 07:54:02 and needrestart restarted the Actions runner five
#     seconds later: the job went to Canceled, and a cancelled job skips even
#     `if: always()`, so the evidence artifact died with it. Keep installing
#     updates, but never let them restart services here:
#         /etc/needrestart/conf.d/90-gitnexus-evolution.conf
#         $nrconf{restart} = 'l';
#     A drop-in, so a needrestart package upgrade cannot clobber it. Nothing
#     is left unpatched in practice — the box is stopped between runs, so the
#     new binaries take effect at the next boot.
# [x] Run workflow_dispatch once and confirm: containment preflight passes,
#     the benchmark completes inside the job timeout, the results artifact
#     uploads, and a promotion (if any) opens a well-formed PR. Run
#     29907431284 (2026-07-22) went green end to end in 14h45m and reached a
#     gate decision (`insufficient_evidence`, no promotion).
# [ ] Confirm a workers=3 dispatch has zero excluded runs (review sessions in
#     33962002890 averaged ~19m serial, well under the 90m session ceiling).
#     Then set GITNEXUS_EVOLUTION_WORKERS=3 and
#     GITNEXUS_EVOLUTION_ENABLED=true for scheduled runs. Scheduled runs
#     require both values, so leaving the var unset is an immediate rollback.
#     Dispatch defaults to 3; pass workers=1 only to debug a contended host.
#     Weekly generations reuse matching incumbent/CE cells from the previous
#     artifact so the paid matrix is the new candidate, not a 54-cell replay.
#     Wall clock is quantised by ceil(cells_per_task / workers), and a review
#     task is 9 cells cold, so 4 costs host contention for exactly the wall
#     clock of 3. The next step up that buys anything is 5 (3 waves -> 2).
name: GitNexus skill evolution

on:
  schedule:
    # Weekly is a deliberate cadence to catch model/harness drift promptly; a
    # no-promotion week only costs one benchmark run (the gate keeps the
    # incumbent unless quality improves). Dial back toward the README's ~90-day
    # re-evaluation guidance if the recurring spend is not worth it.
    - cron: '0 3 * * 6' # weekly, Saturday 03:00 UTC
  workflow_dispatch:
    inputs:
      generations:
        description: 'Propose→bench→gate generations to run'
        required: false
        default: '1'
        type: string
      runs:
        description: 'Runs per arm per task (the gate needs at least 3)'
        required: false
        default: '3'
        type: string
      workers:
        description: 'Benchmark cells of one task to run at once — 3 fits the evolution box; drop to 1 only if siblings hit the session ceiling'
        required: false
        default: '3'
        type: string
      model:
        description: 'Model for the benchmark arms (match the model your skill users run)'
        required: false
        default: 'gpt-5.6-sol'
        type: string
      proposer_model:
        description: 'Model for the proposer/diagnosis session — a stronger model is fine (one session per generation)'
        required: false
        default: 'gpt-5.6-sol'
        type: string
      effort:
        description: 'Reasoning effort for every proposer and benchmark session'
        required: false
        default: xhigh
        type: choice
        options:
          - low
          - medium
          - high
          - xhigh
          - max
      provider:
        description: 'Model backend. auto uses Anthropic when that secret exists; openai forces the loopback OpenAI gateway even if an Anthropic key is also configured.'
        required: false
        default: openai
        type: choice
        options:
          - auto
          - openai
          - anthropic
      include_expensive:
        description: 'Include tasks marked expensive: true'
        required: false
        default: false
        type: boolean
      seed_from_previous:
        description: "Seed the proposer with the previous run's evidence and rejected proposal. Turn off to start from a blank slate — required when the earlier evidence is not trustworthy (e.g. produced before a harness-integrity fix), since a tainted proposal would otherwise propagate into every later generation."
        required: false
        default: true
        type: boolean

