| 1 | // e2ebench runs the committed e2e task suite against a real provider and emits a |
| 2 | // markdown + JSON report (accuracy, cache-hit rate, token use, cost) for a PR. |
| 3 | package main |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "encoding/json" |
| 8 | "flag" |
| 9 | "fmt" |
| 10 | "io" |
| 11 | "os" |
| 12 | "os/exec" |
| 13 | "path/filepath" |
| 14 | "sort" |
| 15 | "strings" |
| 16 | "time" |
| 17 | |
| 18 | "github.com/BurntSushi/toml" |
| 19 | |
| 20 | "reasonix/internal/ablation" |
| 21 | fileencoding "reasonix/internal/fileutil/encoding" |
| 22 | ) |
| 23 | |
| 24 | type task struct { |
| 25 | ID string |
| 26 | Prompt string `toml:"prompt"` |
| 27 | MaxSteps int `toml:"max_steps"` |
| 28 | TimeoutSec int `toml:"timeout_sec"` |
| 29 | dir string |
| 30 | } |
| 31 | |
| 32 | type runMetrics struct { |
| 33 | PromptTokens int `json:"prompt_tokens"` |
| 34 | CompletionTokens int `json:"completion_tokens"` |
| 35 | CacheHitTokens int `json:"cache_hit_tokens"` |
| 36 | CacheMissTokens int `json:"cache_miss_tokens"` |
| 37 | // PrefixChangeReasonCounts mirrors internal/cli.RunMetrics's field of the |
| 38 | // same name: per-run tallies of why the cache prefix changed (e.g. |
| 39 | // "compact_auto", "snip", "tools"), omitempty for older metrics files. |
| 40 | PrefixChangeReasonCounts map[string]int `json:"prefix_change_reason_counts,omitempty"` |
| 41 | Steps int `json:"steps"` |
| 42 | Cost float64 `json:"cost"` |
| 43 | Currency string `json:"currency"` |
| 44 | Compactions int `json:"compactions"` |
| 45 | |
| 46 | // Optional Delivery capability counters (omitempty for baseline/old metrics). |
| 47 | ReadinessChecks int `json:"readiness_checks,omitempty"` |
| 48 | ReadinessRecoveries int `json:"readiness_recoveries,omitempty"` |
| 49 | CapabilityRoutes int `json:"capability_routes,omitempty"` |
| 50 | CapabilityRoutedCandidates int `json:"capability_routed_candidates,omitempty"` |
| 51 | CapabilityRoutedRequire int `json:"capability_routed_require,omitempty"` |
| 52 | CapabilityRoutedPrefer int `json:"capability_routed_prefer,omitempty"` |
| 53 | CapabilityRoutedSuggest int `json:"capability_routed_suggest,omitempty"` |
| 54 | CapabilityDeclines int `json:"capability_declines,omitempty"` |
| 55 | CapabilitySemanticRoutes int `json:"capability_semantic_routes,omitempty"` |
| 56 | CapabilitySkillInvocations int `json:"capability_skill_invocations,omitempty"` |
| 57 | CapabilityMCPCall int `json:"capability_mcp_call,omitempty"` |
| 58 | CapabilityReviewBlocks int `json:"capability_review_blocks,omitempty"` |
| 59 | CapabilityRouterCost float64 `json:"capability_router_cost,omitempty"` |
| 60 | CapabilityRouterLatencyMs int64 `json:"capability_router_latency_ms,omitempty"` |
| 61 | |
| 62 | Complete bool `json:"complete"` |
| 63 | Outcome string `json:"outcome,omitempty"` |
| 64 | ToolCalls int `json:"tool_calls,omitempty"` |
| 65 | ToolFailures int `json:"tool_failures,omitempty"` |
| 66 | SubagentToolCalls int `json:"subagent_tool_calls,omitempty"` |
| 67 | Retries int `json:"retries,omitempty"` |
| 68 | ToolCallsByName map[string]int `json:"tool_calls_by_name,omitempty"` |
| 69 | ToolFailuresByName map[string]int `json:"tool_failures_by_name,omitempty"` |
| 70 | } |
| 71 | |
| 72 | type result struct { |
| 73 | task |
| 74 | runMetrics |
| 75 | Profile string `json:"profile"` |
| 76 | // Arm is the ablation arm the harness requested, not the arm the child |
| 77 | // reported, so a run that died before writing metrics is still attributable. |
