| 1 | package builtin |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "io" |
| 6 | "strings" |
| 7 | "testing" |
| 8 | |
| 9 | "reasonix/internal/event" |
| 10 | "reasonix/internal/evidence" |
| 11 | "reasonix/internal/jobs" |
| 12 | "reasonix/internal/planmode" |
| 13 | ) |
| 14 | |
| 15 | // End-to-end through the actual tools: a background bash job runs under a manager |
| 16 | // injected on the context, the wait tool collects its output, and bash_output |
| 17 | // reads it — the same path the agent drives. |
| 18 | func TestBackgroundBashWaitAndOutput(t *testing.T) { |
| 19 | m := jobs.NewManager(event.Discard) |
| 20 | defer m.Close() |
| 21 | ctx := jobs.WithManager(context.Background(), m) |
| 22 | |
| 23 | start, err := bash{}.Execute(ctx, []byte(`{"command":"printf hello; sleep 0.3","run_in_background":true}`)) |
| 24 | if err != nil { |
| 25 | t.Fatalf("bash background: %v", err) |
| 26 | } |
| 27 | if !strings.Contains(start, "Started background job") { |
| 28 | t.Fatalf("unexpected start message: %q", start) |
| 29 | } |
| 30 | |
| 31 | // The job is registered and running synchronously before Execute returns. |
| 32 | running := m.Running() |
| 33 | if len(running) != 1 { |
| 34 | t.Fatalf("want 1 running job, got %d", len(running)) |
| 35 | } |
| 36 | id := running[0].ID |
| 37 | |
| 38 | // wait blocks until it finishes, then returns its output. |
| 39 | wout, err := waitJob{}.Execute(ctx, []byte(`{"job_ids":["`+id+`"]}`)) |
| 40 | if err != nil { |
| 41 | t.Fatalf("wait: %v", err) |
| 42 | } |
| 43 | if !strings.Contains(wout, "done") || !strings.Contains(wout, "hello") { |
| 44 | t.Errorf("wait output = %q, want it to report done with hello", wout) |
| 45 | } |
| 46 | |
| 47 | // bash_output reads the buffered output (wait doesn't consume the read cursor). |
| 48 | bo, err := bashOutput{}.Execute(ctx, []byte(`{"job_id":"`+id+`"}`)) |
| 49 | if err != nil { |
| 50 | t.Fatalf("bash_output: %v", err) |
| 51 | } |
| 52 | if !strings.Contains(bo, "hello") { |
| 53 | t.Errorf("bash_output = %q, want hello", bo) |
| 54 | } |
| 55 | } |
| 56 | |
| 57 | func TestWaitMergesBackgroundEvidenceExactlyOnce(t *testing.T) { |
| 58 | m := jobs.NewManager(event.Discard) |
| 59 | defer m.Close() |
| 60 | ledger := evidence.NewLedger() |
| 61 | ctx := jobs.WithManager(context.Background(), m) |
| 62 | ctx = jobs.WithSession(ctx, "session") |
| 63 | ctx = evidence.WithLedger(ctx, ledger) |
| 64 | |
| 65 | j := m.StartForSession("session", "task", "writer", func(jobCtx context.Context, _ io.Writer) (string, error) { |
| 66 | jobs.PublishEvidence(jobCtx, evidence.ChildEvidenceSummary{Receipts: []evidence.Receipt{{ |
| 67 | ToolName: "write_file", |
| 68 | Success: true, |
| 69 | Mutation: true, |
| 70 | Write: true, |
| 71 | Paths: []string{"changed.go"}, |
| 72 | }}}) |
| 73 | return "done", nil |
| 74 | }) |
| 75 | |
| 76 | args := []byte(`{"job_ids":["` + j.ID + `"]}`) |
| 77 | if _, err := (waitJob{}).Execute(ctx, args); err != nil { |
| 78 | t.Fatalf("wait: %v", err) |
| 79 | } |
| 80 | if !ledger.Summary().HasMutation() { |
| 81 | t.Fatal("wait did not merge the task job's mutation evidence") |
| 82 | } |
| 83 | firstLen := ledger.Len() |
| 84 | if _, err := (waitJob{}).Execute(ctx, args); err != nil { |
| 85 | t.Fatalf("second wait: %v", err) |
| 86 | } |
| 87 | if got := ledger.Len(); got != firstLen { |
| 88 | t.Fatalf("second wait duplicated evidence: len %d -> %d", firstLen, got) |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | func TestWaitWithoutLedgerDoesNotConsumeBackgroundEvidence(t *testing.T) { |
| 93 | m := jobs.NewManager(event.Discard) |
| 94 | defer m.Close() |
| 95 | baseCtx := jobs.WithSession(jobs.WithManager(context.Background(), m), "session") |
| 96 | j := m.StartForSession("session", "task", "writer", func(jobCtx context.Context, _ io.Writer) (string, error) { |
| 97 | jobs.PublishEvidence(jobCtx, evidence.ChildEvidenceSummary{Receipts: []evidence.Receipt{{ |
| 98 | ToolName: "write_file", Success: true, Mutation: true, Write: true, Paths: []string{"changed.go"}, |
| 99 | }}}) |
| 100 | return "done", nil |
