| 1 | package boot |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "os" |
| 8 | "path/filepath" |
| 9 | "strings" |
| 10 | "testing" |
| 11 | "time" |
| 12 | |
| 13 | "reasonix/internal/config" |
| 14 | "reasonix/internal/extension" |
| 15 | "reasonix/internal/extension/dispatch" |
| 16 | ) |
| 17 | |
| 18 | // Stage 6b1 dispatch wiring tests: the boot builds the dispatcher, runs the |
| 19 | // system_prompt.build strategy before the snapshot freezes, and broadcasts |
| 20 | // the final prompt to observers — all without letting dispatch results flow |
| 21 | // back into the frozen snapshot. |
| 22 | |
| 23 | func TestBootSystemPromptStrategyReplacementLandsInSnapshot(t *testing.T) { |
| 24 | res := bootWithFakePlugin(t, "prompt-owner", map[string]any{ |
| 25 | "replaces": []string{"system_prompt"}, |
| 26 | "env": map[string]string{bootFakeEnvReplacePrompt: "EXTENSION PROMPT"}, |
| 27 | }) |
| 28 | if res.Dispatcher == nil { |
| 29 | t.Fatal("BuildRuntime returned no dispatcher with a live sidecar") |
| 30 | } |
| 31 | if got := res.Snapshot.SystemPrompt(); got != "EXTENSION PROMPT" { |
| 32 | t.Fatalf("snapshot system prompt = %q, want the strategy replacement %q", got, "EXTENSION PROMPT") |
| 33 | } |
| 34 | // Stage 6b2 handoff: the snapshot (and its cache hash) records the |
| 35 | // strategy's final prompt, and the live executor session was swapped to |
| 36 | // the same prompt before the build returned — the session's system |
| 37 | // message is exactly what the agent will send on its first request. |
| 38 | if got := systemMessage(res.Controller.History()); got != "EXTENSION PROMPT" { |
| 39 | t.Fatalf("controller system message = %.80q, want the strategy-replaced prompt", got) |
| 40 | } |
| 41 | } |
| 42 | |
| 43 | // TestBootObservationOnlyKeepsSessionPrompt guards the swap condition: a |
| 44 | // build whose strategy did NOT replace the prompt must leave the executor |
| 45 | // session on the host-composed prompt (byte-identical pre-dispatch path). |
| 46 | func TestBootObservationOnlyKeepsSessionPrompt(t *testing.T) { |
| 47 | res := bootWithFakePlugin(t, "observer", map[string]any{ |
| 48 | "intercepts": []string{"input.receive"}, |
| 49 | }) |
| 50 | if res.Dispatcher == nil { |
| 51 | t.Fatal("BuildRuntime returned no dispatcher with a live sidecar") |
| 52 | } |
| 53 | if got := systemMessage(res.Controller.History()); !strings.Contains(got, "BASE SYSTEM PROMPT") { |
| 54 | t.Fatalf("controller system message = %.80q, want the host-composed prompt", got) |
| 55 | } |
| 56 | } |
| 57 | |
| 58 | func TestBootSystemPromptStrategyAttribution(t *testing.T) { |
| 59 | isolateConfigHome(t) |
| 60 | dir := robustTempDir(t) |
| 61 | t.Chdir(dir) |
| 62 | writeRuntimeFixture(t, dir) |
| 63 | home := config.ReasonixHomeDir() |
| 64 | |
| 65 | // Baseline: no sidecars — no dispatcher, and the golden-path fingerprint. |
| 66 | plain, err := BuildRuntime(context.Background(), Options{}) |
| 67 | if err != nil { |
| 68 | t.Fatalf("BuildRuntime: %v", err) |
| 69 | } |
| 70 | t.Cleanup(plain.Controller.Close) |
| 71 | if plain.Dispatcher != nil || plain.Extensions != nil { |
| 72 | t.Fatal("no-plugin build produced a dispatcher or manager") |
| 73 | } |
| 74 | baseHash := plain.Snapshot.CacheHash() |
| 75 | |
| 76 | // An observation-only sidecar (continue at input.receive) must not |
| 77 | // perturb the provider-visible fingerprint at all. |
| 78 | installBootFakePlugin(t, home, "observer", map[string]any{ |
| 79 | "intercepts": []string{"input.receive"}, |
| 80 | }) |
| 81 | observed, err := BuildRuntime(context.Background(), Options{}) |
| 82 | if err != nil { |
| 83 | t.Fatalf("BuildRuntime with observer: %v", err) |
| 84 | } |
| 85 | t.Cleanup(observed.Controller.Close) |
| 86 | if observed.Dispatcher == nil { |
| 87 | t.Fatal("observer build returned no dispatcher") |
| 88 | } |