| 78 | Arm string `json:"arm"` |
| 79 | Passed bool |
| 80 | Skipped bool |
| 81 | Note string |
| 82 | // WallMs is the harness's own clock, not the agent's self-report, so the |
| 83 | // number stays comparable when the same suite runs against another harness. |
| 84 | WallMs int64 `json:"wall_ms"` |
| 85 | // Unaccounted marks a run whose metrics file never landed — a killed agent |
| 86 | // writes nothing. Its real cost is unknown, so it is kept out of the cost |
| 87 | // and token aggregates instead of being averaged in as zero, which would |
| 88 | // quietly understate every published per-task figure. |
| 89 | Unaccounted bool `json:"unaccounted"` |
| 90 | // Partial marks accounting recovered from an in-flight snapshot after the |
| 91 | // agent was killed. The numbers are real but stop at the last snapshot, so |
| 92 | // they are counted as lower bounds rather than dropped. |
| 93 | Partial bool `json:"partial"` |
| 94 | } |
| 95 | |
| 96 | // class is the published failure taxonomy: solved, the guard that stopped the |
| 97 | // run, or wrong_patch when the agent finished cleanly and the grader still |
| 98 | // failed. outcome carries the agent's own classification when it wrote metrics. |
| 99 | func (r result) class() string { |
| 100 | switch { |
| 101 | case r.Skipped: |
| 102 | return "skipped" |
| 103 | case r.Passed: |
| 104 | return "solved" |
| 105 | case r.Outcome != "" && r.Outcome != "success": |
| 106 | return r.Outcome |
| 107 | case r.Outcome == "": |
| 108 | return "no_metrics" |
| 109 | default: |
| 110 | return "wrong_patch" |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | const defaultSuiteTokenBudget = 800_000 |
| 115 | |
| 116 | func main() { |
| 117 | flag.Usage = func() { |
| 118 | fmt.Fprintf(flag.CommandLine.Output(), "e2ebench — Reasonix end-to-end benchmark.\n\n") |
| 119 | fmt.Fprintf(flag.CommandLine.Output(), "Usage of %s:\n", flag.CommandLine.Name()) |
| 120 | flag.PrintDefaults() |
| 121 | fmt.Fprintf(flag.CommandLine.Output(), "\nExamples:\n") |
| 122 | fmt.Fprintf(flag.CommandLine.Output(), " # Run the committed suite:\n") |
| 123 | fmt.Fprintf(flag.CommandLine.Output(), " %[1]s\n\n", strings.Replace(flag.CommandLine.Name(), "e2ebench", "go run ./cmd/e2ebench", 1)) |
| 124 | fmt.Fprintf(flag.CommandLine.Output(), " # Grade a PR's diff with a retry budget:\n") |
| 125 | fmt.Fprintf(flag.CommandLine.Output(), " %[1]s -mode diff -base origin/main-v2 -repo . -attempts 3 -timeout 1800\n", strings.Replace(flag.CommandLine.Name(), "e2ebench", "go run ./cmd/e2ebench", 1)) |
| 126 | fmt.Fprintf(flag.CommandLine.Output(), "\n # Run the same suite with the delivery contract:\n") |
| 127 | fmt.Fprintf(flag.CommandLine.Output(), " %[1]s -profile delivery\n", strings.Replace(flag.CommandLine.Name(), "e2ebench", "go run ./cmd/e2ebench", 1)) |
| 128 | } |
| 129 | |
| 130 | mode := flag.String("mode", "suite", "suite | diff | swebench") |
| 131 | subset := flag.String("subset", "benchmarks/swebench/subset.json", "swebench mode: instance subset file") |
| 132 | namespace := flag.String("namespace", "swebench", "swebench mode: registry namespace holding the evaluation images") |
| 133 | runID := flag.String("run-id", "reasonix", "swebench mode: run id passed to the official harness") |
| 134 | harnessPy := flag.String("harness-python", "python3", "swebench mode: interpreter with the swebench package installed") |
| 135 | dataset := flag.String("dataset", "princeton-nlp/SWE-bench_Verified", "swebench mode: dataset name") |
