feat(10-02): implement parallel spawning and monitoring

- spawn_parallel_agents: Task tool with run_in_background, XML agent prompt,
  agent-history.json tracking with parallel_group, wave queuing
- monitor_parallel_completion: TaskOutput polling loop, completion report parsing,
  check_and_spawn_dependents, failure handling, progress display
- orchestrator_commit: merge conflict detection failsafe, per-plan staging,
  conflict resolution options
- Total: 956 lines (within 800-1200 target)
This commit is contained in:
Lex Christopherson
2026-01-12 13:07:10 -06:00
parent caf28103e6
commit 511def76a3

View File

@@ -358,7 +358,7 @@ echo "All agents start from commit: $PRE_SPAWN_COMMIT"
**2. Generate parallel group ID:**
```bash
PARALLEL_GROUP="pg-$(date +%Y%m%d%H%M%S)-$(uuidgen | cut -d- -f1)"
PARALLEL_GROUP="pg-$(date +%Y%m%d%H%M%S)-$(openssl rand -hex 4)"
```
**3. Initialize tracking:**
@@ -367,117 +367,340 @@ PARALLEL_GROUP="pg-$(date +%Y%m%d%H%M%S)-$(uuidgen | cut -d- -f1)"
if [ ! -f .planning/agent-history.json ]; then
echo '{"version":"1.2","max_entries":50,"entries":[]}' > .planning/agent-history.json
fi
# Initialize tracking arrays
declare -a RUNNING_AGENTS=()
declare -a QUEUED_PLANS=()
declare -A AGENT_TO_PLAN=()
```
**4. For each wave, spawn independent plans:**
**4. Spawn Wave 1 plans (no dependencies):**
For Wave 1 (no dependencies):
```
For each independent plan in Wave 1:
1. Check max_concurrent_agents limit
2. Spawn via Task tool with run_in_background: true
3. Record in agent-history.json
4. Track spawned agent IDs
For each plan in Wave 1:
# Check concurrent agent limit
if len(RUNNING_AGENTS) >= max_concurrent_agents:
QUEUED_PLANS.append(plan)
continue
# Use Task tool to spawn background agent
Task(
description="Execute {plan_id} (parallel)",
prompt="[Agent prompt below]",
subagent_type="general-purpose",
run_in_background=true
)
# After Task returns, capture agent_id
RUNNING_AGENTS.append(agent_id)
AGENT_TO_PLAN[agent_id] = plan_id
# Record to agent-history.json
add_entry_to_history(...)
```
**Agent spawn prompt (for plans WITHOUT checkpoints):**
```
"You are running as a PARALLEL agent executing plan: {plan_path}
**CRITICAL INSTRUCTIONS:**
1. Execute ALL tasks in the plan
2. DO NOT commit any changes - the orchestrator will handle commits
3. Track all files you create or modify
4. Follow all deviation rules from the plan
5. Create SUMMARY.md when complete (do not commit it)
```xml
<parallel_agent_instructions>
You are executing plan: {plan_path} as part of a PARALLEL phase execution.
**When done, report:**
- Plan name
- Tasks completed (count and names)
- Files created/modified (full paths)
- Deviations encountered
- SUMMARY.md path
- Any issues or blockers
<critical_rules>
1. Execute ALL tasks in the plan following deviation rules
2. DO NOT run git commit - orchestrator handles commits
3. DO NOT run git add - orchestrator stages files
4. Track all files you create or modify
5. Create SUMMARY.md in the phase directory when complete
</critical_rules>
**Do NOT:**
- Run git commit
- Run git add (except to check status)
- Push to remote
- Modify files outside plan scope"
<plan_context>
@{plan_path}
Read the plan for full context, tasks, and deviation rules.
