Enhance P3 Execution with Multi-Turn Conversation and Validation Improvements
- Updated the `RunConfig` to include parameters for multi-turn conversation control, such as `ContinueOnFailure`, `ValidationThreshold`, and `MaxTurnsPerTask`. - Implemented a new multi-turn conversation flow for assistant tasks, allowing for iterative interactions until completion or maximum turns are reached. - Enhanced the `ValidationResult` structure to support multi-turn states, including fields for `Complete`, `NeedReply`, and `ReplyContent`. - Refined the `ExecuteWithRetry` method to accommodate the new conversation flow, ensuring proper handling of task execution and validation. - Revised the `Validator` to include logic for determining when to continue conversations based on validation results. - Updated documentation and tests to reflect the new multi-turn capabilities and validation mechanisms, ensuring comprehensive coverage of the changes.
This commit is contained in:
Max
parent
0c9bdb8000
commit
51e4c6d208
9 changed files with 556 additions and 270 deletions
|
|
@ -372,7 +372,7 @@ type Task struct {
|
|||
```
|
||||
┌─────────────────────────────────────────────────────────────┐
|
||||
│ run.go (P3 Entry) │
|
||||
│ - RunConfig: retries, threshold, continue-on-failure │
|
||||
│ - RunConfig: threshold, continue-on-failure, max-turns │
|
||||
│ - RunExecution: main execution loop │
|
||||
└─────────────────────┬───────────────────────────────────────┘
|
||||
│
|
||||
|
|
@ -383,6 +383,7 @@ type Task struct {
|
|||
│ - Runner │ │ - Validator │
|
||||
│ - Task exec │ │ - Two-layer │
|
||||
│ - Multi-turn │ │ - Rule+Semantic│
|
||||
│ conversation │ │ - NeedReply │
|
||||
└────────┬────────┘ └────────┬────────┘
|
||||
│ │
|
||||
│ ▼
|
||||
|
|
@ -395,8 +396,8 @@ type Task struct {
|
|||
┌─────────────────────────────────────────┐
|
||||
│ Executor Types │
|
||||
│ - assistant: AI Agent (multi-turn) │
|
||||
│ - mcp: MCP Tool (clientID.toolName) │
|
||||
│ - process: Yao Process │
|
||||
│ - mcp: MCP Tool (single-call) │
|
||||
│ - process: Yao Process (single-call) │
|
||||
└─────────────────────────────────────────┘
|
||||
```
|
||||
|
||||
|
|
@ -404,10 +405,15 @@ type Task struct {
|
|||
|
||||
For each task:
|
||||
|
||||
1. **Execute** via appropriate executor (Assistant/MCP/Process)
|
||||
2. **Validate** using two-layer validation
|
||||
3. **Retry** if validation fails (with feedback to expert agent)
|
||||
4. **Update** task status and store result
|
||||
1. **Build Context**: Include previous task results as context
|
||||
2. **Execute**: Call appropriate executor (Assistant/MCP/Process)
|
||||
3. **Validate**: Use two-layer validation (rule-based + semantic)
|
||||
4. **Continue or Complete**:
|
||||
- For Assistant tasks: If `NeedReply`, continue conversation with `ReplyContent`
|
||||
- For MCP/Process tasks: Single-call execution, no multi-turn
|
||||
5. **Update**: Set task status and store result
|
||||
|
||||
**Task Dependency**: Previous task results are automatically passed as context to subsequent tasks via `Runner.BuildTaskContext()` and formatted using `FormatPreviousResultsAsContext()`.
|
||||
|
||||
**Two-Layer Validation:**
|
||||
|
||||
|
|
@ -424,27 +430,42 @@ For each task:
|
|||
| `mcp` | `clientID.toolName` | `filesystem.read_file` |
|
||||
| `process` | Process name | `models.user.Find` |
|
||||
|
||||
**Retry Mechanism:**
|
||||
**Multi-Turn Conversation Flow:**
|
||||
|
||||
- Retries only on validation failure (not execution error)
|
||||
- Validation feedback sent to expert agent on retry
|
||||
- Configurable: `MaxRetries`, `RetryOnValidationFailure`
|
||||
For assistant tasks, P3 uses a multi-turn conversation approach:
|
||||
1. **Call**: Call assistant and get result
|
||||
2. **Validate**: Validate result (determines: passed, complete, needReply, replyContent)
|
||||
3. **Reply**: If needReply, continue conversation with replyContent
|
||||
4. **Repeat**: Until complete or max turns exceeded
|
||||
|
||||
The `Validator.ValidateWithContext()` method determines:
|
||||
- `Complete`: Whether the expected result is obtained
|
||||
- `NeedReply`: Whether to continue conversation
|
||||
- `ReplyContent`: What to send in the next turn (validation feedback, clarification request, etc.)
|
||||
|
||||
This replaces the traditional retry mechanism with intelligent conversation continuation.
|
||||
|
||||
```go
|
||||
// RunConfig configures P3 execution behavior
|
||||
type RunConfig struct {
|
||||
MaxRetries int // default: 3
|
||||
RetryOnValidationFailure bool // default: true
|
||||
ContinueOnFailure bool // default: false
|
||||
ValidationThreshold float64 // default: 0.6
|
||||
MaxTurnsPerTask int // default: 10
|
||||
ContinueOnFailure bool // continue to next task even if current fails (default: false)
|
||||
ValidationThreshold float64 // minimum score to pass validation (default: 0.6)
|
||||
MaxTurnsPerTask int // max conversation turns per task (default: 10)
|
||||
}
|
||||
|
||||
// ValidationResult with multi-turn conversation support
|
||||
type ValidationResult struct {
|
||||
// Basic validation result
|
||||
Passed bool // overall validation passed
|
||||
Score float64 // 0-1 confidence score
|
||||
Issues []string // what failed
|
||||
Suggestions []string // how to improve
|
||||
Details string // detailed report (markdown)
|
||||
|
||||
// Execution state (for multi-turn conversation control)
|
||||
Complete bool // whether expected result is obtained
|
||||
NeedReply bool // whether to continue conversation
|
||||
ReplyContent string // content for next turn (if NeedReply)
|
||||
}
|
||||
```
|
||||
|
||||
|
|
|
|||
|
|
@ -1384,13 +1384,19 @@ type TaskResult struct {
|
|||
Validation *ValidationResult `json:"validation,omitempty"` // P3 validation result
|
||||
}
|
||||
|
||||
// ValidationResult - P3 semantic validation result
|
||||
// ValidationResult - P3 validation result with multi-turn conversation support
|
||||
type ValidationResult struct {
|
||||
// Basic validation result
|
||||
Passed bool `json:"passed"` // overall validation passed
|
||||
Score float64 `json:"score,omitempty"` // 0-1 confidence score
|
||||
Issues []string `json:"issues,omitempty"` // what failed
|
||||
Suggestions []string `json:"suggestions,omitempty"` // how to improve
|
||||
Details string `json:"details,omitempty"` // detailed validation report
|
||||
Details string `json:"details,omitempty"` // detailed validation report (markdown)
|
||||
|
||||
// Execution state (for multi-turn conversation control)
|
||||
Complete bool `json:"complete"` // whether expected result is obtained
|
||||
NeedReply bool `json:"need_reply,omitempty"` // whether to continue conversation
|
||||
ReplyContent string `json:"reply_content,omitempty"` // content for next turn (if NeedReply)
|
||||
}
|
||||
|
||||
// DeliveryResult - P4 delivery output
|
||||
|
|
|
|||
|
|
@ -777,22 +777,30 @@ Each phase test uses different expert combinations:
|
|||
### 9.1 Implementation ✅
|
||||
|
||||
- [x] `executor/run.go` - `RunExecution(ctx, exec, data)` - real implementation
|
||||
- [x] `RunConfig` - configuration for retries, validation threshold, etc.
