Model routing and efficiency setup
Implemented task-level model assignments under the user’s delegated choice.
User direction and implementation
The user asked for Luna, Sol and Astra roles, delegated model selection, and prioritized efficient token use.1 These assignments are the manager’s implementation choices under that direction, not measured project performance claims.
Current role settings are maintained in .codex/agents/*.toml; the checked role table and routing policy are in harness/model-routing.md (repository-relative paths). This record retains rationale and dated history rather than a second editable copy of current model settings. The synchronization change responds to the user’s hardening request.6
Official documentation positions Luna for focused work, Sol for balanced demanding work and Astra for the hardest tasks.2 Custom agent files support explicit model and reasoning settings; documentation suggests high for Luna and low for Astra as starting points.3 Task-to-role assignments are our policy, not an official prescription.
Cost discipline
The user’s follow-up supports Sol implementation, optional Luna assistance and clear task settings.5 The separate harness-builder role and maintain-harness skill are implementation choices under that direction. Routine bounded organization, templates, skills, configuration and tooling work defaults to Sol; Astra owns rule design, consequential ambiguity, substantive policy fixes and final acceptance. Small mechanical tasks need no preliminary Astra consultation. Independent Astra review of substantive Sol implementation satisfies harness review without an automatic additional Sol pass. The canonical maintained policy is harness/model-routing.md; this replaces the earlier Astra-only implementation rule while retaining its historical record below.
Use scripts first, optional Luna only for useful bounded gathering/preprocessing, and compact contracts with exact ownership, constraints, checks and done criteria. Resource bounds and stop conditions are explicit; hard user-approved limits are distinct from soft context/output targets. Missing intent/data calls for clarification or evidence rather than a larger model. Manager coordination, shared index/log ownership, Git checkpoints and worker nondelegation remain required.
Configuration lives in .codex/agents/; the maintained policy is harness/model-routing.md. Skills remain model-independent; coordinate-work explains dispatch and compact handoffs; maintain-harness supplies the harness task contract. Existing research/intake/archive/evaluation skills now link to this routing policy.
The offline scripts/workspace.py helper offers status, packet, route, and check --tests. It does not call models, meter actual tokens, enforce a billing ceiling, or change an already-running chat’s model. Character and return-length targets reduce context overhead without claiming measured savings.
Bounded execution refinement — 2026-10-09
Sol implemented the assigned routing, role and skill refinement. Independent Astra review accepted the implementation after clarifying the mechanical/substantive classification in the canonical routing policy: routine implementation can still require substantive review, and uncertain cases receive that review. The review checked role boundaries, task contracts, resource limits, approval gates and ownership; no blocking harness findings remain. Astra reran all eight role TOML parses, both changed/new skill validators and workspace.py route harness-builder, which returns gpt-6.1-sol at medium effort. Live custom-role discovery was not tested. No scripts changed, so no additional unit tests were run.
The read-only knowledge validator found three missing local targets in the concurrently owned IDEA-002 assessment/completion plan (interface-contract.md and experiment-plan.md). No idea files were edited or relocated. The manager owns the shared index/log updates, final validation and Git checkpoint. Protected skill/role writes required sandbox escalation, which was approved and executed.
Evidence and validation scope
Official pages were captured as HTTP HTML with extracted text; relevant model/agent passages were inspected. Linked assets and dynamic rendering are not included. An earlier failed network attempt is retained alongside each successful capture. API prices are not used to predict Codex account charges.
Prior setup validation (before the Astra ownership update): all six role files parse as TOML; all five skill validators pass; 23 offline tests pass; knowledge validation passes. Independent Sol review found a manager write-permission mismatch, now corrected to workspace-write with ownership limits. Runtime custom-role discovery and Luna/Astra launches were not exercised; the generic Sol worker/reviewer path was exercised. No model calls were made merely to test availability.
Astra ownership update — 2026-10-09
Astra implemented and accepted this policy update. Seven role files parse as TOML, the changed coordinate-work skill passes its validator, the workspace role helper returns harness-maintainer as Astra/low, and knowledge validation passes. An independent Sol review found no blocking routing, approval-boundary or ownership defects; it was a read-only review of the assigned policy/configuration changes. No scripts changed, so no new test run was required. Live custom-role discovery was not tested. The manager owns the subsequent index rebuild and Git checkpoint.
Automatic approval review rejected an initial broader configuration batch before execution. The accepted revision preserves existing approval gates and routes helper requests through the manager, retaining the worker no-recursive-delegation rule.
-
Preserved user message. ↩
-
Model catalog captured 2026-10-10 UTC, Choosing a model and flagship model sections. ↩
-
Subagent documentation captured 2026-10-10 UTC, model choice, reasoning effort and custom-agent configuration sections. ↩
-
Preserved user direction; the dedicated role, low default effort and manager-mediated helper dispatch are implementation choices.
-
Preserved user follow-up. Role, skill, review routing and contract details are implementation choices; no new research or external-action authority is implied. ↩
-
Preserved user direction to keep harness changes synchronized. Historical records below/above describe their original checkpoints; they are not the current role inventory. ↩