v3 workflow: intent, phases, review, submission, enforcement, burn-in

The v3 stack, previously an uncommitted working tree, plus this session's two
units and the burn-in instrument. This commit is the burn-in build identity:
coordinator and worker must both report this revision before a task is created.

Workflow (earlier sessions, uncommitted until now): human decision events and
reduction, source cursors and reconcile-before-launch, turn-boundary
reconciliation, internal/agentctx as the single renderer, ace-fca phases with
sealed artifacts, the trajectory gate, bounded grilling, independent review,
task pr enforcement, and human review reflection.

Capability restrictions at the agent boundary: an authz.Agent surface at
GatedWrite may ask and may not act. It also fixes two bugs the unit exposed --
gated surfaces could not reach the two endpoints written for them, and
RequestHumanDecision would block an unowned task while rejecting a question
from the session that did own it.

Turn-boundary reconcile-failure escalation: a streak of consecutive failures
asks the session to hand off, fenced on the lease epoch, with reconcile_failure
as a real handoff reason. The worker was dropping the coordinator's verdict on
the floor; it now acts on it.

Burn-in: herdr.WriteLaunchContext dumps the exact agentctx.Build result to
<worktree>/.orchestra/launch.md at every launch, local and federated. BURNIN.md
is the runbook. deploy/build.sh stamps both binaries from one commit.

