A portable Codex setup: GPT-6.1 Sol medium coordinates, GPT-6 Luna performs most delegated work, and one Sol xhigh advisor supplies independent critique when warranted. The root owns final engineering approval, Git, and external actions.
Keep this repository generic across personal and work devices. Machine-local workspaces, tenant configuration, personal/business data, and private verification reports belong outside the repository.
AGENTS.md: concise working agreements and plan-completion persistence.config.tomlandagents/: explicit model routing, a shared six-thread ceiling, and delegation up to absolute depth 2.skills/delivery-orchestration/: one planning/delivery workflow and the routing matrix.skills/adversarial-code-review/: one Sol xhigh critic for approach or delivery review; ordinary work uses Sol root review.skills/instruction-learning-loop/: discretionary source-backed instruction maintenance and an explicitly invoked audit.hooks/plan_gap_goal_hook.pyandhooks.json: accepted-plan goal persistence.
Sol defines the outcome and delegates useful independent work to Luna while continuing integration or decisions. Use focused checks for the changed behavior, then review the integrated result. Avoid mandatory ledgers, repeated review stages, and instruction edits that do not address a durable defect.
Luna medium handles discovery, xhigh bounded routine implementation, and max substantial work. Sol low remains an optional speed route for clear critical-path tasks. Sol low handles bounded work needing stronger judgment; Sol high handles difficult implementation; Sol xhigh handles consequential independent review. Prefer Spark for tiny exact checks, bounded instruction checks, and mechanical edits, using the separate allowance reported by the user. Its small context requires fresh packets, narrow reads, and short returns; split or reroute growing work before compaction or a stall. Choose using total completion cost, including handoffs and rework.
Root is depth 0. Luna max and Sol high workers may subdivide independent owned work at depth 1; all agents at depth 2 are terminal. Scanners, fast workers, and Spark are terminal at depths 1–2. Only root dispatches the terminal depth-1 advisor.
All descendants share max_concurrent_threads_per_session = 6; the root is
excluded. This is a ceiling, not a quota. Parent writers suspend edits to child
paths until ownership returns. Children never gain broader authority. Every new
assignment uses a fresh self-contained packet; do not fork threads. Continue or
resume only the same assignment within its 30-minute cache window, then send a
new packet.
The working topology is root → workstream worker → terminal leaf.
Keep max_depth = 2 for runtimes that honor it and apply the same ceiling in
instructions. This matches the demonstrated client behavior without claiming
a universal product maximum. If a worker cannot spawn, it returns the packet
to root for direct dispatch.
See official subagent configuration.
The goal hook is intentionally retained for large accepted plans. A leading
Implement the plan (including the supported polite variants) starts a goal
for full implementation and a final gap analysis when no unfinished goal
already exists. The matching AGENTS.md fallback covers clients that do not
dispatch UserPromptSubmit. Scoped exclusions and explicit reversals remain
binding. Ordinary non-plan work does not automatically create a goal.
Keep exactly one plan handler in the effective hook configuration. Preserve
other installation-specific hooks; do not blindly overwrite a mixed hooks file.
After hook changes, inspect /hooks and use a fresh session to verify activation.
Merge the relevant settings into your user configuration, preserving personal providers, MCP settings, authentication, and project overrides. Install the selected agent profiles and skills alongside it. Configuration layering is described in the official guide.
When upgrading to this routing, replace the previous flagship worker, low worker,
and advisor registrations and files with sol_worker, sol_low_worker, and
sol_advisor. All Sol profiles use gpt-6.1-sol; all Luna profiles use
gpt-6-luna. Root and plan mode both use medium effort. Start a fresh session
to load the updated agent registry and defaults; running sessions retain their
current model. To adopt this setup when resuming a saved session, explicitly
select Sol 6.1 medium and verify the loaded profiles before new delegation. Spark profiles remain pinned to their
existing model and are eligible only when that model is available in the runtime.
The sample disables the redundant Superpowers methodology plugin. Keep specialized domain and artifact skills that supply needed APIs, formats, or validation. The remaining vendor reconciliation guidance resolves scope and approval conflicts without bypassing technical safeguards.
When upgrading an older kit, remove retired agent profiles,
the plan-review-ladder skill, and instruction-learning hook registrations.
Legacy adversarial lifecycle/replay tooling is no longer part of the default
kit. Preserve historical records and check installation-specific callers
before removing old files.
Invoke noninteractive tools directly with no-window child process APIs where
needed. Preserve the user's interactive terminal and focus. Do not use
powershell -WindowStyle Hidden on a shared console. Preserve streams and exit
codes, and terminate only task-owned child shells.
Run the focused routing, plan-hook, and repository contract tests after changing their respective surfaces. Use the instruction audit when maintaining the instruction system. Tests should validate useful behavior and compatibility, not enforce a document's wording or recreate removed workflow ceremonies.
Reusable optional examples under integrations follow the same portability boundary. No private machine report belongs in this kit.