t1118: Add AI self-reflection context — supervisor can now diagnose its own failures

.agents/scripts/supervisor/ai-context.sh

@@ -67,6 +67,11 @@ build_ai_context() {
 		context+="$(build_audit_context "$repo_path")\n\n"
 	fi
 
+	# Section 10: AI Self-Reflection (own execution history, t1118)
+	if [[ "$scope" == "full" ]]; then
+		context+="$(build_self_reflection_context)\n\n"
+	fi
+
 	printf '%b' "$context"
 	return 0
 }
@@ -658,6 +663,198 @@ build_audit_context() {
 	return 0
 }
 
+#######################################
+# Section 10: AI Self-Reflection (t1118)
+# Feeds the AI its own recent action execution results so it can
+# identify recurring failures, prompt issues, and deployment gaps.
+# Sources: action log files + pipeline log
+#######################################
+build_self_reflection_context() {
+	local output="## AI Supervisor Self-Reflection\n\n"
+	local log_dir="${AI_REASON_LOG_DIR:-$HOME/.aidevops/logs/ai-supervisor}"
+
+	if [[ ! -d "$log_dir" ]]; then
+		output+="*No AI supervisor logs found*\n"
+		printf '%b' "$output"
+		return 0
+	fi
+
+	# Collect recent action logs (last 5 cycles)
+	local action_logs
+	action_logs=$(find "$log_dir" -maxdepth 1 -name 'actions-*.md' -print0 2>/dev/null | xargs -0 ls -t 2>/dev/null | head -5)
+
+	if [[ -z "$action_logs" ]]; then
+		output+="*No action execution history yet*\n"
+		printf '%b' "$output"
+		return 0
+	fi
+
+	# Aggregate execution stats across recent cycles
+	local total_executed=0 total_failed=0 total_skipped=0 cycle_count=0
+	local skip_reasons=""
+	local failed_actions=""
+
+	while IFS= read -r log_file; do
+		cycle_count=$((cycle_count + 1))
+
+		# Count outcomes from action log headers
+		local executed failed skipped
+		executed=$(grep -c 'SUCCESS' "$log_file" 2>/dev/null || echo 0)
+		failed=$(grep -c 'FAILED' "$log_file" 2>/dev/null || echo 0)
+		skipped=$(grep -c 'SKIPPED' "$log_file" 2>/dev/null || echo 0)
+		# Ensure clean integers (trim whitespace/newlines)
+		executed="${executed//[^0-9]/}"
+		failed="${failed//[^0-9]/}"
+		skipped="${skipped//[^0-9]/}"
+		: "${executed:=0}" "${failed:=0}" "${skipped:=0}"
+
+		total_executed=$((total_executed + executed))
+		total_failed=$((total_failed + failed))
+		total_skipped=$((total_skipped + skipped))
+
+		# Collect skip reasons (these reveal prompt/code issues)
+		local skips
+		skips=$(grep -E 'SKIPPED' "$log_file" 2>/dev/null || true)
+		if [[ -n "$skips" ]]; then
+			while IFS= read -r skip_line; do
+				# Extract action type and reason from "## Action N: type — SKIPPED (reason)"
+				local skip_type skip_reason
+				skip_type=$(printf '%s' "$skip_line" | sed -E 's/.*: ([a-z_]+) — SKIPPED.*/\1/' 2>/dev/null || echo "unknown")
+				skip_reason=$(printf '%s' "$skip_line" | sed -E 's/.*SKIPPED \(([^)]+)\).*/\1/' 2>/dev/null || echo "unknown")
+				skip_reasons+="$skip_type: $skip_reason\n"
+			done <<<"$skips"
+		fi
+
+		# Collect failed actions
+		local fails
+		fails=$(grep -E 'FAILED' "$log_file" 2>/dev/null || true)
+		if [[ -n "$fails" ]]; then
+			while IFS= read -r fail_line; do
+				local fail_type
+				fail_type=$(printf '%s' "$fail_line" | sed -E 's/.*: ([a-z_]+) — FAILED.*/\1/' 2>/dev/null || echo "unknown")
+				failed_actions+="$fail_type\n"
+			done <<<"$fails"
+		fi
+	done <<<"$action_logs"
+
+	output+="### Execution Summary (last $cycle_count cycles)\n\n"
+	output+="| Metric | Count |\n|--------|-------|\n"
+	output+="| Executed | $total_executed |\n"
+	output+="| Failed | $total_failed |\n"
+	output+="| Skipped | $total_skipped |\n"
+
+	local total_actions=$((total_executed + total_failed + total_skipped))
+	if [[ "$total_actions" -gt 0 ]]; then
+		local exec_rate=$((total_executed * 100 / total_actions))
+		local skip_rate=$((total_skipped * 100 / total_actions))
+		output+="| Execution rate | ${exec_rate}% |\n"
+		output+="| Skip rate | ${skip_rate}% |\n"
+	fi
+	output+="\n"
+
+	# Deduplicate and count skip reasons
+	if [[ -n "$skip_reasons" ]]; then
