fix: explicit slash command detection + document resilience improvements
**Slash Command Detection Fix:** - Check for <command-name> tag explicitly (not pseudocode) - If tag exists: Run in main context (interactive) - If no tag: Delegate to Task agent (internal call) - Reduces task nesting from 3 levels to 1 level **Resilience Documentation:** - Added RESILIENCE-FIX.md documenting state tracking - Proposes state file for resume capability - File read retry logic - Error handling improvements **Expected improvement:** - Slash commands: No Task wrapper (1 level nesting) - Fewer file access issues - Can resume after crashes To be implemented in future release.
This commit is contained in:
Jonah Schulte
parent
df22c71879
commit
b4217ba65f
|
|
@ -17,60 +17,48 @@
|
|||
</WORKFLOW-RULES>
|
||||
|
||||
<llm critical="ABSOLUTE">
|
||||
<mandate>🚨 WORKFLOW EXECUTION ENFORCEMENT 🚨</mandate>
|
||||
<mandate>🚨 WORKFLOW EXECUTION CONTEXT ENFORCEMENT 🚨</mandate>
|
||||
|
||||
<enforcement>
|
||||
<check type="invocation_source">
|
||||
IF this workflow was invoked by USER via slash command (e.g., /bmad_bmm_batch-super-dev):
|
||||
✅ ALLOWED in main context
|
||||
Reason: User-initiated workflows need interactive prompts
|
||||
Continue execution below
|
||||
<check name="interactive_workflow_exemption">
|
||||
CHECK if this conversation has a <command-name> tag (e.g., <command-name>/bmad_bmm_batch-super-dev</command-name>)
|
||||
|
||||
ELSE IF this workflow was invoked by ANOTHER WORKFLOW (internal call):
|
||||
IF <command-name> tag exists:
|
||||
✅ This is a USER-INITIATED slash command
|
||||
✅ EXEMPTED from Task agent delegation
|
||||
✅ Continue execution in main context
|
||||
|
||||
Reason: Interactive workflows need user prompts and responses.
|
||||
Task agents cannot interact with users.
|
||||
|
||||
Proceed to Step 1 below.
|
||||
</check>
|
||||
|
||||
<check name="internal_workflow_delegation" critical="true">
|
||||
IF NO <command-name> tag found:
|
||||
❌ This is an INTERNAL workflow call (invoked by another workflow)
|
||||
❌ FORBIDDEN in main context
|
||||
MUST delegate to workflow-executor Task agent
|
||||
✅ MUST delegate to workflow-executor Task agent
|
||||
|
||||
Display to user:
|
||||
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
|
||||
🔄 Delegating workflow to fresh context agent...
|
||||
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
|
||||
Why: Internal calls should use fresh context to prevent degradation.
|
||||
|
||||
Workflow: {workflow_name}
|
||||
Reason: Internal workflow calls use Task agents (fresh context)
|
||||
Action:
|
||||
1. Display brief message to user:
|
||||
"🔄 Delegating {workflow_name} to fresh context agent..."
|
||||
|
||||
Spawning workflow-executor agent...
|
||||
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
|
||||
|
||||
Then spawn:
|
||||
2. Spawn Task agent:
|
||||
<Task>
|
||||
subagent_type: general-purpose
|
||||
description: Execute {workflow_name}
|
||||
prompt: |
|
||||
You are the workflow-executor agent.
|
||||
|
||||
Execute workflow: {workflow_path}
|
||||
|
||||
LOAD ALL CONTEXT FIRST:
|
||||
1. Read: {workflow_path}/workflow.yaml
|
||||
2. Read: ALL files in {workflow_path}/steps/
|
||||
3. Read: _bmad/core/tasks/workflow.xml
|
||||
4. Read: Story file (if applicable)
|
||||
5. Read: Project context
|
||||
|
||||
THEN execute following workflow.xml rules EXACTLY.
|
||||
|
||||
Report when complete.
|
||||
Load all required files first, then execute following workflow.xml.
|
||||
</Task>
|
||||
|
||||
HALT - Let Task agent handle workflow.
|
||||
</check>
|
||||
3. HALT - Task agent handles workflow
|
||||
|
||||
<detection>
|
||||
How to detect invocation source:
|
||||
- Slash command: User message contains "/bmad_" or conversation has <command-name> tag
|
||||
- Internal call: No slash command, invoked from within another workflow step
|
||||
</detection>
|
||||
</enforcement>
|
||||
DO NOT proceed to Step 1.
