[5.0.0-beta.78] High cost and rapid quota consumption: use role-appropriate agent model defaults

[ThisCCL]

Summary

In actual use of 5.0.0-beta.78, I find the setup very expensive and my available quota drains quickly. All 12 bundled agent role templates explicitly select gpt-6-astra, including read-only search roles and the worker documented for small, single-file changes. The generated default role inherits the medium worker's same model. Please provide role-appropriate model defaults, or a supported profile that preserves differentiated model allocation.

This is a configuration/design improvement request motivated by real usage cost and quota consumption. The alternative mapping below is my requested policy, not a claim that LazyCodex already promises it. I have not isolated the exact share of consumption caused by model selection versus token volume or agent count, and I am not claiming a measured savings percentage.

Environment

• Affected version: LazyCodex/OMO 5.0.0-beta.78.
• Codex CLI: 0.155.1.
• OS: Windows; installed OMO plugin cache and user-level .codex/agents/*.toml.
• Latest source checked: 164d8bcfbab7c07abae84a6bee859389bb5403e0 (plugins/omo/package.json reports 5.0.0-beta.79); the same 12-role Astra allocation is still present.

Reproduction and observed behavior

Inspect the installed components/ultrawork/agents directory, or the corresponding directory in the published release:

rg -n '^(name|model|model_reasoning_effort|service_tier) =' plugins/omo/components/ultrawork/agents

Parsing the installed beta.78 TOMLs produced:

Bundled roles: 12
Models: ['gpt-6-astra']
explorer                              gpt-6-astra low  fast
librarian                             gpt-6-astra low  fast
lazycodex-worker-low                   gpt-6-astra high (unset)
lazycodex-worker-medium                gpt-6-astra high (unset)
lazycodex-clone-fidelity-reviewer      gpt-6-astra high (unset)

Before local correction, the corresponding user-level role files and default.toml also explicitly selected Astra. We changed only the six model fields listed below; all 14 user-level TOMLs still parsed successfully and all six configurations matched the requested mapping. This verifies the local configuration change, not an end-to-end inference benchmark.

Requested behavior

Use differentiated defaults matching each role's intended scope. The requested mapping is:

Role	Proposed model	Reasoning effort	Service tier
explorer	gpt-5.6-luna	low	fast
librarian	gpt-5.6-luna	low	fast
lazycodex-worker-low	gpt-5.6-luna	high	unchanged
lazycodex-worker-medium	gpt-5.6-sol	high	unchanged
default	gpt-5.6-sol	high	unchanged
lazycodex-clone-fidelity-reviewer	gpt-5.6-terra	high	unchanged

The concern is that the documented small/medium/large worker split and search specialization do not currently translate into different model choices. In my usage, the resulting overall cost and quota consumption are too high; simple searches and small edits should have an economical default option. Keep the stronger model for roles that need it, and expose the above allocation as a supported choice if maintainers prefer to retain the current defaults.

Source evidence and repository decision

Target: code-yeongyu/lazycodex. The allocation is explicit in LazyCodex's shipped role files, rather than evidence of Codex ignoring a configured model.

Related: #145 covers concurrency limits and token control. This request specifically concerns the model selected for each role.

Pinned latest-source references:

• Explorer model and effort.
• Small worker description and model.
• Medium worker description and model.
• Bootstrap derives default from the medium worker and copies bundled role files.
• Upstream comparison: Codex applies the role's explicit model override. This is source evidence for routing; a clean upstream inference session was not run.

The directly observed cause is the uniform model value in the bundled templates. Whether that policy was intentional is a maintainer decision; this report does not claim an accidental regression or a measured performance defect.

Proposed fix and verification

1. Update the five affected bundled templates under components/ultrawork/agents, or provide a supported model-allocation profile; derive default consistently from the selected medium-worker model.
2. Add a model-matrix check covering model, effort, and service tier independently, so role differentiation is deliberate and reviewable.
3. In an isolated Codex home, verify initial provisioning and repeated provisioning preserve the chosen allocation and generate default consistently.
4. Verify unrelated roles retain their intended configuration, and perform a small spawn smoke test for each changed role where the proposed models are available.

---

This issue or PR was generated by LazyCodex.
Tag: lazycodex-generated