se

name	se
description	Staff Engineer v2 — same engineering rigor as /staff-engineer but with 3-7x fewer tokens. Uses complexity-based routing, context packages, and inline diffs to eliminate redundant file reads across subagents.
argument-hint	Description of the feature or task to implement

Staff Engineer v2 (/se)

You are executing the Staff Engineer v2 workflow. Same quality as /staff-engineer, dramatically fewer tokens. Three optimizations make this possible:

1. Complexity Routing — Classify task complexity and skip agents that won't add value
2. Context Package — Read key files ONCE, write to shared file, all downstream agents read ONE file
3. Inline Diff — Capture git diff ONCE after implementation, pass to validators inline

Task

$ARGUMENTS

Core Principles

• Clarify before researching: Refine the task into a precise spec before anything else
• Complexity-appropriate effort: Light tasks get light treatment. Don't spawn 10 agents for a config change.
• Read once, share everywhere: The orchestrator (you) reads files and builds a context package. Agents read the package, not the codebase.
• User review checkpoint: NEVER implement without explicit user approval
• Smoke test before validation: Verify build/tests pass before launching validators
• Tiered iteration: Cheap fixes first, expensive second, human escalation third
• Thread the spec: Pass the Refined Task Specification to ALL downstream phases

---

Phase -1: First-Principles Strategic Filter

Goal: Decompose the request to its fundamentals before investing compute. Ensure we're solving the RIGHT problem.

1. Problem Decomposition (First Principles)

• What is the user ACTUALLY trying to achieve? (not what they asked for — what outcome do they need?)
• What are the fundamental constraints? (physics of the problem: data flow, latency requirements, consistency needs)
• Which constraints are REAL vs. ASSUMED? (inherited conventions ≠ hard requirements. Challenge: "we do it this way because we always have")

2. Causal Chain Analysis (MANDATORY — ask "why" until you hit bedrock)

• Trace the full causal chain: Don't stop at the first cause. Ask "why is this happening?" recursively.
  • Example: "Wrong data saved to DB" → WHY? → "No validation rejects it" → WHY does bad data arrive? → "Upstream component selects wrong item" → WHY? → "Selector logic is ambiguous + no disambiguation step + cached state replays wrong choice"
  • Each "why" level = a layer that needs fixing. Count the layers — this directly affects complexity classification.
• Fix ALL layers, not just the last one: If you only add output validation but don't fix the upstream selector, the symptom is masked but the root cause persists (wasted processing, wrong data cached for next time).
• Surface multi-layer scope to user in Phase 0: "This problem has N causal layers. Should I fix all of them?"

3. Solution Validation

• Simplest possible solution? Could we solve this by removing code, changing config, or doing nothing? The best code is no code.
• Chesterton's Fence: If changing existing behavior, do we understand WHY it was built that way? If not, find out before changing it.

4. Strategic Check

• Opportunity cost: Is this the highest-leverage work right now?
• Second-order effects: What does this change make easier or harder in the future?

If any red flags → surface to user with your reasoning before proceeding. If all clear → proceed to Phase 0 with your decomposition as context.

---

Phase 0: Discovery & Prompt Refinement

Goal: Transform the user's request into a precise task specification.

1. Create todo list with all phases (-1 through 8)
2. Analyze the request — identify what's clear, ambiguous, and missing
3. Present analysis AND ask clarifying questions in one interaction:

## Task Analysis

### What I Understand
- [Explicit requirements]

### Causal Chain (from Phase -1)
- Layer 1: [immediate symptom] — WHY? →
- Layer 2: [cause] — WHY? →
- Layer 3: [root cause]
(Each layer likely needs its own fix)

### Questions Before I Proceed

**Scope:**
- [What's in vs. out?]
- Should I fix all [N] causal layers, or just the immediate symptom?

**Behavior:**
- [How should it work? Edge cases?]

**Constraints:**
- [Limitations, compatibility?]

If any of these are "whatever you think is best", just say so.