| 136 | permission := flag.String("permission", "auto", "swebench mode: agent permission posture (auto | yolo)") |
| 137 | network := flag.String("network", "", "swebench mode: docker network for agent containers; must have no off-box route") |
| 138 | proxyURL := flag.String("proxy", "", "swebench mode: the only egress the agent gets, expected to allowlist just the model API") |
| 139 | workers := flag.Int("workers", 4, "swebench mode: parallel grader workers") |
| 140 | keepImages := flag.Bool("keep-images", false, "swebench mode: keep instance images instead of removing them after each run") |
| 141 | suite := flag.String("suite", "benchmarks/e2e", "suite root (contains tasks/<id>/)") |
| 142 | bin := flag.String("bin", "reasonix", "path to the reasonix binary") |
| 143 | model := flag.String("model", "", "provider/model name (default: config default)") |
| 144 | profileFlag := flag.String("profile", benchmarkProfileBaseline, "prompt profile: baseline | delivery") |
| 145 | ablateFlag := flag.String("ablate", "", "ablation arm: subsystems to switch off (evidence, planner, subagent, retrieval, compaction; none|all)") |
| 146 | outMD := flag.String("out", "", "write the markdown report here (default: stdout)") |
| 147 | outJSON := flag.String("json", "", "write the JSON report here (optional)") |
| 148 | budget := flag.Int("budget", defaultSuiteTokenBudget, "abort once total tokens cross this (0 = no cap)") |
| 149 | // diff-mode flags |
| 150 | repo := flag.String("repo", ".", "repo root (diff mode)") |
| 151 | base := flag.String("base", "", "base ref to diff the PR head against (diff mode)") |
| 152 | testCmd := flag.String("test-cmd", "go test", "grader command run on the affected packages (diff mode)") |
| 153 | maxSteps := flag.Int("max-steps", 80, "agent tool-call cap for the diff task") |
| 154 | timeoutSec := flag.Int("timeout", 1200, "agent timeout in seconds (diff mode)") |
| 155 | attempts := flag.Int("attempts", 1, "diff mode: retry up to N times until a run passes (stochastic agent)") |
| 156 | flag.Parse() |
| 157 | profile, err := normalizeBenchmarkProfile(*profileFlag) |
| 158 | if err != nil { |
| 159 | fmt.Fprintln(os.Stderr, err) |
| 160 | os.Exit(2) |
| 161 | } |
| 162 | arm, err := ablation.Parse(*ablateFlag) |
| 163 | if err != nil { |
| 164 | fmt.Fprintln(os.Stderr, err) |
| 165 | os.Exit(2) |
| 166 | } |
| 167 | |
| 168 | if *mode == "swebench" { |
| 169 | if _, err := permissionFlag(*permission); err != nil { |
| 170 | fmt.Fprintln(os.Stderr, err) |
| 171 | os.Exit(2) |
| 172 | } |
| 173 | cwd, _ := os.Getwd() |
| 174 | report := runSwebench(swebenchOpts{ |
| 175 | bin: *bin, subset: *subset, namespace: *namespace, model: *model, |
| 176 | profile: profile, permission: *permission, arm: arm, runID: *runID, workDir: cwd, |
| 177 | harness: *harnessPy, dataset: *dataset, maxSteps: *maxSteps, |
| 178 | timeoutSec: *timeoutSec, workers: *workers, keepImages: *keepImages, |
| 179 | network: *network, proxyURL: *proxyURL, |
| 180 | }) |
| 181 | emit(report, *outMD, "") |
| 182 | if *outJSON != "" { |
| 183 | fmt.Fprintln(os.Stderr, "note: -json is not written in swebench mode; the harness report is authoritative") |
| 184 | } |
| 185 | return |
| 186 | } |
| 187 | |
| 188 | if *mode == "diff" { |
| 189 | report := runDiff(diffOpts{ |
| 190 | bin: *bin, model: *model, repo: *repo, base: *base, |
| 191 | testCmd: *testCmd, profile: profile, ablate: arm, maxSteps: *maxSteps, timeoutSec: *timeoutSec, attempts: *attempts, |