+		output+="### Recurring Skip Reasons\n\n"
+		output+="These actions were proposed by the AI but rejected by the executor.\n"
+		output+="**Fix these by adjusting your output format or the executor code.**\n\n"
+
+		local deduped_skips
+		deduped_skips=$(printf '%b' "$skip_reasons" | sort | uniq -c | sort -rn)
+		output+="| Count | Action Type | Reason |\n|-------|-------------|--------|\n"
+		while IFS= read -r line; do
+			local count type_reason
+			count=$(printf '%s' "$line" | awk '{print $1}')
+			type_reason=$(printf '%s' "$line" | sed 's/^ *[0-9]* *//')
+			local stype sreason
+			stype=$(printf '%s' "$type_reason" | cut -d: -f1)
+			sreason=$(printf '%s' "$type_reason" | cut -d: -f2- | sed 's/^ //')
+			output+="| $count | $stype | $sreason |\n"
+		done <<<"$deduped_skips"
+		output+="\n"
+	fi
+
+	# Deduplicate and count failed actions
+	if [[ -n "$failed_actions" ]]; then
+		output+="### Recurring Failures\n\n"
+		local deduped_fails
+		deduped_fails=$(printf '%b' "$failed_actions" | sort | uniq -c | sort -rn)
+		while IFS= read -r line; do
+			local count ftype
+			count=$(printf '%s' "$line" | awk '{print $1}')
+			ftype=$(printf '%s' "$line" | sed 's/^ *[0-9]* *//')
+			output+="- $ftype: failed $count times\n"
+		done <<<"$deduped_fails"
+		output+="\n"
+	fi
+
+	# Check for repeated identical actions across cycles (dedup detection)
+	output+="### Action Repetition Detection\n\n"
+	local all_actions=""
+	while IFS= read -r log_file; do
+		# Extract issue numbers targeted by actions
+		local issue_actions
+		issue_actions=$(grep -oE '"issue_number":[0-9]+' "$log_file" 2>/dev/null |
+			sort -u || true)
+		if [[ -n "$issue_actions" ]]; then
+			while IFS= read -r ia; do
+				local inum
+				inum=$(printf '%s' "$ia" | grep -oE '[0-9]+')
+				all_actions+="issue:#$inum\n"
+			done <<<"$issue_actions"
+		fi
+		# Extract task IDs targeted by actions
+		local task_actions
+		task_actions=$(grep -oE '"task_id":"t[0-9]+"' "$log_file" 2>/dev/null |
+			sort -u || true)
+		if [[ -n "$task_actions" ]]; then
+			while IFS= read -r ta; do
+				local tid
+				tid=$(printf '%s' "$ta" | grep -oE 't[0-9]+')
+				all_actions+="task:$tid\n"
+			done <<<"$task_actions"
+		fi
+	done <<<"$action_logs"
+
+	if [[ -n "$all_actions" ]]; then
+		local repeated
+		repeated=$(printf '%b' "$all_actions" | sort | uniq -c | sort -rn | awk '$1 > 1')
+		if [[ -n "$repeated" ]]; then
+			output+="The following targets received actions in multiple cycles (possible redundant work):\n\n"
+			while IFS= read -r line; do
+				local count target
+				count=$(printf '%s' "$line" | awk '{print $1}')
+				target=$(printf '%s' "$line" | sed 's/^ *[0-9]* *//')
+				output+="- $target: acted on $count times across $cycle_count cycles\n"
+			done <<<"$repeated"
+			output+="\n"
+			output+="Consider: are these repeated actions necessary, or should you skip targets already addressed in prior cycles?\n"
+		else
+			output+="No repeated targets detected — each action targeted a unique resource.\n"
+		fi
+	else
+		output+="No action targets to analyze.\n"
+	fi
+	output+="\n"
+
+	# Pipeline errors from ai-supervisor.log (last 50 lines)
+	local pipeline_log="${HOME}/.aidevops/.agent-workspace/supervisor/logs/ai-supervisor.log"
+	if [[ -f "$pipeline_log" ]]; then
+		local recent_errors
+		recent_errors=$(tail -100 "$pipeline_log" 2>/dev/null | grep -iE 'error|jq:' | tail -10 || true)
+		if [[ -n "$recent_errors" ]]; then
+			output+="### Pipeline Errors (recent)\n\n"
+			output+="These errors occurred in the action execution pipeline:\n\n"
+			output+="\`\`\`\n"
+			output+="$recent_errors\n"
+			output+="\`\`\`\n\n"
+			output+="If these are recurring, create a \`create_improvement\` task to fix the root cause.\n"
+		fi
+	fi
+
+	printf '%b' "$output"
+	return 0
+}
+
 #######################################
 # CLI entry point for standalone testing
 # Usage: ai-context.sh [--scope full|quick] [--repo /path]