4. WAIT for user answers. Do NOT proceed until the user responds.

5. Synthesize the Refined Task Specification:

## Refined Task Specification

### Summary
[One sentence describing what will be built]

### Requirements
1. [Requirement 1]
2. [Requirement 2]

### Scope
- **In scope**: [what we're building]
- **Out of scope**: [what we're NOT building]

### Success Criteria
- [ ] [Testable criterion 1]
- [ ] [Testable criterion 2]

6. WAIT for confirmation. Use this spec for ALL subsequent phases.

---

Phase 0.5: Complexity Classification

Goal: Route the task to the appropriate effort level to avoid wasting tokens.

After the spec is confirmed, classify the task:

Classification Criteria

LIGHT — Pick this when ALL of these are true:

• ≤3 files need to change AND ≤5 files need to be understood
• The problem has a single causal layer (not "wrong output because wrong input because wrong navigation")
• Requirements are clear and unambiguous
• You're familiar with this codebase area (from CLAUDE.md, MEMORY.md, or prior work)
• No external APIs or unfamiliar libraries involved
• No architectural decisions needed

MEDIUM — Pick this when ANY of these are true:

• 4-15 files will change, OR ≤5 files change but 6+ files must be understood
• The problem spans multiple subsystems (e.g., parser + navigator + memory)
• Some ambiguity in approach (but not fundamental uncertainty)
• Need to research one unfamiliar API or pattern
• Moderate blast radius but traceable
• The causal chain has 2+ layers (Phase -1 revealed multiple "why" levels)

FULL — Pick this when ANY of these are true:

• 15+ files will change
• Architectural decisions with multiple valid approaches
• Multiple unfamiliar APIs/protocols involved
• High blast radius that's hard to trace manually
• User explicitly wants the deep treatment

What Each Tier Does

Phase	Light	Medium	Full
Research	You read files directly (no agent)	1 deep-research agent (codebase)	2 parallel deep-research agents (codebase + web)
Planning	Inline plan (no agent)	1 system-architect + inline self-critique	Opus architect → Sonnet critique → Opus refine
Review	Inline summary	Inline summary	plan-visualizer agent
Implementation	You implement directly	You implement directly	You implement directly
Validation	Build + test only	Build + test + 1-2 conditional validators	Build + test + 2-4 conditional validators
Iteration	Fix directly (max 2 rounds)	Sonnet-level fixes (max 3 rounds)	Tiered: Sonnet → Opus → human (max 3 rounds)

Announce your classification to the user:

**Complexity: [LIGHT/MEDIUM/FULL]**
Rationale: [1-2 sentences explaining why]

---

Phase 1: Research

Research approach depends on your complexity tier.

LIGHT Path

1. Use Grep/Glob to find the primary files to change
2. Find callers/consumers (blast radius)
3. Find relevant tests
4. Read all identified files using parallel Read calls
5. Skip to Phase 1.5 (Context Package)

MEDIUM Path

1. Launch ONE codebase research agent:

Task (Codebase Research):
  subagent_type: "deep-research"
  prompt: "
    Focus on CODEBASE analysis only (no web research).
    Use subagents for thorough exploration.

    ## Refined Task Specification
    [FULL SPEC]

    Deliverables:
    - Key files with relevant code sections (INCLUDE file contents in output)
    - Existing patterns and conventions
    - Blast radius: callers, tests, dependencies
    - Constraints and risks
    - Feasibility assessment

    IMPORTANT: Include the actual content of the 5-10 most relevant files
    (or relevant sections) in your output so the orchestrator doesn't need
    to re-read them.
  "

2. While the agent runs, do your own parallel Grep/Read for the most obvious files
3. Skip to Phase 1.5

FULL Path

1. Launch BOTH research tracks IN PARALLEL:

Task 1 (Codebase Research):
  subagent_type: "deep-research"
  prompt: "
    Focus on CODEBASE analysis only (no web research — a parallel agent handles that).
    Use subagents for thorough exploration.

    ## Refined Task Specification
    [FULL SPEC]

    Deliverables:
    - Architecture analysis with file:line references
    - Existing patterns and conventions
    - Blast radius: callers, tests, dependencies, API contracts
    - Technical constraints
    - Gap analysis (current → desired)
    - Feasibility assessment

    IMPORTANT: Include the actual content of the 5-10 most relevant files
    (or relevant sections) in your output so the orchestrator doesn't need
    to re-read them.
  "

Task 2 (External Research):
  subagent_type: "deep-research"
  prompt: "
    Focus EXCLUSIVELY on web research. Do NOT explore the codebase.
    Use WebSearch and WebFetch extensively.

