mesh-controller/cmd/mesh-controller/secret.go

package main

import (
	"bufio"
	"context"
	"encoding/json"
	"errors"
	"flag"
	"fmt"
	"io"
	"os"
	"strings"
	"time"

	"github.com/novox/mesh-controller/internal/catalogue"
	"github.com/novox/mesh-controller/internal/inventory"
	"github.com/novox/mesh-controller/internal/secrets"
)

// secretCommand gives the mesh a value it must carry and could not have invented.
//
// **Every other secret in this mesh is one the mesh made** — generated, sealed to the machine that
// will use it, and never readable again. That is right for something coming into existence, and
// wrong for something that already exists: a database created last year has the password it was
// created with, and generating a new one puts 32 random bytes where a working credential was.
// The machine applies it, reports success, and whatever reads it fails to authenticate somewhere
// else entirely — with the mesh insisting the secret was delivered, which it was.
//
// So this is the entry point for **adopting** something already running. The store has carried
// the distinction since the beginning: a module secret records whether it was `made` or
// `accepted`, and refuses to invent a replacement for the second. Nothing until now could write
// one, so the only accepted secret in the mesh was the broker account issued to a build machine.
//
// The value is sealed on the way in and the plaintext discarded, exactly as a generated one is.
// **The only difference between the two is where the value came from.**
func secretCommand(ctx context.Context, args []string) error {
	if len(args) == 0 {
		return errors.New(secretUsage)
	}
	switch args[0] {
	case "accept":
	case "rotate":
		return secretRotate(ctx, args[1:])
	case "recover":
		return secretRecover(ctx, args[1:])
	case "export":
		return secretExport(ctx, args[1:])
	default:
		return errors.New(secretUsage)
	}
	rest, flags := split(args[1:])
	set := flag.NewFlagSet("secret accept", flag.ContinueOnError)
	from := set.String("from", "",
		"read the value from this file instead of asking (use - for standard input)")
	provider := set.String("provider", "",
		"the node providing <name>: the value becomes the PAIR credential between <module> on <node> "+
			"and that provider, sealed to both — the vault's operator-delivered secret (ADR 0092)")
	desk := set.String("at-desk", "",
		"ask the operator for the value in a prompt that does not show it, on this machine's desk; the "+
			"answer comes back sealed to this call alone (novox/hq ADR 0259 §10)")
	local := set.String("local", "",
		"with --provider: the name the credential goes by inside <module>, where its manifest keeps "+
			"several for <name> (ADR 0094)")
	if err := set.Parse(flags); err != nil {
		return err
	}
	if len(rest) != 3 {
		return errors.New(secretUsage)
	}
	node, module, name := rest[0], rest[1], rest[2]
	if *desk != "" {
		if *from != "" || *provider != "" {
			return errors.New("--at-desk gives a module's own secret, and takes neither --from nor --provider")
		}
		return giveAtDesk(ctx, node, module, name, *desk)
	}

	value, err := valueFor(node, module, name, *from)
	if err != nil {
		return err
	}
	value = asSupplied(value)
	if value == "" {
		return errors.New("there is nothing to seal")
	}

	open, err := openStores(ctx)
	if err != nil {
		return err
	}
	defer open.Close()

	if *provider != "" {
		// Into the pair, not into the module's own secrets: what the provider is asked to create
		// and what the consumer reads are the same value, and neither end can be told a different
		// one later without the other (novox/hq 04-ISSUES/070).
		if err := open.inventory.AcceptSecretForPair(ctx, name, node, module, *provider, *local, value); err != nil {
			return err
		}
		fmt.Printf("%s on %s now holds %q from %s%s, sealed to both machines.\n", module, node, name, *provider, asLocal(*local))
		fmt.Printf("  the mesh cannot read it back, will not replace it with one of its own, and will not rotate it\n")
		fmt.Printf("  run `push %s` and `push %s` to send it\n", *provider, node)
		return nil
	}
	untilStart, err := open.inventory.AcceptGivenSecret(ctx, node, module, name, value)
	if err != nil {
		return err
	}
	// Not printed back, and there is nowhere it could be printed from: it is sealed to that
	// machine and the mesh cannot read it again.
	fmt.Printf("%s on %s now holds %q, sealed to that machine.\n", module, node, name)
	if untilStart {
		// Accepted, and said what it is for (novox/hq ADR 0228): a value given by hand adopts
		// something that already holds it, and lives only until the module has started on it.
		fmt.Printf("  it lives until %s next starts well under the mesh on it, and is then replaced with a value\n"+
			"  the mesh makes, sealed and sent (ADR 0228): a value given by hand is for adopting something\n"+
			"  already running that holds it\n", module)
		if record, err := open.inventory.NodeByName(ctx, node); err == nil && !record.Adopted {
			fmt.Printf("  %s is not an adopted machine: if %s is installed fresh there, it needs no given value —\n"+
				"  the mesh makes one at the first push\n", node, module)
		}
	} else {
		fmt.Printf("  the mesh cannot read it back, and will not replace it with one of its own\n")
	}
	fmt.Printf("  run `push %s` to send it\n", node)
	return nil
}

