Agent tool ignores `effort:` frontmatter in subagent .md files — inherits global effortLevel instead

Per-agent effort is also an accounting problem.

Each Agent spawn should preserve four fields: agent file path/hash, declared effort from frontmatter, inherited global effort, and effective effort used for the child session. If the child inherits because frontmatter is unsupported or overridden, that reason should be explicit.

That gives cost tooling a clean explanation for "cheap scout ran at xhigh" cases, and it gives users a regression target: parent high + agent low should either run low or report the mismatch before work starts.

---

_Generated with ax._

Still reproduces on 2.1.215 (macOS). Two data points that make this easy to confirm, plus a second affected surface:

1. Transcript-level proof (frontmatter path)

Since 2.1.212 session transcripts record the reasoning effort on each assistant message, so no eyeballing is needed:

• Agent definition: ~/.claude/agents/coder-low.md with model: opus, effort: low (key/value per the sub-agents docs, which say frontmatter effort "overrides the session effort level").
• Session effortLevel in settings.json set higher.
• Spawn the agent via the Agent tool with a trivial prompt, then grep its transcript under <project>/<session-id>/subagents/agent-*.jsonl:

$ grep -o '"effort":"[a-z]*"' agent-<id>.jsonl | sort | uniq -c
   1 "effort":"high"     # = parent session's level, not the frontmatter's "low"

The child records exactly the parent session's effort on every assistant message; the frontmatter value never reaches the request.

2. The Workflow tool's per-call effort opt is dropped too

agent(prompt, {agentType, effort}) documents an effort option, but it's silently ignored the same way. A/B probe: same prompt to the same custom agent at effort: 'low' vs effort: 'max' — a mental-arithmetic task with "verify by an independent method before answering" (48273 × 91457):

| probe | recorded effort | thinking chars | output tokens | answer |
|---|---|---|---|---|
| effort: 'low' | high (inherited) | 0 | 6 | wrong |
| effort: 'max' | high (inherited) | 0 | 6 | wrong (different) |

Identical recorded effort, identical zero-thinking behavior at both extremes — a genuine max would not blurt a wrong 10-digit product with no thinking. So the override is dropped before the request, not just mislabeled in the UI.

Note: the Workflow side was reported in #64033 but auto-closed as a duplicate of #43083, which is a feature request. Since agent()'s effort opt is already documented in the tool schema and silently dropped, it's the same bug as this issue, not a missing feature.

Impact

Any routing setup that pairs a big model with low effort for cheap mechanical work (a common cost/latency pattern) silently runs every such subagent at the session's full effort instead.

I can confirm this on Windows using Claude Code 2.1.219