| 89 | if got := observed.Snapshot.CacheHash(); got != baseHash { |
| 90 | t.Fatalf("observation-only build changed the cache hash: %s vs %s", got, baseHash) |
| 91 | } |
| 92 | |
| 93 | // A strategy replacement lands IN the hash: two builds whose owner |
| 94 | // returns different prompts produce different hashes, and both differ |
| 95 | // from the host-composed baseline. That is attribution — the fingerprint |
| 96 | // covers what the session was actually built with. |
| 97 | installBootFakePlugin(t, home, "strategist", map[string]any{ |
| 98 | "replaces": []string{"system_prompt"}, |
| 99 | "env": map[string]string{bootFakeEnvReplacePrompt: "PROMPT A"}, |
| 100 | }) |
| 101 | buildA, err := BuildRuntime(context.Background(), Options{}) |
| 102 | if err != nil { |
| 103 | t.Fatalf("BuildRuntime with strategist A: %v", err) |
| 104 | } |
| 105 | t.Cleanup(buildA.Controller.Close) |
| 106 | if got := buildA.Snapshot.SystemPrompt(); got != "PROMPT A" { |
| 107 | t.Fatalf("build A prompt = %q, want PROMPT A", got) |
| 108 | } |
| 109 | if buildA.Snapshot.CacheHash() == baseHash { |
| 110 | t.Fatal("strategy-replaced prompt did not change the cache hash") |
| 111 | } |
| 112 | |
| 113 | installBootFakePlugin(t, home, "strategist", map[string]any{ |
| 114 | "replaces": []string{"system_prompt"}, |
| 115 | "env": map[string]string{bootFakeEnvReplacePrompt: "PROMPT B"}, |
| 116 | }) |
| 117 | buildB, err := BuildRuntime(context.Background(), Options{}) |
| 118 | if err != nil { |
| 119 | t.Fatalf("BuildRuntime with strategist B: %v", err) |
| 120 | } |
| 121 | t.Cleanup(buildB.Controller.Close) |
| 122 | if got := buildB.Snapshot.SystemPrompt(); got != "PROMPT B" { |
| 123 | t.Fatalf("build B prompt = %q, want PROMPT B", got) |
| 124 | } |
| 125 | if buildB.Snapshot.CacheHash() == buildA.Snapshot.CacheHash() { |
| 126 | t.Fatal("different strategy replacements produced the same cache hash — attribution is broken") |
| 127 | } |
| 128 | } |
| 129 | |
| 130 | // TestBootStableExtensionCacheGuard proves that an enabled deterministic |
| 131 | // strategy has one stable provider-visible prefix across independent runtime |
| 132 | // generations. Installing the extension may intentionally create one cold |
| 133 | // prefix versus the no-extension baseline; identical reloads must not create |
| 134 | // a new one. |
| 135 | func TestBootStableExtensionCacheGuard(t *testing.T) { |
| 136 | isolateConfigHome(t) |
| 137 | dir := robustTempDir(t) |
| 138 | t.Chdir(dir) |
| 139 | writeRuntimeFixture(t, dir) |
| 140 | installBootFakePlugin(t, config.ReasonixHomeDir(), "stable-strategist", map[string]any{ |
| 141 | "replaces": []string{"system_prompt"}, |
| 142 | "env": map[string]string{bootFakeEnvReplacePrompt: "STABLE EXTENSION PROMPT"}, |
| 143 | }) |
| 144 | |
| 145 | first, err := BuildRuntime(context.Background(), Options{}) |
| 146 | if err != nil { |
| 147 | t.Fatalf("first BuildRuntime: %v", err) |
| 148 | } |
| 149 | firstPrompt := first.Snapshot.SystemPrompt() |
| 150 | firstHash := first.Snapshot.CacheHash() |
| 151 | firstSystemHash, firstToolsHash := first.Snapshot.CacheShape() |
| 152 | firstSchemas, err := json.Marshal(first.Snapshot.ToolSchemas()) |
| 153 | if err != nil { |
| 154 | first.Controller.Close() |
| 155 | t.Fatalf("marshal first tool schemas: %v", err) |
| 156 | } |
| 157 | firstSessionPrompt := systemMessage(first.Controller.History()) |
| 158 | first.Controller.Close() |
| 159 | |
| 160 | second, err := BuildRuntime(context.Background(), Options{}) |
| 161 | if err != nil { |
| 162 | t.Fatalf("second BuildRuntime: %v", err) |
| 163 | } |
| 164 | t.Cleanup(second.Controller.Close) |
| 165 | secondSchemas, err := json.Marshal(second.Snapshot.ToolSchemas()) |
| 166 | if err != nil { |
| 167 | t.Fatalf("marshal second tool schemas: %v", err) |
| 168 | } |
| 169 | secondSystemHash, secondToolsHash := second.Snapshot.CacheShape() |
