mirror of
https://github.com/BaaRai/ARTEX.git
synced 2026-11-16 13:10:53 +08:00
Compare commits
13
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
5463ea4375 | ||
|
|
5a8498bb4f | ||
|
|
2ad8ba16c3 | ||
|
|
09c9e6cfc2 | ||
|
|
59c3002bcc | ||
|
|
68f2db1d0b | ||
|
|
f0be3ae51f | ||
|
|
3b755228d6 | ||
|
|
d147dce344 | ||
|
|
60728092e1 | ||
|
|
ff599c65d3 | ||
|
|
ce096cdba1 | ||
|
|
50f6d07640 |
@@ -48,6 +48,19 @@ jobs:
|
||||
# as rotating failures in the agent package). Serial package execution
|
||||
# keeps the database state deterministic.
|
||||
run: go test -p 1 ./...
|
||||
- name: online index scripts
|
||||
# Apply and roll back scripts/sql on a database built here from
|
||||
# db/schema.sql and db/migrations. The job-level artextest database
|
||||
# never gets the schema (packages use suffixed databases), and cached
|
||||
# test results mean no test database is guaranteed to exist.
|
||||
env:
|
||||
SQL_DSN: postgres://artex:testpass@localhost:5432/artextest_sqlscripts?sslmode=disable
|
||||
run: |
|
||||
psql "$ARTEX_PG_DSN" -X -v ON_ERROR_STOP=1 -c 'CREATE DATABASE artextest_sqlscripts'
|
||||
psql "$SQL_DSN" -X -q -v ON_ERROR_STOP=1 -f db/schema.sql
|
||||
for f in db/migrations/*.sql; do psql "$SQL_DSN" -X -q -v ON_ERROR_STOP=1 -f "$f"; done
|
||||
psql "$SQL_DSN" -X -v ON_ERROR_STOP=1 -f scripts/sql/20261115-task-query-indexes.sql
|
||||
psql "$SQL_DSN" -X -v ON_ERROR_STOP=1 -f scripts/sql/20261115-task-query-indexes.rollback.sql
|
||||
- name: go test -race
|
||||
# Race detector on the packages that pass under -race today. ./server
|
||||
# joined once its suite was checked with -race against postgres and its
|
||||
|
||||
+17
-4
@@ -39,10 +39,17 @@ func enableNoa(opts *agentcore.Options, enabled func() bool, archiveRoot, sessio
|
||||
if opts.OnWarn == nil {
|
||||
opts.OnWarn = onWarn
|
||||
}
|
||||
if err := noaadapter.Enable(opts, noaadapter.Options{
|
||||
ArchiveBaseDir: filepath.Join(archiveRoot, "noa"),
|
||||
SessionID: sessionID,
|
||||
OnWarn: onWarn,
|
||||
// noa sizes its thresholds to the window this agent is configured for, read
|
||||
// before Compaction is cleared below; without it noa assumes its own default.
|
||||
window := defaultWindowK * 1000
|
||||
if opts.Compaction != nil && opts.Compaction.ContextWindow > 0 {
|
||||
window = opts.Compaction.ContextWindow
|
||||
}
|
||||
if err := noaEnable(opts, noaadapter.Options{
|
||||
ArchiveBaseDir: filepath.Join(archiveRoot, "noa"),
|
||||
SessionID: sessionID,
|
||||
OnWarn: onWarn,
|
||||
ModelContextLimit: window,
|
||||
}); err != nil {
|
||||
if onWarn != nil {
|
||||
onWarn("failed to enable noa compression, falling back to built-in compaction: " + err.Error())
|
||||
@@ -52,4 +59,10 @@ func enableNoa(opts *agentcore.Options, enabled func() bool, archiveRoot, sessio
|
||||
// Compactor overrides Compaction, but agentcore warns on every run when both are
|
||||
// set; clear it explicitly.
|
||||
opts.Compaction = nil
|
||||
// noa can only trim what it may drop; a request whose protected part alone
|
||||
// exceeds the window is refused here instead of by the provider.
|
||||
opts.Provider = &noaBudgetProvider{Provider: opts.Provider, window: window}
|
||||
}
|
||||
|
||||
// noaEnable is noaadapter.Enable; tests replace it to observe the options.
|
||||
var noaEnable = noaadapter.Enable
|
||||
|
||||
@@ -0,0 +1,50 @@
|
||||
package agent
|
||||
|
||||
import (
|
||||
"context"
|
||||
"errors"
|
||||
"fmt"
|
||||
"iter"
|
||||
|
||||
"github.com/Autumn-27/artex/sidequestion"
|
||||
"github.com/Autumn-27/norma/llm"
|
||||
)
|
||||
|
||||
// noaRequestOverhead covers request framing the input estimate does not count.
|
||||
const noaRequestOverhead = 512
|
||||
|
||||
// ErrNoaBudget marks a request refused by the NOA budget check. The refusal is
|
||||
// deterministic (the same history fails the same way), so callers must not
|
||||
// retry it.
|
||||
var ErrNoaBudget = errors.New("noa request budget exceeded")
|
||||
|
||||
// noaBudgetProvider refuses, before any transport, a request that cannot fit the
|
||||
// window noa was configured with. noa trims what it is allowed to drop; when
|
||||
// the protected part (system prompt, recent turns, the output reservation)
|
||||
// alone is over the window, sending it only buys a provider rejection.
