complexity-assessment

Complexity Assessment

Evaluate task complexity before planning to ensure appropriate workflow depth and artifact structure.

Contextd Integration

If contextd MCP is available:

• memory_search to find similar past assessments
• memory_record for tier recommendations with rationale
• memory_outcome to track assessment accuracy over time
• branch_create/return for isolated assessment

If contextd is NOT available:

• Assessment runs inline (still works)
• No persistence of tier decisions
• No historical comparison available

When to Use

This skill is called by /brainstorm Phase 2 to determine:

• Question depth (SIMPLE: 3-5, STANDARD: 8-12, COMPLEX: 15+)
• GitHub artifact structure (single issue vs epic + sub-issues vs project board)
• Worktree recommendation
• Context branch budget allocation

---

7 Dimensions

Assess each dimension and aggregate to determine tier:

1. Scope

Question: How many files will be touched?

Files	Score	Indicator
1-2	1	Single component, localized change
3-10	2	Multiple components, cross-cutting
10+	3	System-wide, architectural change

2. Integration

Question: What external dependencies are involved?

Dependencies	Score	Indicator
None	1	Self-contained, no external calls
1-2 APIs/services	2	Moderate integration work
3+ APIs/services	3	Complex orchestration required

3. Infrastructure

Question: Are config/infra changes required?

Changes	Score	Indicator
None	1	Code-only change
Config files	2	Environment variables, feature flags
New infra	3	Database, Docker, cloud resources

4. Knowledge

Question: What domain expertise is required?

Expertise	Score	Indicator
Familiar patterns	1	Standard CRUD, common patterns
Some research	2	New library, unfamiliar domain
Deep expertise	3	Security, performance, compliance

5. Risk

Question: What's the blast radius if something goes wrong?

Risk	Score	Indicator
Low	1	Easily reversible, non-critical path
Medium	2	User-facing, requires testing
High	3	Data integrity, security, payments

6. Testing (New)

Question: What testing effort is required?

Testing	Score	Indicator
Minimal	1	Unit tests only, existing patterns
Moderate	2	Integration tests, new test fixtures
Extensive	3	E2E tests, performance tests, security audits

7. Decomposability (New)

Question: Can this task be split into independent units?

Decomposability	Score	Indicator
Easily decomposable	1	Clear subtasks, no shared state
Partially decomposable	2	Some dependencies between subtasks
Monolithic	3	Tightly coupled, must be done together

---

Risk Multipliers

Apply these multipliers to the base score when present:

Factor	Multiplier	When to Apply
Security-sensitive	1.2x	Auth, encryption, user data, API keys
Data migration	1.3x	Schema changes, data transformations
Breaking changes	1.2x	API changes, deprecated features
Technical debt	1.1x	Working in legacy/poorly-tested code
Time pressure	1.1x	Urgent deadline, reduced review time

Calculation: adjusted_score = base_score * highest_applicable_multiplier

---

Confidence Scoring

Every assessment includes a confidence percentage (0-100%):

Confidence	Range	Criteria
High	80-100%	Clear requirements, familiar domain, similar past tasks
Medium	50-79%	Some ambiguity, partial familiarity
Low	0-49%	Unclear scope, unfamiliar domain, novel problem

When confidence < 60%: Use AskUserQuestion to clarify ambiguous dimensions before finalizing tier.

Confidence factors:

• Requirements clarity (+/- 20%)
• Domain familiarity (+/- 15%)
• Historical data available (+/- 15%)
• Codebase familiarity (+/- 10%)

---

Team Velocity Factors

Adjust expectations based on team context:

Factor	Impact	Consideration
Team size	+/- 1 tier	Solo dev may need COMPLEX timeline adjustment
Experience level	+/- 1 dimension score	Junior devs add +1 to Knowledge dimension
Codebase familiarity	+/- 1 dimension score	New team members add +1 to Scope
Concurrent priorities	+/- 1 tier	High WIP may require elevated treatment

---

Tier Calculation

Sum scores across all 7 dimensions (max 21):

Total Score	Tier	Characteristics
7-11	SIMPLE	Quick change, minimal planning needed
12-16	STANDARD	Moderate complexity, structured approach
17-21	COMPLEX	High complexity, extensive planning required

After multipliers: Round to nearest integer before tier assignment.