|
||||
</check>
|
||||
</llm>
|
||||
|
||||
<flow>
|
||||
|
|
|
|||
|
|
@ -0,0 +1,256 @@
|
|||
# Batch-Super-Dev Resilience Fix
|
||||
|
||||
**Problem:** Agents crash mid-execution, resume fails, no intermediate state saved
|
||||
|
||||
---
|
||||
|
||||
## Issues Observed
|
||||
|
||||
**Story 18-4 → 18-5 Transition:**
|
||||
```
|
||||
✅ Story 18-4: Builder → Inspector → Fixer → Reviewer all complete
|
||||
❌ Story 18-5: Workflow crashed on "Error reading file"
|
||||
```
|
||||
|
||||
**Evidence:**
|
||||
- Task output files empty (0 bytes)
|
||||
- Resume attempts failed (0 tools used, 0 tokens)
|
||||
- No state saved between stories
|
||||
- When agent crashes, all progress lost
|
||||
|
||||
---
|
||||
|
||||
## Root Cause
|
||||
|
||||
**Sequential processing in main context has no resilience:**
|
||||
|
||||
```
|
||||
Story 18-4:
|
||||
├─ Builder agent completes → outputs to temp file
|
||||
├─ Main reads output file → starts Inspector
|
||||
├─ Inspector completes → outputs to temp file
|
||||
├─ Main reads output → starts Fixer
|
||||
└─ Fixer completes → Story 18-4 done
|
||||
|
||||
Story 18-5:
|
||||
├─ Main tries to read Story 18-5 file
|
||||
├─ ❌ "Error reading file" (crash)
|
||||
└─ All progress lost, no state saved
|
||||
```
|
||||
|
||||
**Problem:** Main context doesn't save state between stories. If it crashes, batch starts over.
|
||||
|
||||
---
|
||||
|
||||
## Solution: Save State After Each Story
|
||||
|
||||
### Add state file tracking:
|
||||
|
||||
```yaml
|
||||
# In batch-super-dev/workflow.yaml
|
||||
state_tracking:
|
||||
enabled: true
|
||||
state_file: "{sprint_artifacts}/batch-execution-state-{batch_id}.yaml"
|
||||
save_after_each_story: true
|
||||
```
|
||||
|
||||
### State file format:
|
||||
|
||||
```yaml
|
||||
batch_id: "epic-18-2026-01-26"
|
||||
started: "2026-01-26T18:45:00Z"
|
||||
execution_mode: "fully_autonomous"
|
||||
strategy: "sequential"
|
||||
total_stories: 2
|
||||
|
||||
stories:
|
||||
- story_key: "18-4-billing-worker-retry-logic"
|
||||
status: "completed"
|
||||
started: "2026-01-26T18:46:00Z"
|
||||
completed: "2026-01-26T19:05:00Z"
|
||||
agents:
|
||||
- phase: "builder"
|
||||
agent_id: "ae3bd2b"
|
||||
status: "completed"
|
||||
- phase: "inspector"
|
||||
agent_id: "a9f0d11"
|
||||
status: "completed"
|
||||
- phase: "fixer"
|
||||
agent_id: "abc123"
|
||||
status: "completed"
|
||||
- phase: "reviewer"
|
||||
agent_id: "def456"
|
||||
status: "completed"
|
||||
|
||||
- story_key: "18-5-precharge-payment-validation"
|
||||
status: "in_progress"
|
||||
started: "2026-01-26T19:05:30Z"
|
||||
last_checkpoint: "attempting_to_read_story_file"
|
||||
error: "Error reading file"
|
||||
```
|
||||
|
||||
### Resume logic:
|
||||
|
||||
```bash
|
||||
# At batch-super-dev start, check for existing state file
|
||||
state_file="{sprint_artifacts}/batch-execution-state-*.yaml"
|
||||
|
||||
if ls $state_file 2>/dev/null; then
|
||||
echo "🔄 Found interrupted batch execution"
|
||||
echo "Resume from where it left off? (yes/no)"
|
||||
|
||||
if yes:
|
||||
# Load state file
|
||||
# Skip completed stories
|
||||
# Start from next story
|
||||
# Reuse agent IDs if resumable
|
||||
fi
|
||||
```
|
||||
|
||||
### After each story completes:
|
||||
|
||||
```bash
|
||||
# Update state file
|
||||
update_state_file() {
|
||||
story_key="$1"
|
||||
status="$2" # completed | failed
|
||||
|
||||
# Update YAML
|
||||
# Mark story as completed
|
||||
# Save timestamp
|
||||
# Record agent IDs
|
||||
}
|
||||
|
||||
# After Builder completes
|
||||
update_state_file "$story_key" "builder_complete"
|
||||
|
||||
# After Inspector completes
|
||||
update_state_file "$story_key" "inspector_complete"
|
||||
|
||||
# After Fixer completes
|
||||
update_state_file "$story_key" "fixer_complete"
|
||||
|
||||