|
||||
type noaBudgetProvider struct {
|
||||
llm.Provider
|
||||
window int
|
||||
}
|
||||
|
||||
func (p *noaBudgetProvider) check(req llm.CompletionRequest) error {
|
||||
estimate := sidequestion.EstimateInputTokens(req) + max(req.MaxTokens, 0) + noaRequestOverhead
|
||||
if estimate > p.window {
|
||||
return fmt.Errorf("%w: estimate %d exceeds configured budget %d; protected context cannot fit, increase the configured window or start a new session", ErrNoaBudget, estimate, p.window)
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func (p *noaBudgetProvider) Complete(ctx context.Context, req llm.CompletionRequest) (llm.Message, string, llm.Usage, error) {
|
||||
if err := p.check(req); err != nil {
|
||||
return llm.Message{}, "", llm.Usage{}, err
|
||||
}
|
||||
return p.Provider.Complete(ctx, req)
|
||||
}
|
||||
|
||||
func (p *noaBudgetProvider) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] {
|
||||
if err := p.check(req); err != nil {
|
||||
return func(yield func(llm.StreamEvent, error) bool) { yield(llm.StreamEvent{}, err) }
|
||||
}
|
||||
return p.Provider.Stream(ctx, req)
|
||||
}
|
||||
@@ -0,0 +1,155 @@
|
||||
package agent
|
||||
|
||||
import (
|
||||
"context"
|
||||
"errors"
|
||||
"iter"
|
||||
"strings"
|
||||
"testing"
|
||||
|
||||
"github.com/Autumn-27/norma/agentcore"
|
||||
"github.com/Autumn-27/norma/harness"
|
||||
"github.com/Autumn-27/norma/llm"
|
||||
"github.com/Autumn-27/norma/noaadapter"
|
||||
)
|
||||
|
||||
// A worker whose protected context exceeds a tiny NOA window ends with
|
||||
// model_error and an error that still wraps ErrNoaBudget after the Norma harness
|
||||
// and the capture layer: the engine's retry check relies on that.
|
||||
func TestWorkerNoaBudgetRefusalKeepsSentinelThroughHarness(t *testing.T) {
|
||||
f := newSessionFixture(t)
|
||||
inner := &countingProvider{}
|
||||
w := NewWorker(inner, "test-model", t.TempDir(), nil, 2000, 3)
|
||||
w.SetNoaEnabled(func() bool { return true })
|
||||
reason, _, err := w.Execute(t.Context(), "work#1", f.expID, nil, f.ts, f.intent, passHooks{}, nil, nil, nil)
|
||||
if reason != harness.ReasonModelError || !errors.Is(err, ErrNoaBudget) || inner.calls != 0 {
|
||||
t.Fatalf("reason=%q err=%v provider calls=%d; want model_error wrapping ErrNoaBudget and no call", reason, err, inner.calls)
|
||||
}
|
||||
}
|
||||
|
||||
// countingProvider answers with one text turn and counts the calls it receives.
|
||||
type countingProvider struct{ calls int }
|
||||
|
||||
func (p *countingProvider) Stream(context.Context, llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] {
|
||||
p.calls++
|
||||
return func(yield func(llm.StreamEvent, error) bool) {
|
||||
if yield(llm.StreamEvent{Type: llm.SETextDelta, Text: "ok"}, nil) {
|
||||
yield(llm.StreamEvent{Type: llm.SEMessageStop}, nil)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
func (p *countingProvider) Complete(context.Context, llm.CompletionRequest) (llm.Message, string, llm.Usage, error) {
|
||||
p.calls++
|
||||
return llm.Message{Role: llm.RoleAssistant, Content: []llm.ContentBlock{llm.TextBlock("ok")}}, "end_turn", llm.Usage{}, nil
|
||||
}
|
||||
|
||||
// withNoaEnable swaps the noaadapter.Enable seam for the duration of a test.
|
||||
func withNoaEnable(t *testing.T, fn func(*agentcore.Options, noaadapter.Options) error) {
|
||||
t.Helper()
|
||||
prev := noaEnable
|
||||
noaEnable = fn
|
||||
t.Cleanup(func() { noaEnable = prev })
|
||||
}
|
||||
|
||||
func TestEnableNoaUsesConfiguredWindow(t *testing.T) {
|
||||
for _, tc := range []struct {
|
||||
name string
|
||||
compaction func() *agentcore.Options
|
||||
want int
|
||||
}{
|
||||
{"configured", func() *agentcore.Options { return &agentcore.Options{Compaction: compactionConfig(32000)} }, 32000},
|
||||
{"default", func() *agentcore.Options { return &agentcore.Options{} }, defaultWindowK * 1000},
|
||||
} {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
var got noaadapter.Options
|
||||
withNoaEnable(t, func(opts *agentcore.Options, o noaadapter.Options) error {
|
||||
got = o
|
||||
return noaadapter.Enable(opts, o)
|
||||
})
|
||||
inner := &countingProvider{}
|
||||
opts := tc.compaction()
|
||||
opts.Provider = inner
|
||||
enableNoa(opts, func() bool { return true }, t.TempDir(), "noa-window", nil)
|
||||
|
||||
if got.ModelContextLimit != tc.want {
|
||||
t.Fatalf("noaadapter.Options.ModelContextLimit=%d, want %d", got.ModelContextLimit, tc.want)
|
||||
}
|
||||