---

Structured Output

Return assessment as JSON for downstream consumption:

{
  "tier": "STANDARD",
  "confidence": 75,
  "confidence_justification": "Clear scope but unfamiliar OAuth library",
  "base_score": 14,
  "adjusted_score": 16.8,
  "multipliers_applied": ["security-sensitive"],
  "dimensions": {
    "scope": { "score": 2, "rationale": "Auth routes, middleware, user model, frontend" },
    "integration": { "score": 2, "rationale": "OAuth provider API" },
    "infrastructure": { "score": 2, "rationale": "Environment variables, session config" },
    "knowledge": { "score": 2, "rationale": "OAuth2 flow, security best practices" },
    "risk": { "score": 3, "rationale": "Security-critical, user data" },
    "testing": { "score": 2, "rationale": "Integration tests for OAuth flow" },
    "decomposability": { "score": 1, "rationale": "Backend/frontend can be parallelized" }
  },
  "decomposition_suggestions": null,
  "historical_comparison": "Similar OAuth task took 3 iterations (Issue #45)",
  "team_velocity_adjustments": null,
  "integration_hints": {
    "github_artifact": "epic_with_sub_issues",
    "context_branch_budget": 8192,
    "worktree_recommended": true
  },
  "recommended_agents": {
    "always": ["code-quality-reviewer"],
    "if_go": ["go-reviewer"],
    "if_security_sensitive": ["security-reviewer", "vulnerability-reviewer"],
    "if_api_change": ["user-persona-reviewer", "documentation-reviewer"],
    "all": false
  },
  "intent_confirmation": {
    "required": true,
    "mode": "STANDARD"
  }
}

---

Tier Implications

SIMPLE (7-11)

• Questions: 3-5 quick clarifications
• GitHub: Single Issue with checklist
• Worktree: Optional, can work in main
• Review: Standard consensus review
• Context branch budget: 4096 tokens
• Testing: Unit tests, 70%+ coverage on new code

STANDARD (12-16)

• Questions: 8-12 detailed questions
• GitHub: Epic Issue + sub-Issues
• Worktree: Recommended for isolation
• Review: Full consensus review
• Context branch budget: 8192 tokens
• Testing: Unit + integration tests, 80%+ coverage

COMPLEX (17-21)

• Questions: 15+ comprehensive questions
• GitHub: Epic + sub-Issues + Project board
• Worktree: Required for isolation
• Review: Full consensus + additional scrutiny
• Context branch budget: 16384 tokens
• Testing: Unit + integration + E2E, 90%+ coverage, security audit

---

Intent Confirmation Gating

After tier determination, include an intent_confirmation field in the output to signal the caller whether intent confirmation is recommended:

Tier	Recommendation
SIMPLE (7-11)	"intent_confirmation": "none" - Caller should auto-proceed.
STANDARD (12-16)	"intent_confirmation": "standard" - Caller should invoke intent-confirmation skill.
COMPLEX (17-21)	"intent_confirmation": "detailed" - Caller should invoke intent-confirmation skill in DETAILED mode.

Integration Note

This skill does NOT invoke intent-confirmation itself. It outputs a recommendation that the calling workflow uses to decide whether to invoke intent-confirmation. The typical workflow is:

Caller (e.g., /brainstorm, /plan):
  1. Invoke complexity-assessment → returns tier + intent_confirmation field
  2. If intent_confirmation != "none":
     → Invoke intent-confirmation skill with the recommended mode
     → Wait for user confirmation
  3. Proceed with execution

This avoids a circular dependency between the two skills.

DETAILED Mode (COMPLEX only)

For COMPLEX tasks, the intent preview includes:

• Full decomposition plan with subtask dependencies
• Risk factors and mitigation strategy
• Estimated resource usage (agents, branches, tokens)
• Explicit list of files to be modified
• User must respond with explicit "Yes" (not just acknowledgment)

Why Gate on Complexity?

