phni3j9a/axiom

v0.1.8MIT

Guidance-first Codex engineering with Luna MAX workers, Sol MAX design routing, fresh Sol XHIGH review, context isolation, and safe Git coordination.

Changelog

Added direct-spawn Sol MAX design workers for bounded tasks that require material visual, interaction, or information-design judgment.
Kept Luna MAX as the default for ordinary bounded work and for frontend implementation whose design decisions are already settled.
Defined unresolved interface judgment—not frontend file ownership—as the design-routing boundary.
Allowed Design Sol to implement UI when design and code iteration are inseparable, with mixed Luna lanes for disjoint non-visual work.
Kept independent review on a separate fresh Sol XHIGH and explicitly prohibited reusing a Design Sol as the reviewer.
Added routing examples, trigger evals, validation, and regression coverage for the new role separation.

Scoped same-reviewer continuity to materially stable user intent, acceptance, non-goals, and substantive design while leaving boundary-reset decisions with Main.
Clarified that reviewers provide independent evidence but do not set user risk tolerance or product policy, and that optional hardening should be distinguished from demonstrated failures and requirement gaps.
Clarified that task_complete is a per-turn return signal, not proof of a terminal agent session or Main acceptance.
Extended evidence-aware waiting guidance to shell sessions, tests, builds, benchmarks, and external processes without imposing a fixed polling interval.
Added a context-cost heuristic for overlapping Main/worker inspection without introducing an ownership prohibition.
Added qualitative rollout trace eval guidance and a read-only metrics script; trace signals inform judgment and do not gate agent behavior.

Removed the breaking hide_spawn_agent_metadata = false metadata visibility setting from current setup/config guidance.
Added a migration warning for v0.1.5 users to delete that setting because Codex CLI 0.147.0 returns HTTP 400 for the reserved collaboration.spawn_agent schema mismatch before the first model response.
Clarified that routing verification uses runtime/rollout evidence—the requested spawn args, child turn_context model/effort, and task_complete—rather than child self-report alone.
Added regression coverage for the schema/config change and rollout-based routing verification.

Removed the optional Axiom Dashboard introduced in v0.1.3.
Removed the Rust/Axum backend, React/TypeScript frontend, platform launchers, and explicit Dashboard Skill.
Removed Dashboard-specific Rust/Node CI and multi-platform binary release jobs.
Simplified Plugin packaging to a Core-only distribution.
Removed Dashboard/observability metadata, setup instructions, runtime documentation, and the stale validation report.
Preserved Axiom Core delegation, parallel Luna, same-reviewer continuity, Git safety, direct-spawn policy, and the v0.1.4 Luna economics principle unchanged.

Made the current Luna MAX economics assumption explicit: for ordinary Axiom orchestration decisions, Luna worker compute is treated as almost free.
Added the principle “Main context is expensive; Luna compute is almost free.”
Instructed Main not to suppress useful bounded delegation merely to conserve Luna tokens or model usage.
Shifted the practical spawn-cost model toward coordination, latency, dependency ordering, overlap/write conflicts, integration, and verification burden.
Strengthened proactive delegation and parallel fan-out while retaining natural task boundaries, no fixed worker count, and safeguards against artificial fragmentation.
Preserved the v0.1.3 Dashboard and existing Sol reviewer continuity policies unchanged.
Added validation/tests to prevent accidental removal of the almost-free Luna economics policy.

Bundled Axiom Dashboard v0.1.0 inside the existing Plugin package while keeping Axiom Core and Dashboard runtime responsibilities separate.
Added a Rust/Axum/Tokio read-only collector and local API with an embedded React/TypeScript UI.
Added live Agent Graph, parallel timeline, token and compaction metrics, review/Finding inspection, Git status, recent sessions, and evidence-based Axiom principle checks.
Added direct raw Rollout Trace projection with optional state.json enrichment; the Dashboard does not invoke the reducer or create execution state.
Added localhost-only defaults, explicit non-loopback opt-in, no telemetry, no hooks, and read-only Git inspection.
Added an explicit-only axiom-dashboard Skill so ordinary Axiom engineering requests remain unaffected.
Added platform launchers, binary overlay packaging, source packaging, CI validation, and multi-platform release workflow.
Preserved all v0.1.2 parallel Luna and same-Sol-reviewer policies unchanged.

Made safe parallel Luna MAX execution a first-class default when two or more useful bounded tasks are independent.
Added dependency-driven fleet sizing: Main chooses the natural worker count; Axiom has no fixed fan-out or concurrency target.
Instructed Main to spawn known-independent workers before waiting, avoiding accidental serial spawn -> wait -> spawn execution.
Strengthened proactive parallel read-only investigation and disjoint-write guidance.
Preserved safeguards against artificial task fragmentation, overlapping writes, unstable interfaces, and unnecessary intermediary orchestrators.
Kept the v0.1.1 same-Sol-reviewer continuity and Main adjudication policy unchanged.

Reuse the same Sol reviewer agent for all re-review in a review cycle.
Keep only the initial reviewer context fresh; subsequent checks use follow-up on the same reviewer handle.
Removed the maximum-five-findings rule.
Removed the fixed two-review-round limit.
Made Main adjudication and stable Finding IDs the convergence mechanism.
Allow genuinely new material findings when supported by changed evidence, while preventing rejected or deferred findings from being reopened without new evidence.

Initial Axiom release.