bp, ok := opts.Provider.(*noaBudgetProvider)
|
||||
if !ok || bp.window != tc.want || bp.Provider != inner {
|
||||
t.Fatalf("provider %T %+v, want *noaBudgetProvider{window %d} around the original", opts.Provider, opts.Provider, tc.want)
|
||||
}
|
||||
if opts.Compaction != nil || opts.Compactor == nil {
|
||||
t.Fatalf("Compaction=%v Compactor=%v; noa must replace the built-in compaction", opts.Compaction, opts.Compactor)
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
func irreducibleRequest() llm.CompletionRequest {
|
||||
return llm.CompletionRequest{System: []string{strings.Repeat("protected system prompt ", 8000)}, MaxTokens: 8000}
|
||||
}
|
||||
|
||||
func TestNoaBudgetRefusesIrreducibleRequestBeforeTransport(t *testing.T) {
|
||||
t.Run("atomic", func(t *testing.T) {
|
||||
inner := &countingProvider{}
|
||||
p := &noaBudgetProvider{Provider: inner, window: 32000}
|
||||
_, _, _, err := p.Complete(t.Context(), irreducibleRequest())
|
||||
if err == nil || !strings.Contains(err.Error(), "exceeds configured budget 32000") || inner.calls != 0 {
|
||||
t.Fatalf("err=%v provider calls=%d", err, inner.calls)
|
||||
}
|
||||
if !errors.Is(err, ErrNoaBudget) {
|
||||
t.Fatalf("atomic refusal does not wrap ErrNoaBudget: %v", err)
|
||||
}
|
||||
if _, _, _, err := p.Complete(t.Context(), llm.CompletionRequest{System: []string{"short"}, MaxTokens: 1000}); err != nil || inner.calls != 1 {
|
||||
t.Fatalf("a request that fits was refused or not sent: err=%v calls=%d", err, inner.calls)
|
||||
}
|
||||
})
|
||||
t.Run("stream", func(t *testing.T) {
|
||||
inner := &countingProvider{}
|
||||
p := &noaBudgetProvider{Provider: inner, window: 32000}
|
||||
var events int
|
||||
var err error
|
||||
for ev, e := range p.Stream(t.Context(), irreducibleRequest()) {
|
||||
if e != nil {
|
||||
err = e
|
||||
break
|
||||
}
|
||||
_ = ev
|
||||
events++
|
||||
}
|
||||
if err == nil || !strings.Contains(err.Error(), "exceeds configured budget 32000") || events != 0 || inner.calls != 0 {
|
||||
t.Fatalf("err=%v events=%d provider calls=%d", err, events, inner.calls)
|
||||
}
|
||||
if !errors.Is(err, ErrNoaBudget) {
|
||||
t.Fatalf("stream refusal does not wrap ErrNoaBudget: %v", err)
|
||||
}
|
||||
for _, e := range p.Stream(t.Context(), llm.CompletionRequest{System: []string{"short"}, MaxTokens: 1000}) {
|
||||
if e != nil {
|
||||
t.Fatal(e)
|
||||
}
|
||||
}
|
||||
if inner.calls != 1 {
|
||||
t.Fatalf("a request that fits was not streamed: calls=%d", inner.calls)
|
||||
}
|
||||
})
|
||||
}
|
||||
|
||||
func TestEnableNoaFailureKeepsBuiltInCompaction(t *testing.T) {
|
||||
withNoaEnable(t, func(*agentcore.Options, noaadapter.Options) error { return errors.New("archive dir unusable") })
|
||||
inner := &countingProvider{}
|
||||
compaction := compactionConfig(32000)
|
||||
opts := &agentcore.Options{Compaction: compaction, Provider: inner}
|
||||
var warns []string
|
||||
enableNoa(opts, func() bool { return true }, t.TempDir(), "noa-fail", func(m string) { warns = append(warns, m) })
|
||||
|
||||
if len(warns) != 1 || !strings.Contains(warns[0], "falling back to built-in compaction") || !strings.Contains(warns[0], "archive dir unusable") {
|
||||
t.Fatalf("warnings %q", warns)
|
||||
}
|
||||
if opts.Compaction != compaction || opts.Compactor != nil {
|
||||
t.Fatalf("Compaction=%v Compactor=%v; the built-in compaction must stay", opts.Compaction, opts.Compactor)
|
||||
}
|
||||
if opts.Provider != llm.Provider(inner) {
|
||||
t.Fatalf("provider %T was wrapped although noa is off", opts.Provider)
|
||||
}
|
||||
}
|
||||
+6
-5
@@ -341,11 +341,12 @@ func applyCostEur(d *DB, out []ProfileUsage) error {
|
||||
if !ok || (pr.in == 0 && pr.outp == 0 && pr.cr == 0 && pr.cw == 0) {
|
||||
continue
|
||||
}
|
||||
// InputTokens includes the cached subset (Anthropic and OpenAI usage
|
||||
// semantics), so the fresh share is input minus cache reads; charging
|
||||
// both buckets at full input price double-counts the cache. Clamped at
|
||||
// zero for gateways that report the buckets disjointly.
|
||||
fresh := out[i].InputTokens - out[i].CacheReadTokens
|
||||
// InputTokens already includes both cached subsets: Norma normalizes it to
|
||||
// input + cache read + cache write (OpenAI prompt_tokens semantics). The
|
||||
// fresh share is therefore input minus cache reads minus cache writes;
|
||||
// charging either subset at full input price as well bills it twice.
|
||||
// Clamped at zero for gateways that report the buckets disjointly.
