The 6 Core Superpower Skills

The meta-dispatcher invoked at session start, turn start, and complex task decomposition.

Unguided agents jump directly into modifying files without evaluating task ambiguity. jev-using-superpowers inspects prompt context and repository status in 70ms to select the required workflow:

• Brainstorming: If user requirements contain ambiguities, open-ended architecture questions, or multi-library tradeoffs.
• Planning: If scope touches more than 2 files or requires new crate/npm dependencies.
• Executing: When an approved plan exists, driving disciplined TDD steps with pre-commit gates.
• Systematic Debugging: When tests fail, logs crash, or an unexpected runtime error occurs.

# Session-start hook triggers router evaluation:
jev-axi eval "Route workflow for: 'Refactor SQLite connection pool to support WAL mode'" \
  --options "brainstorm,plan,execute,debug" \
  --context "$GIT_STATUS"

Structured architectural design divergence and trade-off evaluation before code generation.

Forces agents to propose 2 to 3 distinct implementation approaches (e.g. Minimal Stdlib, High-Throughput Batching, or External Engine). Instead of the agent picking arbitrarily based on token probabilities, it executes jev-axi pick:

# Rank 3 design approaches against project constraints:
jev-axi pick "Select optimal cache layer" \
  --candidates "approaches.json" \
  --criteria "Memory budget < 50MB, zero new native deps, sub-millisecond lookups"

Output yields a typed decision with confidence score. If confidence < 0.80, the agent is barred from proceeding without explicit user confirmation.

Implementation plan generation paired with speculative dependency vetting.

Coding agents frequently hallucinate non-existent package names or select abandoned, vulnerable libraries. In jev-writing-plans, every proposed third-party crate or npm package is vetted through jev-scout:

# Scout real open-source crates with Jev System One fan-out scoring:
jev-scout "fast zero-copy trigram search"

# Returns:
# 1. tgrep (score: 9.8, verified real, 12ms cold start)
# 2. fst (score: 9.1, standard finite state transducer)
# Hallucinated entries scored 0.0 and purged from prompt memory.

Step-by-step TDD execution with runtime command guards and git pre-commit reflex checks.

Executes one bite-sized task at a time. Before running shell commands, jev-guard filters out accidental rm -rf, broad find-deletes, or production leaks. Before creating git commits, git-jev gates the diff:

# Shell guard halts destructive commands:
jev-guard "git checkout -f master && git clean -fdx"
# [BLOCKED] Destructive command detected · confidence: 0.99 · requires manual confirmation

# Pre-commit gate validates staged changes:
git jev check
# Evaluates diff against architectural invariants in 110ms.

4-phase root-cause diagnosis without prompt context bloating.

Dumping raw 2,000-line compiler errors or stack traces into agent context causes context rot and hallucinated explanations. jev-systematic-debugging pipes terminal output directly through Jev triage:

# Filter noisy build failure into root cause and exact failure location:
cargo test 2>&1 | jev-axi triage

# Output:
# [ROOT CAUSE] Lifetime mismatch: borrowed value 'pool' does not live long enough
# [LOCATION] src/db/pool.rs:142:18
# [CONFIDENCE] 0.96

Physical gate barring agents from declaring completion without empirical verification.

Agents love to say "All done! I have implemented the requested feature." while unit tests fail or code coverage drops. jev-verification binds to the agent harness's Stop hook via limpet:

# Stop gate executes when agent attempts to terminate turn:
limpet check

# Checks:
# 1. Zero compiler errors
# 2. All unit and integration test assertions green
# 3. Supercov quality score >= 0.85
# If any check fails, limpet exits 1 and feeds failure back to the agent runtime.