|
||||
- [x] `RunConfig` - configuration (ContinueOnFailure, ValidationThreshold, MaxTurnsPerTask)
|
||||
- [x] Sequential task execution with progress tracking
|
||||
- [x] Task status updates (Running → Completed/Failed/Skipped)
|
||||
- [x] `ContinueOnFailure` option for graceful failure handling
|
||||
- [x] Previous task results passed as context to subsequent tasks
|
||||
- [x] `executor/runner.go` - `Runner` struct for task execution
|
||||
- [x] `ExecuteWithRetry()` - retry mechanism for validation failures
|
||||
- [x] `ExecuteTask()` - dispatch to correct executor type
|
||||
- [x] `ExecuteAssistantTask()` - AI assistant execution with multi-turn support
|
||||
- [x] `ExecuteWithRetry()` - multi-turn conversation flow for assistant tasks
|
||||
- [x] `executeNonAssistantTask()` - single-call execution for MCP/Process
|
||||
- [x] `executeAssistantWithMultiTurn()` - AI assistant with conversation support
|
||||
- [x] `ExecuteMCPTask()` - MCP tool execution (format: `clientID.toolName`)
|
||||
- [x] `ExecuteProcessTask()` - Yao process execution
|
||||
- [x] `BuildTaskContext()` - context with previous results
|
||||
- [x] `GenerateAutoReply()` - auto-reply for multi-turn conversations
|
||||
- [x] `FormatValidationFeedback()` - feedback for retry attempts
|
||||
- [x] `BuildAssistantMessages()` - build messages for assistant
|
||||
- [x] `FormatPreviousResultsAsContext()` - format previous results as context
|
||||
- [x] `extractOutput()` - extract output from CallResult
|
||||
- [x] `generateDefaultReply()` - fallback reply generation
|
||||
- [x] `executor/validator.go` - Two-layer validation system
|
||||
- [x] Layer 1: Rule-based validation using `yao/assert`
|
||||
- [x] Layer 2: Semantic validation using Validation Agent
|
||||
- [x] `ValidateWithContext()` - validation with multi-turn support
|
||||
- [x] `isComplete()` - determine if expected result is obtained
|
||||
- [x] `checkNeedReply()` - determine if conversation should continue
|
||||
- [x] `generateFeedbackReply()` - generate validation feedback for next turn
|
||||
- [x] `detectNeedMoreInfo()` - detect if assistant needs clarification
|
||||
- [x] `convertStringRule()` - natural language rules to assertions
|
||||
- [x] `parseRules()` - JSON and string rule parsing
|
||||
- [x] `mergeResults()` - combine rule and semantic results
|
||||
|
|
@ -829,14 +837,16 @@ Created new `yao/assert` package for universal assertion/validation:
|
|||
- [ ] Test: ContinueOnFailure option
|
||||
- [ ] Test: remaining tasks marked as skipped on failure
|
||||
- [ ] `executor/standard/runner_test.go` - Runner tests
|
||||
- [ ] Test: ExecuteWithRetry with validation failures
|
||||
- [ ] Test: ExecuteAssistantTask with multi-turn conversation
|
||||
- [ ] Test: ExecuteWithRetry with multi-turn conversation flow
|
||||
- [ ] Test: executeAssistantWithMultiTurn conversation continuation
|
||||
- [ ] Test: ExecuteMCPTask with correct ID parsing
|
||||
- [ ] Test: ExecuteProcessTask with Yao process
|
||||
- [ ] Test: BuildTaskContext with previous results
|
||||
- [ ] Test: GenerateAutoReply for tool results
|
||||
- [ ] Test: FormatPreviousResultsAsContext formatting
|
||||
- [ ] `executor/standard/validator_test.go` - Validator tests
|
||||
- [ ] Test: two-layer validation (rules + semantic)
|
||||
- [ ] Test: ValidateWithContext with multi-turn state
|
||||
- [ ] Test: isComplete determination logic
|
||||
- [ ] Test: checkNeedReply scenarios (clarification, feedback, incomplete)
|
||||
- [ ] Test: convertStringRule for natural language rules
|
||||
- [ ] Test: parseRules for JSON assertions
|
||||
- [ ] Test: validateSemantic with Validation Agent
|
||||
|
|
@ -846,9 +856,10 @@ Created new `yao/assert` package for universal assertion/validation:
|
|||
|
||||
```
|
||||
┌─────────────────────────────────────────────────────────────┐
|
||||
│ run.go (P3 入口) │
|
||||
│ - RunConfig 配置 │
|
||||
│ - RunExecution 主循环 │
|
||||
│ run.go (P3 Entry) │
|
||||
│ - RunConfig: ContinueOnFailure, ValidationThreshold, │
|
||||
│ MaxTurnsPerTask │
|
||||
│ - RunExecution: main loop with task dependency passing │
|
||||
└─────────────────────┬───────────────────────────────────────┘
|
||||
│
|
||||
┌────────────┴────────────┐
|
||||
|
|
@ -856,34 +867,66 @@ Created new `yao/assert` package for universal assertion/validation:
|
|||
┌─────────────────┐ ┌─────────────────┐
|
||||
│ runner.go │ │ validator.go │
|
||||
│ - Runner │ │ - Validator │
|
||||
│ - 任务执行 │ │ - 两层验证 │
|
||||
│ - 多轮对话 │ │ - 规则 + 语义 │
|
||||
│ - Multi-turn │ │ - Two-layer │
|
||||
│ conversation │ │ - Rule+Semantic│
|
||||
│ - Task context │ │ - NeedReply │
|
||||
│ building │ │ - ReplyContent │
|
||||
└────────┬────────┘ └────────┬────────┘
|
||||
│ │
|
||||
│ ▼
|
||||
│ ┌─────────────────┐
|
||||
│ │ yao/assert │
|
||||
│ │ - Asserter │
|
||||
│ │ - 8种断言类型 │
|
||||
│ │ - 可扩展接口 │
|
||||
│ │ - 8 types │
|
||||
│ │ - Extensible │
|
||||
│ └─────────────────┘
|
||||
▼
|
||||
┌─────────────────────────────────────────┐
|
||||
│ 执行器类型 │
|
||||
│ - ExecutorAssistant → AI 助手 │
|
||||
│ - ExecutorMCP → MCP 工具 │
|
||||
│ - ExecutorProcess → Yao 进程 │
|
||||
│ Executor Types │
|
||||
│ - ExecutorAssistant → Multi-turn AI │
|
||||
│ - ExecutorMCP → Single-call MCP tool │
|
||||
│ - ExecutorProcess → Single-call Process │
|
||||
└─────────────────────────────────────────┘
|
||||
```
|
||||
|
||||
**Multi-Turn Conversation Flow (Assistant Tasks):**
|
||||
|
||||
```
|
||||
┌──────────────────────────────────────────────────────────────┐
|
||||
│ executeAssistantWithMultiTurn │
|
||||
├──────────────────────────────────────────────────────────────┤
|
||||
│ 1. Create Conversation (single instance for entire task) │
|
||||
│ 2. Build initial messages with task context │
|
||||
│ │
|
||||
│ ┌─────────────────── Turn Loop ───────────────────────────┐ │
|
||||
│ │ Phase 1: Call assistant via conv.Turn() │ │
|
||||
│ │ Phase 2: ValidateWithContext() determines: │ │
|
||||
│ │ - Complete: task done? │ │
|
||||
│ │ - NeedReply: continue conversation? │ │
|
||||
│ │ - ReplyContent: what to send next? │ │
|
||||
│ │ Phase 3: If NeedReply, use ReplyContent as next input │ │
|
||||
│ │ Break if: Complete && Passed, or !NeedReply │ │
|
||||
│ └──────────────────────────────────────────────────────────┘ │
|
||||
│ │
|
||||
│ 3. Return output, validation, error │
|
||||
└──────────────────────────────────────────────────────────────┘
|
||||
```
|
||||
|
||||
### 9.5 Notes
|
||||
|
||||
- Validation rules support two formats:
|
||||
1. Natural language: `"output must be valid JSON"`, `"must contain 'field'"`
|
||||
2. Structured JSON: `{"type": "type", "path": "field", "value": "array"}`
|
||||
- Retry mechanism only triggers on validation failures, not execution errors
|
||||
- Multi-turn conversation uses auto-reply generation for tool results