concurrency:
  group: ${{ github.workflow }}
  cancel-in-progress: false

permissions: {}

jobs:
  evolve:
    name: Propose, benchmark, and gate skill candidates
    if: >-
      github.repository == 'abhigyanpatwari/GitNexus' &&
      (
        github.event_name == 'workflow_dispatch' ||
        (
          vars.GITNEXUS_EVOLUTION_ENABLED == 'true' &&
          vars.GITNEXUS_EVOLUTION_WORKERS == '3'
        )
      )
    runs-on: [self-hosted, linux, x64, gitnexus-evolution]
    # Gate promotion runs on a protected Environment. An admin must attach a
    # deployment-branch rule (main only) and ideally scope the model and App
    # secrets to it — server-side enforcement a dispatched non-main ref cannot bypass by
    # editing its own workflow copy. See the activation checklist above.
    environment: gitnexus-evolution
    # Three budgets have to nest, longest first, or the evidence is lost:
    #   EventBridge instance uptime (24h from ~02:45)
    #     > this job timeout (21h)
    #       > the benchmark step timeout (19h, set on the step below)
    # A job-level timeout CANCELS the job, so the upload step never runs and a
    # multi-hour generation's evidence dies with it; a step-level timeout only
    # fails that step, and `if: always()` still uploads what the sweep wrote.
    # The instance must outlive the job for the same reason — when the box
    # stops the runner just disappears mid-step. Scheduled runs can start well
    # after the cron (the 2026-08-01 run was queued 65min late), so the job
    # budget has to absorb that delay and still land inside the uptime window.
    # A Friday workflow_dispatch on a box that already booted for Saturday's
    # cron inherits leftover uptime, not a fresh 24h. Run 33962002890 started
    # Friday 10:57 UTC and vanished at the Saturday 03:00 stop — 51 finished
    # sessions never uploaded. run-evolution.sh therefore passes
    # --max-runtime-from-instance-window, and the CLI derives its cap from
    # /proc/uptime at startup, so the sweep fails in-process and this always()
    # upload still runs.
    timeout-minutes: 1260
    permissions:
      contents: read # The promotion PR uses a short-lived App token minted below.
      actions: read # Read the previous run's evidence artifact to seed the proposer.
    env:
      GENERATIONS: ${{ inputs.generations || '1' }}
      RUNS: ${{ inputs.runs || '3' }}
      # A manual input wins; scheduled runs use the repository rollout knob.
      # Both fall back to serial — see workflow_bench.runner --workers for why.
      WORKERS: ${{ inputs.workers || vars.GITNEXUS_EVOLUTION_WORKERS || '1' }}
      MODEL: ${{ inputs.model || 'gpt-5.6-sol' }}
      PROPOSER_MODEL: ${{ inputs.proposer_model || 'gpt-5.6-sol' }}
      EFFORT: ${{ inputs.effort || 'xhigh' }}
      PROVIDER: ${{ inputs.provider || 'openai' }}
      INCLUDE_EXPENSIVE: ${{ inputs.include_expensive && '1' || '' }}
    steps:
      - name: Require the benchmark auth secret
        env:
          HAS_ANTHROPIC: ${{ secrets.GITNEXUS_BENCH_ANTHROPIC_API_KEY != '' || secrets.GITNEXUS_BENCH_AUTH_TOKEN != '' }}
          HAS_OPENAI: ${{ secrets.GITNEXUS_BENCH_OPENAI_API_KEY != '' }}
        run: |
          set -euo pipefail
          if [[ "${HAS_ANTHROPIC}" != 'true' && "${HAS_OPENAI}" != 'true' ]]; then
            echo '::error::Configure GITNEXUS_BENCH_ANTHROPIC_API_KEY (Anthropic API key, not the Claude Code OAuth token) and/or GITNEXUS_BENCH_OPENAI_API_KEY. The evolution loop runs real benchmark sessions.'
            exit 1
          fi
          case "${PROVIDER}" in
            openai)
              if [[ "${HAS_OPENAI}" != 'true' ]]; then
                echo '::error::provider=openai requires GITNEXUS_BENCH_OPENAI_API_KEY on the gitnexus-evolution environment.'
                exit 1
              fi
              ;;
            anthropic)
              if [[ "${HAS_ANTHROPIC}" != 'true' ]]; then
                echo '::error::provider=anthropic requires GITNEXUS_BENCH_ANTHROPIC_API_KEY on the gitnexus-evolution environment.'
                exit 1
              fi
              ;;
            auto)
              ;;
            *)
              echo "::error::Unknown provider '${PROVIDER}' (expected auto, openai, or anthropic)."
              exit 1
              ;;
          esac

