Fail empty ideas jobs, clear ideas on failure, show each idea as it grounds
CI / ci (push) Successful in 5m24s
CI / ci (pull_request) Successful in 6m11s

- A run that keeps no idea fails with no_ideas_found instead of ending
  done with an empty list.
- Before an ideas job fails it republishes its result with no ideas, so
  GET /api/v1/jobs/{id} never shows ideas on a failed job, even ones a
  previous grounding round already published.
- Each idea is published as soon as its grounding finishes, placed in the
  model's order. recipes.GroundEach reports items as they finish;
  GroundMany is rebuilt on it with unchanged semantics.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-10-08 19:22:51 -04:00
co-authored by Claude Opus 5.5
parent b566be7ddc
commit 4146f31275
6 changed files with 325 additions and 68 deletions
+139 -7
View File
@@ -131,6 +131,16 @@ type recorder struct {
published []Result
}
// latest returns the last result published.
func (r *recorder) latest() (Result, bool) {
r.mu.Lock()
defer r.mu.Unlock()
if len(r.published) == 0 {
return Result{}, false
}
return r.published[len(r.published)-1], true
}
func (r *recorder) publish(res Result) error {
r.mu.Lock()
defer r.mu.Unlock()
@@ -154,6 +164,17 @@ func (e *env) run(t *testing.T, ingredients []string, quickOnly bool) (Result, [
return res, rec.published, err
}
// within reports whether every name in sub appears in list, in list's
// order.
func within(sub, list []string) bool {
for _, name := range list {
if len(sub) > 0 && sub[0] == name {
sub = sub[1:]
}
}
return len(sub) == 0
}
func names(list []Idea) []string {
out := []string{}
for _, i := range list {
@@ -214,9 +235,13 @@ func TestResultFillsAsEachIdeaGrounds(t *testing.T) {
if !reflect.DeepEqual(stages, wantStages) || !reflect.DeepEqual(counts, []int{0, 0, 1, 2, 3, 4, 5}) {
t.Fatalf("published stages %q with %v ideas, want %q with 0, 0, 1, 2, 3, 4, 5", stages, counts, wantStages)
}
// Ideas ground concurrently, so one may appear before an earlier one,
// but each publish keeps every idea already shown, in the model's
// order.
for i := 3; i < len(published); i++ {
if !reflect.DeepEqual(published[i].Ideas[:i-2], published[i-1].Ideas) {
t.Errorf("publish %d changed ideas already published", i+1)
prev, cur := names(published[i-1].Ideas), names(published[i].Ideas)
if !within(prev, cur) || !within(cur, want) {
t.Errorf("publish %d shows %q after %q, want earlier ideas kept in the model's order %q", i+1, cur, prev, want)
}
}
@@ -342,15 +367,122 @@ func TestFailures(t *testing.T) {
}
}
func TestEveryIdeaUngroundableEndsDoneEmpty(t *testing.T) {
func TestEachIdeaShowsAsSoonAsItGrounds(t *testing.T) {
e := newEnv(t)
e.reply(idea("Mystery Stew", "stovetop", 30, "rice"))
res, _, err := e.run(t, pantry, false)
for _, name := range []string{"Burrito Bowls", "Fried Rice", "Black Bean Soup"} {
e.recipe(t, name, name, 35)
}
// The first idea's search waits, so the two after it ground first.