|
||||
- Multi-turn conversation is validator-driven:
|
||||
- `ValidateWithContext()` returns `NeedReply` and `ReplyContent`
|
||||
- Conversation continues until `Complete && Passed` or `!NeedReply`
|
||||
- Max turns controlled by `RunConfig.MaxTurnsPerTask`
|
||||
- Task dependencies handled automatically:
|
||||
- `BuildTaskContext()` collects previous task results
|
||||
- `FormatPreviousResultsAsContext()` formats them for assistant
|
||||
- MCP and Process tasks use single-call execution (no multi-turn)
|
||||
- `yao/assert` is a standalone package, can be used by other modules
|
||||
- Agent context is properly released via `defer agentCtx.Release()` in `AgentCaller.Call()`
|
||||
|
||||
---
|
||||
|
||||
|
|
|
|||
|
|
@ -178,6 +178,7 @@ func (c *AgentCaller) Call(ctx *robottypes.Context, assistantID string, messages
|
|||
|
||||
// Convert robot context to agent context
|
||||
agentCtx := c.buildAgentContext(ctx)
|
||||
defer agentCtx.Release() // IMPORTANT: Release agent context to prevent resource leaks
|
||||
|
||||
// Call assistant with streaming
|
||||
response, err := ast.Stream(agentCtx, messages, opts)
|
||||
|
|
|
|||
|
|
@ -9,27 +9,21 @@ import (
|
|||
|
||||
// RunConfig configures P3 execution behavior
|
||||
type RunConfig struct {
|
||||
// MaxRetries is the maximum number of retry attempts per task (default: 3)
|
||||
MaxRetries int
|
||||
|
||||
// RetryOnValidationFailure enables retry when validation fails (default: true)
|
||||
RetryOnValidationFailure bool
|
||||
|
||||
// ContinueOnFailure continues to next task even if current task fails (default: false)
|
||||
ContinueOnFailure bool
|
||||
|
||||
// ValidationThreshold is the minimum score to pass validation (default: 0.6)
|
||||
ValidationThreshold float64
|
||||
|
||||
// MaxTurnsPerTask is the maximum conversation turns for multi-turn agents (default: 10)
|
||||
// MaxTurnsPerTask is the maximum conversation turns for multi-turn tasks (default: 10)
|
||||
// This controls how many times the assistant can be called for a single task
|
||||
// (including retries with validation feedback)
|
||||
MaxTurnsPerTask int
|
||||
}
|
||||
|
||||
// DefaultRunConfig returns the default P3 configuration
|
||||
func DefaultRunConfig() *RunConfig {
|
||||
return &RunConfig{
|
||||
MaxRetries: 3,
|
||||
RetryOnValidationFailure: true,
|
||||
ContinueOnFailure: false,
|
||||
ValidationThreshold: 0.6,
|
||||
MaxTurnsPerTask: 10,
|
||||
|
|
@ -38,7 +32,7 @@ func DefaultRunConfig() *RunConfig {
|
|||
|
||||
// RunExecution executes P3: Run phase
|
||||
// Executes each task using the appropriate executor (Assistant, MCP, Process)
|
||||
// with validation and retry mechanism
|
||||
// with multi-turn conversation and validation
|
||||
//
|
||||
// Input:
|
||||
// - Tasks (from P2)
|
||||
|
|
@ -46,12 +40,12 @@ func DefaultRunConfig() *RunConfig {
|
|||
// Output:
|
||||
// - TaskResult for each task with output and validation
|
||||
//
|
||||
// Features:
|
||||
// 1. Sequential task execution with progress tracking
|
||||
// 2. Validation after each task using Validation Agent
|
||||
// 3. Retry mechanism with feedback loop to expert agent
|
||||
// 4. Multi-turn conversation support for complex tasks
|
||||
// 5. Previous task results passed as context to next task
|
||||
// Execution Flow (per task):
|
||||
// 1. Call assistant/MCP/process and get result
|
||||
// 2. Validate result using two-layer validation (rule-based + semantic)
|
||||
// 3. If validation.NeedReply, continue conversation with validation.ReplyContent
|
||||
// 4. Repeat until validation.Complete or max turns exceeded
|
||||
// 5. Pass previous task results as context to next task
|
||||
func (e *Executor) RunExecution(ctx *robottypes.Context, exec *robottypes.Execution, _ interface{}) error {
|
||||
robot := exec.GetRobot()
|
||||
if robot == nil {
|
||||
|
|
@ -90,13 +84,16 @@ func (e *Executor) RunExecution(ctx *robottypes.Context, exec *robottypes.Execut
|
|||
// Build task context with previous results
|
||||
taskCtx := runner.BuildTaskContext(exec, i)
|
||||
|
||||
// Execute task with retry
|
||||
// Execute task with multi-turn conversation support
|
||||
result := runner.ExecuteWithRetry(task, taskCtx)
|
||||
|
||||
// Update task status based on result
|
||||
endTime := time.Now()
|
||||
task.EndTime = &endTime
|
||||
if result.Success && (result.Validation == nil || result.Validation.Passed) {
|
||||
|
||||
// Determine task status from result
|
||||
// Note: result.Success is already set to (validation.Complete && validation.Passed) in runner
|
||||
if result.Success {
|
||||
task.Status = robottypes.TaskCompleted
|
||||
} else {
|
||||
task.Status = robottypes.TaskFailed
|
||||
|
|
|
|||
|
|
@ -61,7 +61,11 @@ func (r *Runner) BuildTaskContext(exec *robottypes.Execution, taskIndex int) *Ru
|
|||
return ctx
|
||||
}
|
||||
|
||||
// ExecuteWithRetry executes a task with retry mechanism
|
||||
// ExecuteWithRetry executes a task with the new multi-turn conversation flow:
|
||||
// 1. Call assistant and get result
|
||||
// 2. Validate result (determines: passed, complete, needReply, replyContent)
|
||||
// 3. If needReply, continue conversation with replyContent
|
||||
// 4. Repeat until complete or max turns exceeded
|
||||
func (r *Runner) ExecuteWithRetry(task *robottypes.Task, taskCtx *RunnerContext) *robottypes.TaskResult {
|
||||
startTime := time.Now()
|
||||
|
||||
|
|
@ -69,86 +73,74 @@ func (r *Runner) ExecuteWithRetry(task *robottypes.Task, taskCtx *RunnerContext)
|
|||
TaskID: task.ID,
|
||||
}
|
||||
|
||||
var lastOutput interface{}
|
||||
var lastValidation *robottypes.ValidationResult
|
||||
var allErrors []string // Collect all errors for debugging
|
||||
|
||||
for attempt := 0; attempt <= r.config.MaxRetries; attempt++ {
|
||||
// Execute the task
|
||||
output, err := r.ExecuteTask(task, taskCtx, lastValidation)
|
||||
// For non-assistant tasks (MCP, Process), use simple single-call execution
|
||||
if task.ExecutorType != robottypes.ExecutorAssistant {
|
||||
output, err := r.executeNonAssistantTask(task, taskCtx)
|
||||
if err != nil {
|
||||
// Execution error - don't retry, return immediately
|
||||
// (Retries are only for validation failures, not execution errors)
|
||||
errMsg := fmt.Sprintf("execution failed on attempt %d: %s", attempt+1, err.Error())
|
||||
allErrors = append(allErrors, errMsg)
|
||||
result.Success = false
|
||||
result.Error = strings.Join(allErrors, "; ")
|
||||
result.Error = fmt.Sprintf("execution failed: %s", err.Error())
|
||||
result.Duration = time.Since(startTime).Milliseconds()
|
||||
return result
|
||||
}
|
||||
|
||||
lastOutput = output
|
||||
result.Output = output
|
||||
|
||||
// Validate the result (reuse validator instance)
|
||||
validation := r.validator.Validate(task, output)
|
||||
