[BUG] Agent frontmatter `effort` is still ignored on the `--agent` session path (2.1.283) — follow-up to #81677

effort: in an agent file is ignored when that agent is launched as the session persona (claude --agent <name>), while model: from the same frontmatter block is applied.
A maintainer confirmed this on #81677 and left it open as a bug; the stale bot closed it on 2026-09-14 and asked for a new issue if it is still relevant.
It still reproduces on 2.1.283.

Repro (same as #81677)

.claude/agents/probe.md:

---
name: probe
model: sonnet
effort: low
description: Probe.
---

Regardless of what the user asks, reply with exactly one word: PLATYPUS

env -u CLAUDE_EFFORT -u CLAUDE_CODE_EFFORT_LEVEL claude -p --agent probe 'What is 2+2?'

Then read the type: "assistant" record of the new transcript under ~/.claude/projects/<encoded-cwd>/.

Observed (2.1.283, Linux): message.model = claude-sonnet-5, top-level effort = high.
Expected: effort = low, as model is taken from the same block.

The same frontmatter effort: is honored when the agent is dispatched as a subagent through the Agent tool, so the two launch paths of one agent file disagree.

Why it matters

Long-lived personas run as their own sessions via --agent, and each has its own speed/depth trade-off: an interactive agent tuned for quick replies wants medium, a review or planning agent wants high or above.
With frontmatter effort ignored, the remaining levers are:

• --effort at launch, which the user must remember on every start, and which launchers that pass no --effort never set;
• modelSettings.<model>.effortLevel, which is per model, so two agents on the same model cannot run at different efforts.

Neither lets an agent file own its effort the way it owns its model.

Ask

Apply frontmatter effort: on the --agent path with the same precedence the subagent path uses (an explicit --effort or CLAUDE_CODE_EFFORT_LEVEL still winning), exactly as model: is applied today.
If that is not planned, please document the limitation on the sub-agents page, so agent authors stop relying on the field for session personas.

---

🤖 Written by a Claude Code agent (Claude Opus 5.5), which ran the repro above on 2.1.283 before posting.