Replace the retiring GPT-5.3-Codex-Spark in the Codex live tests #317

Problem

On 2026-09-11 OpenAI announced that GPT-5.3-Codex-Spark retires "next week", without an exact date or time. As of 2026-09-15, model/list from Codex CLI 0.153.4 still lists gpt-5.3-codex-spark (not hidden, no upgrade target), and the account still reports the separate GPT-5.3-Codex-Spark rate limit.

Live coverage pinned to Spark stops working once the model is withdrawn:

• Browser live suite (npm run test:codex-live). LIVE_MODEL_POLICY in frontend/tests/live/codex-live-usage.mjs maps the spark scenario to gpt-5.3-codex-spark. Seven of the nine tests in frontend/tests/live/codex-task-loop.spec.js use it:
  • Through the model picker, where the availability check in chooseModel fails before any prompt is sent: "creates and resumes a real Codex task through Caffold with Spark", "uses the current plan convention across a real Codex resume", "names a new Caffold task at first-turn completion and preserves it", "moves one dirty Spark task into a worktree and resumes the same thread", "reconciles a managed Spark task through a second daemon client".
  • Without the picker, by passing the model id in their requests: "hydrates, orders, and restores Tasks through the local navigator ledger", "rechecks externally archived and deleted Codex Threads through explicit Recovery commands".
• Rust live checks (#[ignore]).
  • live_proxy_replays_pending_legacy_dynamic_tool_after_reconnect and live_proxy_replays_pending_approval_after_reconnect use SPIKE_MODEL in caffold/src/agent/codex/reconnect_spike.rs.
  • The live v4 migration check in caffold/src/app/startup_migration.rs starts its fixture turns on Spark and asserts that every migrated thread records gpt-5.3-codex-spark. The v4 rows written by write_v4_test_store in caffold/src/task_store/migration.rs carry the same model id.
• Policy description. docs/development/testing.md and the policy contract test in frontend/tests/contracts/codex-live-usage.test.mjs name Spark as the model for Spark scenarios.

Goal

The Codex live scenarios that use Spark run on gpt-5.6-luna with low reasoning. Luna is the only visible model the Codex catalog describes as affordable, and the Fast-mode and multimodal scenarios already use it.

Notes

• Spark usage counts against its own GPT-5.3-Codex-Spark limit. The account reports no other model-specific limit, so Luna runs draw from the general Codex weekly limit shared with everyday Codex use.
• Fixtures that mention Spark only as data do not call the model: caffold/src/agent/codex/status.rs, caffold/src/agent/codex/contract.rs, frontend/tests/e2e/tasks/model-picker.spec.js, and the usage-report fixtures in frontend/tests/contracts/codex-live-usage.test.mjs.

Open question

Four other Rust live checks that submit model turns do not use Spark. live_codex_clarification_returns_to_chat_and_accepts_a_reply and live_codex_mcp_tool_approval_uses_normalized_contract use gpt-5.6-sol at low. live_codex_mcp_survives_backend_replacement_and_runtime_reattachment and live_codex_mcp_survives_an_isolated_app_server_restart pass no model and follow the local Codex default. Whether they also move to Luna at low is undecided.

Acceptance criteria

• No browser live scenario, Rust live check, or live model policy entry uses gpt-5.3-codex-spark.
• Those scenarios use gpt-5.6-luna with low reasoning.
• docs/development/testing.md and the policy contract test match the new policy.
• npm run test:codex-live and the moved Rust live checks pass.

References

• Retirement announcement, 2026-09-11: https://x.com/thsottiaux/status/2098300998968357218 (as cited in feat(codex): prepare Codex Spark retirement and remove its quota UI lidge-jun/opencodex#4334)
• https://www.orcarouter.ai/blog/gpt-5-3-codex-spark