gbr-games/cmd/gbrgames/build.go

package main

import (
	"bytes"
	"encoding/json"
	"errors"
	"flag"
	"fmt"
	"os"
	"os/exec"
	"path/filepath"
	"strings"
)

// build turns one researched variant into a template that has really booted: first boot of the image
// untouched (facts), a draft by the model, then the draft deployed with the node-agent's own code, exactly
// as CI's boot test does. What goes wrong in that boot (a crash, an error line, saves outside the volume)
// goes back to the model, a few rounds at most. The result is a games/<id>/ folder ready for a pull request.

type buildResult struct {
	GameID   string   `json:"gameId"`
	Dir      string   `json:"dir"`
	Image    string   `json:"image"`
	Passed   bool     `json:"passed"`
	Rounds   int      `json:"rounds"`
	Verified []string `json:"verified"`
	Problems []string `json:"problems"`
	Summary  string   `json:"summary"`
}

const defaultAgentImage = "www.gitea.gbreed.com/gbreed-public/gbr-node-agent:latest"

func runBuild(args []string) error {
	fs := flag.NewFlagSet("build", flag.ExitOnError)
	gamesDir := fs.String("games", "games", "directory holding one folder per game")
	schemaDir := fs.String("schemas", "schema", "directory holding the JSON Schemas")
	gameName := fs.String("name", "", "the game's name")
	id := fs.String("id", "", "folder/id for the new game (default: from -name)")
	variantArg := fs.String("variant", "", "the chosen variant: a research result variant as JSON, or @file")
	notes := fs.String("notes", "", "what the admin asked for")
	provider := fs.String("provider", "ollama", "ollama or gemini")
	modelName := fs.String("model", "", "model name (default as for draft)")
	attempts := fs.Int("attempts", 3, "fix rounds for validation problems, each time")
	bootRounds := fs.Int("boot-rounds", 3, "boots of the draft before giving up")
	out := fs.String("out", "", "write the result JSON here")
	fs.Parse(args)
	if *gameName == "" || *variantArg == "" {
		return errors.New("build needs -name and -variant")
	}
	raw := []byte(*variantArg)
	if strings.HasPrefix(*variantArg, "@") {
		var err error
		if raw, err = os.ReadFile((*variantArg)[1:]); err != nil {
			return err
		}
	}
	var v variant
	if err := json.Unmarshal(raw, &v); err != nil {
		return fmt.Errorf("-variant: %w", err)
	}
	if *id == "" {
		*id = gameID(*gameName)
	}
	agentImage := os.Getenv("AGENT_IMAGE")
	if agentImage == "" {
		agentImage = defaultAgentImage
	}

	res, err := build(buildOptions{gamesDir: *gamesDir, schemaDir: *schemaDir, name: *gameName, id: *id, variant: v,
		notes: *notes, provider: *provider, model: *modelName, attempts: *attempts, bootRounds: *bootRounds, agentImage: agentImage})
	if res != nil && *out != "" {
		raw, _ := json.MarshalIndent(res, "", "  ")
		if werr := os.WriteFile(*out, raw, 0o644); werr != nil && err == nil {
			err = werr
		}
	}
	if err != nil {
		return err
	}
	if !res.Passed {
		return fmt.Errorf("the draft in %s still has problems:\n- %s", res.Dir, strings.Join(res.Problems, "\n- "))
	}
	fmt.Printf("[build] %s booted with the node-agent and is ready for a pull request\n", res.Dir)
	return nil
}

type buildOptions struct {
	gamesDir, schemaDir, name, id string
	variant                       variant
	notes, provider, model        string
	attempts, bootRounds          int
	agentImage                    string
}

func build(o buildOptions) (*buildResult, error) {
	if len(o.variant.Images) == 0 {
		return nil, errors.New("the variant has no image to build from")
	}
	res := &buildResult{GameID: o.id, Dir: filepath.Join(o.gamesDir, o.id)}
	res.Verified = append(res.Verified, o.variant.Verified...)