| 170 | |
| 171 | if second.Snapshot.Generation() == first.Snapshot.Generation() { |
| 172 | t.Fatal("independent builds reused the same runtime generation") |
| 173 | } |
| 174 | if got := second.Snapshot.SystemPrompt(); got != firstPrompt { |
| 175 | t.Fatalf("stable extension prompt drifted: %q vs %q", got, firstPrompt) |
| 176 | } |
| 177 | if got := systemMessage(second.Controller.History()); got != firstSessionPrompt { |
| 178 | t.Fatalf("controller prompt drifted across stable reload: %q vs %q", got, firstSessionPrompt) |
| 179 | } |
| 180 | if got := second.Snapshot.CacheHash(); got != firstHash { |
| 181 | t.Fatalf("stable extension cache hash drifted: %s vs %s", got, firstHash) |
| 182 | } |
| 183 | if secondSystemHash != firstSystemHash || secondToolsHash != firstToolsHash { |
| 184 | t.Fatalf("stable extension cache shape drifted: system %s/%s tools %s/%s", |
| 185 | firstSystemHash, secondSystemHash, firstToolsHash, secondToolsHash) |
| 186 | } |
| 187 | if string(secondSchemas) != string(firstSchemas) { |
| 188 | t.Fatal("stable extension tool schema bytes drifted across reload") |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | func TestBootSystemPromptStrategyFailureFailsBuild(t *testing.T) { |
| 193 | t.Run("block", func(t *testing.T) { |
| 194 | isolateConfigHome(t) |
| 195 | dir := robustTempDir(t) |
| 196 | t.Chdir(dir) |
| 197 | writeRuntimeFixture(t, dir) |
| 198 | installBootFakePlugin(t, config.ReasonixHomeDir(), "blocker", map[string]any{ |
| 199 | "replaces": []string{"system_prompt"}, |
| 200 | "env": map[string]string{bootFakeEnvBlockEvent: "system_prompt.build"}, |
| 201 | }) |
| 202 | _, err := BuildRuntime(context.Background(), Options{}) |
| 203 | if err == nil { |
| 204 | t.Fatal("BuildRuntime succeeded with a blocking system_prompt owner") |
| 205 | } |
| 206 | var blockErr *dispatch.BlockError |
| 207 | if !errors.As(err, &blockErr) { |
| 208 | t.Fatalf("error %v is not a dispatch.BlockError", err) |
| 209 | } |
| 210 | }) |
| 211 | t.Run("contract violation", func(t *testing.T) { |
| 212 | isolateConfigHome(t) |
| 213 | dir := robustTempDir(t) |
| 214 | t.Chdir(dir) |
| 215 | writeRuntimeFixture(t, dir) |
| 216 | installBootFakePlugin(t, config.ReasonixHomeDir(), "violator", map[string]any{ |
| 217 | "replaces": []string{"system_prompt"}, |
| 218 | "env": map[string]string{bootFakeEnvInvalidEvent: "system_prompt.build"}, |
| 219 | }) |
| 220 | _, err := BuildRuntime(context.Background(), Options{}) |
| 221 | if err == nil { |
| 222 | t.Fatal("BuildRuntime succeeded with a DTO-violating system_prompt owner") |
| 223 | } |
| 224 | var violationErr *dispatch.ViolationError |
| 225 | if !errors.As(err, &violationErr) { |
| 226 | t.Fatalf("error %v is not a dispatch.ViolationError", err) |
| 227 | } |
| 228 | }) |
| 229 | } |
| 230 | |
| 231 | func TestBootSystemPromptBuildEventObserved(t *testing.T) { |
| 232 | isolateConfigHome(t) |
| 233 | dir := robustTempDir(t) |
| 234 | t.Chdir(dir) |
| 235 | writeRuntimeFixture(t, dir) |
| 236 | home := config.ReasonixHomeDir() |
| 237 | eventLog := filepath.Join(dir, "events.log") |
| 238 | |
| 239 | // The slot owner replaces the prompt; the observer subscribes to |
| 240 | // system_prompt.build and records what the broadcast carries. |
| 241 | installBootFakePlugin(t, home, "prompt-owner", map[string]any{ |
| 242 | "replaces": []string{"system_prompt"}, |
| 243 | "env": map[string]string{bootFakeEnvReplacePrompt: "EXTENSION PROMPT"}, |
| 244 | }) |
| 245 | installBootFakePlugin(t, home, "prompt-watcher", map[string]any{ |
| 246 | "intercepts": []string{"system_prompt.build"}, |
| 247 | "env": map[string]string{bootFakeEnvEventLog: eventLog}, |
| 248 | }) |
| 249 | res, err := BuildRuntime(context.Background(), Options{}) |
| 250 | if err != nil { |
| 251 | t.Fatalf("BuildRuntime: %v", err) |
| 252 | } |
| 253 | t.Cleanup(res.Controller.Close) |
