Claude Code's built-in review command runs on whatever model the session is set to

Explore caps at Opus. The general-purpose and Plan agents inherit with no cap ^[12]. In three weeks of one developer's launch log, general-purpose ran 231 times and Plan 10 times, so 241 of 576 launches, 41.8%, had nothing between them and whatever the session was set to ^[8][1]. His settings.json sets the default model to fable ^[4].

The obvious hardening was already in place. All 11 files in his ~/.claude/agents directory carry a model: line, one Fable, four Opus, five Sonnet and one Haiku, and a lint rejects any file that omits it ^[5]. Built-in agents have no frontmatter to write that line into ^[7]. So pinning governs the 203 launches that were not built-ins, 35% of the total ^[2]. The lint is not wrong; it just cannot see the built-ins.

One environment variable is the documented lever, and it is partial. Setting CLAUDE_CODE_SUBAGENT_MODEL changes general-purpose only, and Explore and Plan do not move ^[13]. On these counts that reaches 231 of the 241 uncapped launches and leaves Plan's 10 ^[3].

The same behaviour appears in two public issues. #76514, filed July 10, reported that omitting the per-agent model propagates Fable to every subagent, and it was closed as not planned ^[16]. #93894, filed September 12, reports that one run of /code-review high on Fable 5.1 exhausts the session budget and the review never finishes; it is open ^[17].

The leak has run the other way too. In February, --model opus applied only to the main loop and this developer's subagents ran on Haiku, which he had not intended; the behaviour has changed since ^[14]. "Inheritance defaults leak in both directions," he wrote ^[15].

The second half of the account carries less weight than the first. His plan gives Fable a usage limit separate from Opus's, so Opus keeps taking work after Fable is exhausted, and that is the path by which design decisions moved down a tier ^[3]. The over-engineering he reports is his own reading of his own repositories ^[2]. His launch log does not record which model each agent ran on, so he cannot say how many built-in launches billed Fable ^[10].

For his 63% to describe another setup, those sessions would have to reach for built-in search and planning agents at close to his rate over three weeks ^[9]. The fix he wrote that day was a convention: one line at the end of every implementation plan deciding whether that session implements, with heavy implementation handed to a new Opus session ^[18]. He rejected enforcing it with a hook ^[18].