| name | ai-plan |
| description | Plan abi AI work — model routing, completion, SEA self-learning, and training. Use when the user asks to plan AI orchestration, choose a model, run a completion, drive the SEA learn loop, or schedule training. Routes to complete-base, sea-learn-loop, agent-plan-train, and the abi-superpower-ai/sea/constitution superpowers. |
ai-plan
Entry point for abi's AI subsystem (src/features/ai/). Routes to the
specialist skills instead of duplicating them:
| You want to… | Use |
|---|
Smoke-test base local completion (abi complete) | complete-base |
Drive the SEA self-learning loop (abi complete --learn) | sea-learn-loop |
Plan + train an agent (abi agent plan / train) | agent-plan-train |
| Deep-dive AI routing/constitution/SEA | abi-superpower-ai, abi-superpower-sea, abi-superpower-constitution |
| Toggle SEA mode in the REPL | sea-learning-controller (/learn) |
Facts that constrain any AI plan (from source, not marketing)
- Default model
claude-fable-5; src/features/ai/models.zig is the single
source of truth for ids/aliases/provider routing (mod/stub parity).
- Router
selectBestProfile ties resolve abbey > aviva > abi; neutral input
routes to abi. analyzeSentiment is prefix-only keyword matching.
- Constitution audit is observability-only, not a gate — sets
audit_passed / audit_vetoed / escore; safety+privacy hard-veto only
when either scores < 0.5. Matching is case-insensitive substring (infix),
so "harm" fires on "harmless" — it cannot detect novel harm patterns.
- EMA weights persist only on the
--learn/SEA path; plain complete
re-runs sentiment each turn with no EMA persistence.
Honest boundary
No energy-efficiency, accuracy, or QPS claims without a repo test/benchmark
proving them (see docs/contracts/external-claims-audit.mdx).