ABI Superpower: AI
Maps AI completion, SEA learning, and training to the real CLI. There is no
/abi-superpower-ai binary or slash command.
Actions
complete
Run completion with optional streaming:
./target/debug/abi complete --model claude-fable-5 "explain the Rust workspace"
./target/debug/abi complete --learn "code review"
train
Train agent profiles against WDBX:
./target/debug/abi agent train abbey
learn
Run SEA self-learning loop:
SEA_STORE=$(mktemp -d)
ABI_WDBX_PATH="$SEA_STORE" ./target/debug/abi complete --learn "task"
stream
Stream completion tokens through the live Anthropic transport. Configure
Anthropic credentials first:
./target/debug/abi auth signin anthropic
./target/debug/abi complete --live --stream --model claude-fable-5 "write a function"
status
Show current AI configuration:
./target/debug/abi backends
Profiles
- abbey - Primary empathetic polymath: warm, creative, explanatory, and technically precise
- aviva - Direct expert: concise, candid, analytical, and action-oriented
- abi - Adaptive orchestration/governance: intent, risk, context, policy, and mode selection
These are deterministic local profile routes in the ABI Rust runtime. The
canonical product identity and Current/Partial/Proposed capability mapping live
in docs/spec/abbey-core-identity.mdx; the labels are not model-quality claims.
Implementation
Maps to:
crates/abi-ai/src/completion.rs - deterministic completion and persistence helpers
crates/abi-sea/src/learn_loop.rs - run_learn_loop, evidence recall
crates/abi-ai/src/constitution.rs - 6-principle audit
crates/abi-ai/src/router.rs - sentiment analysis, profile selection
Runtime Boundary
The Rust workspace always links the local AI/SEA paths. Live HTTP completion is
Anthropic-only and requires explicit credentials; local mode is deterministic
persona-template generation, not a production model-quality claim.