The two failures do not share a mechanism, which is why one audit rule will not catch both. clio-coder's version is findable with a text search: the setting sits on three advertised surfaces and has no read site in the execution path the audit traced [19]. That is what issue #324 is in evidentiary terms, a set of file:line citations for what the code actually consults [3].
Codex fails from the other direction, because the key is read. According to the reporter's reading of rust-v0.153.4, `has_explicit_settings` returns true for any token-budget key other than two opt-outs. `TurnContext` construction therefore sets `use_model_token_budget_defaults` to false. `resolve_token_budget` then hands back the user-configured object directly, without merging the fields the user never set [11]. `TokenBudgetConfig::default` leaves `guidance_message`, the fallback prompt and the fallback buffer unset [12]. One scalar intended to move a reminder earlier therefore drops three model-provided fields [17], the ones that make summary-free rollover usable [13].
No search finds that, because the key is present in the read path and behaving as written. What found it was an A/B with a single difference and an observable downstream of the setting: render the prompt twice, and check whether the context window guidance block survives [10]. That is the transferable part. To show a setting is wired you need an artifact that changes when the setting changes, and if the only artifact is a token count buried in a later request, you will not see it.
The dev.to post that collected both instances names the class advertised but never wired, says it turns up everywhere once you look for it, and offers a detection checklist it puts at about ten minutes [16]. What the post documents is two issues in two codebases in one week [1]. That supports a shape to check for, not a rate. Nobody in the material has counted how many advertised keys across agent CLIs are never consumed, and the post does not claim to have.
The codex issue asks that whole-object replacement, if it is intentional, be exposed explicitly, with warnings or rejection for incomplete handoff configurations rather than silent acceptance [15]. That is the more expensive repair. Merging field by field changes behaviour for every config that already sets one token-budget key [11], and the commenter 84dnnvbdvp-debug notes the boolean is captured once at `TurnContext` construction, so a session that switches models keeps suppressing defaults belonging to a model it is no longer running [14]. clio-coder closed its gap by adding a read. Codex cannot close its gap without deciding what a partially specified config means, and that is a semantics change rather than a wiring fix.