      - name: Verify runner survival policy
        run: |
          set -euo pipefail
          needrestart_policy=/etc/needrestart/conf.d/90-gitnexus-evolution.conf
          needrestart_line="\$nrconf{restart} = 'l';"
          if [[ ! -r "${needrestart_policy}" ]] || ! grep -Fqx "${needrestart_line}" "${needrestart_policy}"; then
            echo "::error::${needrestart_policy} must contain: ${needrestart_line}"
            exit 1
          fi
          # The runner sets job processes to 500; the host oom-guard rewrites
          # them to -900. Read once and the check loses that race.
          oom_score_adjustment="$(</proc/self/oom_score_adj)"
          deadline=$((SECONDS + 5))
          while (( oom_score_adjustment > -900 && SECONDS < deadline )); do
            sleep 0.05
            oom_score_adjustment="$(</proc/self/oom_score_adj)"
          done
          if (( oom_score_adjustment > -900 )); then
            echo "::error::Runner.Worker descendants require OOMScoreAdjust=-900 or stronger; effective value is ${oom_score_adjustment}."
            exit 1
          fi

      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
        with:
          persist-credentials: false
          fetch-depth: 0

      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
        with:
          node-version: '22.18.0'
          cache: npm
          cache-dependency-path: |
            gitnexus/package-lock.json
            gitnexus-shared/package-lock.json

      - uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990 # v8.3.2
        with:
          version: '0.11.23'
          python-version: '3.13'
          enable-cache: true
          cache-dependency-glob: eval/uv.lock

      - name: Fetch pinned Compound Engineering review comparator
        env:
          CE_COMMIT: 3ad9b51bceecf0158e590c882034d0398dbb9c5c
        run: |
          set -euo pipefail
          destination="${RUNNER_TEMP}/compound-engineering-plugin"
          rm -rf "${destination}"
          git clone --filter=blob:none --no-checkout \
            https://github.com/EveryInc/compound-engineering-plugin.git "${destination}"
          git -C "${destination}" checkout --detach "${CE_COMMIT}"
          test "$(git -C "${destination}" rev-parse HEAD)" = "${CE_COMMIT}"

      - name: Install sandbox runtime and pinned Claude CLI
        run: |
          set -euo pipefail
          # This box is stopped six days a week, so persistent apt timers can
          # begin their catch-up run shortly after boot. Wait for dpkg instead
          # of racing the same package lock and failing the weekly lane.
          sudo apt-get -o DPkg::Lock::Timeout=600 update
          sudo apt-get -o DPkg::Lock::Timeout=600 install --yes --no-install-recommends bubblewrap ripgrep socat
          apparmor_userns=/proc/sys/kernel/apparmor_restrict_unprivileged_userns
          if [[ -r "${apparmor_userns}" ]] && [[ "$(<"${apparmor_userns}")" == '1' ]]; then
            sudo sysctl -w kernel.apparmor_restrict_unprivileged_userns=0
          fi
          canary_runtime="${RUNNER_TEMP}/claude-canary"
          install -d -m 0700 "${canary_runtime}"
          install -m 0600 \
            .github/claude-canary-runtime/package.json \
            "${canary_runtime}/package.json"
          install -m 0600 \
            .github/claude-canary-runtime/package-lock.json \
            "${canary_runtime}/package-lock.json"
          npm ci \
            --prefix "${canary_runtime}" \
            --ignore-scripts=false \
            --audit=false \
            --fund=false
          node -e \
            "const p=require(process.argv[1]); if(p.version!=='2.1.214') process.exit(1)" \
            "${canary_runtime}/node_modules/@anthropic-ai/claude-code/package.json"
          test "$("${canary_runtime}/node_modules/@anthropic-ai/claude-code-linux-x64/claude" --version)" = \
            '2.1.214 (Claude Code)'