lastValidation = validation
|
||||
validation := r.validator.ValidateWithContext(task, output, nil)
|
||||
result.Validation = validation
|
||||
// For non-assistant tasks (MCP, Process):
|
||||
// - No multi-turn conversation, so Complete is determined by validation alone
|
||||
// - Success if passed OR score meets threshold (for partial success scenarios)
|
||||
result.Success = validation.Complete || (validation.Passed && validation.Score >= r.config.ValidationThreshold)
|
||||
result.Duration = time.Since(startTime).Milliseconds()
|
||||
|
||||
// Check if validation passed (unified logic)
|
||||
validationPassed := validation.Passed || validation.Score >= r.config.ValidationThreshold
|
||||
if validationPassed {
|
||||
result.Success = true
|
||||
if !result.Success && validation != nil {
|
||||
result.Error = fmt.Sprintf("validation failed: %v", validation.Issues)
|
||||
}
|
||||
return result
|
||||
}
|
||||
|
||||
// For assistant tasks, use multi-turn conversation flow
|
||||
output, validation, err := r.executeAssistantWithMultiTurn(task, taskCtx)
|
||||
if err != nil {
|
||||
result.Success = false
|
||||
result.Error = err.Error()
|
||||
result.Output = output // Preserve partial output for debugging
|
||||
result.Validation = validation // Preserve validation result for debugging
|
||||
result.Duration = time.Since(startTime).Milliseconds()
|
||||
return result
|
||||
}
|
||||
|
||||
// Validation failed - check if we should retry
|
||||
if !r.config.RetryOnValidationFailure || attempt >= r.config.MaxRetries {
|
||||
break
|
||||
}
|
||||
|
||||
// Prepare for retry with validation feedback
|
||||
// The next iteration will include validation issues in the context
|
||||
}
|
||||
|
||||
// All retries exhausted
|
||||
result.Success = false
|
||||
result.Output = lastOutput
|
||||
result.Validation = lastValidation
|
||||
result.Output = output
|
||||
result.Validation = validation
|
||||
result.Success = validation.Complete && validation.Passed
|
||||
result.Duration = time.Since(startTime).Milliseconds()
|
||||
|
||||
if len(allErrors) > 0 {
|
||||
result.Error = strings.Join(allErrors, "; ")
|
||||
} else if lastValidation != nil {
|
||||
result.Error = fmt.Sprintf("validation failed after %d attempts: %v",
|
||||
r.config.MaxRetries+1, lastValidation.Issues)
|
||||
if !result.Success && validation != nil {
|
||||
result.Error = fmt.Sprintf("task incomplete: %v", validation.Issues)
|
||||
}
|
||||
|
||||
return result
|
||||
}
|
||||
|
||||
// ExecuteTask executes a single task based on its executor type
|
||||
func (r *Runner) ExecuteTask(task *robottypes.Task, taskCtx *RunnerContext, prevValidation *robottypes.ValidationResult) (interface{}, error) {
|
||||
// executeNonAssistantTask executes MCP or Process tasks (single-call, no multi-turn)
|
||||
func (r *Runner) executeNonAssistantTask(task *robottypes.Task, taskCtx *RunnerContext) (interface{}, error) {
|
||||
switch task.ExecutorType {
|
||||
case robottypes.ExecutorAssistant:
|
||||
return r.ExecuteAssistantTask(task, taskCtx, prevValidation)
|
||||
case robottypes.ExecutorMCP:
|
||||
return r.ExecuteMCPTask(task, taskCtx)
|
||||
case robottypes.ExecutorProcess:
|
||||
return r.ExecuteProcessTask(task, taskCtx)
|
||||
default:
|
||||
return nil, fmt.Errorf("unknown executor type: %s", task.ExecutorType)
|
||||
return nil, fmt.Errorf("unsupported executor type: %s (expected mcp or process)", task.ExecutorType)
|
||||
}
|
||||
}
|
||||
|
||||
// ExecuteAssistantTask executes a task using an AI assistant
|
||||
// Supports multi-turn conversation for complex tasks
|
||||
func (r *Runner) ExecuteAssistantTask(task *robottypes.Task, taskCtx *RunnerContext, prevValidation *robottypes.ValidationResult) (interface{}, error) {
|
||||
// Build messages for the assistant
|
||||
messages := r.BuildAssistantMessages(task, taskCtx, prevValidation)
|
||||
|
||||
// Create conversation for multi-turn support
|
||||
// executeAssistantWithMultiTurn executes an assistant task with multi-turn conversation support
|
||||
// This is the main execution flow for assistant tasks:
|
||||
// 1. Call assistant and get result
|
||||
// 2. Validate result (determines: passed, complete, needReply, replyContent)
|
||||
// 3. If needReply, continue conversation with replyContent
|
||||
// 4. Repeat until complete or max turns exceeded
|
||||
func (r *Runner) executeAssistantWithMultiTurn(task *robottypes.Task, taskCtx *RunnerContext) (interface{}, *robottypes.ValidationResult, error) {
|
||||
// Create conversation for the entire task execution (shared across all turns)
|
||||
chatID := fmt.Sprintf("robot-%s-task-%s", r.robot.MemberID, task.ID)
|
||||
conv := NewConversation(task.ExecutorID, chatID, r.config.MaxTurnsPerTask)
|
||||
|
||||
|
|
@ -157,55 +149,106 @@ func (r *Runner) ExecuteAssistantTask(task *robottypes.Task, taskCtx *RunnerCont
|
|||
conv.WithSystemPrompt(taskCtx.SystemPrompt)
|
||||
}
|
||||
|
||||
// First turn: send the task
|
||||
firstInput := r.FormatMessagesAsText(messages)
|
||||
turnResult, err := conv.Turn(r.ctx, firstInput)
|
||||
// Build initial messages
|
||||
messages := r.BuildAssistantMessages(task, taskCtx)
|
||||
input := r.FormatMessagesAsText(messages)
|
||||
|
||||
// Ensure we have valid input for the first turn
|
||||
if strings.TrimSpace(input) == "" {
|
||||
return nil, nil, fmt.Errorf("no valid input messages for task %s", task.ID)
|
||||
}
|
||||
|
||||
var lastOutput interface{}
|
||||
var lastValidation *robottypes.ValidationResult
|
||||
var lastCallResult *CallResult
|
||||
|
||||
for turn := 1; turn <= r.config.MaxTurnsPerTask; turn++ {
|
||||
// Phase 1: Call assistant
|
||||
turnResult, err := conv.Turn(r.ctx, input)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("assistant call failed: %w", err)
|
||||
return lastOutput, lastValidation, fmt.Errorf("turn %d failed: %w", turn, err)
|
||||
}
|
||||
|
||||
// Check if the assistant needs more information (multi-turn)
|
||||
// We detect this by checking if the response indicates incompleteness
|
||||
// or if there are tool calls that need results
|
||||
response := turnResult.Result
|
||||
lastCallResult = turnResult.Result
|
||||
lastOutput = r.extractOutput(lastCallResult)
|
||||
|
||||
// For simple tasks, return the result directly
|
||||
if response.Response == nil || len(response.Response.Tools) == 0 {
|
||||
// Try to extract structured output
|
||||
if data, err := response.GetJSON(); err == nil {
|
||||
return data, nil
|
||||
}
|
||||
// Return text content
|
||||
return response.GetText(), nil
|
||||
// Phase 2: Validate result
|
||||
lastValidation = r.validator.ValidateWithContext(task, lastOutput, lastCallResult)
|
||||
|
||||
// Check if complete
|
||||
if lastValidation.Complete && lastValidation.Passed {
|
||||
return lastOutput, lastValidation, nil // Success!