	// 1. First boot of each candidate image until one runs as a non-root user
	var report *bootReport
	for _, image := range o.variant.Images {
		steam := isSteamCMDImage(image) && o.variant.SteamAppID != ""
		if isSteamCMDImage(image) && !steam {
			continue // the SteamCMD image alone runs nothing
		}
		res.Image = image
		if steam && o.variant.StartupHint == "" {
			fmt.Printf("[build] no startup command known for Steam app %s; drafting without a first boot\n", o.variant.SteamAppID)
			report = nil
			break
		}
		fmt.Printf("[build] first boot of %s\n", image)
		r, err := firstBoot(o.id+"-probe", image, o.variant.SteamAppID, steam, o.variant.StartupHint)
		if err != nil {
			return nil, err
		}
		if why := imageUnusable(r); why != "" {
			fmt.Printf("[build] %s can't be used: %s\n", image, why)
			res.Problems = append(res.Problems, why)
			res.Image = ""
			continue
		}
		report = r
		res.Verified = append(res.Verified, fmt.Sprintf("first boot of %s: %s", image, r.Outcome))
		break
	}
	if res.Image == "" {
		return res, errors.New("no image of this variant runs as a non-root user")
	}
	res.Problems = nil

	// 2. Draft until validation and the image checks pass
	steamApp := ""
	if isSteamCMDImage(res.Image) {
		steamApp = o.variant.SteamAppID
	}
	d, err := newDrafter(draftJob{GamesDir: o.gamesDir, SchemaDir: o.schemaDir, Name: o.name, ID: o.id, Image: res.Image,
		SteamApp: steamApp, Notes: variantNotes(o.variant, o.notes), Report: report, Provider: o.provider, Model: o.model, NoPull: true,
		Docs: sourceText(o.variant.Sources)})
	if err != nil {
		return nil, err
	}
	problems, err := d.refine(o.attempts, nil)
	res.Rounds = d.rounds
	if err != nil {
		return res, err
	}
	if len(problems) > 0 {
		res.Problems = problems
		res.Summary = buildSummary(o, res)
		return res, nil
	}

	// 3. Boot the draft the way nodes run it; failures go back to the model
	wantOutside := len(outsideDirs(report)) > 0
	for boot := 1; boot <= o.bootRounds; boot++ {
		fmt.Printf("[build] boot %d of the draft with the node-agent\n", boot)
		bootProblems, notes, err := bootCheck(o.gamesDir, o.id, o.agentImage, wantOutside)
		if err != nil {
			return res, err
		}
		if len(bootProblems) == 0 {
			res.Passed = true
			res.Verified = append(res.Verified, notes...)
			break
		}
		fmt.Printf("[build] boot %d failed:\n", boot)
		for _, p := range bootProblems {
			fmt.Println("        - " + firstLine(p))
		}
		res.Problems = bootProblems
		if boot == o.bootRounds {
			break
		}
		problems, err := d.refine(o.attempts, bootProblems)
		res.Rounds = d.rounds
		if err != nil {
			return res, err
		}
		if len(problems) > 0 {
			res.Problems = problems
			break
		}
	}
	if res.Passed {
		res.Problems = nil
	}
	res.Summary = buildSummary(o, res)
	return res, nil
}

func isSteamCMDImage(image string) bool {
	return image == "cm2network/steamcmd" || strings.HasPrefix(image, "cm2network/steamcmd:")
}

func variantNotes(v variant, admin string) string {
	var parts []string
	if v.Title != "" {
		parts = append(parts, "variant: "+v.Title+" ("+v.Description+")")
	}
	if v.StartupHint != "" {
		parts = append(parts, "documented startup: "+v.StartupHint)
	}
	if len(v.Ports) > 0 {
		parts = append(parts, "documented ports: "+strings.Join(v.Ports, ", "))
	}
	if v.Notes != "" {
		parts = append(parts, v.Notes)
	}
	if admin != "" {
		parts = append(parts, "the admin asked for: "+admin)
	}
	return strings.Join(parts, "; ")
}