|
||||
fresh := out[i].InputTokens - out[i].CacheReadTokens - out[i].CacheWriteTokens
|
||||
if fresh < 0 {
|
||||
fresh = 0
|
||||
}
|
||||
|
||||
@@ -0,0 +1,125 @@
|
||||
package db
|
||||
|
||||
import (
|
||||
"fmt"
|
||||
"math"
|
||||
"testing"
|
||||
"time"
|
||||
)
|
||||
|
||||
// costProfiles saves a priced profile (EUR per million tokens: input 2.00,
|
||||
// output 4.00, cache read 0.20, cache write 2.50) and an unpriced one.
|
||||
func costProfiles(t *testing.T, d *DB) (priced, unpriced string) {
|
||||
t.Helper()
|
||||
suffix := time.Now().UnixNano()
|
||||
priced, unpriced = fmt.Sprintf("t-cost-priced-%d", suffix), fmt.Sprintf("t-cost-unpriced-%d", suffix)
|
||||
for _, p := range []*LLMProfile{
|
||||
{Name: priced, Format: "anthropic", Model: "m", APIKey: "k",
|
||||
PriceInputPerM: 2, PriceOutputPerM: 4, PriceCacheReadPerM: 0.2, PriceCacheWritePerM: 2.5},
|
||||
{Name: unpriced, Format: "anthropic", Model: "m", APIKey: "k"},
|
||||
} {
|
||||
id, err := d.SaveProfile(p)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
t.Cleanup(func() { d.Exec(`DELETE FROM llm_profiles WHERE id=$1`, id) })
|
||||
}
|
||||
return priced, unpriced
|
||||
}
|
||||
|
||||
func assertCost(t *testing.T, got *float64, want float64) {
|
||||
t.Helper()
|
||||
if got == nil {
|
||||
t.Fatalf("cost=nil, want %.2f", want)
|
||||
}
|
||||
if math.Abs(*got-want) > 1e-9 {
|
||||
t.Fatalf("cost=%.4f, want %.2f", *got, want)
|
||||
}
|
||||
}
|
||||
|
||||
// InputTokens already includes the cache read and cache write subsets, so only
|
||||
// the remainder is billed at the input price.
|
||||
func TestCostSeparatesFreshReadAndWriteBuckets(t *testing.T) {
|
||||
d, err := Open(testDSN(t))
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
defer d.Close()
|
||||
priced, unpriced := costProfiles(t, d)
|
||||
// Every case adds 100k output tokens: 0.1 × 4.00 = 0.40.
|
||||
for _, tc := range []struct {
|
||||
name string
|
||||
in, read, write int
|
||||
want float64
|
||||
}{
|
||||
{"fresh", 1_000_000, 0, 0, 2.40}, // 1M×2.00 + 0.40
|
||||
{"read_only", 1_000_000, 1_000_000, 0, 0.60}, // 0 fresh + 1M×0.20 + 0.40
|
||||
{"write_only", 1_000_000, 0, 1_000_000, 2.90}, // 0 fresh + 1M×2.50 + 0.40
|
||||
{"mixed", 1_000_000, 200_000, 300_000, 2.19}, // 500k×2.00 + 200k×0.20 + 300k×2.50 + 0.40
|
||||
{"inconsistent_gateway", 100_000, 200_000, 300_000, 1.19}, // fresh clamped to 0 + 0.04 + 0.75 + 0.40
|
||||
} {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
out := []ProfileUsage{
|
||||
{ProfileName: priced, InputTokens: tc.in, CacheReadTokens: tc.read, CacheWriteTokens: tc.write, OutputTokens: 100_000},
|
||||
{ProfileName: unpriced, InputTokens: tc.in, CacheReadTokens: tc.read, CacheWriteTokens: tc.write, OutputTokens: 100_000},
|
||||
{ProfileName: "t-cost-unknown-profile", InputTokens: tc.in, OutputTokens: 100_000},
|
||||
}
|
||||
if err := applyCostEur(d, out); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
assertCost(t, out[0].CostEur, tc.want)
|
||||
if out[1].CostEur != nil || out[2].CostEur != nil {
|
||||
t.Fatalf("unpriced usage got a cost: %v %v", out[1].CostEur, out[2].CostEur)
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
// Several ledger rows of one profile are summed before pricing; with no clamp
|
||||
// involved the profile cost equals the sum of the per-row costs.