      - name: Verify contained review execution before paid sessions
        working-directory: eval
        env:
          GITNEXUS_REQUIRE_BWRAP_CANARY: '1'
          GITNEXUS_REQUIRE_CLAUDE_CANARY: '1'
          CLAUDE_CANARY_BIN: ${{ runner.temp }}/claude-canary/node_modules/@anthropic-ai/claude-code-linux-x64/claude
        run: |
          set -euo pipefail
          uv run --locked --extra dev python -m pytest tests/test_proposer_sandbox.py -q

      - name: Install monorepo root dependencies
        run: |
          set -euo pipefail
          # The benchmark's task bindings sandbox-copy node_modules from the
          # monorepo root as well as gitnexus-shared and gitnexus (see the
          # sandbox_copy entries in tasks.scenarios.yaml). The two steps below
          # install the subpackage trees; the root tree needs its own install
          # or capture_task_dependency_binding aborts at task binding on the
          # missing root node_modules.
          npm ci

      - name: Build pinned shared runtime
        run: |
          set -euo pipefail
          npm ci
          npm run build
        working-directory: gitnexus-shared

      - name: Install and build pinned GitNexus runtime
        run: |
          set -euo pipefail
          npm ci
          npm run build
        working-directory: gitnexus

      - name: Point the benchmark task repo at the checkout
        run: |
          set -euo pipefail
          # tasks.review.scenarios.yaml addresses the target repo as ~/GitNexus (the
          # developer-local convention). On the runner the repo is the checkout
          # at ${GITHUB_WORKSPACE}; link it so runner_tasks.py can resolve the
          # task `repo` path. The benchmark only clones the repo (copy-on-write)
          # and mounts dependencies read-only, so the checkout is never mutated.
          if [[ -e "${HOME}/GitNexus" && ! -L "${HOME}/GitNexus" ]]; then
            echo '::error::~/GitNexus exists and is not a symlink; refusing to place the checkout inside it.'
            exit 1
          fi
          ln -sfn "${GITHUB_WORKSPACE}" "${HOME}/GitNexus"
          # The review corpus pins historical object ids. Fetch main so those
          # objects are present even when actions/checkout selected another ref.
          git -C "${GITHUB_WORKSPACE}" fetch --no-tags --quiet \
            "${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}.git" \
            '+refs/heads/main:refs/remotes/origin/main'
          baseline_sha="$(git -C "${GITHUB_WORKSPACE}" rev-parse --verify 'refs/remotes/origin/main^{commit}')"
          echo "Fetched review corpus history at ${baseline_sha}"

