返回 DeepSeek-Reasonix
longrun.go
根目录 / cmd / e2ebench / longrun.go
1 package main
2
3 import (
4 "fmt"
5 "strings"
6 )
7
8 // faultRecovery splits a run by whether the meter actually failed it. With a
9 // cadence, short tasks never reach a fault and form an in-run control group,
10 // so the cost of failure is measured against the same suite and model rather
11 // than against a separate arm run at a different time.
12 type faultRecovery struct {
13 faulted, faultedSolved int
14 unfaulted, unfaultSolved int
15 keptGoing int // faulted runs that issued another request
16 injected int
17 }
18
19 func gatherFaultRecovery(results []result) faultRecovery {
20 var f faultRecovery
21 for _, r := range results {
22 if r.Skipped || r.Meter == nil || r.Attempt > 1 {
23 continue
24 }
25 if r.Meter.Injected == 0 {
26 f.unfaulted++
27 if r.Passed {
28 f.unfaultSolved++
29 }
30 continue
31 }
32 f.faulted++
33 f.injected += r.Meter.Injected
34 if r.Meter.RequestsAfterFault > 0 {
35 f.keptGoing++
36 }
37 if r.Passed {
38 f.faultedSolved++
39 }
40 }
41 return f
42 }
43
44 // renderFaultRecovery prices injected failure. Two different things are worth
45 // separating: whether the harness kept talking after a failure at all, and
46 // whether the task still landed. A harness can retry forever and still never
47 // finish, and that is not recovery.
48 func renderFaultRecovery(results []result) string {
49 f := gatherFaultRecovery(results)
50 if f.faulted == 0 {
51 return ""
52 }
53 var b strings.Builder
54 fmt.Fprintf(&b, "**Fault recovery** (%d runs failed on purpose, %d injections): **retried** %s (%d) · **still solved** %s (%d/%d)",
55 f.faulted, f.injected, pct(f.keptGoing, f.faulted), f.keptGoing,
56 pct(f.faultedSolved, f.faulted), f.faultedSolved, f.faulted)
57 if f.unfaulted > 0 {
58 fmt.Fprintf(&b, " · in-run control %s (%d/%d never hit a fault)",
59 pct(f.unfaultSolved, f.unfaulted), f.unfaultSolved, f.unfaulted)
60 }
61 return b.String() + "\n\n"
62 }
63
63 lines GO