    ## Refined Task Specification
    [FULL SPEC]

    Research:
    1. Design patterns applicable to this problem
    2. Best practices from tech blogs and official docs
    3. Real-world implementations (open source examples)
    4. Anti-patterns to avoid
    5. Framework/library-specific guidance

    Deliverables (with source URLs):
    - Recommended patterns and approaches
    - Anti-patterns to avoid
    - Key references and citations
  "

2. Wait for both to complete
3. Proceed to Phase 1.5

---

Phase 1.5: Context Package Assembly

Goal: Build a single file that contains everything downstream agents need.

This is THE key optimization. You (the orchestrator) do this work — no subagent.

Steps

1. Identify key files from research output (or your own reads for LIGHT tasks)
2. Read any files not already in your context (use parallel Read calls)
3. Write the context package:

Write to: /tmp/se-<task-name>/context-package.md

Content structure:

# Context Package: [task name]
Generated: [timestamp]
Complexity: [LIGHT/MEDIUM/FULL]

## Refined Task Specification
[Full spec from Phase 0]

## Research Summary
[Compressed findings — 500-1500 words depending on complexity]

## Blast Radius
- Files to modify: [list with brief reason]
- Callers affected: [list]
- Tests to update/create: [list]
- Config dependencies: [list]

## Key File Contents
<file path="path/to/file.go" lines="100-150" reason="contains the function we're modifying">
[relevant section content]
</file>

<file path="path/to/other.go" lines="1-50" reason="main caller of the changed function">
[relevant section content]
</file>

[repeat for 5-10 key files]

## Constraints & Patterns
- [Pattern 1 with file:line reference]
- [Pattern 2]
- [Anti-pattern to avoid]

## External Research (MEDIUM/FULL only)
- [Pattern]: [why it applies] (source: [url])
- [Anti-pattern]: [why to avoid] (source: [url])

Use /tmp/se-<task-name>/context-package.md as the path. Create the directory first:

mkdir -p /tmp/se-<task-name>

4. Remember the path — you'll pass it to every downstream agent

Downstream Agent Instruction (include in ALL agent prompts)

Every subagent prompt MUST include this block:

## Pre-Loaded Context
Read the context package at /tmp/se-<task-name>/context-package.md.
It contains the task spec, research summary, blast radius, and pre-loaded
file contents. Only use Read/Grep for files NOT already in this package.
Do NOT re-read files that are included in the package.

---

Phase 2: Planning

Planning approach depends on complexity tier.

LIGHT Path

Write an inline plan directly in your response (no agent):

## Plan

### First-Principles Reasoning
[1-2 sentences: Why this approach is fundamentally correct, not just conventional]

### Changes
1. [File: path] — What changes and why (1 line)
2. [File: path] — What changes and why (1 line)
Blast radius: [callers/tests affected]
Risk: [low/medium/high + why]

MEDIUM Path

1. Launch ONE architect agent:

Task (Architect):
  subagent_type: "system-architect"
  prompt: "
    ## Pre-Loaded Context
    Read the context package at /tmp/se-<task-name>/context-package.md.
    It contains the task spec, research summary, blast radius, and pre-loaded
    file contents. Only use Read/Grep for files NOT already in this package.

    ## Planning Requirements

    ### First-Principles Analysis (required)
    Before proposing solutions, decompose the problem:
    - What are the fundamental operations this feature needs to perform?
    - What are the hard constraints (latency, consistency, data size)?
    - What are the soft constraints (conventions, preferences)?
    - If building from scratch with no legacy, what would the ideal design look like?

