| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "go/ast" |
| 5 | "go/parser" |
| 6 | "go/token" |
| 7 | "testing" |
| 8 | "time" |
| 9 | ) |
| 10 | |
| 11 | // sessionReset names the fields a new conversation starts from. taskRuntime |
| 12 | // gets this for free — one assignment zeroes anything unlisted — but a struct |
| 13 | // holding atomics and mutexes cannot be assigned, so reset must name each field |
| 14 | // and this list is what keeps it honest. |
| 15 | var sessionReset = map[string]bool{ |
| 16 | "mu": true, |
| 17 | "conversation": true, |
| 18 | "output": true, |
| 19 | "cacheHit": true, |
| 20 | "cacheMiss": true, |
| 21 | "missingReasoning": true, |
| 22 | // A strong-projection repair belongs to the corrupted conversation; a new |
| 23 | // conversation starts without it. |
| 24 | "reasoningReplayStrongProjection": true, |
| 25 | "reasoningReplayStrongProjectionAnchor": true, |
| 26 | "compactionMu": true, |
| 27 | "compactionState": true, |
| 28 | "cacheState": true, |
| 29 | "checkpointState": true, |
| 30 | "pendingModelContextCommit": true, |
| 31 | "compaction": true, |
| 32 | "todoMu": true, |
| 33 | "todoState": true, |
| 34 | "todoWritten": true, |
| 35 | } |
| 36 | |
| 37 | // sessionCarryOver names the fields reset deliberately leaves alone, each with |
| 38 | // an owner that rebinds it. Being on this list is a claim that someone else |
| 39 | // sets the field for the new conversation — not that it does not matter. |
| 40 | var sessionCarryOver = map[string]bool{ |
| 41 | "compactionRunMu": true, // a singleflight latch, not conversation state |
| 42 | "path": true, // preflight rebinds on the next transcript bind |
| 43 | // lastPrefixShape survives the swap today; the next request compares its |
| 44 | // prefix against the replaced conversation's shape. Left as found here. |
| 45 | "lastPrefixShape": true, |
| 46 | "haveLastPrefixShape": true, |
| 47 | } |
| 48 | |
| 49 | func sessionRuntimeFields(t *testing.T) map[string]bool { |
| 50 | t.Helper() |
| 51 | fset := token.NewFileSet() |
| 52 | file, err := parser.ParseFile(fset, "sessionstate.go", nil, 0) |
| 53 | if err != nil { |
| 54 | t.Fatalf("parse sessionstate.go: %v", err) |
| 55 | } |
| 56 | fields := map[string]bool{} |
| 57 | ast.Inspect(file, func(n ast.Node) bool { |
| 58 | spec, ok := n.(*ast.TypeSpec) |
| 59 | if !ok || spec.Name.Name != "sessionRuntime" { |
| 60 | return true |
| 61 | } |
| 62 | st, ok := spec.Type.(*ast.StructType) |
| 63 | if !ok { |
| 64 | return false |
| 65 | } |
| 66 | for _, field := range st.Fields.List { |
| 67 | if len(field.Names) == 0 { |
| 68 | // An embedded type contributes its own name. |
| 69 | if ident, ok := field.Type.(*ast.Ident); ok { |
| 70 | fields[ident.Name] = true |
| 71 | } |
| 72 | continue |
| 73 | } |
| 74 | for _, name := range field.Names { |
| 75 | fields[name.Name] = true |
| 76 | } |
| 77 | } |
| 78 | return false |
| 79 | }) |
| 80 | if len(fields) == 0 { |
| 81 | t.Fatal("sessionRuntime has no fields; the guard would pass vacuously") |
| 82 | } |
| 83 | return fields |
| 84 | } |
| 85 | |
| 86 | func TestSessionRuntimeLifetimeListsCoverTheStruct(t *testing.T) { |
| 87 | fields := sessionRuntimeFields(t) |
| 88 | for _, list := range []map[string]bool{sessionReset, sessionCarryOver} { |
| 89 | for name := range list { |
| 90 | if !fields[name] { |
| 91 | t.Errorf("the lifetime lists name %q, which sessionRuntime no longer has", name) |
| 92 | } |
| 93 | } |
| 94 | } |
| 95 | for name := range fields { |
| 96 | switch { |
| 97 | case sessionReset[name] && sessionCarryOver[name]: |
| 98 | t.Errorf("sessionRuntime.%s is listed as both reset and carried", name) |
| 99 | case !sessionReset[name] && !sessionCarryOver[name]: |
| 100 | t.Errorf("sessionRuntime.%s is on neither list; decide whether a new conversation starts from it", name) |
| 101 | } |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | // The list above is a claim about reset's body, so it is read back from the |
| 106 | // source: a field dropped from reset stops being covered, and the mismatch |
| 107 | // fails here instead of surfacing as state leaking between conversations. |
