From e6ab0f34a7bdc2db20e8fa0017e0a645719d9807 Mon Sep 17 00:00:00 2001 From: BaaRai <130491490+BaaRai@users.noreply.github.com> Date: Sun, 15 Nov 2026 16:14:35 +0100 Subject: [PATCH] fix(server): sum task LLM usage across retries and failover The logical call now returns the usage of every physical attempt (atomic and stream), keeps the usage of failed uncommitted stream attempts without delivering their events, normalizes per-attempt echoes with MergeEcho, stops with ctx.Err() when cancelled during backoff, and treats signed/redacted thinking and tool-use signatures as committing output. The llm_usage ledger stays at one row per physical call. --- server/task_llm.go | 94 ++++++++--- server/task_llm_usage_test.go | 286 ++++++++++++++++++++++++++++++++++ 2 files changed, 361 insertions(+), 19 deletions(-) create mode 100644 server/task_llm_usage_test.go diff --git a/server/task_llm.go b/server/task_llm.go index a04a592..c8a5fed 100644 --- a/server/task_llm.go +++ b/server/task_llm.go @@ -317,10 +317,14 @@ func (r *taskLLMRuntime) Complete(ctx context.Context, req llm.CompletionRequest } func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest, hooks taskLLMStreamHooks) (llm.Message, string, llm.Usage, error) { + // total is the usage of the logical call: the sum of every physical attempt + // (same-provider retries and profile failover). Each attempt is already billed + // once by the recorder below; this sum only feeds the caller's accounting. + var total llm.Usage for { selection, err := hooks.current(ctx) if err != nil { - return llm.Message{}, "", llm.Usage{}, err + return llm.Message{}, "", total, err } var ( msg llm.Message @@ -333,13 +337,14 @@ func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionReque retries, backoffOf := sameProviderRetryPolicy(selection.retry) for attempt := 0; ; attempt++ { msg, sr, usage, callErr = selection.provider.Complete(ctx, req) + total.Add(usage) if callErr != nil && ctx.Err() == nil && attempt < retries && isRetryableStreamError(callErr) { backoff := backoffOf(attempt) log.Printf("[task-llm] task %s non-streaming call failed, retrying on the same provider after %v (%d/%d): %v", taskID, backoff, attempt+1, retries, callErr) if sleepCtx(ctx, backoff) { - break // ctx cancelled during backoff → stop retrying + return llm.Message{}, "", total, ctx.Err() // cancelled during backoff → stop } continue } @@ -347,7 +352,7 @@ func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionReque } if callErr == nil { resetLLMFailStreak(taskID, selection.profileID) - return msg, sr, usage, nil + return msg, sr, total, nil } // Non-quota failures on a chain profile escalate after a consecutive // streak (#99): same advance path as quota exhaust, truthful reason. @@ -356,31 +361,31 @@ func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionReque cause := fmt.Errorf("profile #%d failed %d consecutive calls (last: %w)", selection.profileID, llmFailStreakThreshold, callErr) transition, markErr := hooks.exhaust(selection, cause) if markErr != nil { - return llm.Message{}, "", llm.Usage{}, fmt.Errorf("mark profile exhausted after %v: %w", callErr, markErr) + return llm.Message{}, "", total, fmt.Errorf("mark profile exhausted after %v: %w", callErr, markErr) } if transition.Advanced && !transition.Stale && hooks.transition != nil { hooks.transition(selection, transition, callErr) } if !transition.Stale && transition.NextProfileID == nil { - return llm.Message{}, "", llm.Usage{}, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: callErr} + return llm.Message{}, "", total, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: callErr} } continue // replay the same logical request on the next profile } - return llm.Message{}, "", llm.Usage{}, callErr + return llm.Message{}, "", total, callErr } // profileID=0: the explicit chain was cleared; non-quota error: pass through. Neither changes the task failover state. if selection.profileID == 0 || !isQuotaExhaustedError(callErr) { - return llm.Message{}, "", llm.Usage{}, callErr + return