| 192 | }) |
| 193 | emit(report, *outMD, "") |
| 194 | return |
| 195 | } |
| 196 | |
| 197 | tasks, err := loadTasks(*suite) |
| 198 | if err != nil { |
| 199 | fmt.Fprintln(os.Stderr, "load suite:", err) |
| 200 | os.Exit(1) |
| 201 | } |
| 202 | if len(tasks) == 0 { |
| 203 | dir := filepath.Join(*suite, "tasks") |
| 204 | if _, statErr := os.Stat(dir); statErr != nil { |
| 205 | fmt.Fprintf(os.Stderr, "no tasks found under %s: %v\n", dir, statErr) |
| 206 | } else { |
| 207 | fmt.Fprintf(os.Stderr, "no tasks found under %s (the directory exists but contains no task.toml files)\n", dir) |
| 208 | } |
| 209 | os.Exit(1) |
| 210 | } |
| 211 | |
| 212 | var results []result |
| 213 | total := 0 |
| 214 | for _, t := range tasks { |
| 215 | if *budget > 0 && total >= *budget { |
| 216 | results = append(results, result{task: t, Profile: profile, Skipped: true, Note: "skipped: token budget reached"}) |
| 217 | continue |
| 218 | } |
| 219 | r := runTask(*bin, *model, profile, arm, t) |
| 220 | total += r.PromptTokens + r.CompletionTokens |
| 221 | results = append(results, r) |
| 222 | } |
| 223 | |
| 224 | report := render(results) |
| 225 | if *outMD != "" { |
| 226 | if err := os.WriteFile(*outMD, []byte(report), 0o644); err != nil { |
| 227 | fmt.Fprintln(os.Stderr, "write report:", err) |
| 228 | os.Exit(1) |
| 229 | } |
| 230 | } else { |
| 231 | fmt.Print(report) |
| 232 | } |
| 233 | if *outJSON != "" { |
| 234 | b, err := json.MarshalIndent(results, "", " ") |
| 235 | if err != nil { |
| 236 | fmt.Fprintln(os.Stderr, "marshal json:", err) |
| 237 | os.Exit(1) |
| 238 | } |
| 239 | if err := os.WriteFile(*outJSON, b, 0o644); err != nil { |
| 240 | fmt.Fprintln(os.Stderr, "write json:", err) |
| 241 | os.Exit(1) |
| 242 | } |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | func emit(report, outMD, _ string) { |
| 247 | if outMD != "" { |
| 248 | if err := os.WriteFile(outMD, []byte(report), 0o644); err != nil { |
| 249 | fmt.Fprintln(os.Stderr, "write report:", err) |
| 250 | os.Exit(1) |
| 251 | } |
| 252 | return |
| 253 | } |
| 254 | fmt.Print(report) |
| 255 | } |
| 256 | |
| 257 | func loadTasks(suite string) ([]task, error) { |
| 258 | tasksDir := filepath.Join(suite, "tasks") |
| 259 | entries, err := os.ReadDir(tasksDir) |
| 260 | if err != nil { |
| 261 | return nil, err |
| 262 | } |
| 263 | var tasks []task |
| 264 | for _, e := range entries { |
| 265 | if !e.IsDir() { |
| 266 | continue |
| 267 | } |
| 268 | dir := filepath.Join(tasksDir, e.Name()) |
| 269 | var t task |
| 270 | data, err := fileencoding.ReadFileUTF8(filepath.Join(dir, "task.toml")) |
| 271 | if err != nil { |
| 272 | return nil, fmt.Errorf("%s: %w", e.Name(), err) |
| 273 | } |
| 274 | if _, err := toml.Decode(string(data), &t); err != nil { |
| 275 | return nil, fmt.Errorf("%s: %w", e.Name(), err) |
| 276 | } |
| 277 | t.ID = e.Name() |
| 278 | t.dir = dir |
| 279 | if t.TimeoutSec == 0 { |
| 280 | t.TimeoutSec = 240 |
| 281 | } |
| 282 | tasks = append(tasks, t) |
| 283 | } |
| 284 | sort.Slice(tasks, func(i, j int) bool { return tasks[i].ID < tasks[j].ID }) |
| 285 | return tasks, nil |
| 286 | } |
| 287 | |
| 288 | // runTask copies the task's seed workdir into a temp dir, runs the agent there, |
| 289 | // then drops in verify.sh and runs it as the grader. The grader is added only |
| 290 | // after the run so the agent can't read the answer key. |
| 291 | func runTask(bin, model, profile string, arm ablation.Set, t task) result { |