// firstBoot runs first-boot and reads back its report.
func firstBoot(name, image, steamApp string, steam bool, startup string) (*bootReport, error) {
	outDir := filepath.Join("firstboot-out", name)
	args := []string{"-image", image, "-out", outDir}
	if steam {
		args = append(args, "-steam-app", steamApp, "-startup", startup)
	}
	if err := runFirstBoot(args); err != nil {
		return nil, err
	}
	return readBootReport(filepath.Join(outDir, "report.json"))
}

// bootCheck deploys the template with the node-agent's BootTest, the same check CI runs, and when the
// game wrote files outside its volume on the first boot, boots it once more to see that it no longer does.
func bootCheck(gamesDir, id, agentImage string, wantOutside bool) (problems, verified []string, err error) {
	self, err := os.Executable()
	if err != nil {
		return nil, nil, err
	}
	// A boot test that was killed (worker restarted mid-build) leaves its container and volume behind,
	// and the next deploy under the same name would trip over them
	runCmd("docker", "rm", "-f", "boottest-"+id)
	runCmd("docker", "volume", "rm", "-f", "boottest-"+id+"-data")
	payloadCmd := exec.Command(self, "boottest-payload", "-games", gamesDir, id)
	var stderr bytes.Buffer
	payloadCmd.Stderr = &stderr
	payload, err := payloadCmd.Output()
	if err != nil {
		var exitErr *exec.ExitError
		if errors.As(err, &exitErr) && exitErr.ExitCode() == exitMissingSecret {
			// The builder has no CI secrets: that's the template's problem to fix, not a reason to stop
			return []string{"the boot test can't run: " + strings.TrimSpace(stderr.String()) + ". Remove 'secrets' from " +
				"test.yaml; boot tests fill password settings with a throwaway value"}, nil, nil
		}
		return nil, nil, fmt.Errorf("boottest-payload: %v: %s", err, strings.TrimSpace(stderr.String()))
	}
	if out, err := runCmd("docker", "pull", "-q", agentImage); err != nil {
		return nil, nil, fmt.Errorf("docker pull %s: %v: %s", agentImage, err, out)
	}
	cmd := exec.Command("docker", "run", "--rm", "-i", "-v", "/var/run/docker.sock:/var/run/docker.sock",
		"--entrypoint", "java", agentImage, "-Dloader.main=com.gbreed.gbr.nodeagent.BootTest", "-cp", "app.jar",
		"org.springframework.boot.loader.launch.PropertiesLauncher")
	cmd.Stdin = bytes.NewReader(payload)
	var output bytes.Buffer
	cmd.Stdout = &teeWriter{&output}
	cmd.Stderr = cmd.Stdout
	runErr := cmd.Run()
	if runErr != nil {
		var exitErr *exec.ExitError
		if !errors.As(runErr, &exitErr) {
			return nil, nil, runErr
		}
		return []string{bootFailure(output.String())}, nil, nil
	}
	verified = append(verified, "boot test with the node-agent: "+bootVerdict(output.String()))

	if wantOutside {
		fmt.Println("[build] first boot wrote saves outside the volume; booting the draft to see that it no longer does")
		r, err := firstBootGame(gamesDir, id)
		if err != nil {
			return nil, nil, err
		}
		if dirs := outsideDirs(r); len(dirs) > 0 {
			return []string{fmt.Sprintf("booting your template, the server still wrote its own files outside the server volume, "+
				"in %s, so your startup command or settings did not move them. Check the exact launch option name and "+
				"syntax this server documents for its save/data folder (a wrong option is usually ignored silently) "+
				"and point it at /server", strings.Join(dirs, ", "))}, nil, nil
		}
		verified = append(verified, "the server keeps its files inside /server")
	}
	return nil, verified, nil
}

func firstBootGame(gamesDir, id string) (*bootReport, error) {
	outDir := filepath.Join("firstboot-out", id+"-draft")
	if err := runFirstBoot([]string{"-games", gamesDir, "-game", id, "-out", outDir}); err != nil {
		return nil, err
	}
	return readBootReport(filepath.Join(outDir, "report.json"))
}