const secretUsage = "secret rotate <node> <module> <name> [--why <text> [--cause <word>]]\n" +
	"secret accept <node> <module> <name> [--from <file> | --at-desk <machine>] [--provider <node> [--local <name>]]\n" +
	"secret recover <node> <module> <name> --key <operator-key> [--out <file>] [--from-export <file>] [--provider <node>]\n" +
	"secret export [--out <file>]"

// secretRecover is break-glass: a secret opened with the operator's key, written to a file.
//
// **The mesh cannot show a secret back, and this does not make it able to.** What is opened here
// is the copy sealed to the operator key (novox/hq ADR 0085, amended); the mesh holds that blob and
// no key for it, and this program holds the key for the length of the call and no blob until given
// one. Recovery needs both, which is what keeps the sealing meaningful.
//
// The value goes to a file at 0600, never to the terminal unless asked for with `--out -` — the
// source mesh's secret tools were written after a secret was printed into a transcript, and that
// rule is theirs. `--from-export` reads the blob from a file `secret export` wrote, so recovery
// works with the store gone, which is the case it exists for.
func secretRecover(ctx context.Context, args []string) error {
	rest, flags := split(args)
	set := flag.NewFlagSet("secret recover", flag.ContinueOnError)
	keyFile := set.String("key", "", "the operator's private key, from `operator key make`")
	out := set.String("out", "", "where to write the value (0600); - for standard output. Default <node>.<module>.<name>.secret")
	fromExport := set.String("from-export", "", "read the sealed copy from this `secret export` file instead of the store")
	provider := set.String("provider", "", "for a pair credential held from more than one provider: which one")
	local := set.String("local", "", "for a pair credential the module keeps under a local name (ADR 0094): which one")
	if err := set.Parse(flags); err != nil {
		return err
	}
	if len(rest) != 3 {
		return errors.New(secretUsage)
	}
	node, module, name := rest[0], rest[1], rest[2]
	private, err := readPrivateKey(*keyFile)
	if err != nil {
		return err
	}

	var kept inventory.Kept
	if *fromExport != "" {
		kept, err = keptFromExport(*fromExport, node, module, name, *provider)
		if err != nil {
			return err
		}
	} else {
		open, err := openStores(ctx)
		if err != nil {
			return err
		}
		defer open.Close()
		kept, err = open.inventory.KeptSecret(ctx, node, module, name, *provider, *local)
		if err != nil {
			return err
		}
	}

	value, err := secrets.Open(private, kept.Sealed)
	if err != nil {
		return fmt.Errorf("%s on %s: %q is sealed to operator key %s, and that key does not open it: %w",
			module, node, name, secrets.Fingerprint(kept.Key), err)
	}
	if *out == "-" {
		_, err := os.Stdout.Write(append(value, '\n'))
		return err
	}
	path := *out
	if path == "" {
		path = node + "." + module + "." + name + ".secret"
	}
	if err := writeNew(path, value); err != nil {
		return err
	}
	fmt.Printf("%s on %s: %q recovered to %s (0600) — %d bytes, origin %s\n",
		module, node, name, path, len(value), kept.Origin)
	return nil
}

// An export is what a person keeps beside the operator key — the catalogue's shape, so the vault
// keeps the same document on its disk (Manifest.Keeps).
type export = catalogue.KeptExport