| 108 | func TestSessionRuntimeResetAssignsEveryResetField(t *testing.T) { |
| 109 | fset := token.NewFileSet() |
| 110 | file, err := parser.ParseFile(fset, "sessionstate.go", nil, 0) |
| 111 | if err != nil { |
| 112 | t.Fatalf("parse sessionstate.go: %v", err) |
| 113 | } |
| 114 | touched := map[string]bool{} |
| 115 | ast.Inspect(file, func(n ast.Node) bool { |
| 116 | fn, ok := n.(*ast.FuncDecl) |
| 117 | if !ok || fn.Name.Name != "reset" { |
| 118 | return true |
| 119 | } |
| 120 | ast.Inspect(fn, func(inner ast.Node) bool { |
| 121 | sel, ok := inner.(*ast.SelectorExpr) |
| 122 | if !ok { |
| 123 | return true |
| 124 | } |
| 125 | if recv, ok := sel.X.(*ast.Ident); ok && recv.Name == "r" { |
| 126 | touched[sel.Sel.Name] = true |
| 127 | } |
| 128 | return true |
| 129 | }) |
| 130 | return false |
| 131 | }) |
| 132 | for name := range sessionReset { |
| 133 | if !touched[name] { |
| 134 | t.Errorf("reset never touches sessionRuntime.%s, but the list says a new conversation starts from it", name) |
| 135 | } |
| 136 | } |
| 137 | } |
| 138 | |
| 139 | func TestSetSessionRestartsTheConversationState(t *testing.T) { |
| 140 | a := &Agent{} |
| 141 | a.sess.cacheHit.Store(11) |
| 142 | a.sess.cacheMiss.Store(7) |
| 143 | a.sess.missingReasoning = missingReasoningWatch{active: true, stateRecorded: true, healthyStreak: 2} |
| 144 | a.sess.reasoningReplayStrongProjection = 7 |
| 145 | a.sess.reasoningReplayStrongProjectionAnchor = "anchor" |
| 146 | a.sess.compaction.stuck = true |
| 147 | a.sess.compaction.stuckInputHash = "old-input" |
| 148 | a.sess.compaction.consecutive = 3 |
| 149 | a.sess.compaction.failedTurn.Store(8) |
| 150 | a.sess.compaction.lastTurn.Store(9) |
| 151 | a.sess.compactionState = CompactionState{} |
| 152 | a.sess.checkpointState = "pending" |
| 153 | a.sess.pendingModelContextCommit = &SessionModelContextCommit{OperationID: "old-operation"} |
| 154 | a.unwrittenResolve.at = time.Unix(1, 0) |
| 155 | |
| 156 | next := NewSession("") |
| 157 | a.SetSession(next) |
| 158 | |
| 159 | if a.sess.session() != next { |
| 160 | t.Error("SetSession did not bind the new conversation") |
| 161 | } |
| 162 | if a.sess.cacheHit.Load() != 0 || a.sess.cacheMiss.Load() != 0 { |
| 163 | t.Errorf("cache tallies = %d/%d, want a fresh aggregate", a.sess.cacheHit.Load(), a.sess.cacheMiss.Load()) |
| 164 | } |
| 165 | if a.sess.missingReasoning != (missingReasoningWatch{}) { |
| 166 | t.Errorf("missingReasoning = %+v, want the incident to end with its conversation", a.sess.missingReasoning) |
| 167 | } |
| 168 | if a.sess.reasoningReplayStrongProjection != 0 { |
| 169 | t.Errorf("reasoningReplayStrongProjection = %d, want it restarted", a.sess.reasoningReplayStrongProjection) |
| 170 | } |
| 171 | if a.sess.reasoningReplayStrongProjectionAnchor != "" { |
| 172 | t.Errorf("reasoningReplayStrongProjectionAnchor = %q, want it restarted", a.sess.reasoningReplayStrongProjectionAnchor) |
| 173 | } |
| 174 | if a.sess.compaction.stuck || a.sess.compaction.stuckInputHash != "" || a.sess.compaction.consecutive != 0 || |
| 175 | a.sess.compaction.failedTurn.Load() != 0 || a.sess.compaction.lastTurn.Load() != 0 { |
| 176 | t.Errorf("compaction progress = stuck:%t hash:%q consecutive:%d failedTurn:%d lastTurn:%d, want it restarted", |
| 177 | a.sess.compaction.stuck, a.sess.compaction.stuckInputHash, a.sess.compaction.consecutive, |
| 178 | a.sess.compaction.failedTurn.Load(), a.sess.compaction.lastTurn.Load()) |
| 179 | } |
| 180 | if a.sess.cacheState != CacheStateUnknown { |
| 181 | t.Errorf("cacheState = %q, want %q", a.sess.cacheState, CacheStateUnknown) |
| 182 | } |
| 183 | if a.sess.checkpointState != "none" || a.sess.pendingModelContextCommit != nil { |
| 184 | t.Errorf("pending model context leaked across session reset: state=%q pending=%+v", a.sess.checkpointState, a.sess.pendingModelContextCommit) |
| 185 | } |
| 186 | if a.unwrittenResolve.at.IsZero() { |
| 187 | t.Error("unwrittenResolve was cleared; the retry it owes belongs to the provider configuration, not the conversation") |
| 188 | } |
| 189 | } |
| 190 |