</plan_context>
<execution_protocol>
1. Read plan file and context files
2. Execute each task in order
3. For each task:
- Implement the action
- Run verification
- Track files modified
- Track any deviations
4. After all tasks: create SUMMARY.md
</execution_protocol>
<report_format>
When complete, output this exact format:
PARALLEL_AGENT_COMPLETE
plan_id: {phase}-{plan}
tasks_completed: [count]/[total]
files_modified:
- path/to/file1.ts
- path/to/file2.md
deviations:
- [Rule X] description
summary_path: .planning/phases/{phase-dir}/{phase}-{plan}-SUMMARY.md
issues: [none or list]
END_REPORT
</report_format>
<forbidden_actions>
- git commit
- git add
- git push
- Modifying files outside plan scope
- Running long-blocking network operations
</forbidden_actions>
</parallel_agent_instructions>
```
**5. Record spawn in agent-history.json:**
```json
```bash
# Read current entries
ENTRIES=$(jq '.entries' .planning/agent-history.json)
# Create new entry
NEW_ENTRY=$(cat <<EOF
{
"agent_id": "[from Task response]",
"task_description": "Parallel: Execute {phase}-{plan}-PLAN.md",
"phase": "{phase}",
"plan": "{plan}",
"parallel_group": "{PARALLEL_GROUP}",
"agent_id": "$AGENT_ID",
"task_description": "Parallel: Execute ${PHASE}-${PLAN}-PLAN.md",
"phase": "$PHASE",
"plan": "$PLAN",
"parallel_group": "$PARALLEL_GROUP",
"granularity": "plan",
"wave": 1,
"timestamp": "[ISO]",
"wave": $WAVE_NUM,
"depends_on": $(echo "${PLAN_REQUIRES[$PLAN]}" | jq -R 'split(",") | map(select(. != ""))'),
"timestamp": "$(date -u +"%Y-%m-%dT%H:%M:%SZ")",
"status": "spawned",
"files_modified": [],
"completion_timestamp": null
"completion_timestamp": null,
"deviations": []
}
EOF
)
# Append and write back
echo "$ENTRIES" | jq ". += [$NEW_ENTRY]" > /tmp/entries.json
jq --argjson entries "$(cat /tmp/entries.json)" '.entries = $entries' .planning/agent-history.json > /tmp/history.json
mv /tmp/history.json .planning/agent-history.json
```
**6. Queue dependent plans:**
Plans in Wave 2+ are queued, not spawned yet.
They spawn after their dependencies complete.
**6. Queue remaining plans:**
```bash
# Queue Wave 2+ plans
for wave in $(seq 2 $MAX_WAVE); do
for plan in ${WAVES[$wave]}; do
QUEUED_PLANS+=("$plan:$wave")
done
done
echo "Spawned: ${#RUNNING_AGENTS[@]} agents"
echo "Queued: ${#QUEUED_PLANS[@]} plans"
```
</step>
<step name="monitor_parallel_completion">
**Poll for agent completion and spawn dependents.**
**1. Polling loop:**
```
running_agents = [list of spawned agent IDs]
**1. Polling loop implementation:**
while running_agents.length > 0:
for agent_id in running_agents:
result = TaskOutput(task_id=agent_id, block=false, timeout=5000)
```
declare -A COMPLETED_AGENTS=()
declare -A FAILED_AGENTS=()
while [ ${#RUNNING_AGENTS[@]} -gt 0 ] || [ ${#QUEUED_PLANS[@]} -gt 0 ]; do
# Check each running agent
for agent_id in "${RUNNING_AGENTS[@]}"; do
# Use TaskOutput to check status (non-blocking)
TaskOutput(
task_id=agent_id,
block=false,
timeout=5000
)
if result.status == "completed":
# Capture results
files_modified = parse_files_from_result(result)
# Parse agent's completion report
files_modified = parse_report_files(result.output)
deviations = parse_report_deviations(result.output)
plan_id = AGENT_TO_PLAN[agent_id]
# Update agent-history.json
update_entry(agent_id, status="completed", files_modified=files_modified)
update_history_entry(
agent_id,
status="completed",
