[FEATURE] /code-review: thread a model parameter into the high+ effort review workflow

Resolved 💬 1 comment Opened Jun 25, 2026 by teddyyo-sysfeather Closed Jun 25, 2026

Summary

At high/xhigh/max effort, /code-review runs a workflow-backed multi-agent review. The root cause that the review model can't be controlled is in that workflow: it consumes only level + target, and every reviewer agent() call omits a model override — so all finder/verifier subagents silently inherit the session model. Please thread a model parameter through this workflow and expose it as /code-review … --model <name>.

Root cause (from the generated workflow)

The harness-generated workflow at high/xhigh/max:

  1. Metadata: "Workflow-backed code review … Launched by the /code-review skill at high, xhigh, or max effort when workflows are enabled. Pass args as <level> [target]."
  2. Args parsing: only LEVEL (effort) + TARGET (PR/branch/path/free-form scope). No model field anywhere.
  3. Agent spawns: every reviewer/verifier agent() call omits a model override → inherits the resolved session model. effort maps to API reasoning effort only — orthogonal to model.
  4. Free-form target only scopes what to review; it can't rebind the agent model. Model is bound at subagent instantiation, before any prompt is read.

→ Net: at the exact effort tiers where the heavyweight multi-agent review runs, there is no lever to fix the reviewer model. The only workaround is switching the whole session model via /model first — session-wide, not per-invocation.

Ask

  • Thread a model param into the high+ review workflow (its agent() calls take it as opts.model).
  • Expose it on the command: /code-review <effort> --model <name>. Precedence: --model > session model.
  • Ideally extend to /code-review ultra too.

Use cases

  • Skills/tools driving /code-review that need a fixed, reproducible review model regardless of operator session.
  • Reviewing on a cheaper/faster (or stronger) model than the authoring session without globally switching /model.

Related

  • #67559 — declarative cross-model (different-from-parent) policy (broader goal; this is the simpler explicit-pin case)
  • #69232 — /code-review command/docs collision

Environment

  • Claude Code CLI; observed via locally generated code-review-wf_*.js workflow scripts at high/max effort.

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