• SIMPLE tasks should not slow down developer flow with unnecessary confirmation
• STANDARD tasks benefit from a pause to ensure alignment
• COMPLEX tasks require explicit buy-in because rollback cost is high
• Research (JetBrains NeurIPS 2025) shows observation masking is 52% cheaper than full summarization; confirming intent early avoids wasted work on misunderstood requirements

---

Review Agent Selection

Include a recommended_agents field in the structured output based on the assessment. This allows the consensus-review skill to perform adaptive agent selection without re-analyzing the diff.

Output Field

{
  "recommended_agents": {
    "always": ["code-quality-reviewer"],
    "if_go": ["go-reviewer"],
    "if_security_sensitive": ["security-reviewer", "vulnerability-reviewer"],
    "if_api_change": ["user-persona-reviewer", "documentation-reviewer"],
    "all": false
  }
}

Selection Logic

Tier	all	Base Agents
SIMPLE	false	["code-quality-reviewer"]
STANDARD	false	["code-quality-reviewer"] + conditional agents
COMPLEX	true	All 6 reviewers (overrides conditional logic)

Conditional Agent Triggers

These are derived from the dimension scores:

Condition	Trigger	Agents Added
Risk score >= 2	Security-sensitive	security-reviewer, vulnerability-reviewer
Knowledge score >= 2 AND domain is Go	Go code	go-reviewer
Integration score >= 2	API changes likely	user-persona-reviewer, documentation-reviewer
Infrastructure score >= 2	Config/infra changes	security-reviewer
Scope score == 3	System-wide change	All reviewers (all: true)

Integration with consensus-review

The consensus-review skill reads recommended_agents from the assessment output:

1. Run complexity-assessment → get tier + recommended_agents
2. Pass recommended_agents to consensus-review adaptive agent selection
3. Consensus-review merges recommended_agents with its own override rules
4. Final agent set = union(recommended_agents, override_agents)

---

Integration with Other Skills

github-planning Integration

Tier determines GitHub artifact structure:

Tier	Artifact Type	Details
SIMPLE	Single Issue	Checklist in body, no sub-issues
STANDARD	Epic + Sub-Issues	Parent issue links children
COMPLEX	Epic + Sub-Issues + Project	Board with columns, milestones

context-folding Integration

Tier determines branch budget:

Tier	Token Budget	Max Depth
SIMPLE	4096	1 level
STANDARD	8192	2 levels
COMPLEX	16384	3 levels

---

Decomposition Suggestions

For COMPLEX tasks (score 17+), provide decomposition suggestions:

{
  "decomposition_suggestions": [
    {
      "subtask": "Backend OAuth integration",
      "estimated_tier": "STANDARD",
      "dependencies": [],
      "can_parallelize": true
    },
    {
      "subtask": "Frontend auth UI components",
      "estimated_tier": "SIMPLE",
      "dependencies": [],
      "can_parallelize": true
    },
    {
      "subtask": "Session management middleware",
      "estimated_tier": "SIMPLE",
      "dependencies": ["Backend OAuth integration"],
      "can_parallelize": false
    }
  ],
  "parallelization_potential": "2 of 3 subtasks can run in parallel",
  "recommended_approach": "Split into 3 issues, assign backend and frontend in parallel"
}

---

Ambiguity Handling

When confidence < 60%, use AskUserQuestion to clarify:

AskUserQuestion(
  question: "I'm uncertain about the complexity of this task. To provide an accurate assessment, I need clarification on:",
  options: [
    "What external services/APIs will this integrate with?",
    "Are there existing tests I should reference for patterns?",
    "What's the expected timeline pressure?",
    "Are there schema/database changes involved?"
  ],
  allow_custom: true
)

After clarification: Re-score affected dimensions and recalculate tier.