      - name: Seed the proposer with the previous run's evidence
        id: seed
        # Scheduled runs always seed; a dispatch can opt out to start clean.
        if: github.event_name != 'workflow_dispatch' || inputs.seed_from_previous
        # Best-effort seeding must not consume the benchmark's budget. This
        # step walks up to 10 prior runs and every iteration blocks on network
        # it does not control (`gh run download` of a multi-hundred-megabyte
        # artifact). Unbounded, a wedged download sits here until the 21h job
        # timeout CANCELS the job — and a cancelled job skips even
        # `if: always()`, so the sweep never starts and nothing is uploaded.
        # Bounding the step instead fails it in minutes, which is a loud,
        # cheap, re-runnable failure rather than a silent 21h loss. 15 minutes
        # is an order of magnitude above the observed walk (well under a
        # minute) and a rounding error against the 19h sweep it protects.
        timeout-minutes: 15
        env:
          GH_TOKEN: ${{ github.token }}
        run: |
          set -euo pipefail
          # Without this the weekly lane is memoryless: `--seed-results` is the
          # only way a run sees what already lost (evolve stages the prior
          # proposal when present and summarizes promotion.json when present),
          # and with the default --generations 1 there is no earlier generation
          # in-process to supply it. Every Saturday would otherwise propose
          # from a blank slate and could re-propose the same rejected candidate
          # forever. Best-effort by design: a first run, an expired artifact,
          # or a download failure must not cost a whole generation.
          if ! command -v gh >/dev/null; then
            echo '::warning::gh is not installed on this runner — proposing without prior evidence. The promotion-PR step needs gh too.'
            exit 0
          fi
          if ! previous_runs="$(gh run list \
            --repo "${GITHUB_REPOSITORY}" \
            --workflow gitnexus-skill-evolution.yml \
            --branch main \
            --status completed \
            --limit 10 \
            --json databaseId \
            --jq "map(.databaseId) | map(select(. != ${GITHUB_RUN_ID})) | .[]")"; then
            echo '::warning::Prior workflow runs could not be listed; proposing without prior evidence.'
            exit 0
          fi
          if [[ -z "${previous_runs}" ]]; then
            echo 'No prior completed run to seed from; the proposer starts from the learnings queue only.'
            exit 0
          fi
          seed_root="${RUNNER_TEMP}/wfseed"
          rm -rf "${seed_root}"
          install -d -m 0700 "${seed_root}"
          seed=''
          # Failed sweeps deliberately upload partial evidence, so "completed"
          # is the right population. Walk newest-first until one still-retained
          # artifact actually contains benchmark rows; an empty latest run must
          # not hide an older useful one.
          for previous in ${previous_runs}; do
            if [[ ! "${previous}" =~ ^[0-9]+$ ]]; then
              echo "::warning::Ignoring malformed prior run id: ${previous}"
              continue
            fi
            run_root="${seed_root}/${previous}"
            install -d -m 0700 "${run_root}"
            if ! gh run download "${previous}" --repo "${GITHUB_REPOSITORY}" --dir "${run_root}"; then
              echo "::warning::Evidence from run ${previous} could not be downloaded (expired or absent); trying an older run."
              continue
            fi
            unsafe="$(find "${run_root}" ! -type d ! -type f -print -quit)"
            if [[ -n "${unsafe}" ]]; then
              echo "::warning::Run ${previous} contains a non-regular artifact entry; trying an older run."
              continue
            fi
            # upload-artifact normalizes directories/files to 0755/0644, while
            # the evidence reader deliberately requires transcript paths to be
            # owner-only. Restore that trust-boundary invariant after download.
            if ! chmod -R go-rwx "${run_root}"; then
              echo "::warning::Evidence permissions from run ${previous} could not be restricted; trying an older run."
              continue
            fi
            # The artifact holds gen-N/bench/{results.jsonl,promotion.json,...};
            # the highest generation is the one that actually reached the gate.
            latest="$(find "${run_root}" -type f -path '*/gen-*/bench/results.jsonl' | sort -V | tail -1)"
            if [[ -z "${latest}" || -L "${latest}" || ! -f "${latest}" ]]; then
              echo "::warning::Run ${previous} uploaded no usable benchmark results; trying an older run."
              continue
            fi
            # Existence is insufficient: an interrupted run may leave an empty,
            # malformed, or session/infra-only JSONL. Reuse the same bounded
            # selection and transcript/digest preflight the proposer will use,
            # so an unusable newer run cannot hide an older useful one.
            if uv run --project eval --locked --extra dev python -c \
              'from pathlib import Path; import json, sys, tempfile; from workflow_bench.evolve import load_jsonl, select_evidence, stage_proposer_evidence_bundle, summarize_gate; result = Path(sys.argv[1]); root = result.parent; rows = select_evidence(load_jsonl(result)); rows or sys.exit(10); promotion = root / "promotion.json"; gate = summarize_gate(json.loads(promotion.read_text())) if promotion.is_file() else []; prior = root.parent / "proposal.md"; prior = prior if prior.is_file() and not prior.is_symlink() else None; dest = Path(tempfile.mkdtemp(prefix="wfseed-preflight-")) / "bundle"; stage_proposer_evidence_bundle(dest, results_dir=root, evidence=rows, learnings=[], gate_summary=gate, prior_proposal=prior)' \
              "${latest}"; then
              :
            else
              usability_status=$?
              echo "::warning::Run ${previous} failed evidence preflight (exit ${usability_status}); trying an older run."
              continue
            fi
            seed="$(dirname "${latest}")"
            echo "Seeding the proposer from run ${previous}: ${seed}"
            break
          done
          if [[ -z "${seed}" ]]; then
            echo '::warning::No usable prior benchmark artifact found; proposing without prior evidence.'
            exit 0
          fi
          echo "seed=${seed}" >> "${GITHUB_OUTPUT}"

