1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
|
// Copyright 2025 syzkaller project authors. All rights reserved.
// Use of this source code is governed by Apache 2 LICENSE that can be found in the LICENSE file.
package aflow
import (
"context"
"errors"
"fmt"
"maps"
"os"
"slices"
"strings"
"sync"
"time"
"github.com/google/syzkaller/pkg/aflow/trajectory"
"github.com/google/syzkaller/pkg/osutil"
"google.golang.org/genai"
)
// Execute executes the given AI workflow with provided inputs and returns workflow outputs.
// The model argument sets Gemini model name to execute the workflow.
// The workdir argument should point to a dir owned by aflow to store private data,
// it can be shared across parallel executions in the same process, and preferably
// preserved across process restarts for caching purposes.
func (flow *Flow) Execute(c context.Context, model, workdir string, inputs map[string]any,
cache *Cache, onEvent onEvent) (map[string]any, error) {
if err := flow.checkInputs(inputs); err != nil {
return nil, fmt.Errorf("flow inputs are missing: %w", err)
}
ctx := &Context{
Context: c,
Workdir: osutil.Abs(workdir),
cache: cache,
state: maps.Clone(inputs),
onEvent: onEvent,
}
defer ctx.close()
if s := c.Value(stubContextKey); s != nil {
ctx.stubContext = *s.(*stubContext)
}
if ctx.timeNow == nil {
ctx.timeNow = time.Now
}
if ctx.generateContent == nil {
var err error
ctx.generateContent, err = contentGenerator(c, model)
if err != nil {
return nil, err
}
}
span := &trajectory.Span{
Type: trajectory.SpanFlow,
Name: flow.Name,
}
if err := ctx.startSpan(span); err != nil {
return nil, err
}
flowErr := flow.Root.execute(ctx)
if flowErr == nil {
span.Results = flow.extractOutputs(ctx.state)
}
if err := ctx.finishSpan(span, flowErr); err != nil {
return nil, err
}
if ctx.spanNesting != 0 {
// Since we finish all spans, even on errors, we should end up at 0.
panic(fmt.Sprintf("unbalanced spans (%v)", ctx.spanNesting))
}
return span.Results, nil
}
// FlowError creates an error that denotes failure of the flow itself,
// rather than an infrastructure error. A flow errors mean an expected
// condition in the flow when it cannot continue, and cannot produce
// expected outputs. For example, if we are doing something with the kernel,
// but the kernel build fails. Flow errors shouldn't be flagged in
// infrastructure monitoring.
func FlowError(err error) error {
return &flowError{err}
}
func IsFlowError(err error) bool {
var flowErr *flowError
return errors.As(err, &flowErr)
}
type flowError struct {
error
}
type (
onEvent func(*trajectory.Span) error
generateContentFunc func(*genai.GenerateContentConfig, []*genai.Content) (
*genai.GenerateContentResponse, error)
contextKeyType int
)
var (
createClientOnce sync.Once
createClientErr error
client *genai.Client
modelList = make(map[string]bool)
stubContextKey = contextKeyType(1)
)
func contentGenerator(ctx context.Context, model string) (generateContentFunc, error) {
const modelPrefix = "models/"
createClientOnce.Do(func() {
if os.Getenv("GOOGLE_API_KEY") == "" {
createClientErr = fmt.Errorf("set GOOGLE_API_KEY env var to use with Gemini" +
" (see https://ai.google.dev/gemini-api/docs/api-key)")
return
}
client, createClientErr = genai.NewClient(ctx, nil)
if createClientErr != nil {
return
}
for m, err := range client.Models.All(ctx) {
if err != nil {
createClientErr = err
return
}
modelList[strings.TrimPrefix(m.Name, modelPrefix)] = m.Thinking
}
})
if createClientErr != nil {
return nil, createClientErr
}
thinking, ok := modelList[model]
if !ok {
models := slices.Collect(maps.Keys(modelList))
slices.Sort(models)
return nil, fmt.Errorf("model %q does not exist (models: %v)", model, models)
}
return func(cfg *genai.GenerateContentConfig, req []*genai.Content) (*genai.GenerateContentResponse, error) {
if thinking {
cfg.ThinkingConfig = &genai.ThinkingConfig{
// We capture them in the trajectory for analysis.
IncludeThoughts: true,
// Enable "dynamic thinking" ("the model will adjust the budget based on the complexity of the request").
// See https://ai.google.dev/gemini-api/docs/thinking#set-budget
// However, thoughts output also consumes total output token budget.
// We may consider adjusting ThinkingLevel parameter.
ThinkingBudget: genai.Ptr[int32](-1),
}
}
return client.Models.GenerateContent(ctx, modelPrefix+model, req, cfg)
}, nil
}
type Context struct {
Context context.Context
Workdir string
cache *Cache
cachedDirs []string
state map[string]any
onEvent onEvent
spanSeq int
spanNesting int
stubContext
}
type stubContext struct {
timeNow func() time.Time
generateContent generateContentFunc
}
func (ctx *Context) Cache(typ, desc string, populate func(string) error) (string, error) {
dir, err := ctx.cache.Create(typ, desc, populate)
if err != nil {
return "", err
}
ctx.cachedDirs = append(ctx.cachedDirs, dir)
return dir, nil
}
func (ctx *Context) close() {
for _, dir := range ctx.cachedDirs {
ctx.cache.Release(dir)
}
}
func (ctx *Context) startSpan(span *trajectory.Span) error {
span.Seq = ctx.spanSeq
ctx.spanSeq++
span.Nesting = ctx.spanNesting
ctx.spanNesting++
span.Started = ctx.timeNow()
return ctx.onEvent(span)
}
func (ctx *Context) finishSpan(span *trajectory.Span, spanErr error) error {
ctx.spanNesting--
if ctx.spanNesting < 0 {
panic("unbalanced spans")
}
span.Finished = ctx.timeNow()
if spanErr != nil {
span.Error = spanErr.Error()
}
err := ctx.onEvent(span)
if spanErr != nil {
err = spanErr
}
return err
}
|
