marsai: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)

<cannot_skip>

• MUST add feedback-loop to todo list before any other action
• MUST execute TodoWrite IMMEDIATELY when skill starts
• CANNOT proceed to Step 1 without TodoWrite execution </cannot_skip>

⛔ 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 marsai: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 marsai:dev-feedback-loop: collect metrics, calculate scores, write report"
      status: "completed"
      priority: "high"

Anti-Rationalization:

Rationalization	Why It's WRONG	Required Action
"TodoWrite slows things down"	1 tool call = 2 seconds. Not an excuse.	Execute TodoWrite NOW
"I'll remember to complete it"	Memory is unreliable. Todo is proof.	Execute TodoWrite NOW
"Skill is simple, no tracking needed"	Simple ≠ optional. all skills get tracked.	Execute TodoWrite NOW

You CANNOT proceed to Step 1 without executing TodoWrite above.

---

Severity Calibration

Severity	Criteria	Examples
CRITICAL	Score 0 (rejected), system failure	User rejected validation, complete workflow failure
HIGH	Score < 70, threshold breach	Assertiveness below 70, recurring pattern unresolved
MEDIUM	Score 70-79, pattern emerging	Acceptable but degraded, same issue twice
LOW	Score 80-89, minor improvements	Good but not excellent, optimization opportunity

Report all severities. CRITICAL = mandatory post-mortem. HIGH = root cause analysis. MEDIUM = track pattern. LOW = document.

---

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:

Excuse	Reality
"It was just a spike/experiment"	Spikes produce learnings. Track what worked and what didn't.
"Perfect score, no insights"	Perfect scores reveal what works. Document for replication.
"Reporting my own failures reflects badly"	Unreported failures compound. Self-reporting is professional.
"Round up to passing threshold"	Rounding is falsification. Report exact score.

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:

Bias Pattern	Why It's Wrong	Correct Behavior
"Round up score"	Falsifies data, masks trends	Report exact: 68, not 70
"Skip failed task"	Selection bias, incomplete picture	Report all tasks
"Blame external factors"	Avoids actionable insights	Document factors + still log score
"Report only successes"	Survivorship bias	Success and failure needed

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

<cannot_skip>

• Feedback collection for EVERY completed task
• No exemptions for task complexity, outcome, or time pressure
• "Nothing to report" is still data that must be recorded </cannot_skip>

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

Factor	Still Collect?	Reason
Task complexity	✅ YES	Simple tasks reveal patterns
Outcome quality	✅ YES	100-score tasks need tracking
User satisfaction	✅ YES	Approval ≠ process quality
Time pressure	✅ YES	Metrics take <5 min
"Nothing to report"	✅ YES	Absence of issues is data

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

---

Repeated Feedback Detection

When the same feedback appears multiple times:

Repetition	Classification	Action
2nd occurrence	RECURRING	Flag as recurring issue. Add to patterns.
3rd occurrence	UNRESOLVED	Escalate. Stop current work. Report blocker.

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

<block_condition>

• Task score < 70 → Root cause analysis required
• Gate iterations > 3 → STOP, request human intervention
• Cycle average < 80 → Deep analysis required </block_condition>

When thresholds are breached, response is REQUIRED:

Alert	Threshold	Required Action
Task score	< 70	Document what went wrong. Identify root cause.
Gate iterations	> 3	STOP. Request human intervention. Document blocker.
Cycle average	< 80	Deep analysis required. Pattern identification mandatory.

You CANNOT proceed past threshold without documented response.

Blocker Criteria - STOP and Report

always pause and report blocker for:

Decision Type	Examples	Action
Score interpretation	"Is 65 acceptable?"	STOP. Follow interpretation table.
Threshold override	"Skip analysis for this task"	STOP. Analysis is MANDATORY for low scores.
Pattern judgment	"Is this pattern significant?"	STOP. Document pattern, let user decide significance.
Improvement priority	"Which fix first?"	STOP. Report all findings, let user prioritize.