files_modified=files_modified,
deviations=deviations,
completion_timestamp=now()
)
# Remove from running list
running_agents.remove(agent_id)
# Track completion
COMPLETED_AGENTS[agent_id] = plan_id
RUNNING_AGENTS.remove(agent_id)
# Check if dependents can now start
spawn_ready_dependents()
echo "✓ Agent $agent_id completed plan $plan_id"
echo " Files: ${#files_modified[@]}"
echo " Deviations: ${#deviations[@]}"
# Check if dependents can now spawn
check_and_spawn_dependents()
elif result.status == "failed":
log_failure(agent_id, result.error)
running_agents.remove(agent_id)
# Continue with other agents - don't kill batch
plan_id = AGENT_TO_PLAN[agent_id]
sleep(10) # Poll every 10 seconds
# Log failure
update_history_entry(
agent_id,
status="failed",
error=result.error,
completion_timestamp=now()
)
FAILED_AGENTS[agent_id] = plan_id
RUNNING_AGENTS.remove(agent_id)
echo "✗ Agent $agent_id FAILED on plan $plan_id"
echo " Error: ${result.error}"
# Continue monitoring - don't abort batch
# else: still running, check next agent
done
# Brief pause between polls
sleep 10
done
```
**2. Spawn ready dependents:**
```
For each queued plan in Wave 2+:
if all dependencies completed:
if under max_concurrent_agents limit:
spawn_agent(plan)
move from queue to running_agents
**2. Parse agent completion report:**
```bash
parse_report_files() {
local output="$1"
# Extract files from PARALLEL_AGENT_COMPLETE block
echo "$output" | \
sed -n '/^files_modified:/,/^[a-z_]*:/p' | \
grep '^\s*-' | \
sed 's/^\s*-\s*//'
}
parse_report_deviations() {
local output="$1"
echo "$output" | \
sed -n '/^deviations:/,/^[a-z_]*:/p' | \
grep '^\s*-' | \
sed 's/^\s*-\s*//'
}
```
**3. Handle failures:**
- Log failed agent to agent-history.json with status="failed"
- Continue monitoring other agents
- Report failures in final summary
- Do NOT automatically retry (user decision)
**3. Spawn ready dependents:**
**4. Completion:**
When all agents complete (running_agents empty and queue empty):
- Proceed to orchestrator_commit step
```bash
check_and_spawn_dependents() {
# Get completed plan IDs
local completed_plans=$(printf '%s\n' "${COMPLETED_AGENTS[@]}")
for i in "${!QUEUED_PLANS[@]}"; do
local queued="${QUEUED_PLANS[$i]}"
local plan_id="${queued%%:*}"
local wave="${queued##*:}"
# Get this plan's dependencies
local deps="${PLAN_REQUIRES[$plan_id]}"
# Check if all dependencies are in completed list
local all_deps_met=true
IFS=',' read -ra dep_array <<< "$deps"
for dep in "${dep_array[@]}"; do
[ -z "$dep" ] && continue
if ! echo "$completed_plans" | grep -q "^$dep$"; then
all_deps_met=false
break
fi
done
if [ "$all_deps_met" = true ]; then
# Check concurrent limit
if [ ${#RUNNING_AGENTS[@]} -lt $MAX_CONCURRENT ]; then
# Remove from queue
unset 'QUEUED_PLANS[$i]'
# Spawn agent
spawn_plan_agent "$plan_id" "$wave"
echo "→ Spawned dependent: $plan_id (wave $wave)"
fi
fi
done
# Rebuild array to remove gaps
QUEUED_PLANS=("${QUEUED_PLANS[@]}")
}
```
**4. Handle failures:**
| Failure Type | Action |
|--------------|--------|
| Agent crash | Log status="failed", continue batch |
| Plan error | Same as crash - logged, batch continues |
| All dependents | Plans depending on failed agent also fail |
```bash
# When agent fails, mark its dependents as blocked
mark_dependents_blocked() {
local failed_plan="$1"
for i in "${!QUEUED_PLANS[@]}"; do
local queued="${QUEUED_PLANS[$i]}"