| 254 | |
| 255 | // The notification is delivered when the host writes the frame; the fake |
| 256 | // logs it on its own read loop, so poll instead of reading once. |
| 257 | var line string |
| 258 | waitForCond(t, "system_prompt.build event observation", 10*time.Second, func() bool { |
| 259 | data, err := os.ReadFile(eventLog) |
| 260 | if err != nil { |
| 261 | return false |
| 262 | } |
| 263 | for _, candidate := range strings.Split(strings.TrimSpace(string(data)), "\n") { |
| 264 | if strings.HasPrefix(candidate, "system_prompt.build ") { |
| 265 | line = candidate |
| 266 | return true |
| 267 | } |
| 268 | } |
| 269 | return false |
| 270 | }) |
| 271 | if !strings.Contains(line, `"prompt":"EXTENSION PROMPT"`) { |
| 272 | t.Fatalf("observer saw %q, want the final replaced prompt", line) |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | // TestBootSnapshotImmutableAcrossDispatch is the cache-safety invariant: no |
| 277 | // dispatch call — intercept replace, strategy, or event — ever flows back |
| 278 | // into the frozen snapshot. A replacement at input.receive shapes only that |
| 279 | // call's payload; SystemPrompt, ToolSchemas, and CacheHash stay byte-identical. |
| 280 | func TestBootSnapshotImmutableAcrossDispatch(t *testing.T) { |
| 281 | res := bootWithFakePlugin(t, "input-rewriter", map[string]any{ |
| 282 | "intercepts": []string{"input.receive"}, |
| 283 | "env": map[string]string{bootFakeEnvReplaceInput: "REPLACED INPUT"}, |
| 284 | }) |
| 285 | if res.Dispatcher == nil { |
| 286 | t.Fatal("BuildRuntime returned no dispatcher") |
| 287 | } |
| 288 | snap := res.Snapshot |
| 289 | promptBefore := snap.SystemPrompt() |
| 290 | schemasBefore, err := json.Marshal(snap.ToolSchemas()) |
| 291 | if err != nil { |
| 292 | t.Fatalf("marshal tool schemas: %v", err) |
| 293 | } |
| 294 | hashBefore := snap.CacheHash() |
| 295 | |
| 296 | // A replace at input.receive affects only that call's payload. |
| 297 | payload := dispatch.InputPayload{Text: "original user text"} |
| 298 | result, err := res.Dispatcher.Intercept(context.Background(), extension.PointInputReceive, &payload) |
| 299 | if err != nil { |
| 300 | t.Fatalf("Intercept: %v", err) |
| 301 | } |
| 302 | if result.Blocked || len(result.Applied) != 1 || result.Applied[0] != "input-rewriter" { |
| 303 | t.Fatalf("intercept result = %+v, want one applied replacement", result) |
| 304 | } |
| 305 | if payload.Text != "REPLACED INPUT" { |
| 306 | t.Fatalf("payload text = %q, want the extension replacement", payload.Text) |
| 307 | } |
| 308 | |
| 309 | // Strategy + event traffic flows too (session_policy is unowned here, so |
| 310 | // the strategy is a no-op; events are fire-and-forget). |
| 311 | if err := res.Dispatcher.RunStrategy(context.Background(), extension.SlotSessionPolicy, extension.PointSessionSave, |
| 312 | &dispatch.SessionPayload{SessionPath: "/tmp/s.jsonl", Phase: dispatch.PhaseSave}); err != nil { |
| 313 | t.Fatalf("RunStrategy on an unowned slot: %v", err) |
| 314 | } |
| 315 | res.Dispatcher.Event(extension.PointSessionStart, dispatch.SessionPayload{Phase: dispatch.PhaseStart}) |
| 316 | res.Dispatcher.Event(extension.PointFrontendEvent, dispatch.FrontendEventPayload{Kind: "notice", Text: "hi"}) |
| 317 | |
| 318 | if got := snap.SystemPrompt(); got != promptBefore { |
| 319 | t.Fatal("dispatch mutated the snapshot system prompt") |
| 320 | } |
| 321 | schemasAfter, err := json.Marshal(snap.ToolSchemas()) |
| 322 | if err != nil { |
| 323 | t.Fatalf("marshal tool schemas after: %v", err) |
| 324 | } |
| 325 | if string(schemasAfter) != string(schemasBefore) { |
| 326 | t.Fatal("dispatch mutated the snapshot tool schemas") |
| 327 | } |
| 328 | if got := snap.CacheHash(); got != hashBefore { |
| 329 | t.Fatalf("dispatch mutated the snapshot cache hash: %s vs %s", got, hashBefore) |
| 330 | } |
| 331 | } |
| 332 |