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

Event	Penalty	Max Penalty	Rationale
Extra iteration (beyond 1)	-10 per iteration	-30	Each iteration = rework
Review FAIL verdict	-20	-20	Critical/High issues found
Review NEEDS_DISCUSSION	-10	-10	Uncertainty in implementation
Unmet criterion at validation	-10 per criterion	-40	Requirements gap
User REJECTED validation	-100 (score = 0)	-100	Complete failure

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

Score Range	Rating	Action Required
90-100	Excellent	No action needed
80-89	Good	Minor improvements possible
70-79	Acceptable	Review patterns, optimize
60-69	Needs Improvement	Root cause analysis required
< 60	Poor	Mandatory deep analysis
0	Failed	Full post-mortem required

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:

Gate	Structured Fields	Use For
Gate 0	implementation.standards_compliance, implementation.iterations	Implementation standards patterns
Gate 1	devops.standards_compliance, devops.verification_errors[]	DevOps standards + build/deploy failures
Gate 2	marsai:sre.standards_compliance, marsai:sre.validation_errors[]	SRE standards + observability gaps
Gate 3	testing.standards_compliance, testing.failures[], testing.uncovered_criteria[]	Testing standards + test failures + coverage
Gate 4	review.{reviewer}.standards_compliance, review.{reviewer}.issues[]	Review standards + issues by category/severity

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.marsai: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.marsai: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, marsai: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/marsai:dev-cycle/current-cycle.json or docs/marsai:dev-refactor/current-cycle.json):

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

3.2 Dispatch Prompt Quality Reviewer

<dispatch_required agent="prompt-quality-reviewer"> Analyze prompt quality for all agents executed in this task. </dispatch_required>

Task tool:
  subagent_type: "marsai: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

Alert	Trigger	Action	Report Contents
Score < 70	Individual task assertiveness < 70	Mandatory root cause analysis	Failure events, "5 Whys" per event, Corrective actions, Prevention measures
Iterations > 3	Any gate exceeds 3 iterations	STOP + human intervention	Iteration history, Recurring issue, Options: [Continue/Reassign/Descope/Cancel]
Avg < 80	Cycle average below 80	Deep analysis report	Score distribution, Failure patterns (freq/cause/fix), Improvement plan
Recurring Pattern	Same issue category in 3+ tasks	Pattern alert	Category, frequency, suggested prompt fix

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

Required sections:

Section	Content
Header	Date, Tasks Completed, Average Assertiveness
Task Summary	Table: Task ID, Score, Rating, Key Issue
By Gate	Table: Gate, Avg Iterations, Avg Duration, Pass Rate
By Penalty	Table: Penalty type, Occurrences, Points Lost
Patterns	Positive patterns (what works) + Negative patterns (what needs improvement)
Recommendations	Immediate (this sprint), Short-term (this month), Long-term (this quarter)
Next Review	Date, Target assertiveness, Focus areas

Step 6: Generate Improvement Suggestions

Improvement types based on pattern analysis:

Target	When to Suggest	Format
Agents	Same issue type recurring in reviews	Agent name → Issue → Suggestion → Specific addition to prompt
Skills	Gate consistently needs iterations	Skill name → Issue → Suggestion → Specific change to skill
Process	Pattern spans multiple tasks	Process area → Issue → Suggestion → Implementation

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 marsai: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 docs/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.

Metric	Value
Tasks Analyzed	N
Average Assertiveness	XX.X%
Threshold Alerts	X
Root Cause Analyses	Y
Improvement Suggestions	Z
Report Location	docs/dev-team/feedback/cycle-YYYY-MM-DD.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

Integration	Process
Retrospectives	Share metrics → Discuss trends → Prioritize improvements → Assign actions
Skill Updates	Document gap → Update skill → Track improvement → Iterate
Agent Updates	Identify behavior → Update prompt → Track change → Validate