Claude Code subagents silently inherit opus; echo task hits 101,418 tokens

The engineer runs a blog and an X account on Claude Code from GitHub Actions on a VPS, calling claude -p on a schedule. The audit began when a review pointed out that his model selection was not actually differentiated. He first counted every path that calls Claude — an improvement loop every 4 hours, article writing twice a week, research, chat, and a smoke test — and found three of them left the model unspecified. The subagent definitions had the same problem: of six, only two named a model.

The immediate symptom was practical rather than theoretical. Records from the improvement loop's stop-reason file showed eight consecutive runs from September 20 and 21 all hitting usage_limit, leaving the loop stuck for two days. The engineer is careful not to assert that opus alone caused this, since the flat-rate quota is shared with human sessions and other paths — but he notes that running cheap roles on an expensive model does bring the cap closer before the roles that truly need opus get their turn.

He also documents a hypothesis he nearly reported without testing: that Task was missing from --allowedTools and subagents were structurally impossible to call. A measurement showed subagents spawned and completed with zero permission denials, so the real problem was that nobody was calling them, not that they couldn't be called. The stakes in the rest of the piece flow from that distinction — the fix is a prompt change, not a permissions redesign — and from his choice to keep article writing, public post writing, and the smoke test on opus while pushing no-judgment roles down to haiku.