Overriding reminder_threshold_tokens silently drops model-provided handoff guidance and fallback

What version of Codex CLI is running?

Reproduced with codex-cli 0.153.1 bundled in the macOS app and the official standalone codex-cli 0.153.4 release binary. The latter was downloaded into a temporary directory and its SHA-256 was checked against the GitHub release asset digest.

What subscription do you have?

ChatGPT subscription authentication, using the Codex backend; not API-key authentication or a custom provider.

Which model were you using?

gpt-5.6-luna. Its model catalog supplied nonempty token-budget guidance and fallback prompts. At the time of testing, the catalog specified reminder_threshold_tokens = 6144 and auto_compact_fallback_buffer_tokens = 16384.

What platform is your computer?

Darwin 25.6.0 arm64 arm.

What issue are you seeing?

With experimental context management enabled, changing only features.token_budget.reminder_threshold_tokens disables inheritance of the other model-provided token-budget settings. A user trying to receive an earlier handoff reminder can therefore silently lose the handoff guidance and fallback configuration that make summary-free context rollover usable.

This is a configuration-resolution issue, independently observable without filling a context window or sampling a model response.

What steps can reproduce the bug?

Preconditions: use eligible ChatGPT authentication, a model whose catalog supplies token-budget guidance, and no existing detailed [features.token_budget] overrides.

Run these two prompt-rendering checks. The only configuration difference is the reminder threshold. The filter emits a boolean, not private prompt contents.

codex -m gpt-5.6-luna \
  -c features.context_management.experimental_mode=true \
  debug prompt-input 'Configuration merge probe.' \
  | python3 -c 'import sys; print("<context_window_guidance>" in sys.stdin.read())'

codex -m gpt-5.6-luna \
  -c features.context_management.experimental_mode=true \
  -c features.token_budget.reminder_threshold_tokens=14000 \
  debug prompt-input 'Configuration merge probe.' \
  | python3 -c 'import sys; print("<context_window_guidance>" in sys.stdin.read())'

Observed on both tested versions:

Both checks ran against the same local configuration and model selection. The configuration file hash was unchanged before and after each A/B pair. No long-running task or forced rollover is needed to reproduce this prompt-level difference.

What is the expected behavior?

A scalar timing override should retain unspecified model-provided guidance, reminder text, fallback prompt, and fallback buffer. Changing how early to remind the agent should not implicitly opt out of the rest of the handoff mechanism.

If whole-object replacement is intentional, please expose that explicitly and warn or reject incomplete handoff configurations instead of silently accepting this combination.

Additional information

The code explains the result:

1. has_explicit_settings returns true for any token-budget key other than enabled and use_history_notes_extension.
2. TurnContext construction then sets use_model_token_budget_defaults to false.
3. resolve_token_budget returns the user-configured object directly rather than merging unspecified fields with the model defaults.
4. TokenBudgetConfig::default leaves guidance_message, auto_compact_fallback_prompt, and the fallback buffer unset. The reminder template also falls back to a generic built-in template rather than the model-provided one. Thus a reminder can still exist while the richer guidance and fallback are absent.

The same resolution path is present in the inspected rust-v0.153.4 source and public main.

A separate small-window synthetic subagent test showed the practical risk: with a threshold-only override and no successful checkpoint, automatic rollover discarded the task state and the agent asked what to work on. A later test retaining the full model defaults and explicitly checkpointing early completed two rollovers correctly. Those two task runs changed additional conditions, so they are supporting context, not a single-variable causal experiment. The prompt-rendering A/B above is the isolated reproduction.

Related but broader reports: #36642 (task loss after automatic compaction) and #42449 (handoff requirements without available tools). This report isolates the partial-configuration merge behavior even when the history/notes backend is available. A regression test for a model with nonempty defaults plus only a reminder-threshold override would cover this case.

Start with codex-rs/core/src/session/token_budget.rs and turn_context.rs, then inspect TokenBudgetConfig::default in codex-rs/core/src/config/mod.rs. Run the two debug prompt-input checks from the issue and add a regression test for model defaults plus only a reminder-threshold override. Done means the override changes the threshold while retaining model-provided guidance, fallback prompt, and fallback buffer.