feat: auto-continue plan execution after step completion
The plan nudge was background-only and only fired when no progress was made. Now it fires for all plan executions (foreground included) and whenever unchecked steps remain, with a context-aware message depending on whether progress was recorded. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
dj-oyu
parent
5ae2ece515
commit
627b0b1c32
2 changed files with 256 additions and 14 deletions
|
|
@ -1352,7 +1352,7 @@ func (al *AgentLoop) runLLMIteration(
|
|||
|
||||
// Snapshot unchecked step count before tool loop so we can detect progress.
|
||||
preUnchecked := -1 // -1 = not tracking
|
||||
if opts.Background && agent.ContextBuilder.GetPlanStatus() == "executing" {
|
||||
if agent.ContextBuilder.GetPlanStatus() == "executing" {
|
||||
preUnchecked = strings.Count(agent.ContextBuilder.ReadMemory(), "- [ ]")
|
||||
}
|
||||
|
||||
|
|
@ -1506,31 +1506,36 @@ func (al *AgentLoop) runLLMIteration(
|
|||
|
||||
// Check if no tool calls - we're done
|
||||
if len(response.ToolCalls) == 0 {
|
||||
// Background plan continuation: if unchecked steps remain and
|
||||
// none were marked during this heartbeat, nudge the LLM to
|
||||
// Plan continuation: if unchecked steps remain, nudge the LLM to
|
||||
// either mark completed steps or continue working on them.
|
||||
// This serves as both a marking reminder and a continuation trigger
|
||||
// (otherwise remaining steps wait until the next heartbeat).
|
||||
// This fires for both foreground and background plan execution,
|
||||
// ensuring the loop doesn't exit prematurely after marking a step.
|
||||
curUnchecked := 0
|
||||
if preUnchecked > 0 {
|
||||
curUnchecked = strings.Count(agent.ContextBuilder.ReadMemory(), "- [ ]")
|
||||
}
|
||||
if preUnchecked > 0 && !planMarkNudged &&
|
||||
agent.ContextBuilder.GetPlanStatus() == "executing" &&
|
||||
curUnchecked == preUnchecked {
|
||||
if curUnchecked > 0 && !planMarkNudged &&
|
||||
agent.ContextBuilder.GetPlanStatus() == "executing" {
|
||||
planMarkNudged = true
|
||||
messages = append(messages, providers.Message{
|
||||
Role: "assistant",
|
||||
Content: response.Content,
|
||||
})
|
||||
messages = append(messages, providers.Message{
|
||||
Role: "user",
|
||||
Content: fmt.Sprintf("[System] %d unchecked steps remain in MEMORY.md and none were marked [x] during this session. "+
|
||||
var nudgeMsg string
|
||||
if curUnchecked == preUnchecked {
|
||||
nudgeMsg = fmt.Sprintf("[System] %d unchecked steps remain in MEMORY.md and "+
|
||||
"none were marked [x] during this session. "+
|
||||
"If you completed any steps, use edit_file to mark them [x] now. "+
|
||||
"If steps are still in progress, continue working on them.",
|
||||
curUnchecked),
|
||||
"If steps are still in progress, continue working on them.", curUnchecked)
|
||||
} else {
|
||||
nudgeMsg = fmt.Sprintf("[System] Progress recorded. %d unchecked steps remain. "+
|
||||
"Continue working on the next step.", curUnchecked)
|
||||
}
|
||||
messages = append(messages, providers.Message{
|
||||
Role: "user",
|
||||
Content: nudgeMsg,
|
||||
})
|
||||
logger.InfoCF("agent", "Nudging background task: mark or continue plan steps",
|
||||
logger.InfoCF("agent", "Nudging plan execution: continue plan steps",
|
||||
map[string]any{"agent_id": agent.ID, "iteration": iteration, "unchecked": curUnchecked})
|
||||
continue
|
||||
}
|
||||
|
|
|
|||
|
|
@ -1852,3 +1852,240 @@ func TestSanitizeHistoryForProvider_MultiToolCall(t *testing.T) {
|
|||
t.Errorf("expected 2 tool results, got %d", toolCount)
|
||||
}
|
||||
}
|
||||
|
||||
// ---------- plan nudge tests ----------
|
||||
|
||||
// countingMockProvider counts Chat calls and always returns text-only responses.