# After Reviewer completes
|
||||
update_state_file "$story_key" "reviewer_complete"
|
||||
|
||||
# When entire story done
|
||||
update_state_file "$story_key" "completed"
|
||||
```
|
||||
|
||||
### Error handling:
|
||||
|
||||
```bash
|
||||
# Wrap file reads in try-catch
|
||||
read_with_retry() {
|
||||
file_path="$1"
|
||||
max_attempts=3
|
||||
|
||||
for attempt in {1..$max_attempts}; do
|
||||
if content=$(cat "$file_path" 2>&1); then
|
||||
echo "$content"
|
||||
return 0
|
||||
else
|
||||
echo "⚠️ Failed to read $file_path (attempt $attempt/$max_attempts)" >&2
|
||||
sleep 2
|
||||
fi
|
||||
done
|
||||
|
||||
echo "❌ Cannot read file after $max_attempts attempts: $file_path" >&2
|
||||
return 1
|
||||
}
|
||||
|
||||
# Use in workflow
|
||||
story_content=$(read_with_retry "$story_file") || {
|
||||
echo "❌ Cannot proceed with Story $story_key - file read failed"
|
||||
# Save state
|
||||
# Skip this story
|
||||
# Continue to next story (if continue_on_failure=true)
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Implementation
|
||||
|
||||
Add to batch-super-dev Step 4-Sequential:
|
||||
|
||||
```xml
|
||||
<substep n="4s-0" title="Check for previous execution state">
|
||||
<action>Check for state file: batch-execution-state-*.yaml</action>
|
||||
|
||||
<check if="state file exists">
|
||||
<output>🔄 Found interrupted batch from {state.started}</output>
|
||||
<output>Completed: {state.completed_count} stories</output>
|
||||
<output>Failed: {state.failed_count} stories</output>
|
||||
<output>In progress: {state.current_story}</output>
|
||||
|
||||
<ask>Resume from where it left off? (yes/no)</ask>
|
||||
|
||||
<check if="response == yes">
|
||||
<action>Load state</action>
|
||||
<action>Skip completed stories</action>
|
||||
<action>Start from next story</action>
|
||||
</check>
|
||||
|
||||
<check if="response == no">
|
||||
<action>Archive old state file</action>
|
||||
<action>Start fresh batch</action>
|
||||
</check>
|
||||
</check>
|
||||
</substep>
|
||||
|
||||
<substep n="4s-a" title="Process individual story">
|
||||
<action>Save state: story started</action>
|
||||
|
||||
<try>
|
||||
<action>Read story file with retry</action>
|
||||
<action>Execute super-dev-pipeline</action>
|
||||
<action>Save state: story completed</action>
|
||||
</try>
|
||||
|
||||
<catch error="file_read_error">
|
||||
<output>⚠️ Cannot read story file for {story_key}</output>
|
||||
<action>Save state: story failed (file read error)</action>
|
||||
<action>Add to failed_stories list</action>
|
||||
<action>Continue to next story if continue_on_failure=true</action>
|
||||
</catch>
|
||||
|
||||
<catch error="agent_crash">
|
||||
<output>⚠️ Agent crashed for {story_key}</output>
|
||||
<action>Save state: story failed (agent crash)</action>
|
||||
<action>Record partial progress in state file</action>
|
||||
<action>Continue to next story if continue_on_failure=true</action>
|
||||
</catch>
|
||||
</substep>
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Expected Behavior After Fix
|
||||
|
||||
**If crash happens:**
|
||||
|
||||
```
|
||||
Story 18-4: ✅ Complete (state saved)
|
||||
Story 18-5: ❌ Crashed (state saved with error)
|
||||
|
||||
State file created: batch-execution-state-epic-18.yaml
|
||||
|
||||
User re-runs: /batch-super-dev
|
||||
|
||||
Workflow: "🔄 Found interrupted batch. Resume? (yes/no)"
|
||||
User: "yes"
|
||||
Workflow: "✅ Skipping 18-4 (already complete)"
|
||||
Workflow: "🔄 Retrying 18-5 (was in_progress)"
|
||||
Workflow: Starts 18-5 from beginning
|
||||
```
|
||||
|
||||
**Benefits:**
|
||||
- No lost progress
|
||||
- Can resume after crashes
|
||||
- Intermediate state preserved
|
||||
- Failures don't block batch
|
||||
|
||||
---
|
||||
|
||||
Should I implement this resilience fix now?
|
||||
Loading…
Reference in New Issue