      - name: Run the propose → benchmark → gate loop
        id: loop
        # Kill the sweep with time left in the job to upload what it produced.
        # See the budget nesting on the job above.
        timeout-minutes: 1140
        env:
          GITNEXUS_BENCH_ANTHROPIC_API_KEY: ${{ secrets.GITNEXUS_BENCH_ANTHROPIC_API_KEY || secrets.GITNEXUS_BENCH_AUTH_TOKEN }}
          GITNEXUS_BENCH_OPENAI_API_KEY: ${{ secrets.GITNEXUS_BENCH_OPENAI_API_KEY }}
          # The step's stdout is a pipe, so CPython block-buffers it and a
          # multi-hour generation would report nothing until it exits (run
          # 29907431284 emitted every line at the same timestamp, 14h45m in).
          PYTHONUNBUFFERED: '1'
          SEED_RESULTS: ${{ steps.seed.outputs.seed }}
          EVOLUTION_PROFILE: review
          CE_PLUGIN_DIR: ${{ runner.temp }}/compound-engineering-plugin
          CE_PLUGIN_VERSION: 3.24.0
        run: |
          set -euo pipefail
          ./workflow_bench/run-evolution.sh --apply
        working-directory: eval

      - name: Upload benchmark evidence
        # Unconditional: the sweep writes results.jsonl and transcripts as it
        # goes, so a killed or failed generation still has evidence worth
        # keeping — and that is exactly the run whose evidence is needed.
        if: always()
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
        with:
          name: gitnexus-evolution-${{ github.run_id }}-${{ github.run_attempt }}
          # Addressed directly rather than carried from the sweep step: that is
          # the step whose death is the reason this upload matters, and a value
          # threaded from it would not be there when it counts.
          path: ${{ runner.temp }}/wfevolve
          retention-days: 14
          if-no-files-found: warn

      - name: Detect and bound the applied promotion
        id: promotion
        run: |
          set -euo pipefail
          changed="$(git status --porcelain)"
          if [[ -z "${changed}" ]]; then
            echo 'No promotion this run; the incumbent skills stand.'
            echo "promoted=false" >> "${GITHUB_OUTPUT}"
            exit 0
          fi
          # The apply step may only touch the canonical skill tree and its
          # shipped mirrors. Anything else means the overlay escaped its
          # boundary — refuse to open a PR from it.
          while IFS= read -r line; do
            path="${line:3}"
            case "${path}" in
              .claude/skills/gitnexus-review/*|gitnexus/skills/gitnexus-review/*|gitnexus-claude-plugin/skills/gitnexus-review/*|gitnexus-cursor-integration/skills/gitnexus-review/*) ;;
              *)
                echo "::error::Promotion touched a path outside the skill trees: ${path}"
                exit 1
                ;;
            esac
          done <<< "${changed}"
          echo "promoted=true" >> "${GITHUB_OUTPUT}"
          # The loop returns on the first promotion, so the highest-numbered
          # gen-N/bench/promotion.json is the decision that actually fired.
          # Emit only that one — never every generation's, or a rejected
          # generation's decisions could surface in the PR body. The heredoc
          # uses a per-run random delimiter so a summary value that ever
          # contains the marker cannot close the block early and inject keys.
          promotion_file="$(find "${RUNNER_TEMP}/wfevolve" -name promotion.json | sort -V | tail -1)"
          delim="PROMOTION_EOF_$(openssl rand -hex 16)"
          {
            echo "summary<<${delim}"
            if [[ -n "${promotion_file}" ]]; then
              tail -c 8000 "${promotion_file}"
            fi
            echo
            echo "${delim}"
          } >> "${GITHUB_OUTPUT}"