llm.Message{}, "", total, callErr } transition, markErr := hooks.exhaust(selection, callErr) if markErr != nil { - return llm.Message{}, "", llm.Usage{}, fmt.Errorf("mark profile quota exhausted after %v: %w", callErr, markErr) + return llm.Message{}, "", total, fmt.Errorf("mark profile quota exhausted after %v: %w", callErr, markErr) } if transition.Advanced && !transition.Stale && hooks.transition != nil { hooks.transition(selection, transition, callErr) } if !transition.Stale && transition.NextProfileID == nil { - return llm.Message{}, "", llm.Usage{}, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: callErr} + return llm.Message{}, "", total, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: callErr} } // No output has been delivered to the caller, so replaying the same logical request on the next profile is safe. } @@ -397,6 +402,35 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest committed := false var pending []llm.StreamEvent var streamErr error + // The caller's llm.Accumulator sums usage across the whole logical call + // (every attempt and profile): message_start is added, message_delta is + // folded with MergeEcho. Folding one attempt's echo against the sum of + // earlier attempts would lose input/cache tokens, so each attempt's echoes + // are normalized here against that attempt alone and only the increment + // is passed on, as a message_start. The recorder below already billed the + // physical attempt; this only shapes what the caller adds up. + var attemptUsage llm.Usage + forward := func(event llm.StreamEvent) bool { + switch event.Type { + case llm.SEMessageStart: + attemptUsage.Add(event.Usage) + case llm.SEMessageDelta: + next := attemptUsage + next.MergeEcho(event.Usage) + inc := llm.Usage{ + InputTokens: next.InputTokens - attemptUsage.InputTokens, + OutputTokens: next.OutputTokens - attemptUsage.OutputTokens, + CacheReadTokens: next.CacheReadTokens - attemptUsage.CacheReadTokens, + CacheWriteTokens: next.CacheWriteTokens - attemptUsage.CacheWriteTokens, + } + attemptUsage = next + if inc != (llm.Usage{}) && !yield(llm.StreamEvent{Type: llm.SEMessageStart, Usage: inc}, nil) { + return false + } + event.Usage = llm.Usage{} + } + return yield(event, nil) + } retries, backoffOf := sameProviderRetryPolicy(selection.retry) // Same-provider safe-window retry: a transient failure before committed // (nothing delivered to the caller yet) can be replayed verbatim without @@ -405,6 +439,7 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest committed = false pending = nil streamErr = nil + attemptUsage = llm.Usage{} for event, err := range selection.provider.Stream(ctx, req) { if err != nil { streamErr = err @@ -416,24 +451,34 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest } if !committed { for _, buffered := range pending { - if !yield(buffered, nil) { + if !forward(buffered) { return } } pending = nil committed = true } - if !yield(event, nil) { + if !forward(event) { return } } + if streamErr != nil && !committed { + // A failed uncommitted attempt delivers none of its events (they are + // never replayed), but the usage it reported survives as one + // message_start so the logical call still counts it. + if u := pendingUsage(pending); u != (llm.Usage{}) && !yield(llm.StreamEvent{Type: llm.SEMessageStart, Usage: u}, nil) { + return + } + pending = nil + } if streamErr != nil && !committed && ctx.Err() == nil && attempt < retries && isRetryableStreamError(streamErr) { backoff := backoffOf(attempt) log.Printf("[task-llm] task %s pre-commit stream failure, retrying on the same provider after %v (%d/%d): %v", taskID, backoff, attempt+1, retries, streamErr) if sleepCtx(ctx, backoff) { - break // ctx cancelled during backoff → stop retrying + yield(llm.StreamEvent{}, ctx.Err()) // cancelled during backoff → stop + return } continue } @@ -442,7 +487,7 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest if streamErr == nil { resetLLMFailStreak(taskID, selection.profileID) for _, buffered := range pending { - if !yield(buffered, nil) { + if !forward(buffered) { return } } @@ -474,11 +519,6 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest // bundle was still in use. Agent/global fallback errors follow the legacy // behavior and never mutate task failover state. if selection.profileID == 0 || !isQuotaExhaustedError(streamErr) { - for _, buffered := range pending { - if !yield(buffered, nil) { - return - } - } yield(llm.StreamEvent{}, streamErr) return } @@ -508,9 +548,25 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest } } +// pendingUsage is the usage the buffered (uncommitted) events of one attempt +// reported. Only message_start can be buffered: message_delta commits. +func pendingUsage(pending []llm.StreamEvent) llm.Usage { + var u llm.Usage + for _, ev := range pending { + if ev.Type == llm.SEMessageStart { + u.Add(ev.Usage) + } + } + return u +} + +// streamEventCommitsOutput reports whether delivering event makes the attempt +// unreplayable. Signed and redacted thinking and tool-use signatures are opaque +// output the model bound to this exact response, so they commit like text. func streamEventCommitsOutput(event llm.StreamEvent) bool { switch event.Type { - case llm.SETextDelta, llm.SEThinkingDelta, llm.SEToolInputJSON: + case llm.SETextDelta, llm.SEThinkingDelta, llm.SEToolInputJSON, + llm.SEThinkingSignature, llm.SERedactedThinking, llm.SEToolUseSignature: return event.Text != "" case llm.SEToolUseStart, llm.SEMessageDelta, llm.SEMessageStop: return true diff --git a/server/task_llm_usage_test.go b/server/task_llm_usage_test.go new file mode 100644 index 0000000..fb743a6 --- /dev/null +++ b/server/task_llm_usage_test.go @@ -0,0 +1,286 @@ +package server + +import ( + "context" + "errors" + "fmt" + "iter" + "testing" + "time" + + "github.com/Autumn-27/artex/agent" + "github.com/Autumn-27/artex/db" + "github.com/Autumn-27/artex/llmrec" + "github.com/Autumn-27/norma/llm" +) + +// Usage reported by one failed attempt, and by a successful one. A stream +// reports the input side in message_start and echoes it, with the output, in +// message_delta; the atomic path returns the attempt's normalized usage. +var ( + failedAttemptUsage = llm.Usage{InputTokens: 11, OutputTokens: 2, CacheReadTokens: 3, CacheWriteTokens: 2} + okStartUsage = llm.Usage{InputTokens: 13, CacheReadTokens: 3, CacheWriteTokens: 2} + okDeltaUsage = llm.Usage{InputTokens: 17, OutputTokens: 5, CacheReadTokens: 4, CacheWriteTokens: 2} + okAttemptUsage = llm.Usage{InputTokens: 17, OutputTokens: 5, CacheReadTokens: 4, CacheWriteTokens: 2} +) + +// usageAttemptProvider fails its first fail attempts (always, when fail < 0) +// after reporting failedAttemptUsage, then answers "done". +type usageAttemptProvider struct { + fail int + failErr error + onFail func() + // extra events a failed stream attempt emits before its error. + failEvents []llm.StreamEvent + calls int +} + +func (p *usageAttemptProvider) failing() bool { + p.calls++ + failed := p.fail < 0 || p.calls <= p.fail + if failed && p.onFail != nil { + p.onFail() + } + return failed +} + +func (p *usageAttemptProvider) Complete(context.Context, llm.CompletionRequest) (llm.Message, string, llm.Usage, error) { + if p.failing() { + return llm.Message{}, "", failedAttemptUsage, p.failErr + } + return llm.Message{Role: llm.RoleAssistant, Content: []llm.ContentBlock{llm.TextBlock("done")}}, "end_turn", okAttemptUsage, nil +} + +func (p *usageAttemptProvider) Stream(context.Context, llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] { + return func(yield func(llm.StreamEvent, error) bool) { + if p.failing() { + events := append([]llm.StreamEvent{{Type: llm.SEMessageStart, Usage: failedAttemptUsage}}, p.failEvents...) + for _, ev := range events { + if !yield(ev, nil) { + return + } + } + yield(llm.StreamEvent{}, p.failErr) + return + } + for _, ev := range []llm.StreamEvent{ + {Type: llm.SEMessageStart, Usage: okStartUsage}, + {Type: llm.SETextDelta, Text: "done"}, + {Type: llm.SEMessageDelta, StopReason: "end_turn", Usage: okDeltaUsage}, + {Type: llm.SEMessageStop}, + } { + if !yield(ev, nil) { + return + } + } + } +} + +// logicalResult is what the SDK sees for one logical call. +type logicalResult struct { + usage llm.Usage + text string + events []llm.StreamEvent + err error +} + +func runLogicalCall(ctx context.Context, stream bool, taskID string, hooks taskLLMStreamHooks) logicalResult { + if !stream { + msg, _, usage, err := completeTaskLLM(ctx, taskID, llm.CompletionRequest{}, hooks) + return logicalResult{usage: usage, text: msg.Text(), err: err} + } + var r logicalResult + acc := llm.NewAccumulator() + for ev, err := range streamTaskLLM(ctx, taskID, llm.CompletionRequest{}, hooks) { + if err != nil { + r.err = err + break + } + r.events = append(r.events, ev) + acc.Add(ev) + } + r.usage, r.text = acc.Usage, acc.Message().Text() + return r +} + +func singleProfileHooks(p llm.Provider, retry agent.RetryConfig) taskLLMStreamHooks { + return taskLLMStreamHooks{ + current: func(context.Context) (taskLLMSelection, error) { + return taskLLMSelection{provider: p, retry: retry}, nil + }, + exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) { + return db.TaskLLMTransition{}, nil + }, + } +} + +func sumUsage(us ...llm.Usage) llm.Usage { + var t llm.Usage + for _, u := range us { + t.Add(u) + } + return t +} + +var usageModes = []struct { + name string + stream bool +}{{"atomic", false}, {"stream", true}} + +func TestTaskLLMRetrySumsAttemptUsage(t *testing.T) { + for _, m := range usageModes { + t.Run(m.name, func(t *testing.T) { + p := &usageAttemptProvider{fail: 1, failErr: errors.New("connection reset by peer")} + r := runLogicalCall(t.Context(), m.stream, "p12-retry", singleProfileHooks(p, agent.RetryConfig{StreamAttempts: 1, StreamInterval: time.Millisecond})) + if r.err != nil { + t.Fatal(r.err) + } + want := sumUsage(failedAttemptUsage, okAttemptUsage) // {28 7 7 4} + if r.usage != want || p.calls != 2 || r.text != "done" { + t.Fatalf("usage=%+v want %+v, calls=%d, text=%q", r.usage, want, p.calls, r.text) + } + }) + } +} + +func TestTaskLLMExhaustedRetriesReturnSummedUsage(t *testing.T) { + for _, m := range usageModes { + t.Run(m.name, func(t *testing.T) { + p := &usageAttemptProvider{fail: -1, failErr: errors.New("connection reset by peer")} + r := runLogicalCall(t.Context(), m.stream, "p12-exhausted", singleProfileHooks(p, agent.RetryConfig{StreamAttempts: 1, StreamInterval: time.Millisecond})) + if r.err == nil || r.err.Error() != "connection reset by peer" { + t.Fatalf("err=%v", r.err) + } + want := sumUsage(failedAttemptUsage, failedAttemptUsage) + if r.usage != want || p.calls != 2 { + t.Fatalf("usage=%+v want %+v, calls=%d", r.usage, want, p.calls) + } + }) + } +} + +func TestTaskLLMProfileFailoverSumsUsage(t *testing.T) { + for _, m := range usageModes { + t.Run(m.name, func(t *testing.T) { + first := &usageAttemptProvider{fail: -1, failErr: errors.New("openai: status 402: insufficient_quota")} + second := &usageAttemptProvider{} + active := int64(11) + hooks := taskLLMStreamHooks{ + current: func(context.Context) (taskLLMSelection, error) { + if active == 11 { + return taskLLMSelection{profileID: 11, provider: first}, nil + } + return taskLLMSelection{profileID: 22, provider: second}, nil + }, + exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) { + active = 22 + next := active + return db.TaskLLMTransition{PreviousProfileID: 11, NextProfileID: &next, Advanced: true}, nil + }, + } + r := runLogicalCall(t.Context(), m.stream, "p12-failover-"+m.name, hooks) + if r.err != nil { + t.Fatal(r.err) + } + want := sumUsage(failedAttemptUsage, okAttemptUsage) + if r.usage != want || first.calls != 1 || second.calls != 1 || r.text != "done" { + t.Fatalf("usage=%+v want %+v, calls first=%d second=%d, text=%q", r.usage, want, first.calls, second.calls, r.text) + } + }) + } +} + +func TestTaskLLMCancelDuringBackoffStops(t *testing.T) { + for _, m := range usageModes { + t.Run(m.name, func(t *testing.T) { + ctx, cancel := context.WithCancel(t.Context()) + defer cancel() + p := &usageAttemptProvider{fail: -1, failErr: errors.New("connection reset by peer")} + // Cancel once the retry is waiting out its (long) backoff. + p.onFail = func() { time.AfterFunc(20*time.Millisecond, cancel) } + r := runLogicalCall(ctx, m.stream, "p12-cancel", singleProfileHooks(p, agent.RetryConfig{StreamAttempts: 2, StreamInterval: time.Minute})) + if !errors.Is(r.err, context.Canceled) { + t.Fatalf("err=%v, want context.Canceled", r.err) + } + if p.calls != 1 || r.usage != failedAttemptUsage { + t.Fatalf("calls=%d usage=%+v", p.calls, r.usage) + } + }) + } +} + +func TestTaskLLMUncommittedStreamAttemptKeepsUsageWithoutReplay(t *testing.T) { + p := &usageAttemptProvider{ + fail: 1, + failErr: errors.New("anthropic: overloaded_error"), + // Buffered, non-committing events of the failed attempt. + failEvents: []llm.StreamEvent{{Type: llm.SEThinkingDelta}, {Type: llm.SEToolInputJSON}}, + } + r := runLogicalCall(t.Context(), true, "p12-uncommitted", singleProfileHooks(p, agent.RetryConfig{StreamAttempts: 1, StreamInterval: time.Millisecond})) + if r.err != nil { + t.Fatal(r.err) + } + if want := sumUsage(failedAttemptUsage, okAttemptUsage); r.usage != want { + t.Fatalf("usage=%+v want %+v", r.usage, want) + } + if r.text != "done" { + t.Fatalf("text=%q, want the second attempt's only", r.text) + } + // The failed attempt contributes one usage-only message_start, nothing else. + if len(r.events) == 0 || r.events[0].Type != llm.SEMessageStart || r.events[0].Usage != failedAttemptUsage { + t.Fatalf("first event %+v", r.events) + } + for _, ev := range r.events { + if ev.Type == llm.SEThinkingDelta || ev.Type == llm.SEToolInputJSON { + t.Fatalf("buffered event of the failed attempt was delivered: %+v", r.events) + } + } +} + +func TestTaskLLMOpaqueThinkingCommitsStream(t *testing.T) { + defer withZeroRetryBackoff()() + for _, kind := range []llm.StreamEventType{llm.SEThinkingSignature, llm.SERedactedThinking, llm.SEToolUseSignature} { + t.Run(string(kind), func(t *testing.T) { + if !streamEventCommitsOutput(llm.StreamEvent{Type: kind, Text: "opaque"}) { + t.Fatalf("%s does not commit", kind) + } + // A retryable failure after the opaque block is surfaced, not replayed. + p := &scriptedLLMProvider{events: []llm.StreamEvent{{Type: kind, Text: "opaque"}}, err: errors.New("connection reset by peer")} + events, err := collectTaskLLMStream(streamTaskLLM(t.Context(), "p12-opaque", llm.CompletionRequest{}, singleProfileHooks(p, agent.RetryConfig{}))) + if err == nil || p.calls != 1 || len(events) != 1 || events[0].Type != kind { + t.Fatalf("calls=%d events=%+v err=%v", p.calls, events, err) + } + }) + } +} + +// The recorder below the task router writes one llm_usage row per physical +// attempt; summing the attempts for the logical call adds no row. +func TestTaskLLMRetryKeepsOneLedgerRowPerPhysicalCall(t *testing.T) { + m, err := NewManager(t.TempDir(), "") + if err != nil { + t.Skipf("postgres unavailable (%v) — skipping", err) + } + defer m.Close() + pg := m.PG() + for _, mode := range usageModes { + t.Run(mode.name, func(t *testing.T) { + taskID := fmt.Sprintf("p12-ledger-%s-%d", mode.name, time.Now().UnixNano()) + t.Cleanup(func() { _, _ = pg.Exec(`DELETE FROM llm_usage WHERE task_id=$1`, taskID) }) + p := &usageAttemptProvider{fail: 1, failErr: errors.New("connection reset by peer")} + rec := llmrec.Wrap(p, pg, "test-model", "test-profile", "", "", func() bool { return false }) + ctx := llmrec.WithTaskID(t.Context(), taskID) + r := runLogicalCall(ctx, mode.stream, taskID, singleProfileHooks(rec, agent.RetryConfig{StreamAttempts: 1, StreamInterval: time.Millisecond})) + if r.err != nil { + t.Fatal(r.err) + } + var rows, failed int + if err := pg.QueryRow(`SELECT count(*), count(*) FILTER (WHERE status='error') FROM llm_usage WHERE task_id=$1`, taskID).Scan(&rows, &failed); err != nil { + t.Fatal(err) + } + if rows != 2 || failed != 1 || p.calls != 2 { + t.Fatalf("llm_usage rows=%d (error %d), physical calls=%d; want 2 rows, 1 error", rows, failed, p.calls) + } + }) + } +}