|
||||
}
|
||||
|
||||
// Handle multi-turn conversation with auto-reply simulation
|
||||
// Similar to the test framework's dynamic runner
|
||||
for turn := 2; turn <= r.config.MaxTurnsPerTask; turn++ {
|
||||
// Check if we have a complete response
|
||||
if r.IsResponseComplete(response) {
|
||||
break
|
||||
// Phase 3: Check if we should continue conversation
|
||||
if !lastValidation.NeedReply {
|
||||
// No need to continue, but not complete either
|
||||
// This could be a validation failure that can't be fixed by conversation
|
||||
if lastValidation.Passed {
|
||||
// Passed but not complete (e.g., empty output)
|
||||
return lastOutput, lastValidation, nil
|
||||
}
|
||||
// Failed and can't retry
|
||||
return lastOutput, lastValidation, fmt.Errorf("validation failed: %v", lastValidation.Issues)
|
||||
}
|
||||
|
||||
// Generate auto-reply based on tool results or context
|
||||
autoReply := r.GenerateAutoReply(response, task)
|
||||
if autoReply == "" {
|
||||
break // No more input needed
|
||||
// Prepare next turn input
|
||||
input = lastValidation.ReplyContent
|
||||
if input == "" {
|
||||
// Fallback: generate default reply
|
||||
input = r.generateDefaultReply(lastValidation, task)
|
||||
}
|
||||
}
|
||||
|
||||
// Continue conversation
|
||||
turnResult, err = conv.Turn(r.ctx, autoReply)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("assistant turn %d failed: %w", turn, err)
|
||||
// Max turns exceeded
|
||||
if lastValidation == nil {
|
||||
lastValidation = &robottypes.ValidationResult{
|
||||
Passed: false,
|
||||
Complete: false,
|
||||
Issues: []string{fmt.Sprintf("max turns (%d) exceeded without completion", r.config.MaxTurnsPerTask)},
|
||||
}
|
||||
response = turnResult.Result
|
||||
} else {
|
||||
lastValidation.Issues = append(lastValidation.Issues,
|
||||
fmt.Sprintf("max turns (%d) exceeded without completion", r.config.MaxTurnsPerTask))
|
||||
}
|
||||
|
||||
// Extract final output
|
||||
if data, err := response.GetJSON(); err == nil {
|
||||
return data, nil
|
||||
return lastOutput, lastValidation, fmt.Errorf("max turns (%d) exceeded without completion", r.config.MaxTurnsPerTask)
|
||||
}
|
||||
|
||||
// extractOutput extracts the output from a CallResult
|
||||
func (r *Runner) extractOutput(result *CallResult) interface{} {
|
||||
if result == nil {
|
||||
return nil
|
||||
}
|
||||
return response.GetText(), nil
|
||||
|
||||
// Try to extract structured JSON output
|
||||
if data, err := result.GetJSON(); err == nil {
|
||||
return data
|
||||
}
|
||||
|
||||
// Fall back to text content
|
||||
return result.GetText()
|
||||
}
|
||||
|
||||
// generateDefaultReply generates a default reply when validation doesn't provide one
|
||||
func (r *Runner) generateDefaultReply(validation *robottypes.ValidationResult, task *robottypes.Task) string {
|
||||
var sb strings.Builder
|
||||
|
||||
if len(validation.Issues) > 0 {
|
||||
sb.WriteString("Please address the following issues:\n")
|
||||
for _, issue := range validation.Issues {
|
||||
sb.WriteString(fmt.Sprintf("- %s\n", issue))
|
||||
}
|
||||
sb.WriteString("\n")
|
||||
}
|
||||
|
||||
if task.ExpectedOutput != "" {
|
||||
sb.WriteString(fmt.Sprintf("Expected output: %s\n", task.ExpectedOutput))
|
||||
}
|
||||
|
||||
sb.WriteString("\nPlease provide an improved response.")
|
||||
|
||||
return sb.String()
|
||||
}
|
||||
|
||||
// ExecuteMCPTask executes a task using an MCP tool
|
||||
|
|
@ -269,7 +312,8 @@ func (r *Runner) ExecuteProcessTask(task *robottypes.Task, taskCtx *RunnerContex
|
|||
}
|
||||
|
||||
// BuildAssistantMessages builds messages for an assistant task
|
||||
func (r *Runner) BuildAssistantMessages(task *robottypes.Task, taskCtx *RunnerContext, prevValidation *robottypes.ValidationResult) []agentcontext.Message {
|
||||
// Note: In the new multi-turn flow, validation feedback is handled via ValidateWithContext.ReplyContent
|
||||
func (r *Runner) BuildAssistantMessages(task *robottypes.Task, taskCtx *RunnerContext) []agentcontext.Message {
|
||||
messages := make([]agentcontext.Message, 0)
|
||||
|
||||
// Add context from previous tasks if available
|
||||
|
|
@ -286,15 +330,6 @@ func (r *Runner) BuildAssistantMessages(task *robottypes.Task, taskCtx *RunnerCo
|
|||
// Add task messages
|
||||
messages = append(messages, task.Messages...)
|
||||
|
||||
// Add validation feedback if this is a retry
|
||||
if prevValidation != nil && !prevValidation.Passed {
|
||||
feedbackMsg := r.FormatValidationFeedback(prevValidation)
|
||||
messages = append(messages, agentcontext.Message{
|
||||
Role: agentcontext.RoleUser,
|
||||
Content: feedbackMsg,
|
||||
})
|
||||
}
|
||||
|
||||
return messages
|
||||
}
|
||||
|
||||
|
|
@ -357,89 +392,3 @@ func (r *Runner) FormatPreviousResultsAsContext(results []robottypes.TaskResult)
|
|||
|
||||
return sb.String()
|
||||
}
|
||||
|
||||
// FormatValidationFeedback formats validation feedback for retry
|
||||
func (r *Runner) FormatValidationFeedback(validation *robottypes.ValidationResult) string {
|
||||
var sb strings.Builder
|
||||
sb.WriteString("## Validation Feedback\n\n")
|
||||
sb.WriteString("Your previous response did not pass validation. Please address the following issues:\n\n")
|
||||
|
||||
if len(validation.Issues) > 0 {
|
||||
sb.WriteString("### Issues\n")
|
||||
for _, issue := range validation.Issues {
|
||||
sb.WriteString(fmt.Sprintf("- %s\n", issue))
|
||||
}
|
||||
sb.WriteString("\n")
|
||||
}
|
||||
|
||||
if len(validation.Suggestions) > 0 {
|
||||
sb.WriteString("### Suggestions\n")
|
||||
for _, suggestion := range validation.Suggestions {
|
||||
sb.WriteString(fmt.Sprintf("- %s\n", suggestion))
|
||||
}
|
||||
sb.WriteString("\n")
|
||||
}
|
||||
|
||||
sb.WriteString("Please provide an improved response that addresses these issues.\n")
|
||||
|
||||
return sb.String()
|
||||
}
|
||||
|
||||
// IsResponseComplete checks if an assistant response is complete
|
||||
// (no pending tool calls, has content)
|
||||
func (r *Runner) IsResponseComplete(result *CallResult) bool {
|
||||
if result == nil || result.Response == nil {
|
||||
return true
|
||||
}
|
||||
|
||||
// If there are tool calls, check if all have results
|
||||
if len(result.Response.Tools) > 0 {
|
||||
for _, tool := range result.Response.Tools {
|
||||
if tool.Result == nil {
|
||||
return false // Still waiting for tool results
|
||||
}
|
||||
}
|
||||
// All tools have results - response is complete
|
||||
return true
|
||||
}
|
||||
|
||||
// No tools - check if there's content
|
||||
return result.Content != "" || result.Next != nil
|
||||
}
|
||||
|
||||
// GenerateAutoReply generates an automatic reply for multi-turn conversation
|
||||
// This simulates user responses when the assistant needs more information
|
||||
func (r *Runner) GenerateAutoReply(result *CallResult, task *robottypes.Task) string {
|
||||
if result == nil || result.Response == nil {
|
||||
return ""
|
||||
}
|
||||
|
||||
// If there are tool results, format them as the reply
|
||||
if len(result.Response.Tools) > 0 {
|
||||
var replies []string
|
||||
for _, tool := range result.Response.Tools {
|
||||
if tool.Result != nil {
|
||||
resultJSON, err := json.Marshal(tool.Result)
|
||||
if err == nil {
|
||||
replies = append(replies, fmt.Sprintf("Tool %s result: %s", tool.Tool, string(resultJSON)))
|
||||
}
|
||||
}
|
||||
}
|
||||
if len(replies) > 0 {
|
||||
return fmt.Sprintf("Tool execution results:\n%s\n\nPlease continue with the task.", strings.Join(replies, "\n"))
|
||||
}
|
||||
}
|
||||
|
||||
// If the response asks for clarification, provide generic guidance
|
||||
// Use case-insensitive matching
|
||||
content := strings.ToLower(result.GetText())
|
||||
clarificationKeywords := []string{"need more", "clarify", "please provide", "what", "which"}
|
||||
for _, keyword := range clarificationKeywords {
|
||||
if strings.Contains(content, keyword) {
|
||||
return fmt.Sprintf("Please proceed with the task as best as you can based on the available information. "+
|
||||
"The expected output is: %s", task.ExpectedOutput)
|
||||
}
|
||||
}
|
||||
|
||||
return ""
|
||||
}
|
||||
|
|
|
|||
|
|
@ -36,15 +36,30 @@ func NewValidator(ctx *robottypes.Context, robot *robottypes.Robot, config *RunC
|
|||
return v
|
||||
}
|
||||
|
||||
// Validate validates task output using two-layer validation:
|
||||
// 1. First, run rule-based assertions (fast, deterministic)
|
||||
// 2. Then, if ExpectedOutput is set, run semantic validation via Agent
|
||||
// Validate validates task output using two-layer validation (without multi-turn context)
|
||||
// Equivalent to ValidateWithContext(task, output, nil)
|
||||
// Use ValidateWithContext when you have a CallResult for better multi-turn support
|
||||
func (v *Validator) Validate(task *robottypes.Task, output interface{}) *robottypes.ValidationResult {
|
||||
// If no validation rules and no expected output, return passed
|
||||
return v.ValidateWithContext(task, output, nil)
|
||||
}
|
||||
|
||||
// ValidateWithContext validates task output and determines execution state for multi-turn conversation.
|
||||
// It extends basic validation with:
|
||||
// - Complete: whether expected result is obtained
|
||||
// - NeedReply: whether to continue conversation
|
||||
// - ReplyContent: content for next turn
|
||||
//
|
||||
// Parameters:
|
||||
// - task: the task being executed
|
||||
// - output: the output from assistant/mcp/process
|
||||
// - callResult: the full call result (for detecting assistant's need for more info)
|
||||
func (v *Validator) ValidateWithContext(task *robottypes.Task, output interface{}, callResult *CallResult) *robottypes.ValidationResult {
|
||||
// If no validation rules and no expected output, return passed and complete
|
||||
if task.ExpectedOutput == "" && len(task.ValidationRules) == 0 {
|
||||
return &robottypes.ValidationResult{
|
||||
Passed: true,
|
||||
Score: 1.0,
|
||||
Complete: v.hasValidOutput(output),
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -57,6 +72,9 @@ func (v *Validator) Validate(task *robottypes.Task, output interface{}) *robotty
|
|||
if len(task.ValidationRules) > 0 {
|
||||
ruleResult := v.validateRules(task.ValidationRules, output)
|
||||
if !ruleResult.Passed {
|
||||
// Rule validation failed - check if we should retry with feedback
|
||||
ruleResult.Complete = false
|
||||
ruleResult.NeedReply, ruleResult.ReplyContent = v.checkNeedReplyOnFailure(task, ruleResult)
|
||||
return ruleResult
|
||||
}
|
||||
// Merge rule validation results
|
||||
|
|
@ -71,9 +89,186 @@ func (v *Validator) Validate(task *robottypes.Task, output interface{}) *robotty
|
|||
result = v.mergeResults(result, semanticResult)
|
||||
}
|
||||
|
||||
// Determine execution state
|
||||
result.Complete = v.isComplete(task, output, result)
|
||||
result.NeedReply, result.ReplyContent = v.checkNeedReply(task, output, callResult, result)
|
||||
|
||||
return result
|
||||
}
|
||||
|
||||
// hasValidOutput checks if output is non-empty and valid
|
||||
func (v *Validator) hasValidOutput(output interface{}) bool {
|
||||
if output == nil {
|
||||
return false
|
||||
}
|
||||
switch o := output.(type) {
|
||||
case string:
|
||||
return strings.TrimSpace(o) != ""
|
||||
case []interface{}:
|
||||
return len(o) > 0
|
||||
case map[string]interface{}:
|
||||
return len(o) > 0
|
||||
default:
|
||||
return true
|
||||
}
|
||||
}
|
||||
|
||||
// isComplete determines if the expected result has been obtained
|
||||
func (v *Validator) isComplete(task *robottypes.Task, output interface{}, result *robottypes.ValidationResult) bool {
|
||||
// If validation failed, not complete
|
||||
if !result.Passed {
|
||||
return false
|
||||
}
|
||||
|
||||
// Must have valid output
|
||||
if !v.hasValidOutput(output) {
|
||||
return false
|
||||
}
|
||||
|
||||
// If score is below threshold, consider incomplete
|
||||
if result.Score < v.config.ValidationThreshold {
|
||||
return false
|
||||
}
|
||||
|
||||
return true
|
||||
}
|
||||
|
||||
// checkNeedReply determines if conversation should continue and generates reply content
|
||||
func (v *Validator) checkNeedReply(task *robottypes.Task, output interface{}, callResult *CallResult, result *robottypes.ValidationResult) (bool, string) {
|
||||
// If already complete, no need to reply
|
||||
if result.Complete {
|
||||
return false, ""
|
||||
}
|
||||
|
||||
// Scenario 1: Assistant explicitly asks for more information
|
||||
if callResult != nil {
|
||||
text := callResult.GetText()
|
||||
if v.detectNeedMoreInfo(text) {
|
||||
return true, v.generateClarificationReply(task, text)
|
||||
}
|
||||
}
|
||||
|
||||
// Scenario 2: Validation passed but output is incomplete/empty
|
||||
if result.Passed && !v.hasValidOutput(output) {
|
||||
return true, "Please continue and provide the complete result as specified in the task."