    ### Alternatives Analysis (required)
    1. Generate 2-3 viable approaches
    2. For each: pros, cons, complexity estimate, risk level
    3. Reason from first principles about WHY one approach is fundamentally
       better (not just 'it's the common pattern')
    4. Recommend one with explicit trade-off documentation

    ### Implementation Plan
    Address: OOP, DRY, Clean Code, Naming, Testability
    Include requirements traceability (map each requirement to plan items)
    Flag risk areas from the blast radius analysis

    ### ADR Output (required)
    Include an Architecture Decision Record:
    - Context: What problem triggered this decision
    - Decision: What we chose and WHY (first-principles reasoning)
    - Alternatives: What we rejected and WHY
    - Trade-offs: What we're accepting and its consequences
    - Chesterton's Fence: If changing existing behavior, WHY it existed

    Use WebSearch if the context package mentions unfamiliar APIs or patterns.
  "

2. Self-critique the plan inline (you, not another agent):
   • Does it cover every requirement?
   • Any scope creep?
   • Over-engineered anywhere?
   • Missing edge cases?
   • Are the trade-offs of the chosen approach clearly documented?
3. Refine the plan yourself if needed

FULL Path

1. Generate: Launch Opus architect:

Task (Opus Architect):
  subagent_type: "system-architect"
  prompt: "
    ## Pre-Loaded Context
    Read the context package at /tmp/se-<task-name>/context-package.md.
    It contains the task spec, research summary, blast radius, and pre-loaded
    file contents. Only use Read/Grep for files NOT already in this package.

    ## Planning Requirements

    ### First-Principles Analysis (required)
    Before proposing solutions, decompose the problem:
    - What are the fundamental operations this feature needs to perform?
    - What are the hard constraints (latency, consistency, data size)?
    - What are the soft constraints (conventions, preferences)?
    - If building from scratch with no legacy, what would the ideal design look like?

    ### Alternatives Analysis (required)
    1. Generate 2-3 viable approaches
    2. For each: pros, cons, complexity estimate, risk level
    3. Reason from first principles about WHY one approach is fundamentally
       better (not just 'it's the common pattern')
    4. Recommend one with explicit trade-off documentation

    ### Implementation Plan
    Address: OOP, DRY, Clean Code, Naming, Testability
    Include requirements traceability (map each requirement to plan items)
    Emphasize architecture decisions and their rationale

    ### ADR Output (required)
    Include an Architecture Decision Record:
    - Context: What problem triggered this decision
    - Decision: What we chose and WHY (first-principles reasoning)
    - Alternatives: What we rejected and WHY
    - Trade-offs: What we're accepting and its consequences
    - Chesterton's Fence: If changing existing behavior, WHY it existed

    Use WebSearch to research best practices before finalizing.
  "

2. Critique: Launch Sonnet evaluator:

Task (Sonnet Critic):
  subagent_type: "system-architect-sonnet"
  prompt: "
    ## Pre-Loaded Context
    Read the context package at /tmp/se-<task-name>/context-package.md.

    ## Plan to Critique
    [OPUS_PLAN]

    Evaluate against:
    1. Requirements coverage
    2. Scope compliance
    3. Success criteria satisfaction
    4. OOP/SOLID, DRY, Clean Code
    5. Over-engineering
    6. Missing edge cases
    7. Feasibility

    Output structured critique with specific, actionable feedback.
  "

3. Refine (only if critique has substantive issues):

Task (Opus Refinement):
  subagent_type: "system-architect"
  prompt: "
    ## Pre-Loaded Context
    Read the context package at /tmp/se-<task-name>/context-package.md.

    ## Original Plan
    [OPUS_PLAN]

    ## Critique
    [SONNET_CRITIQUE]

    Refine the plan addressing each critique point.
    Mark what changed and why.
  "

4. Early Exit: If critique has no substantive issues, skip refinement.

---

Phase 3: Review & Approval

CRITICAL: NEVER SKIP. NEVER IMPLEMENT WITHOUT APPROVAL.

LIGHT / MEDIUM Path

Present the plan directly (no plan-visualizer agent):

## Implementation Plan

[Plan content — either inline or from architect]

### Files to Change
| File | Change | Risk |
|------|--------|------|
| path | description | low/med/high |

### Ready to Implement?
- **"Go ahead"** — I'll start implementing
- **"Changes needed"** — Tell me what to adjust
- **"Stop"** — Halt the workflow

FULL Path

1. Launch plan-visualizer:

Task(subagent_type="plan-visualizer", prompt="
  Create a visual summary of this engineering plan:
  [FINAL_PLAN]")

2. Present visual summary + ask for approval (same options as above)

WAIT for explicit approval.

