Merge pull request #247 from BaaRai/fix/noa-budget-no-retry

fix(server): do not rerun intents refused by the NOA budget
This commit is contained in:
BaaRai
2026-11-15 22:54:09 +01:00
committed by GitHub
4 changed files with 63 additions and 2 deletions
+7 -1
View File
@@ -2,6 +2,7 @@ package agent
import (
"context"
"errors"
"fmt"
"iter"
@@ -12,6 +13,11 @@ import (
// 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)
@@ -24,7 +30,7 @@ type noaBudgetProvider struct {
func (p *noaBudgetProvider) check(req llm.CompletionRequest) error {
estimate := sidequestion.EstimateInputTokens(req) + max(req.MaxTokens, 0) + noaRequestOverhead
if estimate > p.window {
return fmt.Errorf("Norma NOA request estimate %d exceeds configured budget %d; protected context cannot fit, increase the configured window or start a new session", 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
}
+21
View File
@@ -8,10 +8,25 @@ import (
"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 }
@@ -83,6 +98,9 @@ func TestNoaBudgetRefusesIrreducibleRequestBeforeTransport(t *testing.T) {
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)
}
@@ -103,6 +121,9 @@ func TestNoaBudgetRefusesIrreducibleRequestBeforeTransport(t *testing.T) {
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)
+3 -1
View File
@@ -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
+32
View File
@@ -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)
}
})
}
}