| 101 | }) |
| 102 | args := []byte(`{"job_ids":["` + j.ID + `"]}`) |
| 103 | if _, err := (waitJob{}).Execute(baseCtx, args); err != nil { |
| 104 | t.Fatalf("wait without ledger: %v", err) |
| 105 | } |
| 106 | |
| 107 | ledger := evidence.NewLedger() |
| 108 | if _, err := (waitJob{}).Execute(evidence.WithLedger(baseCtx, ledger), args); err != nil { |
| 109 | t.Fatalf("wait with ledger: %v", err) |
| 110 | } |
| 111 | if !ledger.Summary().HasMutation() { |
| 112 | t.Fatal("wait without a ledger consumed evidence before a collecting turn could merge it") |
| 113 | } |
| 114 | } |
| 115 | |
| 116 | func TestWaitInPlanModeDefersBackgroundEvidence(t *testing.T) { |
| 117 | m := jobs.NewManager(event.Discard) |
| 118 | defer m.Close() |
| 119 | baseCtx := jobs.WithSession(jobs.WithManager(context.Background(), m), "session") |
| 120 | j := m.StartForSession("session", "task", "writer", func(jobCtx context.Context, _ io.Writer) (string, error) { |
| 121 | jobs.PublishEvidence(jobCtx, evidence.ChildEvidenceSummary{Receipts: []evidence.Receipt{{ |
| 122 | ToolName: "write_file", Success: true, Mutation: true, Write: true, Paths: []string{"changed.go"}, |
| 123 | }}}) |
| 124 | return "done", nil |
| 125 | }) |
| 126 | args := []byte(`{"job_ids":["` + j.ID + `"]}`) |
| 127 | |
| 128 | // A planning turn may wait on jobs, but merging mutation receipts there |
| 129 | // would arm delivery sign-off demands the read-only turn cannot satisfy. |
| 130 | planLedger := evidence.NewLedger() |
| 131 | planCtx := planmode.WithActive(evidence.WithLedger(baseCtx, planLedger), true) |
| 132 | if _, err := (waitJob{}).Execute(planCtx, args); err != nil { |
| 133 | t.Fatalf("wait in plan mode: %v", err) |
| 134 | } |
| 135 | if planLedger.Summary().HasMutation() { |
| 136 | t.Fatal("plan-mode wait merged mutation evidence into the planning turn") |
| 137 | } |
| 138 | |
| 139 | // The evidence stays on the job for the first normal turn to collect. |
| 140 | ledger := evidence.NewLedger() |
| 141 | normalCtx := planmode.WithActive(evidence.WithLedger(baseCtx, ledger), false) |
| 142 | if _, err := (waitJob{}).Execute(normalCtx, args); err != nil { |
| 143 | t.Fatalf("wait after plan mode: %v", err) |
| 144 | } |
| 145 | if !ledger.Summary().HasMutation() { |
| 146 | t.Fatal("plan-mode wait consumed the background evidence instead of deferring it") |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | // kill_shell terminates a long-running background job. |
| 151 | func TestBackgroundKill(t *testing.T) { |
| 152 | m := jobs.NewManager(event.Discard) |
| 153 | defer m.Close() |
| 154 | ctx := jobs.WithManager(context.Background(), m) |
| 155 | |
| 156 | if _, err := (bash{}).Execute(ctx, []byte(`{"command":"sleep 120","run_in_background":true}`)); err != nil { |
| 157 | t.Fatalf("bash background: %v", err) |
| 158 | } |
| 159 | id := m.Running()[0].ID |
| 160 | |
| 161 | kout, err := killShell{}.Execute(ctx, []byte(`{"job_id":"`+id+`"}`)) |
| 162 | if err != nil { |
| 163 | t.Fatalf("kill_shell: %v", err) |
| 164 | } |
| 165 | if !strings.Contains(kout, "Killed") { |
| 166 | t.Errorf("kill_shell = %q, want it to report Killed", kout) |
| 167 | } |
| 168 | // 120s natural duration keeps the job far from finishing on its own, so the |
| 169 | // reap window is the only thing this measures: a loaded machine's slow |
| 170 | // process-tree teardown (up to ~bashWaitDelay) still fits, while a genuinely |
| 171 | // broken kill trips the 40s timeout. Pairing the sleep with the timeout (as |
| 172 | // 10/10 did) raced natural completion against the reap. |
| 173 | res := m.Wait(ctx, []string{id}, 40) |
| 174 | if len(res) != 1 || res[0].Status != jobs.Killed { |
| 175 | t.Fatalf("want killed, got %+v", res) |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | // kill_shell flips a job's status to Killed synchronously, well before its |
| 180 | // cancelled run goroutine actually unwinds, flushes PublishEvidence, and closes |
| 181 | // done. A bash_output poll that lands in that window must not note an empty |
| 182 | // lease: the ledger's lease is idempotent per turn, so noting it early would |