| 292 | r := result{task: t, Profile: profile} |
| 293 | r.Arm = arm.Arm() |
| 294 | |
| 295 | work, err := os.MkdirTemp("", "e2ebench-"+t.ID+"-") |
| 296 | if err != nil { |
| 297 | r.Note = "mktemp: " + err.Error() |
| 298 | return r |
| 299 | } |
| 300 | defer os.RemoveAll(work) |
| 301 | |
| 302 | if seed := filepath.Join(t.dir, "workdir"); dirExists(seed) { |
| 303 | if err := copyDir(seed, work); err != nil { |
| 304 | r.Note = "copy seed: " + err.Error() |
| 305 | return r |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | ctx, cancel := context.WithTimeout(context.Background(), time.Duration(t.TimeoutSec)*time.Second) |
| 310 | defer cancel() |
| 311 | |
| 312 | metricsPath := filepath.Join(work, ".run-metrics.json") |
| 313 | args := buildRunTaskArgs(metricsPath, model, profile, arm, t.MaxSteps, t.Prompt) |
| 314 | |
| 315 | cmd := exec.CommandContext(ctx, bin, args...) |
| 316 | cmd.Dir = work |
| 317 | cmd.Stdout = os.Stderr // stream the run to the job log, keep stdout clean for the report |
| 318 | cmd.Stderr = os.Stderr |
| 319 | cmd.WaitDelay = 10 * time.Second // bound the wait for a stuck child after ctx timeout |
| 320 | startedAt := time.Now() |
| 321 | runErr := cmd.Run() |
| 322 | r.WallMs = time.Since(startedAt).Milliseconds() |
| 323 | |
| 324 | if m, err := readMetrics(metricsPath); err == nil { |
| 325 | r.runMetrics = m |
| 326 | } |
| 327 | // A killed child never writes metrics, so the deadline is the only place |
| 328 | // this failure mode is still observable. |
| 329 | if ctx.Err() == context.DeadlineExceeded { |
| 330 | r.Outcome = "timeout" |
| 331 | } |
| 332 | if runErr != nil { |
| 333 | r.Note = "run: " + runErr.Error() |
| 334 | // still grade — a non-zero exit may just be a max-steps notice |
| 335 | } |
| 336 | |
| 337 | r.Passed = grade(work, t.dir) |
| 338 | return r |
| 339 | } |
| 340 | |
| 341 | func buildRunTaskArgs(metricsPath, model, profile string, arm ablation.Set, maxSteps int, prompt string) []string { |
| 342 | // Benchmarks are unattended and their fixtures require ordinary workspace |
| 343 | // writes. Auto still honors explicit ask/deny rules and the sandbox boundary. |
| 344 | args := []string{"run", "--auto", "--metrics", metricsPath} |
| 345 | if model != "" { |
| 346 | args = append(args, "--model", model) |
| 347 | } |
| 348 | if maxSteps > 0 { |
| 349 | args = append(args, "--max-steps", fmt.Sprint(maxSteps)) |
| 350 | } |
| 351 | args = appendBenchmarkProfileArgs(args, profile) |
| 352 | // The control arm must produce a byte-identical command line to the one the |
| 353 | // suite ran before ablation existed, so its numbers stay comparable. |
| 354 | if !arm.Empty() { |
| 355 | args = append(args, "--ablate", arm.String()) |
| 356 | } |
| 357 | return append(args, prompt) |
| 358 | } |
| 359 | |
| 360 | func grade(work, taskDir string) bool { |
| 361 | verify := filepath.Join(taskDir, "verify.sh") |
| 362 | if !fileExists(verify) { |
| 363 | return false |
| 364 | } |
| 365 | dst := filepath.Join(work, "verify.sh") |
| 366 | if err := copyFile(verify, dst); err != nil { |
| 367 | return false |
| 368 | } |
| 369 | cmd := exec.Command("bash", "verify.sh") |
| 370 | cmd.Dir = work |
| 371 | cmd.Stdout = os.Stderr |
| 372 | cmd.Stderr = os.Stderr |
| 373 | return cmd.Run() == nil |
| 374 | } |
| 375 | |
| 376 | func render(results []result) string { |
| 377 | profile := benchmarkProfileBaseline |
| 378 | arm := "full" |
| 379 | if len(results) > 0 { |