      - name: Mint GitHub App token
        id: app-token
        if: steps.promotion.outputs.promoted == 'true'
        uses: actions/create-github-app-token@bcd2ba49218906704ab6c1aa796996da409d3eb1 # v3.2.0
        with:
          # `client-id` supersedes the deprecated `app-id` in v3.x (the action
          # accepts the numeric App ID here, as publish.yml does). Request only
          # the permissions this job needs — push a branch and open a PR — so
          # the minted token drops the installation's other grants (e.g.
          # Workflows: write).
          client-id: ${{ secrets.RELEASE_APP_ID }}
          private-key: ${{ secrets.RELEASE_APP_PRIVATE_KEY }}
          permission-contents: write
          permission-pull-requests: write

      - name: Open the promotion PR
        if: steps.promotion.outputs.promoted == 'true'
        env:
          APP_TOKEN: ${{ steps.app-token.outputs.token }}
          GH_TOKEN: ${{ steps.app-token.outputs.token }}
          PROMOTION_SUMMARY: ${{ steps.promotion.outputs.summary }}
          RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
        run: |
          set -euo pipefail
          # Include the run attempt: GITHUB_RUN_ID is stable across re-runs, so
          # a re-run after a push-succeeds/PR-create-fails partial failure needs
          # a fresh branch to push (a non-force push to the existing branch
          # would be rejected non-fast-forward and wedge the lane).
          branch="evolution/skills-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}"
          git config user.name 'gitnexus-evolution[bot]'
          git config user.email 'gitnexus-evolution[bot]@users.noreply.github.com'
          git checkout -b "${branch}"
          git add .claude/skills gitnexus/skills gitnexus-claude-plugin/skills gitnexus-cursor-integration/skills/gitnexus-review
          git commit -m 'feat(skills): promoted evolution overlay (gate-passed)'

          # The App token reaches git through GIT_ASKPASS reading step env at
          # push time — it never appears in argv, git config, or the checkout.
          askpass="${RUNNER_TEMP}/evolution-askpass"
          cat > "${askpass}" <<'ASKPASS_EOF'
          #!/usr/bin/env bash
          printf '%s\n' "${APP_TOKEN}"
          ASKPASS_EOF
          chmod 0700 "${askpass}"
          GIT_ASKPASS="${askpass}" GIT_TERMINAL_PROMPT=0 git push \
            "https://x-access-token@github.com/${GITHUB_REPOSITORY}.git" \
            "HEAD:refs/heads/${branch}"

          {
            cat <<'BODY_HEAD'
          Automated skill-evolution promotion. The deterministic gate passed; this PR is the human-review step — inspect the diff and the evidence before merging.
          BODY_HEAD
            printf '\n%s\n\n' "Benchmark evidence: ${RUN_URL} (artifact gitnexus-evolution-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT})."
            cat <<'BODY_OPEN'
          <details><summary>Promotion decisions</summary>

          ```json
          BODY_OPEN
            printf '%s\n' "${PROMOTION_SUMMARY}"
            cat <<'BODY_CLOSE'
          ```

          </details>
          BODY_CLOSE
          } > "${RUNNER_TEMP}/pr-body.md"
          gh pr create \
            --repo "${GITHUB_REPOSITORY}" \
            --base main \
            --head "${branch}" \
            --title 'feat(skills): promoted evolution overlay' \
            --body-file "${RUNNER_TEMP}/pr-body.md"