|
||||
func TestUsageByProfileCostAggregatesRows(t *testing.T) {
|
||||
d, err := Open(testDSN(t))
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
defer d.Close()
|
||||
priced, unpriced := costProfiles(t, d)
|
||||
t.Cleanup(func() { d.Exec(`DELETE FROM llm_usage WHERE profile_name IN ($1,$2)`, priced, unpriced) })
|
||||
rows := []LLMUsage{
|
||||
// fresh 400k×2.00 + 200k×0.20 + 40k×4.00 = 0.80 + 0.04 + 0.16 = 1.00
|
||||
{InputTokens: 600_000, CacheRead: 200_000, OutputTokens: 40_000},
|
||||
// fresh 100k×2.00 + 300k×2.50 + 60k×4.00 = 0.20 + 0.75 + 0.24 = 1.19
|
||||
{InputTokens: 400_000, CacheWrite: 300_000, OutputTokens: 60_000},
|
||||
}
|
||||
for _, profile := range []string{priced, unpriced} {
|
||||
for _, r := range rows {
|
||||
r.ProfileName, r.Model, r.Status = profile, "m", "ok"
|
||||
if err := d.InsertLLMUsage(&r); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
}
|
||||
}
|
||||
usage, err := d.UsageByProfile()
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
found := 0
|
||||
for _, p := range usage {
|
||||
switch p.ProfileName {
|
||||
case priced:
|
||||
found++
|
||||
if p.Calls != 2 || p.InputTokens != 1_000_000 || p.CacheReadTokens != 200_000 || p.CacheWriteTokens != 300_000 || p.OutputTokens != 100_000 {
|
||||
t.Fatalf("aggregate %+v", p)
|
||||
}
|
||||
assertCost(t, p.CostEur, 2.19) // = 1.00 + 1.19, the mixed case
|
||||
case unpriced:
|
||||
found++
|
||||
if p.Calls != 2 || p.CostEur != nil {
|
||||
t.Fatalf("unpriced aggregate %+v", p)
|
||||
}
|
||||
}
|
||||
}
|
||||
if found != 2 {
|
||||
t.Fatalf("profiles missing from UsageByProfile: found %d of 2", found)
|
||||
}
|
||||
}
|
||||
@@ -3,7 +3,7 @@ module github.com/Autumn-27/artex
|
||||
go 1.26.10
|
||||
|
||||
require (
|
||||
github.com/Autumn-27/norma v0.4.3
|
||||
github.com/Autumn-27/norma v0.4.10
|
||||
github.com/golang-jwt/jwt/v5 v5.3.1
|
||||
github.com/google/uuid v1.6.0
|
||||
github.com/jackc/pgx/v5 v5.10.0
|
||||
@@ -61,4 +61,4 @@ require (
|
||||
modernc.org/memory v1.11.0 // indirect
|
||||
)
|
||||
|
||||
replace github.com/Autumn-27/norma => github.com/BaaRai/Norma v0.4.8
|
||||
replace github.com/Autumn-27/norma => github.com/BaaRai/Norma v0.4.10
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
github.com/BaaRai/Norma v0.4.8 h1:k3at2aYSJ6aIDjEo7tC4VDSLBHXw2HobYjpKChVe6hI=
|
||||
github.com/BaaRai/Norma v0.4.8/go.mod h1:BudWrf7iHl96FgHqFr/HwN3mKYmGjJDt+xL2ZHXfFZw=
|
||||
github.com/BaaRai/Norma v0.4.10 h1:+Sj6m8gZQIcVoOr5x9b3y7T1jI+95u43M07MLuvdaZk=
|
||||
github.com/BaaRai/Norma v0.4.10/go.mod h1:BudWrf7iHl96FgHqFr/HwN3mKYmGjJDt+xL2ZHXfFZw=
|
||||
github.com/andybalholm/brotli v1.2.1 h1:R+f5xP285VArJDRgowrfb9DqL18yVK0gKAW/F+eTWro=
|
||||
github.com/andybalholm/brotli v1.2.1/go.mod h1:rzTDkvFWvIrjDXZHkuS16NPggd91W3kUSvPlQ1pLaKY=
|
||||
github.com/asaskevich/govalidator v0.0.0-20230301143203-a9d515a09cc2 h1:DklsrG3dyBCFEj5IhUbnKptjxatkF07cF2ak3yi77so=
|
||||
|
||||
+6
-6
@@ -195,11 +195,11 @@ func (r *Recorder) Stream(ctx context.Context, req llm.CompletionRequest) iter.S
|
||||
return
|
||||
}
|
||||
// Always track usage (cheap); accumulate text/thinking only for bodies.
|
||||
// Anthropic (and the other providers) split token usage across events:
|
||||
// message_start carries ONLY the input side (input + cache), message_delta
|
||||
// ONLY the output. They must be FOLDED with Add — overwriting on delta
|
||||
// would zero out the input already counted at start (mirrors the SDK's own
|
||||
// llm.Accumulator; see norma/llm/accumulate.go).
|
||||
// message_start carries the input side (input + cache) and is added.
|
||||
// Since Norma v0.4.7 message_delta also echoes that input side next to the
|
||||
// output, so it is folded with llm.Usage.MergeEcho (max on echoed fields,
|
||||
// sum on output), as llm.Accumulator does; Add would count the input and
|
||||
// cache twice. An output-only delta folds the same either way.
|
||||
switch ev.Type {
|
||||
case llm.SETextDelta:
|
||||
if recordBodies {
|
||||
@@ -215,7 +215,7 @@ func (r *Recorder) Stream(ctx context.Context, req llm.CompletionRequest) iter.S
|
||||
if ev.StopReason != "" {
|
||||
stopReason = ev.StopReason
|
||||
}
|
||||
usage.Add(ev.Usage)
|
||||
usage.MergeEcho(ev.Usage)
|
||||
}
|
||||
if !yield(ev, err) {
|
||||
return
|
||||
|
||||
@@ -0,0 +1,118 @@
|
||||
package llmrec
|
||||
|
||||
import (
|
||||
"context"
|
||||
"iter"
|
||||
"testing"
|
||||
|
||||
"github.com/Autumn-27/artex/db"
|
||||
"github.com/Autumn-27/norma/llm"
|
||||
"github.com/google/uuid"
|
||||
)
|
||||
|
||||
// echoStreamProvider streams one message_start and one message_delta with the
|
||||
// given (already normalized) usage, as Norma's providers emit them.