local plan_id="${queued%%:*}"
local deps="${PLAN_REQUIRES[$plan_id]}"
if echo "$deps" | grep -q "$failed_plan"; then
echo "⚠ Plan $plan_id blocked (depends on failed $failed_plan)"
# Mark in history as blocked
update_history_entry("queued-$plan_id", status="blocked", blocked_by="$failed_plan")
fi
done
}
```
**5. Completion conditions:**
```
while true:
if RUNNING_AGENTS is empty AND QUEUED_PLANS is empty:
break # All done
if RUNNING_AGENTS is empty AND QUEUED_PLANS is not empty:
# All queued plans are blocked by failed dependencies
echo "All remaining plans blocked by failures"
break
poll_and_check()
```
**6. Progress display:**
```
During execution, show:
═══════════════════════════════════════════════════
Parallel Execution Status
═══════════════════════════════════════════════════
Running: [agent-1: 10-01] [agent-2: 10-04]
Queued: 10-02, 10-03
Complete: 0
Failed: 0
═══════════════════════════════════════════════════
[Update periodically as agents complete]
```
</step>
<step name="orchestrator_commit">
@@ -496,18 +719,73 @@ for entry in $(jq -r ".entries[] | select(.parallel_group==\"$PARALLEL_GROUP\")
done
```
**2. Check for merge conflicts:**
```bash
# Files modified by multiple agents = potential conflict
CONFLICTS=$(echo "${ALL_FILES[@]}" | tr ' ' '\n' | sort | uniq -d)
**2. Check for merge conflicts (failsafe):**
if [ -n "$CONFLICTS" ]; then
echo "WARNING: Multiple agents modified same files:"
echo "$CONFLICTS"
echo "Manual conflict resolution may be needed."
```bash
# Build file-to-agent mapping
declare -A FILE_AGENTS
for entry in $(jq -c ".entries[] | select(.parallel_group==\"$PARALLEL_GROUP\" and .status==\"completed\")" .planning/agent-history.json); do
agent_id=$(echo "$entry" | jq -r '.agent_id')
for file in $(echo "$entry" | jq -r '.files_modified[]'); do
FILE_AGENTS["$file"]="${FILE_AGENTS[$file]} $agent_id"
done
done
# Detect conflicts (file modified by multiple agents)
CONFLICTS=()
for file in "${!FILE_AGENTS[@]}"; do
agents=(${FILE_AGENTS[$file]})
if [ ${#agents[@]} -gt 1 ]; then
CONFLICTS+=("$file: ${agents[*]}")
fi
done
# Handle conflicts
if [ ${#CONFLICTS[@]} -gt 0 ]; then
echo ""
echo "═══════════════════════════════════════════════════"
echo "⚠ MERGE CONFLICT DETECTED"
echo "═══════════════════════════════════════════════════"
echo ""
echo "Multiple agents modified the same files:"
for conflict in "${CONFLICTS[@]}"; do
echo " - $conflict"
done
echo ""
echo "Options:"
echo "1. Review and merge manually"
echo "2. Re-run sequential with /gsd:execute-plan"
echo "3. Accept last-write-wins (risky)"
echo ""
# Present to user
AskUserQuestion(
header="Merge Conflict",
question="How to resolve file conflicts?",
options=[
"Manual review",
"Abort and retry sequential",
"Accept last-write-wins"
]
)
# Handle based on response
if response == "Abort":
git checkout -- . # Discard changes
exit 1
elif response == "Manual review":
echo "Review files and run: git add <resolved-files> && /gsd:execute-phase --resume"
exit 0
fi
```
**Conflict prevention (dependency analysis should catch this):**
The analyze_plan_dependencies step should detect file conflicts and add dependencies.
This step is a failsafe for edge cases where:
- Agents create new files with same name
- File patterns weren't caught during analysis
- Context files are modified unexpectedly
**3. Stage and commit per-plan:**
```bash
# For each completed agent (in execution order):