feat(dispatcher): enforce depends_on with .blocked marker (Phase 2/G2)

Before claiming a session for a task, the dispatcher now:
1. Parses the task's frontmatter
2. If `depends_on: [project:task_id]` is non-empty, checks each entry
   against `<projectsDir>/<project>/.agent-queue/done/<task_id>.md`
3. If any dep is unresolved -> skip the task and write
   `<task>.md.blocked` next to it. The watchdog (G1) will resolve
   this marker on its next tick.

The `.blocked` marker is idempotent: re-running the dispatcher does not
refresh its mtime, so the watchdog can compute the blocked-since
timestamp from the FIRST detection (timeout precision).

Path-traversal hardening: project / task_id segments must match
`[A-Za-z0-9._-]+` and cannot be `.` or `..`. A malicious frontmatter
like `depends_on: ../../tmp:foo` is rejected before any filesystem
lookup.

assignNextTask (the doneChan path) applies the same gate so that a
session freed mid-cycle cannot bypass enforcement.

Tests (-race clean):
- DependsOnUnresolved -> .blocked marker, no dispatch
- DependsOnResolved -> normal dispatch, no marker
- PartialResolution -> stay blocked
- RejectPathTraversal -> blocked, not dispatched
- BlockedMarker idempotent (mtime stable across passes)
- NoDependsOn regression guard
This commit is contained in:
Ubuntu 2026-04-16 20:30:17 +00:00
parent 47ab86eef9
commit 5cfb58c202
3 changed files with 439 additions and 9 deletions