| 380 | if results[0].Profile != "" { |
| 381 | profile = results[0].Profile |
| 382 | } |
| 383 | if results[0].Arm != "" { |
| 384 | arm = results[0].Arm |
| 385 | } |
| 386 | } |
| 387 | return fmt.Sprintf("## 🤖 Reasonix e2e benchmark (%s · arm `%s`)\n\n", profile, arm) + renderBody(results) |
| 388 | } |
| 389 | |
| 390 | // renderBody is the report without a heading, so a caller that supplies its own |
| 391 | // (SWE-bench mode) does not stack two titles. |
| 392 | func renderBody(results []result) string { |
| 393 | var b strings.Builder |
| 394 | passed, ran := 0, 0 |
| 395 | accounted, accountedSolved, unaccounted, unaccountedSolved, partial := 0, 0, 0, 0, 0 |
| 396 | var pTok, cTok, hit, miss, compacts, tools, toolFails int |
| 397 | var cost float64 |
| 398 | var walls []int64 |
| 399 | currency := "" |
| 400 | classes := map[string]int{} |
| 401 | prefixChangeReasons := map[string]int{} |
| 402 | for _, r := range results { |
| 403 | if r.Skipped { |
| 404 | continue |
| 405 | } |
| 406 | ran++ |
| 407 | if r.Passed { |
| 408 | passed++ |
| 409 | } |
| 410 | classes[r.class()]++ |
| 411 | walls = append(walls, r.WallMs) |
| 412 | if r.Unaccounted { |
| 413 | unaccounted++ |
| 414 | if r.Passed { |
| 415 | unaccountedSolved++ |
| 416 | } |
| 417 | continue |
| 418 | } |
| 419 | accounted++ |
| 420 | if r.Passed { |
| 421 | accountedSolved++ |
| 422 | } |
| 423 | if r.Partial { |
| 424 | partial++ |
| 425 | } |
| 426 | pTok += r.PromptTokens |
| 427 | cTok += r.CompletionTokens |
| 428 | hit += r.CacheHitTokens |
| 429 | miss += r.CacheMissTokens |
| 430 | compacts += r.Compactions |
| 431 | tools += r.ToolCalls |
| 432 | toolFails += r.ToolFailures |
| 433 | cost += r.Cost |
| 434 | if r.Currency != "" { |
| 435 | currency = r.Currency |
| 436 | } |
| 437 | for reason, n := range r.PrefixChangeReasonCounts { |
| 438 | prefixChangeReasons[reason] += n |
| 439 | } |
| 440 | } |
| 441 | |
| 442 | // Cost and tokens are divided by the solved instances we actually have |
| 443 | // accounting for. Dividing by every solve would treat a lost metrics file as |
| 444 | // a free solve and understate the published figure. |
| 445 | fmt.Fprintf(&b, "**Solved:** %d/%d (%s) · **Cost per solved:** %s · **Tokens per solved:** %s · **Median wall time:** %s\n\n", |
| 446 | passed, ran, pct(passed, ran), |
| 447 | costPerSolved(cost, accountedSolved, currency), tokensPerSolved(pTok+cTok, accountedSolved), dur(median(walls))) |
| 448 | fmt.Fprintf(&b, "**Cache hit:** %s · **Tokens:** %s (prompt %s / completion %s) · **Tool calls:** %s (%s failed) · **Compactions:** %d · **Cost:** %s%.4f\n\n", |
| 449 | pct(hit, hit+miss), comma(pTok+cTok), comma(pTok), comma(cTok), |
| 450 | comma(tools), comma(toolFails), compacts, currencySym(currency), cost) |
| 451 | if unaccounted > 0 { |
| 452 | fmt.Fprintf(&b, "> **Accounting incomplete for %d of %d instances** (%d of them solved): the agent was killed before it wrote any metrics, so their cost and tokens are unknown. Totals above cover the %d accounted instances only, and per-solved figures divide by the %d accounted solves — the true totals are higher.\n\n", |
| 453 | unaccounted, ran, unaccountedSolved, accounted, accountedSolved) |
| 454 | } |
| 455 | if partial > 0 { |
| 456 | fmt.Fprintf(&b, "> **%d of %d instances contributed partial accounting**: the agent was killed mid-run and its numbers were recovered from the last in-flight snapshot. What is counted is real but stops at that snapshot, so every total above is a lower bound.\n\n", |