gate := make(chan struct{})
e.search.Gates = map[string]chan struct{}{"Burrito Bowls": gate}
e.reply(
idea("Burrito Bowls", "stovetop", 35, "rice", "black beans"),
idea("Fried Rice", "stovetop", 20, "rice"),
idea("Black Bean Soup", "stovetop", 40, "black beans"),
)
ings, err := skills.CleanIngredients(pantry)
if err != nil {
t.Fatal(err)
}
if res.Stage != StageDone || len(res.Ideas) != 0 || res.Dropped != 1 {
t.Fatalf("result = %+v, want done with none and 1 dropped", res)
var rec recorder
type outcome struct {
res Result
err error
}
done := make(chan outcome, 1)
go func() {
res, err := e.svc.run(t.Context(), FromIngredientsJob{Ingredients: ings}, rec.publish)
done <- outcome{res, err}
}()
deadline := time.Now().Add(10 * time.Second)
for {
if latest, ok := rec.latest(); ok && len(latest.Ideas) == 2 {
if got := names(latest.Ideas); !reflect.DeepEqual(got, []string{"Fried Rice", "Black Bean Soup"}) {
t.Fatalf("while the first idea grounds, ideas = %q, want the two grounded", got)
}
if latest.Stage != StageGrounding {
t.Fatalf("stage = %s, want grounding", latest.Stage)
}
break
}
if time.Now().After(deadline) {
t.Fatal("the ideas that grounded were not shown while the first idea was still grounding")
}
time.Sleep(5 * time.Millisecond)
}
close(gate)
out := <-done
if out.err != nil {
t.Fatal(out.err)
}
want := []string{"Burrito Bowls", "Fried Rice", "Black Bean Soup"}
if got := names(out.res.Ideas); !reflect.DeepEqual(got, want) || out.res.Stage != StageDone {
t.Fatalf("result = %s %q, want done with %q in the model's order", out.res.Stage, got, want)
}
}
func TestNoIdeaKeptFailsTheJob(t *testing.T) {
for _, tc := range []struct {
name string
quickOnly bool
setup func(t *testing.T, e *env)
}{
{"no recipe grounds any", false, func(t *testing.T, e *env) {
e.reply(idea("Mystery Stew", "stovetop", 30, "rice"), idea("Unknown Casserole", "oven", 50, "rice"))
}},
{"no grounded recipe is quick", true, func(t *testing.T, e *env) {
e.recipe(t, "Roast Chicken", "Roast Chicken", 90)
e.reply(idea("Roast Chicken", "oven", 25, "Chicken thighs"))
}},
} {
t.Run(tc.name, func(t *testing.T) {
e := newEnv(t)
tc.setup(t, e)
_, published, err := e.run(t, pantry, tc.quickOnly)
var failure *jobs.Failure
if !errors.As(err, &failure) || failure.Code != ReasonNoIdeasFound {
t.Fatalf("err = %v, want failure %s", err, ReasonNoIdeasFound)
}
for _, p := range published {
if len(p.Ideas) != 0 || p.Stage == StageDone {
t.Fatalf("a job that kept no idea published %+v", p)
}
}
})
}
}
func TestLaterRoundFailureClearsPublishedIdeas(t *testing.T) {
e := newEnv(t)
kept := []string{"Burrito Bowls", "Fried Rice", "Black Bean Soup", "Chicken Tacos"}
for _, name := range kept {
e.recipe(t, name, name, 35)
}
// The first round keeps four ideas; the second round's search fails.
e.search.Fail = map[string]error{"Rice Pilaf": errors.New("boom")}
e.reply(
idea("Burrito Bowls", "stovetop", 35, "rice"),
idea("Mystery Stew", "stovetop", 35, "rice"),
idea("Fried Rice", "stovetop", 20, "rice"),
idea("Black Bean Soup", "stovetop", 40, "black beans"),
idea("Chicken Tacos", "stovetop", 25, "Chicken thighs"),
idea("Rice Pilaf", "stovetop", 30, "rice"),
)
_, published, err := e.run(t, pantry, false)
var failure *jobs.Failure
if !errors.As(err, &failure) || failure.Code != ReasonSearchFailed {
t.Fatalf("err = %v, want failure %s", err, ReasonSearchFailed)
}
most := 0
for _, p := range published {
most = max(most, len(p.Ideas))
}
if most != len(kept) {
t.Fatalf("at most %d ideas published, want the first round's %d before the failure", most, len(kept))
}
if last := published[len(published)-1]; len(last.Ideas) != 0 {
t.Fatalf("last publish before the failure shows %q, want no ideas", names(last.Ideas))
}
}