// bootFailure turns BootTest's output into feedback: its verdict and the server's last lines, which hold
// the actual error (an unknown option, a missing file, a port in use).
func bootFailure(output string) string {
	verdict, tail := bootVerdict(output), output
	if i := strings.Index(output, "---- last "); i >= 0 {
		tail = output[i:]
	}
	lines := strings.Split(strings.TrimSpace(tail), "\n")
	if len(lines) > 60 {
		lines = append(lines[:1], lines[len(lines)-59:]...)
	}
	return "booting your template with the node-agent (the real deploy code) failed: " + verdict +
		". Find the cause in the server's output below and fix the template (startup command, settings, " +
		"ready pattern, memory):\n" + strings.Join(lines, "\n")
}

func bootVerdict(output string) string {
	for _, l := range strings.Split(output, "\n") {
		for _, k := range []string{"[boot-test] FAILED: ", "[boot-test] PASSED: "} {
			if i := strings.Index(l, k); i >= 0 {
				return strings.TrimSpace(l[i+len("[boot-test] "):])
			}
		}
	}
	return "no verdict"
}

func firstLine(s string) string {
	line, _, _ := strings.Cut(s, "\n")
	return line
}

// teeWriter shows the boot as it happens (the worker streams it to the panel) while keeping it for the verdict.
type teeWriter struct{ buf *bytes.Buffer }

func (t *teeWriter) Write(p []byte) (int, error) {
	os.Stdout.Write(p)
	return t.buf.Write(p)
}

// buildSummary is the pull request's description: where the facts came from and what was really checked.
func buildSummary(o buildOptions, res *buildResult) string {
	var b strings.Builder
	fmt.Fprintf(&b, "Drafted by the GBR AI builder. **Review before merging**: check the settings, "+
		"ports and descriptions against the game's documentation.\n\n")
	fmt.Fprintf(&b, "- Game: **%s**", o.name)
	if o.variant.Title != "" {
		fmt.Fprintf(&b, " — %s", o.variant.Title)
	}
	fmt.Fprintf(&b, "\n- Image: `%s`", res.Image)
	if o.variant.SteamAppID != "" && isSteamCMDImage(res.Image) {
		fmt.Fprintf(&b, ", Steam app %s", o.variant.SteamAppID)
	}
	fmt.Fprintf(&b, "\n- Model rounds: %d\n", res.Rounds)
	if len(res.Verified) > 0 {
		b.WriteString("\n**Checked for real**\n")
		for _, v := range res.Verified {
			fmt.Fprintf(&b, "- %s\n", v)
		}
	}
	if len(res.Problems) > 0 {
		b.WriteString("\n**Still open** (the builder could not fix these)\n")
		for _, p := range res.Problems {
			fmt.Fprintf(&b, "- %s\n", firstLine(p))
		}
	}
	if o.variant.Notes != "" {
		fmt.Fprintf(&b, "\n**Notes from research:** %s\n", o.variant.Notes)
	}
	if len(o.variant.Sources) > 0 {
		b.WriteString("\n**Sources**\n")
		for _, s := range o.variant.Sources {
			fmt.Fprintf(&b, "- %s\n", s)
		}
	}
	return b.String()
}

// sourceText reads the pages research based the variant on, in full, so the drafter can check that every
// launch option the model uses is documented there. Empty (check skipped) without a reader.
func sourceText(sources []string) string {
	reader := os.Getenv("READER_URL")
	if reader == "" || len(sources) == 0 {
		fmt.Println("[build] no research sources to check launch options against (READER_URL or sources missing)")
		return ""
	}
	var b strings.Builder
	read := 0
	for _, s := range sources {
		text, err := readPageLimit(reader, s, 200_000)
		if err != nil {
			fmt.Printf("[build] could not read %s: %v\n", s, err)
			continue
		}
		read++
		b.WriteString(text)
		b.WriteString("\n")
	}
	fmt.Printf("[build] read %d of %d research sources to check launch options against\n", read, len(sources))
	return b.String()
}