| 457 | partial, ran) |
| 458 | } |
| 459 | |
| 460 | fmt.Fprintf(&b, "| Task | Result | Class | Steps | Tools | Time | Prompt | Completion | Cache hit | Cost |\n") |
| 461 | fmt.Fprintf(&b, "|------|--------|-------|------:|------:|-----:|-------:|-----------:|----------:|-----:|\n") |
| 462 | for _, r := range results { |
| 463 | switch { |
| 464 | case r.Skipped: |
| 465 | fmt.Fprintf(&b, "| `%s` | ⏭️ skipped | — | — | — | — | — | — | — | — |\n", r.ID) |
| 466 | default: |
| 467 | res := "❌ fail" |
| 468 | if r.Passed { |
| 469 | res = "✅ pass" |
| 470 | } |
| 471 | fmt.Fprintf(&b, "| `%s` | %s | %s | %d | %d | %s | %s | %s | %s | %s%.4f |\n", |
| 472 | r.ID, res, r.class(), r.Steps, r.ToolCalls, dur(r.WallMs), |
| 473 | comma(r.PromptTokens), comma(r.CompletionTokens), |
| 474 | pct(r.CacheHitTokens, r.CacheHitTokens+r.CacheMissTokens), |
| 475 | currencySym(r.Currency), r.Cost) |
| 476 | } |
| 477 | } |
| 478 | fmt.Fprintf(&b, "\n<sub>Real provider run. Cache-hit %% is cached prompt tokens / total prompt tokens. Wall time is measured by the harness and includes process startup.</sub>\n") |
| 479 | |
| 480 | if breakdown := failureBreakdown(classes); breakdown != "" { |
| 481 | fmt.Fprintf(&b, "\n**Failures by class:** %s\n", breakdown) |
| 482 | } |
| 483 | if breakdown := reasonBreakdown(prefixChangeReasons); breakdown != "" { |
| 484 | fmt.Fprintf(&b, "\n**Cache resets by cause:** %s\n", breakdown) |
| 485 | } |
| 486 | |
| 487 | notes := false |
| 488 | for _, r := range results { |
| 489 | if r.Note != "" { |
| 490 | if !notes { |
| 491 | fmt.Fprintf(&b, "\n<details><summary>Notes</summary>\n\n") |
| 492 | notes = true |
| 493 | } |
| 494 | fmt.Fprintf(&b, "- `%s`: %s\n", r.ID, r.Note) |
| 495 | } |
| 496 | } |
| 497 | if notes { |
| 498 | fmt.Fprintf(&b, "\n</details>\n") |
| 499 | } |
| 500 | return b.String() |
| 501 | } |
| 502 | |
| 503 | func pct(n, d int) string { |
| 504 | if d == 0 { |
| 505 | return "n/a" |
| 506 | } |
| 507 | return fmt.Sprintf("%.0f%%", 100*float64(n)/float64(d)) |
| 508 | } |
| 509 | |
| 510 | func costPerSolved(cost float64, solved int, currency string) string { |
| 511 | if solved == 0 { |
| 512 | return "n/a" |
| 513 | } |
| 514 | return fmt.Sprintf("%s%.4f", currencySym(currency), cost/float64(solved)) |
| 515 | } |
| 516 | |
| 517 | func tokensPerSolved(tokens, solved int) string { |
| 518 | if solved == 0 { |
| 519 | return "n/a" |
| 520 | } |
| 521 | return comma(tokens / solved) |
| 522 | } |
| 523 | |
| 524 | func median(ms []int64) int64 { |
| 525 | if len(ms) == 0 { |
| 526 | return 0 |
| 527 | } |
| 528 | sorted := append([]int64(nil), ms...) |
| 529 | sort.Slice(sorted, func(i, j int) bool { return sorted[i] < sorted[j] }) |
| 530 | return sorted[len(sorted)/2] |
| 531 | } |
| 532 | |
| 533 | func dur(ms int64) string { |
| 534 | if ms <= 0 { |
| 535 | return "—" |
| 536 | } |
| 537 | d := time.Duration(ms) * time.Millisecond |
| 538 | if d < time.Minute { |
| 539 | return fmt.Sprintf("%.1fs", d.Seconds()) |
| 540 | } |
| 541 | return fmt.Sprintf("%dm%02ds", int(d.Minutes()), int(d.Seconds())%60) |
| 542 | } |
| 543 | |
| 544 | func failureBreakdown(classes map[string]int) string { |
| 545 | names := make([]string, 0, len(classes)) |
| 546 | for name := range classes { |
| 547 | if name != "solved" { |
| 548 | names = append(names, name) |
| 549 | } |
| 550 | } |
| 551 | if len(names) == 0 { |
| 552 | return "" |
| 553 | } |
| 554 | sort.Strings(names) |