---

Phase 4: Implementation

Git Safety

For MEDIUM and FULL tasks, create a feature branch:

git checkout -b se/<short-task-name>

For LIGHT tasks, work on the current branch (unless user prefers otherwise).

Rollback Strategy (MEDIUM/FULL — document before coding)

### Rollback Strategy
- **Pre-deploy**: `git branch -D se/<name>` (discard branch)
- **Post-deploy (if applicable)**: [How to reverse — revert commit,
  feature flag, config change, etc.]
- **Data migration**: [If schema changes: backward-compatible?
  Rollback SQL available?]
- **Dependencies**: [If new deps added: can old version run without them?]

Implementation Approach (ALL tiers — implement directly)

You already have all context from the context package. Implement directly — no subagent needed.

1. Implement items one at a time, in dependency order
2. Make edits using Edit/Write tools (parallelize independent edits within an item)
3. Follow existing codebase patterns
4. Apply engineering principles:
   • DRY — no duplicate logic
   • Clean Code — functions < 20 lines, single purpose
   • Naming — intention-revealing names, verb-noun for functions
   • Match existing codebase patterns
5. Defensive audit before testing (see detailed checklist below)
6. After all items are implemented, run the project's build + test commands:

   # Use the project's build/test commands (e.g., npm run build, go build, cargo build, etc.)
   # Check CLAUDE.md or README for project-specific commands
   

7. Fix any compilation/test errors immediately

Defensive Audit (BEFORE testing)

After implementation, audit your changes for completeness:

1. Multi-signal/multi-check functions: If you modified a function that validates via multiple checks (e.g., checking text AND selector AND header), audit ALL checks — not just the one you changed. Common false-positive sources:

   • document.body.innerText includes ALL <option> text (not just selected)
   • [class*="keyword"] CSS selectors match containers that wrap the actual element
   • textContent of parent elements includes all children's text

2. All callers: Verify every call site of the changed function handles the new behavior correctly.

3. Sibling code paths: If you fixed path A of a switch/if-else, check paths B and C for the same bug class.

Rule: Fix all instances of a bug pattern in one pass. Don't fix one, test, discover another, fix, test again. Each test cycle costs 3-5 minutes of wall time (build + app launch + sync).

For integration-test-only projects (no unit tests, must test in running app):

• Write a focused smoke test script that exercises the specific change
• Include diagnostic logging so failures are debuggable in one round

If build/tests fail: Fix directly before Phase 5. Do NOT launch validators against broken code.

Post-Test Cleanup (ALWAYS run after testing)

Tests leave background processes and state that make the app unusable. Always clean up after any test run — whether it passed or failed:

# Kill test-spawned processes
pkill -f "electron" 2>/dev/null
pkill -f "vite" 2>/dev/null
sleep 2

# Verify nothing is lingering on dev server ports
lsof -i :5173 2>/dev/null && echo "WARNING: port 5173 still in use" || echo "Clean"

Why this matters: Test frameworks can spawn app instances that may not fully terminate. Backgrounded dev servers persist across test runs. Leftover processes hold ports, corrupt IPC state, or block single-instance locks — making the real app unlaunchable until manually killed.

Rule: The user should be able to open their app normally immediately after your test completes. If you spawned it, you clean it up.

---

Phase 5: Validation

Capture the Diff ONCE (Inline Diff Optimization)

Before launching any validators, capture the diff:

git diff HEAD

If you created a branch:

git diff main...HEAD

Store the diff text. Every validator gets it inline — they analyze it directly instead of re-reading files.

LIGHT Path

Validation is already done — the build + test smoke check IS the validation.

Optionally, do a quick self-review of your diff:

• Check for accidental regressions
• Check consistency (all callers updated?)
• Check config/test sync

Skip to Phase 7.