.agents/scripts/supervisor/ai-reason.sh

@@ -358,6 +358,7 @@ For each analysis, consider:
 6. **Health**: Are there concerning patterns in worker outcomes? High failure rates? Recurring errors?
 7. **Efficiency**: Are tokens being wasted? Are tasks assigned to models that are too powerful (opus for simple tasks) or too weak (sonnet for complex reasoning)? Are there repeated failures that indicate a systemic issue rather than a task-specific one? Look at pattern data for model tier success rates.
 8. **Self-improvement**: What automation gaps exist? Are there manual steps that could be automated? Missing test coverage? Processes that break repeatedly? Documentation gaps that cause worker confusion? Create improvement tasks to fix these — the goal is maximum utility from minimal token use.
+9. **Self-reflection**: Review the "AI Supervisor Self-Reflection" section. Are your own actions being skipped or failing? If an action type is repeatedly skipped (e.g., missing required fields), create a `create_improvement` task to fix the prompt or executor. If you keep acting on the same targets across cycles, stop repeating those actions. If pipeline errors appear, diagnose the root cause and create a fix task. Your goal is to make yourself more effective over time.
 
 ## Output Format
 
@@ -397,6 +398,12 @@ Respond with ONLY a JSON array of actions. Each action is an object with:
     "to_tier": "opus",
     "reasoning": "Task failed 2/3 retries at sonnet. Pattern data shows similar tasks succeed at opus."
   },
+  {
+    "type": "adjust_priority",
+    "task_id": "t1234",
+    "new_priority": "high",
+    "reasoning": "This task is blocking 3 others and should be dispatched next"
+  },
   {
     "type": "flag_for_review",
     "issue_number": 456,

MODELS.md

@@ -3,9 +3,9 @@
 Live performance data from pattern-tracker and response-scoring databases.
 Auto-generated by `generate-models-md.sh` — do not edit manually.
 
-**Last updated**: 2026-02-18T14:28:26Z
+**Last updated**: 2026-02-18T15:33:40Z
 
-- **Pattern data points**: 865
+- **Pattern data points**: 876
 - **Scored responses**: 18
 - **Date range**: 2026-02-05 to 2026-02-18
 
@@ -49,17 +49,17 @@ Success rates from autonomous task execution (pattern-tracker data):
 
 | Model | Tasks | Successes | Failures | Success Rate | Last Used |
 |-------|-------|-----------|----------|--------------|-----------|
-| opus | 497 | 492 | 5 | 98% | 2026-02-18 |
-| sonnet | 154 | 154 | 0 | 100% | 2026-02-18 |
+| opus | 502 | 496 | 6 | 98% | 2026-02-18 |
+| sonnet | 160 | 160 | 0 | 100% | 2026-02-18 |
 | pro | 2 | 2 | 0 | 100% | 2026-02-18 |
 | haiku | 1 | 0 | 1 | 0% | 2026-02-05 |
 
 ### By Task Type
 
 | Task Type | Tasks | Successes | Failures | Success Rate |
 |-----------|-------|-----------|----------|--------------|
-| feature | 524 | 508 | 16 | 96% |
-| bugfix | 8 | 5 | 3 | 62% |
+| feature | 533 | 516 | 17 | 96% |
+| bugfix | 10 | 7 | 3 | 70% |
 | refactor | 1 | 0 | 1 | 0% |
 | testing | 1 | 1 | 0 | 100% |
 | security | 1 | 1 | 0 | 100% |