|
||||
type echoStreamProvider struct {
|
||||
completeProvider
|
||||
start, delta llm.Usage
|
||||
}
|
||||
|
||||
func (p echoStreamProvider) Stream(context.Context, llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] {
|
||||
return func(yield func(llm.StreamEvent, error) bool) {
|
||||
for _, ev := range []llm.StreamEvent{
|
||||
{Type: llm.SEMessageStart, Usage: p.start},
|
||||
{Type: llm.SETextDelta, Text: "answer"},
|
||||
{Type: llm.SEMessageDelta, StopReason: "end_turn", Usage: p.delta},
|
||||
{Type: llm.SEMessageStop},
|
||||
} {
|
||||
if !yield(ev, nil) {
|
||||
return
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Norma normalizes Anthropic usage so InputTokens includes the cache read and
|
||||
// write subsets, and since v0.4.7 message_delta echoes that input side.
|
||||
var usageEchoCases = []struct {
|
||||
name string
|
||||
start, delta llm.Usage
|
||||
want llm.Usage
|
||||
}{
|
||||
{
|
||||
name: "full_echo",
|
||||
start: llm.Usage{InputTokens: 23, CacheReadTokens: 5, CacheWriteTokens: 3},
|
||||
delta: llm.Usage{InputTokens: 23, CacheReadTokens: 5, CacheWriteTokens: 3, OutputTokens: 5},
|
||||
want: llm.Usage{InputTokens: 23, OutputTokens: 5, CacheReadTokens: 5, CacheWriteTokens: 3},
|
||||
},
|
||||
{
|
||||
// Raw start: input 500, cache read 200, cache creation 40 → normalized 740.
|
||||
// Raw delta: input 500, output 77, no cache fields → normalized 500.
|
||||
name: "partial_echo",
|
||||
start: llm.Usage{InputTokens: 740, CacheReadTokens: 200, CacheWriteTokens: 40},
|
||||
delta: llm.Usage{InputTokens: 500, OutputTokens: 77},
|
||||
want: llm.Usage{InputTokens: 740, OutputTokens: 77, CacheReadTokens: 200, CacheWriteTokens: 40},
|
||||
},
|
||||
{
|
||||
name: "no_echo",
|
||||
start: llm.Usage{InputTokens: 23, CacheReadTokens: 5, CacheWriteTokens: 3},
|
||||
delta: llm.Usage{OutputTokens: 5},
|
||||
want: llm.Usage{InputTokens: 23, OutputTokens: 5, CacheReadTokens: 5, CacheWriteTokens: 3},
|
||||
},
|
||||
}
|
||||
|
||||
func drainRecorderStream(t *testing.T, ctx context.Context, r *Recorder) {
|
||||
t.Helper()
|
||||
for _, err := range r.Stream(ctx, llm.CompletionRequest{}) {
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// The metered usage of a streamed call counts an echoed input side once.
|
||||
func TestStreamUsageEchoCountedOnce(t *testing.T) {
|
||||
for _, tc := range usageEchoCases {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
var got []llm.Usage
|
||||
ctx := WithUsageSink(t.Context(), func(u llm.Usage) { got = append(got, u) })
|
||||
drainRecorderStream(t, ctx, Wrap(echoStreamProvider{start: tc.start, delta: tc.delta}, nil, "fixture", "profile", "", "", func() bool { return false }))
|
||||
if len(got) != 1 || got[0] != tc.want {
|
||||
t.Fatalf("metered %+v, want one call with %+v", got, tc.want)
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
// The llm_usage ledger gets one row per streamed call, with the echo counted once.
|
||||
func TestStreamUsageEchoLedgerRow(t *testing.T) {
|
||||
dsn, _, err := db.DSN()
|
||||
if err != nil {
|
||||
t.Skipf("no database config (%v)", err)
|
||||
}
|
||||
pg, err := db.Open(dsn)
|
||||
if err != nil {
|
||||
t.Skipf("postgres unavailable (%v)", err)
|
||||
}
|
||||
defer pg.Close()
|
||||
if err := pg.EnsureLLMUsageTable(); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
for _, tc := range usageEchoCases {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
profile := "usage-echo-" + uuid.NewString()
|
||||
t.Cleanup(func() { _, _ = pg.Exec(`DELETE FROM llm_usage WHERE profile_name=$1`, profile) })
|
||||
drainRecorderStream(t, t.Context(), Wrap(echoStreamProvider{start: tc.start, delta: tc.delta}, pg, "fixture", profile, "", "", func() bool { return false }))
|
||||
var rows int
|
||||
var got llm.Usage
|
||||
err := pg.QueryRow(`SELECT count(*), COALESCE(sum(input_tokens),0), COALESCE(sum(output_tokens),0), COALESCE(sum(cache_read),0), COALESCE(sum(cache_write),0)
|
||||
FROM llm_usage WHERE profile_name=$1`, profile).Scan(&rows, &got.InputTokens, &got.OutputTokens, &got.CacheReadTokens, &got.CacheWriteTokens)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if rows != 1 || got != tc.want {
|
||||
t.Fatalf("ledger rows=%d usage=%+v, want 1 row with %+v", rows, got, tc.want)
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,6 @@
|
||||
\set ON_ERROR_STOP on
|
||||
-- Removes only the two indexes introduced by this change; retains table data.
|
||||
SET lock_timeout = '5s';
|
||||
SET statement_timeout = '10min';
|
||||
DROP INDEX CONCURRENTLY IF EXISTS public.idx_activity_tool_result_lookup;
|
||||
DROP INDEX CONCURRENTLY IF EXISTS public.idx_llm_usage_task_model_tokens;
|
||||
Executable
+22
@@ -0,0 +1,22 @@
|
||||
\set ON_ERROR_STOP on
|
||||
-- Run with psql outside a transaction. Adds indexes without blocking normal DML.