---

Historical Comparison

When contextd is available, search for similar past assessments:

# Search for similar tasks
mcp__contextd__memory_search(
  project_id: "<project>",
  query: "<task description keywords>",
  tags: ["complexity-assessment"]
)

Include in output:

• Similar past tasks and their actual outcomes
• Iteration count (how many rounds to complete)
• Assessment accuracy (predicted vs actual complexity)

Example output:

Historical comparison: "Similar OAuth task (Issue #45) was assessed as STANDARD,
took 3 iterations to complete. Assessment was accurate."

---

Contextd Recording

Store assessment in contextd with full rationale:

mcp__contextd__memory_record(
  project_id: "<project>",
  title: "Complexity assessment: <task-summary>",
  content: JSON.stringify({
    tier: "<tier>",
    confidence: <0-100>,
    base_score: <sum>,
    adjusted_score: <after multipliers>,
    dimensions: { <all 7 with scores and rationale> },
    multipliers: [<applied multipliers>],
    decomposition: <suggestions if COMPLEX>,
    integration_hints: { <github, context-folding settings> }
  }),
  outcome: "success",
  tags: ["complexity-assessment", "<tier>", "confidence-<high|medium|low>"]
)

---

Example Assessment

Task: "Add user authentication with OAuth2"

Dimension	Score	Rationale
Scope	2	Auth routes, middleware, user model, frontend components
Integration	2	OAuth provider API integration
Infrastructure	2	Environment variables, session config
Knowledge	2	OAuth2 flow, security best practices
Risk	3	Security-critical, user data
Testing	2	Integration tests for OAuth flow needed
Decomposability	1	Backend/frontend can be parallelized

Base Score: 14/21 Multiplier: 1.2x (security-sensitive) Adjusted Score: 16.8 -> 17 = COMPLEX (borderline) Confidence: 75% (clear scope, unfamiliar library)

JSON Output:

{
  "tier": "COMPLEX",
  "confidence": 75,
  "confidence_justification": "Clear scope but unfamiliar OAuth library",
  "base_score": 14,
  "adjusted_score": 17,
  "multipliers_applied": ["security-sensitive"],
  "dimensions": {
    "scope": { "score": 2, "rationale": "Auth routes, middleware, user model, frontend" },
    "integration": { "score": 2, "rationale": "OAuth provider API" },
    "infrastructure": { "score": 2, "rationale": "Environment variables, session config" },
    "knowledge": { "score": 2, "rationale": "OAuth2 flow, security best practices" },
    "risk": { "score": 3, "rationale": "Security-critical, user data" },
    "testing": { "score": 2, "rationale": "Integration tests for OAuth flow" },
    "decomposability": { "score": 1, "rationale": "Backend/frontend can be parallelized" }
  },
  "decomposition_suggestions": [
    { "subtask": "Backend OAuth routes", "estimated_tier": "STANDARD", "can_parallelize": true },
    { "subtask": "Frontend auth UI", "estimated_tier": "SIMPLE", "can_parallelize": true }
  ],
  "historical_comparison": "Similar OAuth task took 3 iterations (Issue #45)",
  "integration_hints": {
    "github_artifact": "epic_with_sub_issues_and_project",
    "context_branch_budget": 16384,
    "worktree_recommended": true
  },
  "recommended_agents": {
    "always": ["code-quality-reviewer"],
    "if_go": [],
    "if_security_sensitive": ["security-reviewer", "vulnerability-reviewer"],
    "if_api_change": ["user-persona-reviewer", "documentation-reviewer"],
    "all": true
  },
  "intent_confirmation": {
    "required": true,
    "mode": "DETAILED"
  }
}

---

Integration

This skill is designed to be called by other commands/skills:

# In /brainstorm command Phase 2:

## Phase 2: Complexity Assessment (Context Folded)

mcp__contextd__branch_create(
  session_id: "<session>",
  description: "Assess task complexity",
  budget: 4096
)

1. Search contextd for similar past assessments
2. Apply 7-dimension analysis to user's task description
3. Calculate confidence score
4. Apply risk multipliers if applicable
5. Calculate tier based on adjusted score
6. Generate decomposition suggestions if COMPLEX
7. Record assessment in contextd with full JSON

mcp__contextd__branch_return(
  branch_id: "<branch>",
  message: JSON.stringify({
    tier: "<tier>",
    confidence: <0-100>,
    score: "<adjusted>/<max>",
    key_factors: ["<top 2 dimensions>"],
    decomposition: <if COMPLEX>,
    integration_hints: { github_artifact, context_branch_budget, worktree_recommended }
  })
)