|
||||
type countingMockProvider struct {
|
||||
callCount int
|
||||
}
|
||||
|
||||
func (m *countingMockProvider) Chat(
|
||||
ctx context.Context,
|
||||
messages []providers.Message,
|
||||
tools []providers.ToolDefinition,
|
||||
model string,
|
||||
opts map[string]any,
|
||||
) (*providers.LLMResponse, error) {
|
||||
m.callCount++
|
||||
return &providers.LLMResponse{
|
||||
Content: fmt.Sprintf("Response %d", m.callCount),
|
||||
ToolCalls: []providers.ToolCall{},
|
||||
}, nil
|
||||
}
|
||||
|
||||
func (m *countingMockProvider) GetDefaultModel() string {
|
||||
return "mock-counting-model"
|
||||
}
|
||||
|
||||
func TestPlanNudge_ForegroundExecution(t *testing.T) {
|
||||
tmpDir, err := os.MkdirTemp("", "agent-nudge-test-*")
|
||||
if err != nil {
|
||||
t.Fatalf("Failed to create temp dir: %v", err)
|
||||
}
|
||||
defer os.RemoveAll(tmpDir)
|
||||
|
||||
cfg := &config.Config{
|
||||
Agents: config.AgentsConfig{
|
||||
Defaults: config.AgentDefaults{
|
||||
Workspace: tmpDir,
|
||||
Model: "test-model",
|
||||
MaxTokens: 4096,
|
||||
MaxToolIterations: 10,
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
provider := &countingMockProvider{}
|
||||
msgBus := bus.NewMessageBus()
|
||||
al := NewAgentLoop(cfg, msgBus, provider)
|
||||
|
||||
agent := al.registry.GetDefaultAgent()
|
||||
if agent == nil {
|
||||
t.Fatal("no default agent")
|
||||
}
|
||||
|
||||
// Write a plan in executing status with unchecked steps
|
||||
plan := "# Active Plan\n\n> Task: Test\n> Status: executing\n> Phase: 1\n\n## Phase 1: Setup\n- [ ] Step one\n- [ ] Step two\n\n## Context\n"
|
||||
agent.ContextBuilder.WriteMemory(plan)
|
||||
|
||||
// Process a foreground message (no background metadata)
|
||||
ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
|
||||
defer cancel()
|
||||
|
||||
msg := bus.InboundMessage{
|
||||
Channel: "test",
|
||||
SenderID: "user1",
|
||||
ChatID: "chat1",
|
||||
Content: "continue working",
|
||||
SessionKey: "nudge-test",
|
||||
}
|
||||
_, err = al.processMessage(ctx, msg)
|
||||
if err != nil {
|
||||
t.Fatalf("processMessage failed: %v", err)
|
||||
}
|
||||
|
||||
// The provider should have been called at least 2 times:
|
||||
// 1st call: returns text-only → nudge fires (unchecked steps remain)
|
||||
// 2nd call: returns text-only → nudge already fired, loop exits
|
||||
if provider.callCount < 2 {
|
||||
t.Errorf("expected at least 2 provider calls (nudge should trigger continuation), got %d", provider.callCount)
|
||||
}
|
||||
}
|
||||
|
||||
func TestPlanNudge_NoNudgeWhenAllStepsComplete(t *testing.T) {
|
||||
tmpDir, err := os.MkdirTemp("", "agent-nudge-test-*")
|
||||
if err != nil {
|
||||
t.Fatalf("Failed to create temp dir: %v", err)
|
||||
}
|
||||
defer os.RemoveAll(tmpDir)
|
||||
|
||||
cfg := &config.Config{
|
||||
Agents: config.AgentsConfig{
|
||||
Defaults: config.AgentDefaults{
|
||||
Workspace: tmpDir,
|
||||
Model: "test-model",
|
||||
MaxTokens: 4096,
|
||||
MaxToolIterations: 10,
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
provider := &countingMockProvider{}
|
||||
msgBus := bus.NewMessageBus()
|
||||
al := NewAgentLoop(cfg, msgBus, provider)
|
||||
|
||||
agent := al.registry.GetDefaultAgent()
|
||||
if agent == nil {
|
||||
t.Fatal("no default agent")
|
||||
}
|
||||
|
||||
// Write a plan where all steps are already checked
|
||||
plan := "# Active Plan\n\n> Task: Test\n> Status: executing\n> Phase: 1\n\n## Phase 1: Setup\n- [x] Step one\n- [x] Step two\n\n## Context\n"
|
||||
agent.ContextBuilder.WriteMemory(plan)
|
||||
|
||||
ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
|
||||
defer cancel()
|
||||
|
||||
msg := bus.InboundMessage{
|
||||
Channel: "test",
|
||||
SenderID: "user1",
|
||||
ChatID: "chat1",
|
||||
Content: "all done",
|
||||
SessionKey: "nudge-test-complete",
|
||||
}
|
||||
_, err = al.processMessage(ctx, msg)
|
||||
if err != nil {
|
||||
t.Fatalf("processMessage failed: %v", err)
|
||||
}
|
||||
|
||||