func secretExport(ctx context.Context, args []string) error {
	set := flag.NewFlagSet("secret export", flag.ContinueOnError)
	out := set.String("out", "", "where to write the export (0600); - or empty for standard output")
	if err := set.Parse(args); err != nil {
		return err
	}
	open, err := openStores(ctx)
	if err != nil {
		return err
	}
	defer open.Close()
	inv := open.inventory
	key, err := inv.OperatorKey(ctx)
	if err != nil {
		return err
	}
	if key == "" {
		return errors.New("the mesh has no operator key, so nothing is sealed to one; `operator key make` and `operator key set` first")
	}
	doc, err := inv.OperatorExport(ctx)
	if err != nil {
		return err
	}
	body, err := json.MarshalIndent(doc, "", "  ")
	if err != nil {
		return err
	}
	body = append(body, '\n')
	if *out == "" || *out == "-" {
		_, err := os.Stdout.Write(body)
		return err
	}
	// Replaced whole, and made 0600 whether or not it existed: an export is ciphertext and a public
	// key, but it is also the list of every secret the mesh has, and a file left at an earlier mode
	// while the command says 0600 is a lie in the one place a person checks.
	if err := writeReplacing(*out, body); err != nil {
		return err
	}
	fmt.Printf("%d secret(s) exported to %s (0600), sealed to operator key %s — ciphertext, keep it with the key\n",
		len(doc.Kept), *out, doc.Fingerprint)
	if len(doc.EarlierKey) > 0 {
		fmt.Printf("  %d secret(s) are sealed to an EARLIER operator key: recoverable with that key only\n", len(doc.EarlierKey))
	}
	if len(doc.Unrecoverable) > 0 {
		fmt.Printf("  %d secret(s) are NOT in it: made before the mesh had an operator key\n", len(doc.Unrecoverable))
	}
	return nil
}

// writeNew writes a file that must not exist yet, atomically: create-exclusive, 0600. A check
// followed by a write is a window in which a key somebody still needs can be overwritten.
func writeNew(path string, content []byte) error {
	f, err := os.OpenFile(path, os.O_WRONLY|os.O_CREATE|os.O_EXCL, 0o600)
	if err != nil {
		if os.IsExist(err) {
			return fmt.Errorf("%s already exists; not overwriting it", path)
		}
		return err
	}
	if _, err := f.Write(content); err != nil {
		f.Close()
		return err
	}
	return f.Close()
}

// writeReplacing writes a file whole, creating or truncating it, and leaves it at 0600 either way.
func writeReplacing(path string, content []byte) error {
	f, err := os.OpenFile(path, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o600)
	if err != nil {
		return err
	}
	if _, err := f.Write(content); err != nil {
		f.Close()
		return err
	}
	if err := f.Chmod(0o600); err != nil {
		f.Close()
		return err
	}
	return f.Close()
}

func keptFromExport(path, node, module, name, provider string) (inventory.Kept, error) {
	raw, err := os.ReadFile(path)
	if err != nil {
		return inventory.Kept{}, err
	}
	var e export
	if err := json.Unmarshal(raw, &e); err != nil {
		return inventory.Kept{}, fmt.Errorf("%s is not a secret export: %w", path, err)
	}
	// Sealed to the current key or to an earlier one: both are copies the given key might open,
	// and Open says which. Not the unrecoverable list, which holds no copy at all.
	var found []inventory.Kept
	for _, k := range append(append([]inventory.Kept{}, e.Kept...), e.EarlierKey...) {
		if k.Node == node && k.Module == module && k.Name == name && (provider == "" || k.Provider == provider) {
			found = append(found, k)
		}
	}
	switch len(found) {
	case 0:
		return inventory.Kept{}, fmt.Errorf("%s holds no copy of %s's %q on %s", path, module, name, node)
	case 1:
		return found[0], nil
	default:
		providers := make([]string, 0, len(found))
		for _, f := range found {
			providers = append(providers, f.Provider)
		}
		return inventory.Kept{}, fmt.Errorf("%s holds %s's %q on %s from more than one provider (%s); say which with --provider",
			path, module, name, node, strings.Join(providers, ", "))
	}
}

// split separates what this command is about from how it was asked.
//
// **Because the standard library stops parsing at the first non-flag argument.** With the
// positionals first — which is the order that reads correctly — everything after them is left
// sitting in the arguments, so `secret accept a b c --from -` arrives as five positionals and the
// flag is never seen. The host's own parser carries the same note, and the fault it names is
// worse than this one: there, a flag somebody passed was silently ignored and the command
// succeeded anyway.
func split(args []string) (positional, flags []string) {
	for i, arg := range args {
		if strings.HasPrefix(arg, "-") {
			return args[:i], args[i:]
		}
	}
	return args, nil
}

// asSupplied is the value with its line ending removed and nothing else.
//
// **A file has a trailing newline and a password does not**, so the ending goes — a credential
// wrong by one byte fails in a way nobody connects to how it was supplied.
//
// **And only the ending.** Trimming both ends is the obvious thing and it is wrong: a password
// chosen with a leading space is one the mesh would then deliver as a different password, silently,
// with the operator certain they had supplied it correctly.
func asSupplied(raw string) string {
	return strings.TrimRight(raw, "\r\n")
}

