Axiom v0.1.8
Axiom is a guidance-first Codex plugin for non-trivial software engineering.
Sol thinks and designs. Luna executes. Sol reviews.
Main context is expensive; Luna compute is almost free.
- Main Sol owns intent, architecture, design direction and constraints, integration, and acceptance.
- Axiom v0.1.8 preserves the v0.1.4 economics principle that ordinary Luna MAX worker usage is treated as almost free for orchestration decisions.
- Direct-spawn Luna MAX performs bounded exploration and implementation; useful spawns should not be suppressed merely to conserve Luna usage.
- Direct-spawn Sol MAX performs bounded work that requires material visual, interaction, or information-design judgment; frontend files alone do not trigger this route.
- A Sol MAX design worker may implement its design when the feedback loop is inseparable, but it is never reused as the independent reviewer.
- Independent useful bounded work fans out to parallel Luna MAX workers; there is no fixed fleet size, and coordination/integration cost—not Luna token cost—limits fan-out.
- Fresh direct-spawn Sol XHIGH performs meaningful independent review.
- The initial Sol reviewer is reused for re-review while the review boundary remains materially stable; Main adjudicates every finding and any boundary reset.
- Reviewers provide evidence but do not set user risk tolerance or product policy.
task_completeis treated as a child-turn return signal, not Main acceptance or terminal agent state.- Evidence-aware wait and optional rollout metrics inform Main without imposing fixed cadence or workflow gates.
- No fixed workflow, no Terra default, and no custom agent installation.
allow_implicit_invocation: truemakes the core Axiom skill available proactively.
See the repository root README.md for installation and configuration.
Codex CLI 0.147.0 migration: v0.1.5 users must delete hide_spawn_agent_metadata = false from their existing configuration. Leaving it in place causes an HTTP 400 reserved collaboration.spawn_agent schema mismatch before the first model response.
Routing verification uses runtime/rollout evidence—the requested spawn args, the child turn_context model/effort, and the corresponding child-turn task_complete—and never a child self-report alone. A returned child turn is not by itself Main acceptance or a terminal agent session.
Optional read-only rollout metrics:
python3 skills/axiom/scripts/audit_rollout.py /path/to/rollout.jsonl
Metrics support qualitative evaluation; they do not gate or control Main.