|
||||
SET lock_timeout = '5s';
|
||||
SET statement_timeout = '10min';
|
||||
SET maintenance_work_mem = '128MB';
|
||||
SET max_parallel_maintenance_workers = 1;
|
||||
|
||||
-- db/commands.go joins by tool_use_id without an exploration_id constraint.
|
||||
-- The existing (exploration_id, tool_use_id, id) index has the wrong leading key.
|
||||
CREATE INDEX CONCURRENTLY idx_activity_tool_result_lookup
|
||||
ON public.activity (tool_use_id) INCLUDE (is_error)
|
||||
WHERE kind = 'tool_result';
|
||||
|
||||
-- Match TokenByModel's actual COALESCE(task_id, '') predicate and cover its sums.
|
||||
CREATE INDEX CONCURRENTLY idx_llm_usage_task_model_tokens
|
||||
ON public.llm_usage ((COALESCE(task_id, '')), model)
|
||||
INCLUDE (task_id, input_tokens, output_tokens, cache_read, cache_write);
|
||||
|
||||
-- Mark committed append-only pages visible so the covering index can avoid heap reads.
|
||||
SET vacuum_cost_delay = '2ms';
|
||||
SET vacuum_cost_limit = 1000;
|
||||
VACUUM (ANALYZE) public.llm_usage;
|
||||
Executable
+38
@@ -0,0 +1,38 @@
|
||||
# Online task-query indexes
|
||||
|
||||
`20261115-task-query-indexes.sql` adds two PostgreSQL indexes for existing queries:
|
||||
|
||||
- `activity(tool_use_id) INCLUDE (is_error) WHERE kind = 'tool_result'` supports the tool-result join in `db/commands.go` (`LEFT JOIN activity r ON r.tool_use_id = u.tool_use_id AND r.kind = 'tool_result'`). That join does not constrain the result row's exploration ID, so the existing `idx_act_tool_call (exploration_id, tool_use_id, id)` has the wrong leading key for it.
|
||||
- `llm_usage((COALESCE(task_id, '')), model) INCLUDE (task_id, input_tokens, output_tokens, cache_read, cache_write)` covers the per-task model aggregation in `db/llm_usage.go` (`TokenByModel`). That query filters on `COALESCE(task_id,'') = $1` and groups by `model`. The expression matches that predicate; the existing `idx_llm_usage_model (task_id, model)` cannot serve it.
|
||||
|
||||
## Apply
|
||||
|
||||
```sh
|
||||
psql -X -v ON_ERROR_STOP=1 -d <database> -f scripts/sql/20261115-task-query-indexes.sql
|
||||
```
|
||||
|
||||
Run it outside a transaction: `CREATE INDEX CONCURRENTLY` cannot run inside one, so do not wrap the file in `BEGIN`/`COMMIT` or use `psql -1`. Use the deployment's normal PostgreSQL authentication.
|
||||
|
||||
Both indexes are created concurrently, so normal reads and writes continue. The script sets a 5-second lock timeout, a 10-minute statement timeout and bounded maintenance resources. Its final `VACUUM (ANALYZE)` is an ordinary vacuum, not `VACUUM FULL`; it updates visibility and statistics so the covering index can serve index-only scans.
|
||||
|
||||
This is an explicit, one-time maintenance step, not a numbered migration. `db/migrate.go` runs each migration file inside a transaction, where `CONCURRENTLY` is not allowed. A plain `CREATE INDEX` at startup would lock `activity` against writes for the whole build. No application or database restart is required.
|
||||
|
||||
## After an interruption
|
||||
|
||||
The script intentionally fails if an index with the same name already exists. A partial run may have created the first index before the second failed. An interrupted concurrent build can also leave an invalid index behind. Before retrying, check the state of both indexes:
|
||||
|
||||
```sql
|
||||
SELECT c.relname, i.indisvalid, i.indisready
|
||||
FROM pg_index i JOIN pg_class c ON c.oid = i.indexrelid
|
||||
WHERE c.relname IN ('idx_activity_tool_result_lookup', 'idx_llm_usage_task_model_tokens');
|
||||
```
|
||||
|
||||
Both must show `indisvalid = true` and `indisready = true`. Drop an index that is invalid or not ready with `DROP INDEX CONCURRENTLY`, or run the rollback below, then rerun the script.
|
||||
|
||||
## Rollback
|
||||
|
||||
```sh
|
||||
psql -X -v ON_ERROR_STOP=1 -d <database> -f scripts/sql/20261115-task-query-indexes.rollback.sql
|
||||
```
|
||||
|
||||
The rollback drops only these two indexes, concurrently and only if they exist. Table data is kept. Review the current index definitions before applying it. No application restart is required.
|
||||
+3
-1
@@ -79,8 +79,10 @@ var errWorkControlConflict = errors.New("work control conflict")
|
||||
// retryableWorkerModelError excludes errors already handled by the task router.
|
||||
// In particular, a quota error after partial streaming advances the task cursor
|
||||
// for the next LLM call but must not replay this whole intent on the backup.
|
||||
// A NOA budget refusal is deterministic and never reached the provider, so a
|
||||
// rerun would only fail the same way.
