| 1 | package boot |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "fmt" |
| 6 | "os/exec" |
| 7 | "strings" |
| 8 | "sync" |
| 9 | "testing" |
| 10 | |
| 11 | "reasonix/internal/control" |
| 12 | "reasonix/internal/event" |
| 13 | "reasonix/internal/provider" |
| 14 | ) |
| 15 | |
| 16 | // readThenMutateProvider reads a large file, then runs git commit, then lists |
| 17 | // capabilities and finishes. Truncated/windowed reads must not freeze commits |
| 18 | // or finals through the real Build stack. |
| 19 | type readThenMutateProvider struct { |
| 20 | mu sync.Mutex |
| 21 | reqs []provider.Request |
| 22 | round int |
| 23 | } |
| 24 | |
| 25 | func (p *readThenMutateProvider) Name() string { return "boot-read-then-mutate" } |
| 26 | |
| 27 | func (p *readThenMutateProvider) Stream(_ context.Context, req provider.Request) (<-chan provider.Chunk, error) { |
| 28 | p.mu.Lock() |
| 29 | p.reqs = append(p.reqs, req) |
| 30 | p.round++ |
| 31 | round := p.round |
| 32 | p.mu.Unlock() |
| 33 | switch round { |
| 34 | case 1: |
| 35 | return streamChunks( |
| 36 | provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ |
| 37 | ID: "read-1", Name: "read_file", Arguments: `{"path":"big.txt"}`, |
| 38 | }}, |
| 39 | provider.Chunk{Type: provider.ChunkDone}, |
| 40 | ), nil |
| 41 | case 2: |
| 42 | return streamChunks( |
| 43 | provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ |
| 44 | ID: "commit-1", Name: "bash", Arguments: `{"command":"git commit -m test-commit"}`, |
| 45 | }}, |
| 46 | provider.Chunk{Type: provider.ChunkDone}, |
| 47 | ), nil |
| 48 | case 3: |
| 49 | return streamChunks( |
| 50 | provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ |
| 51 | ID: "list-1", Name: "use_capability", Arguments: `{"action":"list"}`, |
| 52 | }}, |
| 53 | provider.Chunk{Type: provider.ChunkDone}, |
| 54 | ), nil |
| 55 | default: |
| 56 | return streamChunks( |
| 57 | provider.Chunk{Type: provider.ChunkText, Text: "committed after the partial read"}, |
| 58 | provider.Chunk{Type: provider.ChunkDone}, |
| 59 | ), nil |
| 60 | } |
| 61 | } |
| 62 | |
| 63 | func (p *readThenMutateProvider) requests() []provider.Request { |
| 64 | p.mu.Lock() |
| 65 | defer p.mu.Unlock() |
| 66 | out := make([]provider.Request, len(p.reqs)) |
| 67 | copy(out, p.reqs) |
| 68 | return out |
| 69 | } |
| 70 | |
| 71 | func TestEffectTruncatedReadDoesNotBlockCommitOrFinalThroughRealBuild(t *testing.T) { |
| 72 | isolateConfigHome(t) |
| 73 | dir := robustTempDir(t) |
| 74 | t.Chdir(dir) |
| 75 | for _, args := range [][]string{ |
| 76 | {"init"}, |
| 77 | {"config", "user.email", "test@example.com"}, |
| 78 | {"config", "user.name", "test"}, |
| 79 | } { |
| 80 | cmd := exec.Command("git", args...) |
| 81 | cmd.Dir = dir |
| 82 | if out, err := cmd.CombinedOutput(); err != nil { |
| 83 | t.Fatalf("git %v: %v\n%s", args, err, out) |
| 84 | } |
| 85 | } |
| 86 | writeFile(t, dir, "tracked.txt", "tracked\n") |
| 87 | cmd := exec.Command("git", "add", "tracked.txt") |
| 88 | cmd.Dir = dir |
| 89 | if out, err := cmd.CombinedOutput(); err != nil { |
| 90 | t.Fatalf("git add: %v\n%s", err, out) |
| 91 | } |
| 92 | |
| 93 | var body strings.Builder |
| 94 | for i := 1; i <= 4500; i++ { |
| 95 | fmt.Fprintf(&body, "line-%04d-content-for-pagination-test\n", i) |
| 96 | } |
| 97 | writeFile(t, dir, "big.txt", body.String()) |
| 98 | |
| 99 | rec := &readThenMutateProvider{} |
| 100 | provider.Register("boot-read-then-mutate", func(provider.Config) (provider.Provider, error) { |
| 101 | return rec, nil |
| 102 | }) |