|
||||
}
|
||||
|
||||
// Scenario 3: Validation failed with suggestions - can retry with feedback
|
||||
if !result.Passed && len(result.Suggestions) > 0 {
|
||||
return true, v.generateFeedbackReply(result)
|
||||
}
|
||||
|
||||
// Scenario 4: Low confidence score - ask for improvement
|
||||
if result.Passed && result.Score < v.config.ValidationThreshold {
|
||||
return true, fmt.Sprintf("The result is partially correct (score: %.2f), but needs improvement. Please refine your response to better match the expected output: %s", result.Score, task.ExpectedOutput)
|
||||
}
|
||||
|
||||
// No need to continue
|
||||
return false, ""
|
||||
}
|
||||
|
||||
// checkNeedReplyOnFailure handles the case when rule validation fails
|
||||
func (v *Validator) checkNeedReplyOnFailure(task *robottypes.Task, result *robottypes.ValidationResult) (bool, string) {
|
||||
// If there are suggestions, we can try to fix
|
||||
if len(result.Suggestions) > 0 {
|
||||
return true, v.generateFeedbackReply(result)
|
||||
}
|
||||
|
||||
// If there are issues, provide feedback
|
||||
if len(result.Issues) > 0 {
|
||||
var sb strings.Builder
|
||||
sb.WriteString("Your response did not pass validation. Please fix the following issues:\n\n")
|
||||
for _, issue := range result.Issues {
|
||||
sb.WriteString(fmt.Sprintf("- %s\n", issue))
|
||||
}
|
||||
sb.WriteString(fmt.Sprintf("\nExpected output: %s", task.ExpectedOutput))
|
||||
return true, sb.String()
|
||||
}
|
||||
|
||||
return false, ""
|
||||
}
|
||||
|
||||
// detectNeedMoreInfo checks if assistant's response indicates need for more information
|
||||
func (v *Validator) detectNeedMoreInfo(text string) bool {
|
||||
if text == "" {
|
||||
return false
|
||||
}
|
||||
|
||||
textLower := strings.ToLower(text)
|
||||
keywords := []string{
|
||||
"need more information",
|
||||
"please clarify",
|
||||
"could you provide",
|
||||
"can you specify",
|
||||
"what is the",
|
||||
"which one",
|
||||
"please provide",
|
||||
"i need to know",
|
||||
"could you tell me",
|
||||
"what do you mean",
|
||||
}
|
||||
|
||||
for _, kw := range keywords {
|
||||
if strings.Contains(textLower, kw) {
|
||||
return true
|
||||
}
|
||||
}
|
||||
|
||||
// Check for question marks at the end (likely asking for clarification)
|
||||
// Note: We require 2+ question marks to avoid false positives from rhetorical questions
|
||||
// or questions that are part of the output (e.g., "How can I help you?")
|
||||
// Single questions are often just conversational and don't need clarification
|
||||
trimmed := strings.TrimSpace(text)
|
||||
if strings.HasSuffix(trimmed, "?") {
|
||||
if strings.Count(text, "?") >= 2 {
|
||||
return true
|
||||
}
|
||||
}
|
||||
|
||||
return false
|
||||
}
|
||||
|
||||
// generateClarificationReply generates a reply when assistant asks for clarification
|
||||
func (v *Validator) generateClarificationReply(task *robottypes.Task, assistantText string) string {
|
||||
var sb strings.Builder
|
||||
sb.WriteString("Please proceed with the task based on the available information.\n\n")
|
||||
|
||||
if task.ExpectedOutput != "" {
|
||||
sb.WriteString(fmt.Sprintf("**Expected Output**: %s\n\n", task.ExpectedOutput))
|
||||
}
|
||||
|
||||
sb.WriteString("If you need to make assumptions, please state them clearly and proceed with the most reasonable interpretation.")
|
||||
|
||||
return sb.String()
|
||||
}
|
||||
|
||||
// generateFeedbackReply generates a reply with validation feedback
|
||||
func (v *Validator) generateFeedbackReply(result *robottypes.ValidationResult) string {
|
||||
var sb strings.Builder
|
||||
sb.WriteString("## Validation Feedback\n\n")
|
||||
sb.WriteString("Your previous response needs improvement. Please address the following:\n\n")
|
||||
|
||||
if len(result.Issues) > 0 {
|
||||
sb.WriteString("### Issues\n")
|
||||
for _, issue := range result.Issues {
|
||||
sb.WriteString(fmt.Sprintf("- %s\n", issue))
|
||||
}
|
||||
sb.WriteString("\n")
|
||||
}
|
||||
|
||||
if len(result.Suggestions) > 0 {
|
||||
sb.WriteString("### Suggestions\n")
|
||||
for _, suggestion := range result.Suggestions {
|
||||
sb.WriteString(fmt.Sprintf("- %s\n", suggestion))
|
||||
}
|
||||
sb.WriteString("\n")
|
||||
}
|
||||
|
||||
sb.WriteString("Please provide an improved response that addresses these points.")