| 183 | // dedupe away every later retry in this same turn while the job's real |
| 184 | // mutation evidence is still forthcoming — and a turn that goes on to deliver |
| 185 | // would then commit (permanently drain) evidence nobody ever merged or |
| 186 | // reviewed. This deterministically drives that exact window with channels |
| 187 | // instead of a timing race. |
| 188 | func TestKilledJobBashOutputDoesNotNoteLeaseBeforeEvidenceIsReady(t *testing.T) { |
| 189 | m := jobs.NewManager(event.Discard) |
| 190 | defer m.Close() |
| 191 | ledger := evidence.NewLedger() |
| 192 | ctx := jobs.WithManager(context.Background(), m) |
| 193 | ctx = jobs.WithSession(ctx, "session") |
| 194 | ctx = evidence.WithLedger(ctx, ledger) |
| 195 | |
| 196 | cancelSeen := make(chan struct{}) |
| 197 | release := make(chan struct{}) |
| 198 | j := m.StartForSession("session", "task", "writer", func(jobCtx context.Context, _ io.Writer) (string, error) { |
| 199 | <-jobCtx.Done() |
| 200 | close(cancelSeen) |
| 201 | // Simulate a job that keeps unwinding (e.g. a subprocess still tearing |
| 202 | // down) after cancellation is requested but before it actually returns. |
| 203 | <-release |
| 204 | jobs.PublishEvidence(jobCtx, evidence.ChildEvidenceSummary{Receipts: []evidence.Receipt{{ |
| 205 | ToolName: "write_file", Success: true, Mutation: true, Write: true, Paths: []string{"changed.go"}, |
| 206 | }}}) |
| 207 | return "", context.Canceled |
| 208 | }) |
| 209 | |
| 210 | if _, err := (killShell{}).Execute(ctx, []byte(`{"job_id":"`+j.ID+`"}`)); err != nil { |
| 211 | t.Fatalf("kill_shell: %v", err) |
| 212 | } |
| 213 | <-cancelSeen // the goroutine observed the cancellation but has not returned |
| 214 | |
| 215 | // bash_output lands in the unwinding window: status already reports Killed, |
| 216 | // but the job's done channel is not closed yet and no evidence exists. |
| 217 | bo, err := bashOutput{}.Execute(ctx, []byte(`{"job_id":"`+j.ID+`"}`)) |
| 218 | if err != nil { |
| 219 | t.Fatalf("bash_output during unwind: %v", err) |
| 220 | } |
| 221 | if !strings.Contains(bo, "killed") { |
| 222 | t.Fatalf("bash_output during unwind = %q, want killed status", bo) |
| 223 | } |
| 224 | if ledger.Summary().HasMutation() { |
| 225 | t.Fatal("bash_output merged mutation evidence before the job was ready") |
| 226 | } |
| 227 | if leases := ledger.BackgroundLeases(); len(leases) != 0 { |
| 228 | t.Fatalf("bash_output noted a lease before the job was ready: %+v", leases) |
| 229 | } |
| 230 | |
| 231 | close(release) |
| 232 | if res := m.WaitForSession(context.Background(), "session", []string{j.ID}, 5); len(res) != 1 || res[0].Status != jobs.Killed { |
| 233 | t.Fatalf("post-unwind wait = %+v, want one killed result", res) |
| 234 | } |
| 235 | |
| 236 | // A later retry — the model calling bash_output again, or the next turn's |
| 237 | // automatic re-lease — must still find the evidence, not a dead dedupe entry. |
| 238 | if _, err := (bashOutput{}).Execute(ctx, []byte(`{"job_id":"`+j.ID+`"}`)); err != nil { |
| 239 | t.Fatalf("bash_output after unwind: %v", err) |
| 240 | } |
| 241 | if !ledger.Summary().HasMutation() { |
| 242 | t.Fatal("bash_output did not collect the killed job's evidence once it became ready") |
| 243 | } |
| 244 | if leases := ledger.BackgroundLeases(); len(leases) != 1 { |
| 245 | t.Fatalf("leases = %+v, want exactly one lease recorded", leases) |
| 246 | } |
| 247 | } |
| 248 | |
| 249 | // Without a manager on the context the background tools degrade to a clear error |
| 250 | // rather than panicking. |
| 251 | func TestBackgroundToolsNoManager(t *testing.T) { |
| 252 | ctx := context.Background() |
| 253 | if _, err := (bashOutput{}).Execute(ctx, []byte(`{"job_id":"bash-1"}`)); err == nil { |
| 254 | t.Error("bash_output without a manager should error") |
| 255 | } |
| 256 | if _, err := (bash{}).Execute(ctx, []byte(`{"command":"echo hi","run_in_background":true}`)); err == nil { |
| 257 | t.Error("background bash without a manager should error") |
| 258 | } |
| 259 | } |
| 260 |