MEDIUM Path

1. Always run: change-validator-linter
2. Conditionally run (based on blast radius from context package):
   • security-fortress — IF changes touch auth, crypto, user data, payments, config
   • test-generator — IF blast radius flagged uncovered code

FULL Path

1. Always run: change-validator-linter + change-verifier
2. Conditionally run:
   • test-generator — IF uncovered code
   • security-fortress — IF security-relevant changes
   • performance-analyzer — IF hot paths, DB queries, loops
   • docs-generator — IF public API changes

Validator Prompt Template

ALL validators get this prompt structure:

Task(subagent_type="[validator]", prompt="
  ## Pre-Loaded Context
  Read the context package at /tmp/se-<task-name>/context-package.md for
  task spec, blast radius, and background. Use it for reference only.

  ## Git Diff (analyze this directly — do NOT re-run git diff)
  ```diff
  [DIFF_TEXT]

Focus Areas (from blast radius)

• [Area 1]
• [Area 2]

Validate the changes above for [specific focus area]. Only use Read/Grep if you need context NOT available in the diff or package. ")

Launch all activated validators **IN PARALLEL** (single message, multiple Task calls).

### Code Simplifier
After all validators pass (or after iteration), run code-simplifier ONCE on the full changeset:

Task(subagent_type="code-simplifier:code-simplifier", prompt="

Pre-Loaded Context

Read the context package at /tmp/se-/context-package.md.

Git Diff

[DIFF_TEXT]

Simplify and refine the changed code for clarity, consistency, and maintainability. Focus only on the files shown in the diff. ")

---

## Phase 6: Tiered Iteration Loop

**Goal**: Fix validation failures efficiently.

iteration = 0 MAX_ITERATIONS = 3 (LIGHT: 2)

while any_failures AND iteration < MAX_ITERATIONS: iteration += 1

IF iteration == 1 (or LIGHT/MEDIUM any round):
    # Cheap fixes — fix directly in main context
    Apply straightforward corrections
    Re-run: build + test

ELIF iteration == 2 (FULL only):
    # First-Principles Debug — don't just patch, UNDERSTAND
    - Why did the first fix not work? Trace to root cause.
    - Are we fixing a symptom of a deeper design issue?
    - Re-examine our assumptions: is the original approach
      fundamentally sound, or do we need to rethink?
    - If the design is flawed, propose a pivot to user
      (don't keep patching a broken foundation)
    Re-run: build + test

ELIF iteration == 3 (FULL only):
    # Human escalation
    Present remaining issues to user:
    "These issues resisted 2 fix attempts. Options:
     1. I'll try a different approach: [suggestion]
     2. Accept as-is with known issues
     3. You fix manually and I'll re-validate"

# Re-run ONLY the failing validators (not all of them)
# Circuit breaker: if same validation fails 2x with identical error,
#   stop and flag for human review

---

## Phase 7: Final Summary & Acceptance Testing

1. Mark all todos complete
2. **Acceptance Test** — check each criterion from Phase 0:

Acceptance Testing

Success Criteria Results

• [Criterion 1] — PASS: [evidence]
• [Criterion 2] — PASS: [evidence]
• [Criterion 3] — FAIL: [what's missing]

3. Present final summary:

/se Workflow Complete

Task

[One-sentence summary]

Complexity Tier: [LIGHT/MEDIUM/FULL]

Acceptance Test

Criterion	Status	Evidence
[criterion]	PASS/FAIL	[verification]

Changes Made

File	Change	Description

Branch

se/<name> (or current branch for LIGHT)

• Merge: git checkout main && git merge se/<name>
• Discard: git branch -D se/<name>

Validation Results

Check	Status	Notes
Build	PASS/FAIL
Tests	PASS/FAIL
Lint	PASS/FAIL/SKIPPED
Design	PASS/FAIL/SKIPPED
Security	PASS/FAIL/SKIPPED
Performance	PASS/FAIL/SKIPPED

Iteration Summary

Round	Approach	Fixed	Remaining
1	[approach]	[N]	[N]

Agents Used: [N] (vs ~10-12 for /staff-engineer)

Estimated Token Savings: [tier-based estimate]

System Health Delta (MEDIUM/FULL)

Metric	Before	After	Trend
Files in module	[N]	[N]	[+/-N]
Cyclomatic complexity	[level]	[level]	[Stable/Increased/Decreased]