View file

@ -22,10 +22,16 @@ import (
// TaskFrontmatter is the YAML header parsed from task .md files.
type TaskFrontmatter struct {
Title string `yaml:"title"`
Priority string `yaml:"priority"` // critical, high, default, low
Tags []string `yaml:"tags"`
NeedsClaude bool `yaml:"needs_claude_code"`
Title string `yaml:"title"`
Priority string `yaml:"priority"` // critical, high, default, low
Tags []string `yaml:"tags"`
NeedsClaude bool `yaml:"needs_claude_code"`
// Phase 2/G2: cross-project dependencies. Each entry is
// "project:task_id" (e.g. "filesecure:FIX-0123"). The dispatcher
// refuses to launch a task whose deps aren't all in the target
// project's done/, and drops a `<task>.md.blocked` marker so the
// watchdog can resolve it later (Phase 2/G1).
DependsOn []string `yaml:"depends_on"`
}
// Dispatcher watches project inbox directories and assigns tasks to idle sessions.
@ -121,15 +127,26 @@ func (d *Dispatcher) dispatchProject(inboxDir string) {
projectDir := filepath.Dir(filepath.Dir(inboxDir)) // inboxDir/.agent-queue/inbox → project
for _, e := range entries {
name := e.Name()
if !strings.HasSuffix(name, ".md") || strings.Contains(name, ".dispatched") {
// Skip non-md files, .dispatched markers, .blocked markers, and any
// other sidecar (.stuck, .tmp, etc).
if !strings.HasSuffix(name, ".md") {
continue
}
taskPath := filepath.Join(inboxDir, name)
// Phase 2/G2: enforce depends_on before claiming a session. The
// session pool is a precious resource — we don't want to burn an
// idle slot on a task that can't proceed.
if d.taskBlocked(taskPath) {
d.touchBlockedMarker(taskPath)
continue
}
session := d.findFreeSession()
if session == "" {
d.logger.Printf("[dispatcher] no free session for task in %s", inboxDir)
return
}
taskPath := filepath.Join(inboxDir, name)
if err := d.launchAgent(session, projectDir, taskPath); err != nil {
d.logger.Printf("[dispatcher] launchAgent error: %v", err)
continue
@ -140,6 +157,106 @@ func (d *Dispatcher) dispatchProject(inboxDir string) {
}
}
// taskBlocked returns true when taskPath declares `depends_on` entries that
// are not yet present in the target project's `.agent-queue/done/`. The
// frontmatter is parsed best-effort; on parse failure we treat the task as
// non-blocked (current behaviour preserved — bad frontmatter is the
// agent's problem, not the dispatcher's).
func (d *Dispatcher) taskBlocked(taskPath string) bool {
content, err := os.ReadFile(taskPath)
if err != nil {
return false
}
fm, _ := parseFrontmatter(content)
if len(fm.DependsOn) == 0 {
return false
}
for _, dep := range fm.DependsOn {
if !d.dependencyResolved(dep) {
d.logger.Printf("[dispatcher] task %s blocked by unresolved dep %q",
filepath.Base(taskPath), dep)
return true
}
}
return false
}
// dependencyResolved checks whether `project:task_id` is in
// `<projectsDir>/<project>/.agent-queue/done/<task_id>.md`. Accepts an
// exact match or a `<task_id>*.md` prefix match (some queues append a
// timestamp). Path segments are validated (no `..`, no `/` inside the
// segment) — see isSafeSegment.
func (d *Dispatcher) dependencyResolved(dep string) bool {
parts := strings.SplitN(dep, ":", 2)
if len(parts) != 2 {
d.logger.Printf("[dispatcher] malformed depends_on %q (expected 'project:task_id')", dep)
return false
}
project, taskID := parts[0], parts[1]
if !isSafeSegment(project) || !isSafeSegment(taskID) {
d.logger.Printf("[dispatcher] unsafe depends_on segment %q — refusing to look up", dep)
return false
}
doneDir := filepath.Join(d.projectsDir, project, ".agent-queue", "done")
exact := filepath.Join(doneDir, taskID+".md")
if _, err := os.Stat(exact); err == nil {
return true
}
entries, err := os.ReadDir(doneDir)
if err != nil {
return false
}
for _, e := range entries {
if e.IsDir() {
continue
}
name := e.Name()
if strings.HasPrefix(name, taskID) && strings.HasSuffix(name, ".md") &&
!strings.HasSuffix(name, ".dispatched") {
return true
}
}
return false
}
// touchBlockedMarker creates `<taskPath>.blocked` (or refreshes its mtime)
// so the watchdog (Phase 2/G1) sees the task as blocked and tracks its
// timeout. We use mtime as the "blocked since" timestamp; we do NOT
// refresh it on subsequent ticks — the operator wants the timeout to be
// counted from the FIRST detection, not the last.
func (d *Dispatcher) touchBlockedMarker(taskPath string) {
marker := taskPath + ".blocked"
if _, err := os.Stat(marker); err == nil {
return // marker already exists, leave mtime alone
}
if err := os.WriteFile(marker, []byte(""), 0o644); err != nil {
d.logger.Printf("[dispatcher] write .blocked marker for %s: %v",
filepath.Base(taskPath), err)
return
}
d.logger.Printf("[dispatcher] task %s marked .blocked (waiting on depends_on)",
filepath.Base(taskPath))
}
// isSafeSegment guards the project / task_id pair against path traversal.
// Same rule as the watchdog (`[A-Za-z0-9._-]+`, no `.` / `..`).
func isSafeSegment(s string) bool {
if s == "" || s == "." || s == ".." {
return false
}
for _, r := range s {
switch {
case r >= 'a' && r <= 'z',
r >= 'A' && r <= 'Z',
r >= '0' && r <= '9',
r == '_', r == '-', r == '.':
default:
return false
}
}
return true
}
// findFreeSession returns the name of an idle, live, cooldown-free session
// from the autonomous pool. Dedicated sessions are intentionally NOT
// considered: those host the operator's manual interactive work. Routing a
@ -184,10 +301,17 @@ func (d *Dispatcher) assignNextTask(session string) {
continue
}
for _, e := range entries {
if !strings.HasSuffix(e.Name(), ".md") || strings.Contains(e.Name(), ".dispatched") {
name := e.Name()
if !strings.HasSuffix(name, ".md") {
continue
}
taskPath := filepath.Join(inbox, name)
// Phase 2/G2: respect depends_on here too, otherwise a
// session freed mid-cycle would still bypass the gate.
if d.taskBlocked(taskPath) {
d.touchBlockedMarker(taskPath)
continue
}
taskPath := filepath.Join(inbox, e.Name())
if err := d.launchAgent(session, ds.Project, taskPath); err == nil {
os.Rename(taskPath, taskPath+".dispatched")
return