| 555 | parts := make([]string, 0, len(names)) |
| 556 | for _, name := range names { |
| 557 | parts = append(parts, fmt.Sprintf("%s ×%d", name, classes[name])) |
| 558 | } |
| 559 | return strings.Join(parts, " · ") |
| 560 | } |
| 561 | |
| 562 | // reasonBreakdown renders cache-prefix-change reason counts (compact_auto, |
| 563 | // snip, prune, tools, ...) the same way failureBreakdown renders failure |
| 564 | // classes, so a hit-rate regression in a PR shows which operation caused it. |
| 565 | func reasonBreakdown(reasons map[string]int) string { |
| 566 | names := make([]string, 0, len(reasons)) |
| 567 | for name := range reasons { |
| 568 | names = append(names, name) |
| 569 | } |
| 570 | if len(names) == 0 { |
| 571 | return "" |
| 572 | } |
| 573 | sort.Strings(names) |
| 574 | parts := make([]string, 0, len(names)) |
| 575 | for _, name := range names { |
| 576 | parts = append(parts, fmt.Sprintf("%s ×%d", name, reasons[name])) |
| 577 | } |
| 578 | return strings.Join(parts, " · ") |
| 579 | } |
| 580 | |
| 581 | func comma(n int) string { |
| 582 | s := fmt.Sprint(n) |
| 583 | if len(s) <= 3 { |
| 584 | return s |
| 585 | } |
| 586 | var out []byte |
| 587 | for i, c := range []byte(s) { |
| 588 | if i > 0 && (len(s)-i)%3 == 0 { |
| 589 | out = append(out, ',') |
| 590 | } |
| 591 | out = append(out, c) |
| 592 | } |
| 593 | return string(out) |
| 594 | } |
| 595 | |
| 596 | func currencySym(c string) string { |
| 597 | if c == "" { |
| 598 | return "" |
| 599 | } |
| 600 | return c + " " |
| 601 | } |
| 602 | |
| 603 | func readMetrics(path string) (runMetrics, error) { |
| 604 | var m runMetrics |
| 605 | b, err := fileencoding.ReadFileUTF8(path) |
| 606 | if err != nil { |
| 607 | return m, err |
| 608 | } |
| 609 | return m, json.Unmarshal(b, &m) |
| 610 | } |
| 611 | |
| 612 | func dirExists(p string) bool { |
| 613 | fi, err := os.Stat(p) |
| 614 | return err == nil && fi.IsDir() |
| 615 | } |
| 616 | |
| 617 | func fileExists(p string) bool { |
| 618 | fi, err := os.Stat(p) |
| 619 | return err == nil && !fi.IsDir() |
| 620 | } |
| 621 | |
| 622 | func copyDir(src, dst string) error { |
| 623 | return filepath.Walk(src, func(p string, info os.FileInfo, err error) error { |
| 624 | if err != nil { |
| 625 | return err |
| 626 | } |
| 627 | // Skip symlinks so a seed link can't leak a file from outside the seed tree. |
| 628 | if info.Mode()&os.ModeSymlink != 0 { |
| 629 | return nil |
| 630 | } |
| 631 | rel, _ := filepath.Rel(src, p) |
| 632 | target := filepath.Join(dst, rel) |
| 633 | if info.IsDir() { |
| 634 | return os.MkdirAll(target, 0o755) |
| 635 | } |
| 636 | return copyFile(p, target) |
| 637 | }) |
| 638 | } |
| 639 | |
| 640 | func copyFile(src, dst string) error { |
| 641 | if err := os.MkdirAll(filepath.Dir(dst), 0o755); err != nil { |
| 642 | return err |
| 643 | } |
| 644 | in, err := os.Open(src) |
| 645 | if err != nil { |
| 646 | return err |
| 647 | } |
| 648 | defer in.Close() |
| 649 | info, err := in.Stat() |
| 650 | if err != nil { |
| 651 | return err |
| 652 | } |
| 653 | out, err := os.OpenFile(dst, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, info.Mode().Perm()) |
| 654 | if err != nil { |
| 655 | return err |
| 656 | } |
| 657 | defer out.Close() |
| 658 | if _, err := io.Copy(out, in); err != nil { |
| 659 | return err |
| 660 | } |
| 661 | // Mirror the source mode so a seed's read-only / exec bit survives the copy. |
| 662 | return os.Chmod(dst, info.Mode().Perm()) |
| 663 | } |
| 664 |