|
||||
func retryableWorkerModelError(reason harness.TerminalReason, err error) bool {
|
||||
return reason == harness.ReasonModelError && !isTaskLLMRuntimeError(err)
|
||||
return reason == harness.ReasonModelError && !isTaskLLMRuntimeError(err) && !errors.Is(err, agent.ErrNoaBudget)
|
||||
}
|
||||
|
||||
// Engine drives the event-driven exploration loop with real LLM agents
|
||||
|
||||
@@ -0,0 +1,32 @@
|
||||
package server
|
||||
|
||||
import (
|
||||
"errors"
|
||||
"fmt"
|
||||
"testing"
|
||||
|
||||
"github.com/Autumn-27/artex/agent"
|
||||
"github.com/Autumn-27/norma/harness"
|
||||
)
|
||||
|
||||
// A NOA budget refusal is deterministic: the engine must not rerun the intent,
|
||||
// whether the sentinel arrives bare or wrapped.
|
||||
func TestWorkerModelErrorRetrySkipsNoaBudget(t *testing.T) {
|
||||
for _, tc := range []struct {
|
||||
name string
|
||||
err error
|
||||
want bool
|
||||
}{
|
||||
{"noa budget", agent.ErrNoaBudget, false},
|
||||
{"noa budget wrapped", fmt.Errorf("%w: estimate 64000 exceeds configured budget 32000", agent.ErrNoaBudget), false},
|
||||
{"noa budget wrapped twice", fmt.Errorf("worker: %w", fmt.Errorf("%w: estimate", agent.ErrNoaBudget)), false},
|
||||
{"generic model error", errors.New("connection reset by peer"), true},
|
||||
{"task router error", &taskLLMError{taskID: "7", cause: errors.New("insufficient_quota")}, false},
|
||||
} {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
if got := retryableWorkerModelError(harness.ReasonModelError, tc.err); got != tc.want {
|
||||
t.Fatalf("retryableWorkerModelError(model_error, %v) = %v, want %v", tc.err, got, tc.want)
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
@@ -32,8 +32,11 @@ func (s SideQuestionService) Answer(ctx context.Context, req llm.CompletionReque
|
||||
out.Text += ev.Text
|
||||
case llm.SEToolUseStart:
|
||||
out.ToolUse = true
|
||||
case llm.SEMessageStart, llm.SEMessageDelta:
|
||||
case llm.SEMessageStart:
|
||||
out.Usage.Add(ev.Usage)
|
||||
case llm.SEMessageDelta:
|
||||
// message_delta may echo message_start's input side: MergeEcho, not Add.
|
||||
out.Usage.MergeEcho(ev.Usage)
|
||||
case llm.SEMessageStop:
|
||||
complete = true
|
||||
}
|
||||
|
||||
@@ -0,0 +1,52 @@
|
||||
package sidequestion
|
||||
|
||||
import (
|
||||
"context"
|
||||
"iter"
|
||||
"testing"
|
||||
|
||||
"github.com/Autumn-27/norma/llm"
|
||||
)
|
||||
|
||||
// A streamed side answer counts message_delta's echo of the input side once.
|
||||
func TestAnswerStreamUsageEchoCountedOnce(t *testing.T) {
|
||||
for _, tc := range []struct {
|
||||
name string
|
||||
start, delta llm.Usage
|
||||
want llm.Usage
|
||||
}{
|
||||
{"full_echo",
|
||||
llm.Usage{InputTokens: 23, CacheReadTokens: 5, CacheWriteTokens: 3},
|
||||
llm.Usage{InputTokens: 23, CacheReadTokens: 5, CacheWriteTokens: 3, OutputTokens: 5},
|
||||
llm.Usage{InputTokens: 23, OutputTokens: 5, CacheReadTokens: 5, CacheWriteTokens: 3}},
|
||||
{"partial_echo",
|
||||
llm.Usage{InputTokens: 740, CacheReadTokens: 200, CacheWriteTokens: 40},
|
||||
llm.Usage{InputTokens: 500, OutputTokens: 77},
|
||||
llm.Usage{InputTokens: 740, OutputTokens: 77, CacheReadTokens: 200, CacheWriteTokens: 40}},
|
||||
{"no_echo",
|
||||
llm.Usage{InputTokens: 23, CacheReadTokens: 5, CacheWriteTokens: 3},
|
||||
llm.Usage{OutputTokens: 5},
|
||||
llm.Usage{InputTokens: 23, OutputTokens: 5, CacheReadTokens: 5, CacheWriteTokens: 3}},
|
||||
} {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
p := fakeProvider{stream: func(context.Context, llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] {
|
||||
return func(yield func(llm.StreamEvent, error) bool) {
|
||||
for _, ev := range []llm.StreamEvent{
|
||||
{Type: llm.SEMessageStart, Usage: tc.start},
|
||||
{Type: llm.SETextDelta, Text: "answer"},
|
||||
{Type: llm.SEMessageDelta, StopReason: "end_turn", Usage: tc.delta},
|
||||
{Type: llm.SEMessageStop},
|
||||
} {
|
||||
if !yield(ev, nil) {
|
||||
return
|
||||
}
|
||||
}
|
||||
}
|
||||
}}
|
||||
out, err := (SideQuestionService{Provider: p}).Answer(t.Context(), llm.CompletionRequest{}, true, nil)
|
||||
if err != nil || out.Text != "answer" || out.Usage != tc.want {
|
||||
t.Fatalf("answer=%+v err=%v, want usage %+v", out, err, tc.want)
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user