action.yml

name: Commit summary
description: Summarises a GameTracking commit with Claude Code and posts it as a commit comment. Reads the token from the CLAUDE_CODE_OAUTH_TOKEN environment variable (made with claude setup-token).

inputs:
    sha:
        description: 'Commit to summarise'
        required: true
    model:
        description: 'Claude model'
        required: false
        default: 'opus'
    worker-model:
        description: 'Claude model for the workers that read the diff'
        required: false
        default: 'sonnet'

runs:
    using: composite
    steps:
        # Older builds are needed for revert checks and "is this new" lookups, history searches in the prompt are limited to the same six months
        - name: Fetch older builds
          shell: bash
          working-directory: tracked_game
          run: git fetch --quiet --shallow-since="6 months ago" origin

        # Other Source 2 games with recent history, to spot shared engine changes and when they first appeared
        - name: Clone GameTracking-CS2
          if: github.repository != 'SteamTracking/GameTracking-CS2'
          shell: bash
          run: git clone --quiet --shallow-since="2 months ago" https://github.com/SteamTracking/GameTracking-CS2.git "$GITHUB_WORKSPACE/siblings/GameTracking-CS2"

        - name: Clone GameTracking-Dota2
          if: github.repository != 'SteamTracking/GameTracking-Dota2'
          shell: bash
          run: git clone --quiet --shallow-since="2 months ago" https://github.com/SteamTracking/GameTracking-Dota2.git "$GITHUB_WORKSPACE/siblings/GameTracking-Dota2"

        - name: Clone GameTracking-Deadlock
          if: github.repository != 'SteamTracking/GameTracking-Deadlock'
          shell: bash
          run: git clone --quiet --shallow-since="2 months ago" https://github.com/SteamTracking/GameTracking-Deadlock.git "$GITHUB_WORKSPACE/siblings/GameTracking-Deadlock"

        # The main session orchestrates, worker agents read the diff and a writer agent turns their notes into the comment
        - name: Prepare prompts
          shell: bash
          run: |
            # Forward slashes on Windows, so the paths survive unquoted in shell commands
            norm() { if command -v cygpath > /dev/null 2>&1; then cygpath -m "$1"; else printf '%s' "$1"; fi; }
            workspace=$(norm "$GITHUB_WORKSPACE")
            tools=$workspace/tools/exe
            notes=$(norm "$RUNNER_TEMP")/commit_summary_notes
            out=$workspace/commit_summary.md

            run=$(
              echo '## This run'
              echo "- The game repository is checked out at \`$workspace/tracked_game\` with its last six months of history; run git as \`git -C <that path> …\`."
              echo '- The commit to summarise is `<sha>` = `${{ inputs.sha }}`.'
              echo "- Repos of the other games are checked out at their latest commit under \`$workspace/siblings/\`, with their last two months of history."
              echo "- \`<tools>\` is \`$tools\`."
              echo "- \`<notes>\` is \`$notes\`."
              echo "- The comment body (or \`NO_COMMENT\`) goes to \`$out\`, this exact absolute path, and only the writer writes it. Apart from that file, notes and temp files from \`mktemp\`, nobody writes anything. Only that file is posted, not the final reply."
            )

            mkdir -p "$RUNNER_TEMP/commit_summary_notes"

            printf '%s\n\n%s\n\n%s\n' "$(cat "$GITHUB_ACTION_PATH/common.md")" "$(cat "$GITHUB_ACTION_PATH/orchestrator.md")" "$run" > "$RUNNER_TEMP/commit_summary_prompt.md"

            role() { printf '%s\n\n%s\n\n%s\n' "$(cat "$GITHUB_ACTION_PATH/common.md")" "$(cat "$GITHUB_ACTION_PATH/$1.md")" "$run"; }

            jq -n \
              --arg worker "$(role worker)" \
              --arg writer "$(role writer)" \
              --arg worker_model "${{ inputs.worker-model }}" \
              '{
                worker: {description: "Reads one area of a GameTracking commit diff completely and writes notes with findings and evidence", prompt: $worker, model: $worker_model},
                writer: {description: "Writes the commit comment from the workers notes and checks it against the rules", prompt: $writer}
              }' \
              > "$RUNNER_TEMP/commit_summary_agents.json"

