| 1 | package control |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "fmt" |
| 7 | "sync/atomic" |
| 8 | "testing" |
| 9 | |
| 10 | "reasonix/internal/agent" |
| 11 | "reasonix/internal/event" |
| 12 | "reasonix/internal/provider" |
| 13 | ) |
| 14 | |
| 15 | // workingGoalTool is a writer, so every round counts as acting on the |
| 16 | // investigation rather than only looking at it. |
| 17 | type workingGoalTool struct{ name string } |
| 18 | |
| 19 | func (t workingGoalTool) Name() string { return t.name } |
| 20 | func (workingGoalTool) Description() string { return "apply an edit" } |
| 21 | func (workingGoalTool) ReadOnly() bool { return false } |
| 22 | func (workingGoalTool) Schema() json.RawMessage { |
| 23 | return json.RawMessage(`{"type":"object","properties":{"path":{"type":"string"}}}`) |
| 24 | } |
| 25 | func (workingGoalTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 26 | return "applied", nil |
| 27 | } |
| 28 | |
| 29 | // steadyWorkProvider makes real progress every round: a distinct mutation, so |
| 30 | // neither the zero-evidence ladder nor the exploration decay ever escalates. |
| 31 | type steadyWorkProvider struct { |
| 32 | calls atomic.Int32 |
| 33 | max int32 |
| 34 | } |
| 35 | |
| 36 | func (p *steadyWorkProvider) Name() string { return "steady-work" } |
| 37 | |
| 38 | func (p *steadyWorkProvider) Stream(context.Context, provider.Request) (<-chan provider.Chunk, error) { |
| 39 | round := p.calls.Add(1) |
| 40 | ch := make(chan provider.Chunk, 4) |
| 41 | if round > p.max { |
| 42 | if round == p.max+1 { |
| 43 | ch <- provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "finished", Name: "update_goal", Arguments: `{"status":"complete","reason":"All edits applied"}`}} |
| 44 | } |
| 45 | ch <- provider.Chunk{Type: provider.ChunkText, Text: "All done."} |
| 46 | ch <- provider.Chunk{Type: provider.ChunkDone} |
| 47 | close(ch) |
| 48 | return ch, nil |
| 49 | } |
| 50 | ch <- provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ |
| 51 | ID: fmt.Sprintf("edit-%d", round), |
| 52 | Name: "apply_edit", |
| 53 | Arguments: fmt.Sprintf(`{"path":"pkg%d/file.go"}`, round), |
| 54 | }} |
| 55 | ch <- provider.Chunk{Type: provider.ChunkDone} |
| 56 | close(ch) |
| 57 | return ch, nil |
| 58 | } |
| 59 | |
| 60 | // A Goal turn used to be cut at 16 model rounds. Rounds were the wrong unit |
| 61 | // there for the same reason they were wrong for chat: a turn that reaches a |
| 62 | // high count while still producing new work is the case least worth |
| 63 | // interrupting. Goal now ends on its own terms — completion, a block, or a |
| 64 | // structural no-progress loop — none of which is a round count. |
| 65 | func TestGoalTurnRunsPastTheOldRoundCeiling(t *testing.T) { |
| 66 | prov := &steadyWorkProvider{max: 24} |
| 67 | reg := goalRegistry() |
| 68 | reg.Add(workingGoalTool{name: "apply_edit"}) |
| 69 | exec := agent.New(prov, reg, agent.NewSession("sys"), agent.Options{}, event.Discard) |
| 70 | c, done := newChatBudgetController(t, exec) |
| 71 | |
| 72 | c.SetGoal("apply every pending edit") |
| 73 | c.Submit("start") |
| 74 | <-done |
| 75 | |
| 76 | if got := prov.calls.Load(); got <= 16 { |
| 77 | t.Fatalf("provider rounds = %d, want productive work to run past the retired 16-round ceiling", got) |
| 78 | } |
| 79 | } |
| 80 |