// valueFor gets the secret without putting it somewhere it can be read afterwards.
//
// **Not an argument, and there is no flag that takes one.** A value on the command line is in the
// shell's history, in the process list for as long as it runs, and in whatever collects either.
// The paths here are a file the operator already has, or a prompt that does not echo — the same
// two ways a model-access key is supplied (novox/hq ADR 0024).
func valueFor(node, module, name, from string) (string, error) {
	switch {
	case from == "-":
		body, err := io.ReadAll(os.Stdin)
		if err != nil {
			return "", err
		}
		return string(body), nil

	case from != "":
		body, err := os.ReadFile(from)
		if err != nil {
			return "", err
		}
		return string(body), nil

	default:
		// The same path a model-access key takes, and for the same reason: a value given as an
		// argument is in the shell's history and in the process list. Read from standard input,
		// echoed nowhere by this program.
		fmt.Fprintf(os.Stderr,
			"reading %s's %q for %s from standard input; it is not echoed anywhere\n",
			module, name, node)
		line, err := bufio.NewReader(os.Stdin).ReadString('\n')
		if err != nil && line == "" {
			return "", fmt.Errorf("nothing was given on standard input: %w", err)
		}
		return line, nil
	}
}

// secretRotate makes a module's own secret anew and sends the machine, so the module starts again on
// the new value (novox/hq ADR 0114, issue 180). A pair credential rotates with `rotate <provision>`;
// this is the secret with one party. Said in the log with who asked and when, never the value.
//
// **A value given to the mesh rotates the same way** (novox/hq ADR 0228): what a module reads at start
// is held by nobody else, so the old value is not needed to replace it. Refused for a value an
// outside party issued, which no value of the mesh's would replace. Why it was rotated is recorded in
// the hand-act log when given — a rotation asked by a person is an act by hand, and a leak is a cause
// worth counting.
func secretRotate(ctx context.Context, args []string) error {
	rest, flags := split(args)
	set := flag.NewFlagSet("secret rotate", flag.ContinueOnError)
	why := addHandActFlags(set)
	if err := set.Parse(flags); err != nil {
		return err
	}
	if len(rest) != 3 || set.NArg() != 0 {
		return errors.New(secretUsage)
	}
	node, module, name := rest[0], rest[1], rest[2]
	open, err := openStores(ctx)
	if err != nil {
		return err
	}
	defer open.Close()
	origin, given, err := open.inventory.OwnSecretOrigin(ctx, node, module, name)
	if err != nil {
		return err
	}
	if err := open.inventory.RotateModuleSecret(ctx, node, module, name); err != nil {
		var refused inventory.ErrNotRotatable
		if errors.As(err, &refused) {
			return fmt.Errorf("not rotated: %s", refused.Why)
		}
		return err
	}
	why.record(ctx, "secret rotate", []string{node, module, name})
	fmt.Printf("rotated %q of %s on %s at %s, asked by %s; the value is sealed and not shown\n",
		name, module, node, time.Now().UTC().Format(time.RFC3339), whoAsked())
	if origin == inventory.OriginAccepted {
		fmt.Printf("  it replaces the value given to the mesh on %s: the mesh made this one, so `secret rotate` "+
			"replaces it again whenever asked (ADR 0228)\n", given.UTC().Format("2006-01-02"))
	}
	// A shared credential (ADR 0158) has as many holders as the provision has consumers, and all
	// of them are sent in one act, so no machine is left reading a value the provider no longer takes.
	machines, err := open.inventory.SharedHolders(ctx, node, module, name)
	if err != nil {
		return err
	}
	if len(machines) == 0 {
		machines = []string{node}
	}
	if len(machines) == 1 {
		fmt.Printf("sending %s, so %s starts again on the new value:\n", node, module)
	} else {
		fmt.Printf("shared with every consumer; sending %s together:\n", strings.Join(machines, ", "))
	}
	if err := sendTo(ctx, open, machines); err != nil {
		return fmt.Errorf("%w\n\nThe new value is sealed and not yet delivered; what runs keeps the old "+
			"one until the machines next apply. Fix the cause and run `push --behind`", err)
	}
	return nil
}

// whoAsked names the caller for the log: the account the command runs as, which for a tool call
// through the console is the mesh's own.
func whoAsked() string {
	if u := os.Getenv("SUDO_USER"); u != "" {
		return u
	}
	if u := os.Getenv("USER"); u != "" {
		return u
	}
	return "the mesh"
}