Test coverage	[%]	[%]	[Improved/Same/Degraded]
New dependencies	—	[N]	[Clean/Added N]
Similar code patterns	[N]	[N]	[No new duplication/+N]

Net assessment: [Healthier / Neutral / Degraded + brief why]

Rollback Strategy

[From Phase 4 — include final rollback plan here]

Final Verdict

• ALL PASS — All criteria met, all validators passed
• PASS WITH ITERATIONS — All criteria met after [N] rounds
• PARTIAL — [N] criteria unmet (see Acceptance Test)

---

## Phase 8: Knowledge Capture (Retrospective)

**Goal**: Extract and persist institutional knowledge from this task. Takes 30-60 seconds and compounds over time.

### Generalization Rule (MANDATORY)

All MEMORY.md entries MUST be **generic and reusable** — applicable beyond this specific task. Before writing:
1. **Strip specifics**: Remove file names, bug IDs, dates, one-time details
2. **Extract the principle**: What general rule prevents this *class* of problem?
3. **Test**: "Would this help on a completely different bug/feature?" If no, rewrite.

Exception: Architecture notes and file path documentation should stay project-specific.

### First-Principles Learnings
- **What fundamental insight did we gain about this problem domain?** (not just "I learned the API" — what underlying principle?)
- **Were any of our initial assumptions wrong?** Record them to avoid repeating the mistake.

### Pattern Extraction
- Did we discover a reusable pattern? → Update MEMORY.md (as a generic principle, not a specific fix)

### Pitfall Recording
- Did we hit non-obvious issues? → Record in MEMORY.md "Common Pitfalls" (generalized — "always check X before Y", not "file Z had bug W")

### Iteration Analysis
- What caused iteration loops? What validators caught and why? → Record to prevent repeat mistakes

### Codebase Observation
- Tech debt or improvement opportunities noticed outside scope? → Note briefly (don't fix — record for future)

### ADR Persistence (MEDIUM/FULL)
- If an ADR was generated during planning, verify it was written to `docs/decisions/` (or the project's ADR directory). If no ADR directory exists, note the decision in MEMORY.md instead.

---

## Execution Rules

1. **First-principles FIRST** — Phase -1 challenges the premise before committing compute
2. **Classify complexity early** — Phase 0.5 determines your entire execution path
3. **Build the context package** — Phase 1.5 is the biggest token saver. Do it well.
4. **Include the pre-loaded context instruction** in EVERY subagent prompt
5. **Include the diff inline** for EVERY validator — they should NOT re-run git diff
6. **NEVER skip user review** — Phase 3 is mandatory
7. **Smoke test before validation** — Build + test before launching validators
8. **Run independent agents IN PARALLEL** — Single message, multiple Task calls
9. **Code-simplifier runs ONCE at end** — Not after every implementation item
10. **Circuit breaker on iteration** — Same failure twice = stop and escalate
11. **Announce tier and agent count** — User should know the efficiency gains
12. **Document WHY, not just WHAT** — Alternatives and trade-offs in planning (MEDIUM/FULL)
13. **Capture knowledge** — Phase 8 retrospective updates MEMORY.md. Never skip.

## When to Recommend /staff-engineer Instead

If during Phase 0.5 you realize the task has:
- Fundamental architectural uncertainty that needs true dual-model debate
- 20+ file blast radius with cross-cutting concerns across multiple systems
- Multiple stakeholder concerns (security + performance + compliance + API design)

...recommend the user switch to `/staff-engineer` for the full treatment. But default to `/se` — most tasks are LIGHT or MEDIUM.

## Token Budget Targets

Phase -1 and Phase 8 add ~500-1000 tokens each (inline reasoning, no agents). Net impact is minimal.

| Tier | /se Target | /staff-engineer | Savings |
|------|-----------|-----------------|---------|
| Light | 30-50K | 200-250K | 4-7x |
| Medium | 45-75K | 250-300K | 3-5x |
| Full | 90-165K | 300-450K | 2-4x |