---

Mandatory Checklist

EVERY assessment MUST complete ALL steps:

• Search contextd for similar past assessments (if available)
• Analyze Scope dimension with score and rationale
• Analyze Integration dimension with score and rationale
• Analyze Infrastructure dimension with score and rationale
• Analyze Knowledge dimension with score and rationale
• Analyze Risk dimension with score and rationale
• Analyze Testing dimension with score and rationale
• Analyze Decomposability dimension with score and rationale
• Calculate base score (sum of 7 dimensions)
• Apply risk multipliers if applicable
• Calculate confidence score with justification
• Determine tier from adjusted score
• Generate decomposition suggestions if COMPLEX
• Include integration hints (github-planning, context-folding)
• Include recommended_agents field based on tier and dimension scores
• Include intent_confirmation field (required + mode)
• Gate on intent confirmation (STANDARD: show plan; COMPLEX: require explicit "Yes")
• RECORD to contextd with full JSON (see Recording section)
• Return structured JSON output

The assessment is NOT complete until contextd recording is done.

---

Red Flags - STOP and Reconsider

If you're thinking any of these, you're about to violate the skill:

Thought	Reality
"This is obviously simple, no need to score"	Score anyway. Intuition misses edge cases.
"I'll just say STANDARD to be safe"	Calculate the actual score. Don't guess.
"Recording to contextd is optional"	Recording is MANDATORY. No exceptions.
"Risk doesn't matter for this task"	Risk ALWAYS matters. Score it.
"I can skip dimensions for simple tasks"	ALL 7 dimensions. Every time.
"The user already knows the complexity"	Document for future sessions. Record it.
"Confidence scoring is extra work"	Confidence enables better decisions. Calculate it.
"Decomposition is only for COMPLEX"	Consider it for STANDARD too if helpful.
"Historical comparison is optional"	Past data improves accuracy. Search for it.
"Intent confirmation slows things down"	It prevents wasted work on misunderstood requirements.
"I don't need to include recommended_agents"	Downstream consensus-review depends on it. Always include.
"SIMPLE tasks don't need intent logging"	Log anyway. Calibration data improves future accuracy.

---

Common Mistakes

Mistake	Correct Approach
Inflating simple tasks	Typo fix = SIMPLE. Don't over-engineer.
Deflating complex tasks	Payment/security = COMPLEX. Don't underestimate.
Gut-feel tier assignment	Show your math. 7 dimensions, explicit scores.
Skipping contextd	Recording enables future efficiency. Always record.
Ignoring high-risk dimensions	Risk=3 deserves attention regardless of total.
Not applying multipliers	Security/migrations/breaking changes need multipliers.
Low confidence without clarification	Use AskUserQuestion when confidence < 60%.
Missing decomposition for COMPLEX	Always suggest how to break down COMPLEX tasks.
Returning text instead of JSON	Downstream tools expect structured output.
Skipping intent confirmation for STANDARD	STANDARD requires plan approval before proceeding.
Not including recommended_agents	consensus-review uses this for adaptive agent selection.
Setting all: true for SIMPLE	Only COMPLEX warrants all 6 reviewers.

---

Tracking Assessment Accuracy

After task completion, update the original assessment:

mcp__contextd__memory_outcome(
  memory_id: "<original assessment id>",
  outcome: "success|partial|failure",
  notes: "Predicted STANDARD, actual was COMPLEX due to unexpected API rate limits"
)

Track metrics over time:

• Tier prediction accuracy
• Iteration count vs predicted
• Common underestimation patterns
• Team velocity calibration

---

Attribution

Adapted from Auto-Claude complexity_assessor.md by AndyMik90 (MIT). Enhanced with testing dimension, confidence scoring, decomposability analysis, and structured output. See CREDITS.md for full attribution.