go build, go vet and go test ./... pass, 20 packages.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
2026-08-26 18:31:20 +04:00
parent 97a9c65302
commit 7f12c7fc37
78 changed files with 16417 additions and 352 deletions
+336
View File
@@ -0,0 +1,336 @@
package herdr
import (
"context"
"crypto/sha256"
"errors"
"fmt"
"os"
"os/exec"
"path/filepath"
"strings"
"time"
)
// TmuxBackend runs one Claude Code process per isolated tmux session. It is
// intentionally Claude-only for now: Codex and OpenCode keep using the
// verified herdr protocol until their terminal behavior has been exercised
// against a live installation.
type TmuxBackend struct {
// Socket is a tmux socket name (-L) or an absolute socket path (-S).
// An empty value uses the isolated socket name "orchestra".
Socket string
// Command is the Claude Code executable. An empty value resolves "claude"
// through PATH.
Command string
// Binary is test/packaging override for tmux itself.
Binary string
}
func NewTmuxBackend(socket, command string) *TmuxBackend {
return &TmuxBackend{Socket: socket, Command: command}
}
func (b *TmuxBackend) Kind() string { return "tmux" }
func (b *TmuxBackend) binary() string {
if b.Binary != "" {
return b.Binary
}
return "tmux"
}
func (b *TmuxBackend) socketArgs() []string {
socket := b.Socket
if socket == "" {
socket = "orchestra"
}
if filepath.IsAbs(socket) {
return []string{"-S", socket}
}
return []string{"-L", socket}
}
func (b *TmuxBackend) command(ctx context.Context, args ...string) ([]byte, error) {
all := append(b.socketArgs(), args...)
out, err := exec.CommandContext(ctx, b.binary(), all...).CombinedOutput()
if err != nil {
return out, fmt.Errorf("tmux %s: %s: %w", strings.Join(args, " "), strings.TrimSpace(string(out)), err)
}
return out, nil
}
func (b *TmuxBackend) Check(ctx context.Context) error {
if _, err := exec.LookPath(b.binary()); err != nil {
return fmt.Errorf("tmux backend: %w", err)
}
// tmux -V does not require a server to exist. An idle backend is healthy
// and will create its isolated server with the first session.
if out, err := exec.CommandContext(ctx, b.binary(), "-V").CombinedOutput(); err != nil {
return fmt.Errorf("tmux backend: %s: %w", strings.TrimSpace(string(out)), err)
}
command := b.Command
if command == "" {
command = "claude"
}
if _, err := exec.LookPath(command); err != nil {
return fmt.Errorf("tmux backend Claude command: %w", err)
}
return nil
}
func (b *TmuxBackend) Worktree(_ context.Context, _ string, path, _ string) (string, error) {
info, err := os.Stat(path)
if err != nil {
return "", fmt.Errorf("tmux backend worktree: %w", err)
}
if !info.IsDir() {
return "", fmt.Errorf("tmux backend worktree %s is not a directory", path)
}
return path, nil
}
func tmuxSessionName(taskID string) string {
id := strings.Trim(invalidAgentName.ReplaceAllString(strings.ToLower(taskID), "-"), "-_")
if id == "" {
id = "session"
}
if len(id) > 36 {
id = strings.TrimRight(id[:36], "-_")
}
sum := sha256.Sum256([]byte(taskID))
return fmt.Sprintf("orchestra-%s-%x", id, sum[:4])
}
func tmuxSession(paneID string) string {
if before, _, ok := strings.Cut(paneID, ":"); ok {
return before
}
return paneID
}
func tmuxTarget(paneID string) string { return "=" + paneID }
func (b *TmuxBackend) hasSession(ctx context.Context, session string) (bool, error) {
out, err := b.command(ctx, "has-session", "-t", "="+session)
if err == nil {
return true, nil
}
message := strings.ToLower(string(out) + " " + err.Error())
if strings.Contains(message, "can't find session") || strings.Contains(message, "no server running") || strings.Contains(message, "no sessions") || (strings.Contains(message, "error connecting to") && strings.Contains(message, "no such file")) {
return false, nil
}
return false, err
}
func (b *TmuxBackend) StartAgent(ctx context.Context, _, path, _, harness, taskID string) (Session, error) {
if !strings.EqualFold(harness, "claude") {
return Session{}, fmt.Errorf("tmux backend: harness %q is unsupported; only claude is enabled", harness)
}
if _, err := b.Worktree(ctx, "", path, ""); err != nil {
return Session{}, err
}
command := b.Command
if command == "" {
command = "claude"
}
resolved, err := exec.LookPath(command)
if err != nil {
return Session{}, fmt.Errorf("tmux backend Claude command: %w", err)
}
session := tmuxSessionName(taskID)
existing, err := b.hasSession(ctx, session)
if err != nil {
return Session{}, err
}
if !existing {
if _, err := b.command(ctx, "new-session", "-d", "-s", session, "-c", path, resolved); err != nil {
return Session{}, err
}
}
// The exact pane id is resolved below. Deployments may configure tmux
// base-index/base-pane-index, so neither index is assumed to be zero.
paneOut, err := b.command(ctx, "list-panes", "-t", "="+session, "-F", "#{session_name}:#{window_index}.#{pane_index}")
if err != nil {
return Session{}, err
}
paneID := strings.TrimSpace(strings.SplitN(string(paneOut), "\n", 2)[0])
if paneID == "" {
return Session{}, fmt.Errorf("tmux backend: session %s has no pane", session)
}
if existing {
out, err := b.command(ctx, "display-message", "-p", "-t", tmuxTarget(paneID), "#{pane_current_path}")