        # The installers are per platform, both install to ~/.local/bin. Retried as downloads can be rate limited.
        - name: Install Claude Code (Windows)
          if: runner.os == 'Windows'
          shell: pwsh
          run: |
            for ($attempt = 1; $true; $attempt++) {
              try {
                & ([scriptblock]::Create((Invoke-RestMethod https://claude.ai/install.ps1)))
                if (Test-Path "$env:USERPROFILE\.local\bin\claude.exe") { break }
                throw "claude.exe was not installed"
              } catch {
                if ($attempt -ge 3) { throw }
                Write-Host "Install attempt $attempt failed: $_"
                Start-Sleep -Seconds (10 * $attempt)
              }
            }
            "$env:USERPROFILE\.local\bin" | Out-File -FilePath $env:GITHUB_PATH -Append -Encoding utf8

        - name: Install Claude Code (Linux)
          if: runner.os != 'Windows'
          shell: bash
          run: |
            set -o pipefail
            for attempt in 1 2 3; do
              if curl -fsSL https://claude.ai/install.sh | bash; then
                break
              fi
              if [[ $attempt == 3 ]]; then
                exit 1
              fi
              echo "Install attempt $attempt failed"
              sleep $((10 * attempt))
            done
            echo "$HOME/.local/bin" >> "$GITHUB_PATH"

        # The transcript and the workers' notes are kept as an artifact, so skipped steps can be audited
        - name: Run Claude Code
          shell: bash
          run: |
            set -o pipefail
            claude -p "$(cat "$RUNNER_TEMP/commit_summary_prompt.md")" \
              --model "${{ inputs.model }}" \
              --agents "$RUNNER_TEMP/commit_summary_agents.json" \
              --permission-mode auto \
              --output-format stream-json \
              --verbose < /dev/null > "$RUNNER_TEMP/commit_summary_transcript.jsonl"
            jq -r 'select(.type == "result") | .result' "$RUNNER_TEMP/commit_summary_transcript.jsonl" | tee -a "$GITHUB_STEP_SUMMARY"
          # Set-diffing big string dumps can take minutes, so commands get longer than the default 2 minute timeout
          env:
            BASH_DEFAULT_TIMEOUT_MS: '600000'
            BASH_MAX_TIMEOUT_MS: '1800000'

        - name: Upload transcript
          if: always()
          uses: actions/upload-artifact@v7
          with:
            name: commit-summary-transcript
            path: |
              ${{ runner.temp }}/commit_summary_transcript.jsonl
              ${{ runner.temp }}/commit_summary_notes/
            if-no-files-found: ignore
            retention-days: 7

        - name: Post comment
          shell: bash
          run: |
            if [[ ! -s commit_summary.md ]]; then
              echo "::error::Claude Code did not write commit_summary.md"
              find . -name commit_summary.md -not -path '*/.git/*'
              exit 1
            fi

            if [[ "$(tr -d '[:space:]`' < commit_summary.md)" == "NO_COMMENT" ]]; then
              echo "Nothing worth commenting"
              exit 0
            fi

            if grep -nE '\ba value\b|(^|[^A-Za-z])I (did|have|read|checked|could|was)\b|\b(leak|unreleased|upcoming)\b' commit_summary.md; then
              echo "::warning::The comment matched a banned phrase (see above)"
            fi

            printf '\n\n---\n<sub>This summary was generated by AI and may contain mistakes.</sub>\n' >> commit_summary.md

            gh api "repos/$GITHUB_REPOSITORY/commits/${{ inputs.sha }}/comments" \
              -F body=@commit_summary.md
          env:
            GH_TOKEN: ${{ github.token }}