// No unchecked steps → preUnchecked=0 → no nudge → only 1 provider call
|
||||
if provider.callCount != 1 {
|
||||
t.Errorf("expected exactly 1 provider call (no nudge needed), got %d", provider.callCount)
|
||||
}
|
||||
}
|
||||
|
||||
func TestPlanNudge_ProgressMessage(t *testing.T) {
|
||||
tmpDir, err := os.MkdirTemp("", "agent-nudge-test-*")
|
||||
if err != nil {
|
||||
t.Fatalf("Failed to create temp dir: %v", err)
|
||||
}
|
||||
defer os.RemoveAll(tmpDir)
|
||||
|
||||
cfg := &config.Config{
|
||||
Agents: config.AgentsConfig{
|
||||
Defaults: config.AgentDefaults{
|
||||
Workspace: tmpDir,
|
||||
Model: "test-model",
|
||||
MaxTokens: 4096,
|
||||
MaxToolIterations: 10,
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
// Provider that checks the nudge message content on the 2nd call
|
||||
var nudgeContent string
|
||||
provider := &nudgeCaptureMockProvider{onSecondCall: func(msgs []providers.Message) {
|
||||
// The last user message should be the nudge
|
||||
for i := len(msgs) - 1; i >= 0; i-- {
|
||||
if msgs[i].Role == "user" {
|
||||
nudgeContent = msgs[i].Content
|
||||
break
|
||||
}
|
||||
}
|
||||
}}
|
||||
msgBus := bus.NewMessageBus()
|
||||
al := NewAgentLoop(cfg, msgBus, provider)
|
||||
|
||||
agent := al.registry.GetDefaultAgent()
|
||||
if agent == nil {
|
||||
t.Fatal("no default agent")
|
||||
}
|
||||
|
||||
// Write a plan with 3 unchecked steps; the provider edits memory to
|
||||
// mark one step between calls (simulated by the first-call hook).
|
||||
plan := "# Active Plan\n\n> Task: Test\n> Status: executing\n> Phase: 1\n\n## Phase 1: Setup\n- [ ] Step one\n- [ ] Step two\n- [ ] Step three\n\n## Context\n"
|
||||
agent.ContextBuilder.WriteMemory(plan)
|
||||
|
||||
// After the first LLM response (no tool calls), simulate that
|
||||
// one step was marked [x] externally (as if the AI did it via tool).
|
||||
// We do this by hooking the provider's first call to mutate memory.
|
||||
provider.onFirstCall = func() {
|
||||
updated := strings.Replace(agent.ContextBuilder.ReadMemory(), "- [ ] Step one", "- [x] Step one", 1)
|
||||
agent.ContextBuilder.WriteMemory(updated)
|
||||
}
|
||||
|
||||
ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
|
||||
defer cancel()
|
||||
|
||||
msg := bus.InboundMessage{
|
||||
Channel: "test",
|
||||
SenderID: "user1",
|
||||
ChatID: "chat1",
|
||||
Content: "work on the plan",
|
||||
SessionKey: "nudge-progress-test",
|
||||
}
|
||||
_, err = al.processMessage(ctx, msg)
|
||||
if err != nil {
|
||||
t.Fatalf("processMessage failed: %v", err)
|
||||
}
|
||||
|
||||
// Should have gotten the "Progress recorded" nudge (not the "none were marked" one)
|
||||
if !strings.Contains(nudgeContent, "Progress recorded") {
|
||||
t.Errorf("expected 'Progress recorded' nudge, got %q", nudgeContent)
|
||||
}
|
||||
if !strings.Contains(nudgeContent, "2 unchecked steps remain") {
|
||||
t.Errorf("expected '2 unchecked steps remain' in nudge, got %q", nudgeContent)
|
||||
}
|
||||
}
|
||||
|
||||
// nudgeCaptureMockProvider calls hooks on 1st and 2nd Chat invocations.
|
||||
type nudgeCaptureMockProvider struct {
|
||||
callCount int
|
||||
onFirstCall func()
|
||||
onSecondCall func([]providers.Message)
|
||||
}
|
||||
|
||||
func (m *nudgeCaptureMockProvider) Chat(
|
||||
ctx context.Context,
|
||||
messages []providers.Message,
|
||||
tools []providers.ToolDefinition,
|
||||
model string,
|
||||
opts map[string]any,
|
||||
) (*providers.LLMResponse, error) {
|
||||
m.callCount++
|
||||
if m.callCount == 1 && m.onFirstCall != nil {
|
||||
m.onFirstCall()
|
||||
}
|
||||
if m.callCount == 2 && m.onSecondCall != nil {
|
||||
m.onSecondCall(messages)
|
||||
}
|
||||
return &providers.LLMResponse{
|
||||
Content: fmt.Sprintf("Response %d", m.callCount),
|
||||
ToolCalls: []providers.ToolCall{},
|
||||
}, nil
|
||||
}
|
||||
|
||||
func (m *nudgeCaptureMockProvider) GetDefaultModel() string {
|
||||
return "mock-nudge-model"
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue