Dev Feedback Loop

Dev Feedback Loop

Overview

See CLAUDE.md for canonical validation and gate requirements. This skill collects metrics and generates improvement reports.

Continuous improvement system that tracks development cycle effectiveness through assertiveness scores, identifies recurring failure patterns, and generates actionable improvement suggestions.

Core principle: What gets measured gets improved. Track every gate transition to identify systemic issues.

Step 0: TodoWrite Tracking (MANDATORY FIRST ACTION)

⛔ HARD GATE: Before ANY other action, you MUST add feedback-loop to todo list.

Execute this TodoWrite call IMMEDIATELY when this skill starts:

TodoWrite tool:
  todos:
    - id: "feedback-loop-execution"
      content: "Execute dev-feedback-loop: collect metrics, calculate scores, write report"
      status: "in_progress"
      priority: "high"

Why this is mandatory:

• TodoWrite creates visible tracking of feedback-loop execution
• Prevents skill from being "forgotten" mid-execution
• Creates audit trail that feedback was collected
• Hook enforcer checks for this todo item

After completing ALL feedback-loop steps, mark as completed:

TodoWrite tool:
  todos:
    - id: "feedback-loop-execution"
      content: "Execute dev-feedback-loop: collect metrics, calculate scores, write report"
      status: "completed"
      priority: "high"

Anti-Rationalization:

You CANNOT proceed to Step 1 without executing TodoWrite above.

Pressure Resistance

See shared-patterns/shared-pressure-resistance.md for universal pressure scenarios.

Feedback-specific note: Feedback MUST be collected for EVERY completed task, regardless of outcome or complexity. "Simple tasks" and "perfect scores" still need tracking.

Common Rationalizations - REJECTED

See shared-patterns/shared-anti-rationalization.md for universal anti-rationalizations.

Feedback-specific rationalizations:

Red Flags - STOP

See shared-patterns/shared-red-flags.md for universal red flags.

If you catch yourself thinking ANY of those patterns, STOP immediately. Collect metrics for EVERY task.

Self-Preservation Bias Prevention

Agents must report accurately, even when scores are low:

Reporting protocol:

1. Calculate score using exact formula (no rounding)
2. Report score regardless of value
3. If score < 70, mandatory root cause analysis
4. Document honestly - "I made this mistake" not "mistake was made"
5. Patterns in low scores = improvement opportunities, not blame

Self-interest check: If you're tempted to adjust a score, ask: "Would I report this score if someone else achieved it?" If yes, report as-is.

Mandatory Feedback Collection

Non-negotiable: Feedback MUST be collected for EVERY completed task, regardless of:

Consequence: Skipping feedback breaks continuous improvement loop and masks systemic issues.

Repeated Feedback Detection

When the same feedback appears multiple times:

Recurring feedback indicates systemic issue not being addressed.

Escalation format:

## RECURRING ISSUE - Escalation Required

**Issue:** [Description]
**Occurrences:** [Count] times across [N] tasks
**Pattern:** [What triggers this issue]
**Previous Responses:** [What was tried]

**Recommendation:** [Systemic fix needed]
**Awaiting:** User decision on root cause resolution

Threshold Alerts - MANDATORY RESPONSE

When thresholds are breached, response is REQUIRED:

You CANNOT proceed past threshold without documented response.

Blocker Criteria - STOP and Report

ALWAYS pause and report blocker for:

Before skipping ANY feedback collection:

1. Check if task is complete (feedback required for ALL completed tasks)
2. Check threshold status (alerts are mandatory)
3. If in doubt → STOP and report blocker

You CANNOT skip feedback collection. Period.

Assertiveness Score Calculation

Base score of 100 points, with deductions for inefficiencies:

Penalty Matrix

Score Calculation Formula

score = 100 - min(30, extra_iterations*10) - review_fail*20 - needs_discussion*10 - min(40, unmet_criteria*10) | User rejected → score = 0

Score Interpretation

Step 1: Collect Cycle Metrics

MANDATORY: Execute this step for ALL tasks, regardless of:

• Score value (even 100%)
• User satisfaction (even immediate approval)
• Outcome quality (even perfect)

Anti-exemption check: If you're thinking "perfect outcome, skip metrics" → STOP. This is Red Flag at line 75 ("Perfect outcome, skip the metrics").

After task completion, gather from agent_outputs in state file:

Structured Data Fields (NEW)

The state file now contains structured error/issue data for direct analysis:

All gates have standards_compliance with:

• total_sections, compliant, not_applicable, non_compliant
• gaps[] - array of non-compliant sections with details

Reading Structured Data

# From state file, extract standards compliance from ALL gates:
all_standards_gaps = [
  ...agent_outputs.implementation.standards_compliance.gaps,
  ...agent_outputs.devops.standards_compliance.gaps,
  ...agent_outputs.sre.standards_compliance.gaps,
  ...agent_outputs.testing.standards_compliance.gaps,
  ...agent_outputs.review.code_reviewer.standards_compliance.gaps,
  ...agent_outputs.review.business_logic_reviewer.standards_compliance.gaps,
  ...agent_outputs.review.security_reviewer.standards_compliance.gaps
]

# Gate-specific errors/issues:
devops_errors = agent_outputs.devops.verification_errors
sre_errors = agent_outputs.sre.validation_errors
test_failures = agent_outputs.testing.failures
uncovered_acs = agent_outputs.testing.uncovered_criteria
review_issues = [
  ...agent_outputs.review.code_reviewer.issues,
  ...agent_outputs.review.business_logic_reviewer.issues,
  ...agent_outputs.review.security_reviewer.issues
]

# Aggregate standards compliance metrics:
total_standards_sections = sum(all_gates.standards_compliance.total_sections)
total_compliant = sum(all_gates.standards_compliance.compliant)
overall_compliance_rate = total_compliant / total_standards_sections * 100

Iteration Penalty Calculation

# Total extra iterations across all gates:
extra_iterations = (
  max(0, implementation.iterations - 1) +
  max(0, devops.iterations - 1) +
  max(0, sre.iterations - 1) +
  max(0, testing.iterations - 1) +
  max(0, review.iterations - 1)
)

Step 2: Calculate Assertiveness Score

Apply formula: Base 100 - deductions (extra iterations, review failures, unmet criteria) = Final Score / 100. Map to rating per interpretation table.

Step 3: Analyze Prompt Quality (Agents Only)

After calculating assertiveness, analyze prompt quality for all agents that executed in the task.

3.1 Load Agent Outputs

Read agent_outputs from state file (docs/dev-cycle/current-cycle.json or docs/dev-refactor/current-cycle.json):

Agents to analyze (if executed, not null):
  - implementation: backend-engineer-golang | backend-engineer-typescript
  - devops: devops-engineer
  - sre: sre
  - testing: qa-analyst
  - review: code-reviewer, business-logic-reviewer, security-reviewer

3.2 Dispatch Prompt Quality Reviewer

Task tool:
  subagent_type: "prompt-quality-reviewer"
  prompt: |
    Analyze prompt quality for agents in task [task_id].

    Agent outputs from state:
    [agent_outputs]

    For each agent:
    1. Load definition from dev-team/agents/ or default/agents/
    2. Extract rules: MUST, MUST NOT, ask_when, output_schema
    3. Compare output vs rules
    4. Calculate score
    5. Identify gaps with evidence
    6. Generate improvements

    Return structured analysis per agent.

3.3 Write Feedback Files

Directory: docs/feedbacks/cycle-YYYY-MM-DD/

One file per agent, accumulating all tasks that used that agent.

File: docs/feedbacks/cycle-YYYY-MM-DD/{agent-name}.md

# Prompt Feedback: {agent-name}

**Cycle:** YYYY-MM-DD
**Total Executions:** N
**Average Score:** XX%

---

## Task T-001 (Gate X)

**Score:** XX/100
**Rating:** {rating}

### Gaps Found

| Category | Rule | Evidence | Impact |
|----------|------|----------|--------|
| MUST | [rule text] | [quote from output] | -X |

### What Went Well

- [positive observation]

---

## Task T-002 (Gate X)

**Score:** XX/100
...

---

## Consolidated Improvements

### Priority 1: [Title]

**Occurrences:** X/Y tasks
**Impact:** +X points expected
**File:** dev-team/agents/{agent}.md

**Current text (line ~N):**

[existing prompt]

**Suggested addition:**
```markdown
[new prompt text]

Priority 2: [Title]

...

### 3.4 Append to Existing File

If file already exists (from previous task in same cycle), **append** the new task section before "## Consolidated Improvements" and update:
- Total Executions count
- Average Score
- Consolidated Improvements (re-analyze patterns)

## Step 3.5: Pattern Analysis from Structured Data (NEW)

**Analyze structured error/issue data to identify recurring patterns:**

### 3.5.1 Standards Compliance Patterns

```yaml
# Group gaps by section name
standards_gaps_by_section = group(all_implementation_gaps, by: "section")

# Identify recurring gaps (same section fails across tasks)
recurring_standards_gaps = filter(standards_gaps_by_section, count >= 2)

# Output pattern
For each recurring gap:
  - Section: [section name]
  - Occurrences: [N] tasks
  - Common reason: [most frequent reason]
  - Recommendation: [agent prompt improvement]

3.5.2 Review Issue Patterns

# Group review issues by category
issues_by_category = group(all_review_issues, by: "category")

# Group by severity for prioritization
issues_by_severity = group(all_review_issues, by: "severity")

# Identify top recurring categories
top_categories = sort(issues_by_category, by: count, descending).take(5)

# Output pattern
For each top category:
  - Category: [category name]
  - Occurrences: [N] issues across [M] tasks
  - Severity breakdown: [CRITICAL: X, HIGH: Y, MEDIUM: Z]
  - Most common: [most frequent description pattern]
  - Fix rate: [fixed_count / total_count]%

3.5.3 Test Failure Patterns

# Group failures by error_type
failures_by_type = group(all_test_failures, by: "error_type")

# Identify flaky tests (same test fails multiple times)
flaky_tests = filter(all_test_failures, count_by_test_name >= 2)

# Output pattern
For each error type:
  - Type: [assertion|panic|timeout|compilation]
  - Occurrences: [N] failures
  - Fix iterations: avg [X] iterations to fix

3.5.4 Cross-Gate Pattern Correlation

# Correlate: Do standards gaps predict review issues?
correlation_standards_review = correlate(
  implementation.standards_compliance.gaps[].section,
  review.*.issues[].category
)

# Output insight
If correlation > 0.5:
  "Standards gap in [section] correlates with review issues in [category]"
  → Recommendation: Strengthen agent prompt for [section]

Step 4: Threshold Alerts

Report formats: RCA = Score → Failure Events → 5 Whys → Root cause → Corrective action | Gate Blocked = History → Issue → BLOCKED UNTIL human decision | Deep Analysis = Distribution → Patterns → Improvement Plan

Step 5: Write Feedback Report

Location: .ring/dev-team/feedback/cycle-YYYY-MM-DD.md

Required sections:

Step 6: Generate Improvement Suggestions

Improvement types based on pattern analysis:

Step 7: Complete TodoWrite Tracking (MANDATORY FINAL ACTION)

⛔ HARD GATE: After ALL steps complete, you MUST mark feedback-loop todo as completed.

Execute this TodoWrite call to finalize:

TodoWrite tool:
  todos:
    - id: "feedback-loop-execution"
      content: "Execute dev-feedback-loop: collect metrics, calculate scores, write report"
      status: "completed"
      priority: "high"

Verification before marking complete:

• Step 1: Metrics collected from state file
• Step 2: Assertiveness score calculated
• Step 3: Prompt quality analyzed (if agents executed)
• Step 4: Threshold alerts checked
• Step 5: Feedback report written to .ring/dev-team/feedback/
• Step 6: Improvement suggestions generated

You CANNOT mark todo as completed until ALL steps above are done.

Execution Report

Base metrics per shared-patterns/output-execution-report.md.

Anti-Patterns

Never:

• Skip feedback collection for "simple" tasks
• Ignore threshold alerts
• Accept low scores without analysis
• Generate suggestions without data
• Blame individuals instead of process

Always:

• Track every gate transition
• Calculate score for every task
• Investigate scores < 70
• Document root causes
• Generate actionable improvements
• Review trends over time

Feedback Loop Integration