if err != nil {
return Session{}, err
}
current, currentErr := filepath.Abs(strings.TrimSpace(string(out)))
want, wantErr := filepath.Abs(path)
if currentErr != nil || wantErr != nil || current != want {
return Session{}, fmt.Errorf("tmux backend: existing session %s belongs to %q, not %q", session, current, want)
}
} else {
if _, err := b.command(ctx, "set-window-option", "-t", tmuxTarget(paneID), "remain-on-exit", "on"); err != nil {
return Session{}, err
}
}
s := Session{PaneID: paneID, Worktree: path, Harness: "claude", AgentName: session}
if err := b.confirmClaudeWorkspaceTrust(ctx, s); err != nil {
return Session{}, err
}
status, err := b.AgentStatus(ctx, s)
if err != nil {
return Session{}, err
}
if status == "exited" || status == "dead" {
return Session{}, fmt.Errorf("tmux backend: Claude exited while starting session %s", session)
}
return s, nil
}
func (b *TmuxBackend) confirmClaudeWorkspaceTrust(ctx context.Context, s Session) error {
deadline := time.Now().Add(15 * time.Second)
accepted := false
for {
text, err := b.PaneCapture(ctx, s, "recent")
if err == nil && claudeWorkspaceTrustPrompt(text) {
if !accepted {
if err := b.SendText(ctx, s, "1"); err != nil {
return fmt.Errorf("tmux backend: accept Claude workspace trust: %w", err)
}
if err := b.SendKeys(ctx, s, []string{"Enter"}); err != nil {
return fmt.Errorf("tmux backend: accept Claude workspace trust: %w", err)
}
accepted = true
}
}
// Claude's input prompt is the readiness boundary. A banner or partially
// painted fullscreen UI is not enough: input sent there can be lost.
if err == nil && !claudeWorkspaceTrustPrompt(text) && strings.Contains(text, "") {
return nil
}
if time.Now().After(deadline) {
return fmt.Errorf("tmux backend: Claude input prompt did not become ready in session %s", tmuxSession(s.PaneID))
}
select {
case <-ctx.Done():
return ctx.Err()
case <-time.After(100 * time.Millisecond):
}
}
}
func (b *TmuxBackend) Prompt(ctx context.Context, pane, text string, _ time.Duration) error {
s := Session{PaneID: pane}
status, err := b.AgentStatus(ctx, s)
if err != nil {
return err
}
if status == "blocked" {
return fmt.Errorf("tmux backend: refusing prompt while pane %s shows a permission dialog", pane)
}
if err := b.SendText(ctx, s, text); err != nil {
return err
}
return b.SendKeys(ctx, s, []string{"Enter"})
}
func (b *TmuxBackend) Kill(ctx context.Context, s Session) error {
exists, err := b.hasSession(ctx, tmuxSession(s.PaneID))
if err != nil {
return err
}
if !exists {
return nil
}
_, err = b.command(ctx, "kill-session", "-t", "="+tmuxSession(s.PaneID))
return err
}
func (b *TmuxBackend) paneState(ctx context.Context, s Session) (dead bool, command string, err error) {
out, err := b.command(ctx, "display-message", "-p", "-t", tmuxTarget(s.PaneID), "#{pane_dead}\t#{pane_current_command}")
if err != nil {
return false, "", err
}
parts := strings.SplitN(strings.TrimSpace(string(out)), "\t", 2)
dead = len(parts) > 0 && parts[0] == "1"
if len(parts) == 2 {
command = parts[1]
}
return dead, command, nil
}
func (b *TmuxBackend) AgentStatus(ctx context.Context, s Session) (string, error) {
exists, err := b.hasSession(ctx, tmuxSession(s.PaneID))
if err != nil {
return "", err
}
if !exists {
return "exited", nil
}
dead, command, err := b.paneState(ctx, s)
if err != nil {
return "", err
}
if dead || command == "" {
return "exited", nil
}
text, err := b.PaneCapture(ctx, s, "recent")
if err != nil {
return "", err
}
if permissionPrompt(text) {
return "blocked", nil
}
lower := strings.ToLower(text)
for _, marker := range []string{"esc to interrupt", "ctrl+c to interrupt", "press esc to interrupt"} {
if strings.Contains(lower, marker) {
return "busy", nil
}
}
return "idle", nil
}
func (b *TmuxBackend) PaneCapture(ctx context.Context, s Session, source string) (string, error) {
start := "-200"
if source != "" && source != "recent" {
start = "-1000"
}
out, err := b.command(ctx, "capture-pane", "-p", "-J", "-S", start, "-t", tmuxTarget(s.PaneID))
return string(out), err
}
func (b *TmuxBackend) SendText(ctx context.Context, s Session, text string) error {
_, err := b.command(ctx, "send-keys", "-t", tmuxTarget(s.PaneID), "-l", "--", text)
return err
}
func (b *TmuxBackend) SendKeys(ctx context.Context, s Session, keys []string) error {
if len(keys) == 0 {
return nil
}
for _, key := range keys {
if strings.TrimSpace(key) == "" {
return errors.New("tmux backend: empty key name")
}
}
args := []string{"send-keys", "-t", tmuxTarget(s.PaneID)}
args = append(args, keys...)
_, err := b.command(ctx, args...)
return err
}
// tmux has no separate agent binding to release. Keeping the session alive is
// the tmux equivalent of herdr's split-then-close protocol; the worker kills
// it only after successor pickup has been validated.
func (b *TmuxBackend) ReleaseAgent(ctx context.Context, s Session, _ string) error {
exists, err := b.hasSession(ctx, tmuxSession(s.PaneID))
if err != nil {
return err
}
if !exists {
return fmt.Errorf("tmux backend: session %s is not running", tmuxSession(s.PaneID))
}
return nil
}
var _ Backend = (*TmuxBackend)(nil)