| 103 | // This case proves the read-evidence gate releases a later commit, so git must |
| 104 | // really run. An enforced sandbox fails closed wherever the host lacks a backend |
| 105 | // (coverage runners), which is the sandbox's own contract, not this one. |
| 106 | writeUserConfig(t, "[sandbox]\nbash = \"off\"\n") |
| 107 | writeFile(t, dir, "reasonix.toml", ` |
| 108 | default_model = "test-model" |
| 109 | |
| 110 | [agent] |
| 111 | system_prompt = "BASE" |
| 112 | |
| 113 | [environment] |
| 114 | enabled = false |
| 115 | |
| 116 | [[providers]] |
| 117 | name = "test-model" |
| 118 | kind = "boot-read-then-mutate" |
| 119 | model = "x" |
| 120 | `) |
| 121 | approveWorkspace(t, dir) |
| 122 | |
| 123 | ctrl, err := Build(context.Background(), Options{ |
| 124 | Sink: event.Discard, |
| 125 | PermissionAllow: []string{"Bash(git *)"}, |
| 126 | }) |
| 127 | if err != nil { |
| 128 | t.Fatalf("Build: %v", err) |
| 129 | } |
| 130 | defer ctrl.Close() |
| 131 | ctrl.ApplyHeadlessApprovalMode(control.ToolApprovalDangerFullAccess) |
| 132 | |
| 133 | if err := ctrl.Run(context.Background(), "read big.txt then commit"); err != nil { |
| 134 | t.Fatalf("Run: %v", err) |
| 135 | } |
| 136 | |
| 137 | reqs := rec.requests() |
| 138 | if len(reqs) < 4 { |
| 139 | t.Fatalf("provider rounds=%d, want at least read + commit + list + final", len(reqs)) |
| 140 | } |
| 141 | |
| 142 | readResult := lastToolContentByID(reqs, "read-1") |
| 143 | if readResult == "" { |
| 144 | t.Fatal("read_file result never reached a later provider request") |
| 145 | } |
| 146 | if strings.Contains(readResult, "do not answer, modify state") { |
| 147 | t.Fatalf("read_file result still forces completion:\n%.400s", readResult) |
| 148 | } |
| 149 | if !strings.Contains(readResult, "PARTIAL view") && !strings.Contains(readResult, "next_offset=") { |
| 150 | t.Fatalf("read_file result missing pagination hint:\n%.400s", readResult) |
| 151 | } |
| 152 | |
| 153 | commitResult := lastToolContentByID(reqs, "commit-1") |
| 154 | if commitResult == "" { |
| 155 | t.Fatal("git commit result never reached a later provider request") |
| 156 | } |
| 157 | if strings.Contains(commitResult, "unread content retained") || |
| 158 | strings.Contains(commitResult, "restricted search/read mode") || |
| 159 | strings.Contains(commitResult, "cannot declare which files it changes") { |
| 160 | t.Fatalf("git commit was blocked by leftover read gate:\n%.400s", commitResult) |
| 161 | } |
| 162 | |
| 163 | verify := exec.Command("git", "log", "-1", "--format=%s") |
| 164 | verify.Dir = dir |
| 165 | if out, err := verify.CombinedOutput(); err != nil || strings.TrimSpace(string(out)) != "test-commit" { |
| 166 | t.Fatalf("commit did not land: %s %v", out, err) |
| 167 | } |
| 168 | |
| 169 | listResult := lastToolContentByID(reqs, "list-1") |
| 170 | if listResult == "" { |
| 171 | t.Fatal("use_capability list result never reached a later provider request") |
| 172 | } |
| 173 | if !strings.Contains(listResult, "session:tool_result") { |
| 174 | t.Fatalf("capability list dropped session:tool_result:\n%.400s", listResult) |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | func lastToolContentByID(reqs []provider.Request, callID string) string { |
| 179 | var found string |
| 180 | for _, req := range reqs { |
| 181 | for _, msg := range req.Messages { |
| 182 | if msg.Role == provider.RoleTool && msg.ToolCallID == callID { |
| 183 | found = msg.Content |
| 184 | } |
| 185 | } |
| 186 | } |
| 187 | return found |
| 188 | } |
| 189 |