Models and effort

/emet picks a model and a reasoning effort per role automatically; there is nothing to switch by hand mid-run. Every key is in the configuration reference.

Role	Model	Effort	Set by
Planning (Phases 1–4)	Opus, via `opusplan`	your session’s	you, once
Coordinator, Phase 5 on	Sonnet, via `opusplan`	your session’s	you, once
Workers (Phase 7)	Sonnet	`medium`	`agents/emet-worker.md`, or `models.worker`
Review, `passive` tier (docs-only diff)	none	—	`emet assess`
Review, `medium` tier (ordinary source change)	Opus	`medium`	`models.reviewer.medium`
Review, `high` tier (risky paths or sinks)	Opus	`medium` by default	`models.reviewer.high`
CI-log triage (Phase 8)	Haiku	—	`models.triage`

The two agent definitions (installed by install.sh to ~/.claude/agents/) pin their own model/effort in frontmatter, so the coordinator’s own model/effort choice never leaks into a worker or the reviewer. emet doctor reports (never blocks on) either file going missing. emet preflight does block dispatch when emet-worker or emet-reviewer is missing, because Phase 6 and Phase 8 dispatch by subagent_type.

Risk-tier routing

Phase 8 runs emet assess --base <trunk> --head <integration branch>, which classifies the diff deterministically and prints JSON with the tier, the reasons and the recommended reviewer model/effort:

passive: docs-only diff; no reviewer is dispatched.
medium: the reviewer runs on models.reviewer.medium.
high: the diff touches auth, security, payments, webhooks, CI, the installer, lockfiles, migrations, a dangerous sink (raw SQL, pipe-to-shell and the like) or a path matching review.highRiskPaths. The reviewer runs on models.reviewer.high, and auto-merge lands the PR only with green checks and no open review finding.

A ticket that touches high-risk paths can carry a Model: line (for example Model: opus or Model: opus/high), which overrides models.worker for that worker only.

Haiku triage

The coordinator never reads a CI log whole. A failing check is handed to an agent on models.triage (default haiku), which returns at most about 2k tokens with file:line pointers for the fix.

Config

Every key is optional, and the defaults reproduce the table above:

{
  "models": {
    "worker": { "model": "sonnet", "effort": "medium" },
    "reviewer": {
      "medium": { "model": "opus", "effort": "medium" },
      "high": { "model": "opus", "effort": "medium" }
    },
    "triage": "haiku"
  },
  "review": { "highRiskPaths": [] }
}

review.highRiskPaths is a list of extra case-insensitive path substrings that force the high tier. To get more scrutiny on risky waves, raise models.reviewer.high.effort (for example to high).

Why a single reviewer

Why a single fresh-context reviewer instead of a multi-angle fan-out: no public benchmark compares the two directly, but a fan-out re-reads the whole diff once per angle and has hung for over an hour in this repo, while Anthropic and third-party reports on Opus-tier models suggest the model upgrade buys as much as the extra angles would. For a high-stakes wave, run /code-review ultra by hand afterward; /emet never dispatches that on its own.

What you set once, yourself

/emet can’t switch these mid-run:

model: "opusplan" in ~/.claude/settings.json: this is what makes Phases 1–4 run on Opus and Phase 5 onward run on Sonnet, with no /model needed. emet doctor warns (never edits your settings) if it’s unset.
A session effort of medium or high rather than xhigh: the workers and reviewer already carry their own effort regardless of yours, and a high global default mostly adds cost to the coordinator’s own turns (dispatch bookkeeping, merges, reports) that don’t need it.
permissions.defaultMode able to run unattended (auto, acceptEdits, or bypassPermissions), plus permissions.deny rules blocking git push * main|master and git push --force *. emet doctor warns on both, since a wave that stalls on a permission prompt or that can force its way past the trunk branch defeats the point of running unattended.

These are starting points, not settled numbers. Opus 5.5 launched 2026-09-22, and effort/model choice for coding agents is still an active area; re-tune after a few real waves rather than treating this table as final.