|
||||
|
||||
return sb.String()
|
||||
}
|
||||
|
||||
// validateRules validates output against rule-based assertions
|
||||
func (v *Validator) validateRules(rules []string, output interface{}) *robottypes.ValidationResult {
|
||||
result := &robottypes.ValidationResult{
|
||||
|
|
@ -233,16 +428,21 @@ func (v *Validator) validateSemantic(task *robottypes.Task, output interface{})
|
|||
}
|
||||
|
||||
// BuildSemanticPrompt builds the prompt for semantic validation
|
||||
// Format matches the Validation Agent's expected input structure:
|
||||
// 1. Task: task definition with expected_output and validation_rules
|
||||
// 2. Result: actual output from task execution
|
||||
// 3. Success Criteria: overall criteria (optional)
|
||||
func (v *Validator) BuildSemanticPrompt(task *robottypes.Task, output interface{}) string {
|
||||
var sb strings.Builder
|
||||
|
||||
sb.WriteString("## Task Definition\n\n")
|
||||
// Section 1: Task (matches Agent's expected "Task" input)
|
||||
sb.WriteString("## Task\n\n")
|
||||
sb.WriteString(fmt.Sprintf("**Task ID**: %s\n", task.ID))
|
||||
sb.WriteString(fmt.Sprintf("**Executor**: %s (%s)\n\n", task.ExecutorID, task.ExecutorType))
|
||||
|
||||
// Task description
|
||||
// Task description (instructions)
|
||||
if len(task.Messages) > 0 {
|
||||
sb.WriteString("**Task Instructions**:\n")
|
||||
sb.WriteString("**Instructions**:\n")
|
||||
for _, msg := range task.Messages {
|
||||
if content, ok := msg.Content.(string); ok {
|
||||
sb.WriteString(content + "\n")
|
||||
|
|
@ -253,21 +453,21 @@ func (v *Validator) BuildSemanticPrompt(task *robottypes.Task, output interface{
|
|||
|
||||
// Expected output (primary criterion for semantic validation)
|
||||
if task.ExpectedOutput != "" {
|
||||
sb.WriteString(fmt.Sprintf("**Expected Output**: %s\n\n", task.ExpectedOutput))
|
||||
sb.WriteString(fmt.Sprintf("**expected_output**: %s\n\n", task.ExpectedOutput))
|
||||
}
|
||||
|
||||
// Semantic validation rules (rules that couldn't be converted to assertions)
|
||||
// Validation rules
|
||||
semanticRules := v.getSemanticRules(task.ValidationRules)
|
||||
if len(semanticRules) > 0 {
|
||||
sb.WriteString("**Validation Criteria**:\n")
|
||||
sb.WriteString("**validation_rules**:\n")
|
||||
for _, rule := range semanticRules {
|
||||
sb.WriteString(fmt.Sprintf("- %s\n", rule))
|
||||
}
|
||||
sb.WriteString("\n")
|
||||
}
|
||||
|
||||
// Actual output
|
||||
sb.WriteString("## Actual Output\n\n")
|
||||
// Section 2: Result (matches Agent's expected "Result" input)
|
||||
sb.WriteString("## Result\n\n")
|
||||
if output != nil {
|
||||
outputJSON, err := json.MarshalIndent(output, "", " ")
|
||||
if err == nil {
|
||||
|
|
@ -279,10 +479,14 @@ func (v *Validator) BuildSemanticPrompt(task *robottypes.Task, output interface{
|
|||
sb.WriteString("(no output)\n")
|
||||
}
|
||||
|
||||
sb.WriteString("\n## Validation Request\n\n")
|
||||
sb.WriteString("Please validate the actual output against the expected output and validation criteria. ")
|
||||
sb.WriteString("Focus on semantic correctness and completeness. ")
|
||||
sb.WriteString("Return a JSON object with: passed (bool), score (0-1), issues (array), suggestions (array), details (markdown report).\n")
|
||||
// Section 3: Success Criteria (optional, from goals if available)
|
||||
// Note: This could be extended to include criteria from exec.Goals if needed
|
||||
sb.WriteString("\n## Success Criteria\n\n")
|
||||
if task.ExpectedOutput != "" {
|
||||
sb.WriteString(fmt.Sprintf("The task should produce: %s\n", task.ExpectedOutput))
|
||||
} else {
|
||||
sb.WriteString("Complete the task successfully with valid output.\n")
|
||||
}
|
||||
|
||||
return sb.String()
|
||||
}
|
||||
|
|
|
|||
|
|
@ -249,11 +249,17 @@ type TaskResult struct {
|
|||
|
||||
// ValidationResult - P3 semantic validation result
|
||||
type ValidationResult struct {
|
||||
// Basic validation result
|
||||
Passed bool `json:"passed"` // overall validation passed
|
||||
Score float64 `json:"score,omitempty"` // 0-1 confidence score
|
||||
Issues []string `json:"issues,omitempty"` // what failed
|
||||
Suggestions []string `json:"suggestions,omitempty"` // how to improve
|
||||
Details string `json:"details,omitempty"` // detailed validation report (markdown)
|
||||
|
||||
// Execution state (for multi-turn conversation control)
|
||||
Complete bool `json:"complete"` // whether expected result is obtained
|
||||
NeedReply bool `json:"need_reply,omitempty"` // whether to continue conversation
|
||||
ReplyContent string `json:"reply_content,omitempty"` // content for next turn (if NeedReply)
|
||||
}
|
||||
|
||||
// DeliveryResult - P4 delivery output
|
||||
|
|
|
|||
|
|
@ -474,6 +474,65 @@ func TestValidationResultStructure(t *testing.T) {
|
|||
assert.Len(t, validation.Suggestions, 2)
|
||||
}
|
||||
|
||||
func TestValidationResultMultiTurnFields(t *testing.T) {
|
||||
// Test new multi-turn conversation control fields
|
||||
t.Run("complete and passed", func(t *testing.T) {
|
||||
validation := &types.ValidationResult{
|
||||
Passed: true,
|
||||
Score: 0.95,
|
||||
Complete: true,
|
||||
}
|
||||
assert.True(t, validation.Passed)
|
||||
assert.True(t, validation.Complete)
|
||||
assert.False(t, validation.NeedReply)
|
||||
assert.Empty(t, validation.ReplyContent)
|
||||
})
|
||||
|
||||
t.Run("passed but not complete - needs reply", func(t *testing.T) {
|
||||
validation := &types.ValidationResult{
|
||||
Passed: true,
|
||||
Score: 0.7,
|
||||
Complete: false,
|
||||
NeedReply: true,
|
||||
ReplyContent: "Please continue and provide the complete result.",
|
||||
}
|
||||
assert.True(t, validation.Passed)
|
||||
assert.False(t, validation.Complete)
|
||||
assert.True(t, validation.NeedReply)
|
||||
assert.NotEmpty(t, validation.ReplyContent)
|
||||
})
|
||||
|
||||
t.Run("failed with suggestions - needs reply", func(t *testing.T) {
|
||||
validation := &types.ValidationResult{
|
||||
Passed: false,
|
||||
Score: 0.3,
|
||||
Complete: false,
|
||||
Issues: []string{"Missing required field"},
|
||||
Suggestions: []string{"Add the field"},
|
||||
NeedReply: true,
|
||||
ReplyContent: "## Validation Feedback\n\nPlease fix: Missing required field",
|
||||
}
|
||||
assert.False(t, validation.Passed)
|
||||
assert.False(t, validation.Complete)
|
||||
assert.True(t, validation.NeedReply)
|
||||
assert.Contains(t, validation.ReplyContent, "Validation Feedback")
|
||||
})
|
||||
|
||||
t.Run("failed without suggestions - no reply", func(t *testing.T) {
|
||||
validation := &types.ValidationResult{
|
||||
Passed: false,
|
||||
Score: 0.0,
|
||||
Complete: false,
|
||||
Issues: []string{"Critical error: invalid format"},
|
||||
NeedReply: false,
|
||||
}
|
||||
assert.False(t, validation.Passed)
|
||||
assert.False(t, validation.Complete)
|
||||
assert.False(t, validation.NeedReply)
|
||||
assert.Empty(t, validation.ReplyContent)
|
||||
})
|
||||
}
|
||||
|
||||
func TestDeliveryResultStructure(t *testing.T) {
|
||||
sentAt := time.Now()
|
||||
delivery := &types.DeliveryResult{
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue