feat: sync deployed v1.0.0-rc.38-custom.67
This commit is contained in:
1 parent
d09216bbde
commit
8e6ed0317b
154 files changed
+39540
-20509
No files matched your search
@@ -21,7 +21,7 @@ func TestAstraSamplingThirtyMinuteSchedulePersistsAcrossRestart(t *testing.T) {
|
||||
handler := &anchoredSamplingTestHandler{stubScheduledHandler{taskType: AstraSamplingTaskType, enabled: true, interval: 30 * time.Minute}}
|
||||
withSystemTaskRegistry(t, handler)
|
||||
now := common.GetTimestamp()
|
||||
state, _ := common.Marshal(AstraSamplingState{Rounds: []AstraSamplingRound{{ChannelID: 42, StartedAt: now - 1900, FinishedAt: now - 1800, Result: fingerprint.Result{Status: "match", ValidSamples: 3, Closest: fingerprint.Model}}}})
|
||||
state, _ := common.Marshal(AstraSamplingState{Rounds: []AstraSamplingRound{{ChannelID: 42, StartedAt: now - 1900, FinishedAt: now - 1800, Result: fingerprint.Result{Engine: fingerprint.Version, Status: "match", ValidSamples: 3, Closest: fingerprint.Model}}}})
|
||||
row := model.SystemTask{TaskID: "sampling-old", Type: AstraSamplingTaskType, Status: model.SystemTaskStatusSucceeded, CreatedAt: now - 1900, UpdatedAt: now - 10, State: string(state)}
|
||||
if err := model.DB.Create(&row).Error; err != nil {
|
||||
t.Fatal(err)
|
||||
@@ -61,9 +61,9 @@ func TestAstraSamplingAggregationPrivacyAndFailures(t *testing.T) {
|
||||
if status == "mismatch" {
|
||||
closest = "gpt-5.6-luna"
|
||||
}
|
||||
rounds = append(rounds, AstraSamplingRound{ChannelID: 1, StartedAt: now.Unix(), FinishedAt: now.Unix(), Result: fingerprint.Result{Status: status, ValidSamples: 3, Closest: closest}})
|
||||
rounds = append(rounds, AstraSamplingRound{ChannelID: 1, StartedAt: now.Unix(), FinishedAt: now.Unix(), Result: fingerprint.Result{Engine: fingerprint.Version, Status: status, ValidSamples: 3, Closest: closest}})
|
||||
}
|
||||
rounds = append(rounds, AstraSamplingRound{ChannelID: 99, StartedAt: now.Unix(), Result: fingerprint.Result{Status: "mismatch"}})
|
||||
rounds = append(rounds, AstraSamplingRound{ChannelID: 99, StartedAt: now.Unix(), Result: fingerprint.Result{Engine: fingerprint.Version, Status: "mismatch"}})
|
||||
state, _ := common.Marshal(AstraSamplingState{Rounds: rounds})
|
||||
summaries := AggregateAstraSampling(channels, []*model.SystemTask{{Status: model.SystemTaskStatusFailed, State: string(state)}}, now)
|
||||
first := summaries["first"]
|
||||
@@ -88,17 +88,17 @@ func TestAstraSamplingAggregationPrivacyAndFailures(t *testing.T) {
|
||||
}
|
||||
}
|
||||
|
||||
func TestAstraSamplingAggregationCountsLunaLikeInconclusiveWithinSixHours(t *testing.T) {
|
||||
func TestAstraSamplingExcludesInconclusiveFromRateWithinSixHours(t *testing.T) {
|
||||
now := time.Now()
|
||||
channel := model.Channel{Id: 1, Group: "example-variable", Models: fingerprint.Model}
|
||||
rounds := []AstraSamplingRound{
|
||||
{ID: "recent-match", ChannelID: 1, StartedAt: now.Add(-time.Hour).Unix(), FinishedAt: now.Unix(), Result: fingerprint.Result{Status: "match", ValidSamples: 3, Closest: fingerprint.Model}},
|
||||
{ID: "recent-luna", ChannelID: 1, StartedAt: now.Add(-2 * time.Hour).Unix(), FinishedAt: now.Unix(), Result: fingerprint.Result{Status: "inconclusive", ValidSamples: 2, Closest: "gpt-5.6-luna", Gap: .9}},
|
||||
{ID: "old-luna", ChannelID: 1, StartedAt: now.Add(-7 * time.Hour).Unix(), FinishedAt: now.Unix(), Result: fingerprint.Result{Status: "inconclusive", ValidSamples: 2, Closest: "gpt-5.6-luna", Gap: .9}},
|
||||
{ID: "recent-match", ChannelID: 1, StartedAt: now.Add(-time.Hour).Unix(), FinishedAt: now.Unix(), Result: fingerprint.Result{Engine: fingerprint.Version, Status: "match", ValidSamples: 3, Closest: fingerprint.Model}},
|
||||
{ID: "recent-luna", ChannelID: 1, StartedAt: now.Add(-2 * time.Hour).Unix(), FinishedAt: now.Unix(), Result: fingerprint.Result{Engine: fingerprint.Version, Status: "inconclusive", ValidSamples: 2, Closest: "gpt-5.6-luna", Gap: .9}},
|
||||
{ID: "old-luna", ChannelID: 1, StartedAt: now.Add(-7 * time.Hour).Unix(), FinishedAt: now.Unix(), Result: fingerprint.Result{Engine: fingerprint.Version, Status: "inconclusive", ValidSamples: 2, Closest: "gpt-5.6-luna", Gap: .9}},
|
||||
}
|
||||
state, _ := common.Marshal(AstraSamplingState{Rounds: rounds})
|
||||
summary := AggregateAstraSampling([]model.Channel{channel}, []*model.SystemTask{{Status: model.SystemTaskStatusSucceeded, State: string(state)}}, now)["example-variable"]
|
||||
if summary.Rate == nil || *summary.Rate != .5 || summary.Matches != 1 || summary.Mismatches != 1 || summary.Inconclusive != 0 {
|
||||
if summary.Rate == nil || *summary.Rate != 0 || summary.Matches != 1 || summary.Mismatches != 0 || summary.Inconclusive != 1 {
|
||||
t.Fatalf("%+v", summary)
|
||||
}
|
||||
if summary.WindowHours != 6 || summary.IntervalMinutes != 30 || len(summary.History) != 2 {
|
||||
@@ -107,6 +107,7 @@ func TestAstraSamplingAggregationCountsLunaLikeInconclusiveWithinSixHours(t *tes
|
||||
}
|
||||
|
||||
func TestAstraSamplingAggregationMarksNonAstraGroupsAvailabilityOnly(t *testing.T) {
|
||||
t.Setenv("CHANNEL_AVAILABILITY_INTERVAL_MINUTES", "30")
|
||||
now := time.Now()
|
||||
channels := []model.Channel{{Id: 1, Group: "opencode", Models: "gpt-5.6-sol"}}
|
||||
summary := AggregateAstraSampling(channels, nil, now)["opencode"]
|
||||
@@ -124,13 +125,12 @@ func TestAstraSamplingAggregationMarksNonAstraGroupsAvailabilityOnly(t *testing.
|
||||
}
|
||||
}
|
||||
|
||||
func TestAstraSamplingCountsAllValidNonAstraFingerprints(t *testing.T) {
|
||||
func TestAstraSamplingExcludesOldEngineAndUncertainCandidates(t *testing.T) {
|
||||
now := time.Now()
|
||||
results := []fingerprint.Result{
|
||||
{Status: "error", ValidSamples: 1, Closest: "gpt-5.6-luna", Gap: .99},
|
||||
{Status: "inconclusive", ValidSamples: 3, Closest: "gpt-5.5", Gap: .17},
|
||||
{Status: "inconclusive", ValidSamples: 2, Closest: "gpt-5.6-sol", Gap: .09},
|
||||
{Status: "error", ValidSamples: 0},
|
||||
{Engine: "meow-4.5.4", Status: "mismatch", ValidSamples: 32, Closest: "other"},
|
||||
{Engine: fingerprint.Version, Status: "inconclusive", ValidSamples: 2, Closest: "other"},
|
||||
{Engine: fingerprint.Version, Status: "match", ValidSamples: 3, Closest: fingerprint.Model},
|
||||
}
|
||||
var rounds []AstraSamplingRound
|
||||
for _, result := range results {
|
||||
@@ -140,16 +140,15 @@ func TestAstraSamplingCountsAllValidNonAstraFingerprints(t *testing.T) {
|
||||
require.NoError(t, err)
|
||||
summary := AggregateAstraSampling([]model.Channel{{Id: 1, Group: "test", Models: fingerprint.Model}}, []*model.SystemTask{{Status: model.SystemTaskStatusSucceeded, State: string(state)}}, now)["test"]
|
||||
require.NotNil(t, summary.Rate)
|
||||
assert.Equal(t, 1.0, *summary.Rate)
|
||||
assert.Equal(t, 3, summary.Mismatches)
|
||||
assert.Equal(t, 0, summary.Inconclusive)
|
||||
assert.Equal(t, 1, summary.Errors)
|
||||
for _, h := range summary.History {
|
||||
if h.ValidSamples > 0 {
|
||||
assert.Equal(t, "mismatch", h.Status)
|
||||
}
|
||||
assert.Equal(t, 0.0, *summary.Rate)
|
||||
assert.Equal(t, 1, summary.Matches)
|
||||
assert.Equal(t, 0, summary.Mismatches)
|
||||
assert.Equal(t, 1, summary.Inconclusive)
|
||||
assert.Len(t, summary.History, 2)
|
||||
}
|
||||
|
||||
func TestModelTraceVerdictDoesNotPromoteClosestCandidate(t *testing.T) {
|
||||
for _, status := range []string{"match", "mismatch", "inconclusive", "error"} {
|
||||
assert.Equal(t, status, AstraSamplingResultStatus(fingerprint.Result{Engine: fingerprint.Version, Status: status, ValidSamples: 32, Closest: fingerprint.Model, Weight: .63436}))
|
||||
}
|
||||
assert.Equal(t, "match", AstraSamplingResultStatus(fingerprint.Result{Status: "inconclusive", ValidSamples: 1, Closest: fingerprint.Model}))
|
||||
assert.Equal(t, "inconclusive", AstraSamplingResultStatus(fingerprint.Result{Status: "mismatch", ValidSamples: 0, Closest: "gpt-5.6-sol"}))
|
||||
assert.Equal(t, "running", AstraSamplingResultStatus(fingerprint.Result{Status: "running", ValidSamples: 1, Closest: "gpt-5.6-sol"}))
|
||||
}
|
||||
Reference in new issue
Block a user