Critic options

Options under the top-level critics key — the critic harnesses a session may hand its work to for a second opinion (issue-108, decision-043), run by the-loop critic and read by the review loop.

Declared here, in the operator's config, since issue #352 (decision-123): a critic is a harness and a model installed on the machine that runs the round, and executable configuration belongs in a file no pull request to a repository can edit. How many rounds run is the reviews block beside this one (moved here from the harness config in the same change); which critics exist is yours.

critics:
  - name: cursor-gpt          # a harness the-loop has an adapter for needs nothing else
    harness: cursor           # built-in: claude | cursor
    model: gpt-5.5
  - name: aider-review        # any other CLI: the executable and its argv, spelled out
    harness: aider
    model: gpt-5.5
    command: aider            # argv[0] — NOT a shell line
    args: ["--message-file", "{promptFile}", "--model", "{model}", "--no-auto-commits"]
    outputFormat: text
    timeoutSeconds: 900

This is executable config

A critic entry becomes an argv the CLI spawns, with your credentials in scope. Review one the way you would review code, and never put a secret in env — name a variable and keep its value in env.file or the ambient environment.

Each entry's keys

critics[].name

• Type: string
• Default: none — required, unique across the list

The critic's id: the-loop critic run <name> runs it, and the review loop records rounds under it.

critics[].harness

• Type: string
• Default: ""

The critic harness/CLI. Free-form so any harness works — it is also the [<harness>/<model>] attribution prefix on every finding. When it names a harness the-loop has an adapter for (claude, cursor), the invocation is derived from that adapter and command/args may be omitted entirely.

critics[].model

• Type: string
• Default: ""

The model the critic runs. Passed via the harness's own model flag for a built-in harness, or wherever {model} appears in args.

critics[].command

• Type: string
• Default: ""

The executable to spawn (argv[0]) — not a shell line: arguments go in args, and nothing is ever run through a shell. Required for a harness the-loop has no adapter for; wins over harness when both are set.

critics[].args

• Type: array of string
• Default: []

argv after the executable. Placeholders are substituted element-wise, never via a shell: {prompt}, {promptFile}, {model}, {workItem}, {specDir}, {cwd}. With an explicit command the list must carry {prompt} or {promptFile} — otherwise the critic is handed nothing to review. With a built-in harness, these are appended as extra flags to the derived invocation.

critics[].env

• Type: object of string
• Default: {}

Environment overlaid on the inherited environment for this critic. Never a secret's value — name the variable; the child inherits the ambient environment.

critics[].cwd

• Type: string
• Default: the checkout the round is run from (--root)

Directory to run the critic in.

critics[].outputFormat

• Type: string — text | json
• Default: text

How to read the critic's stdout. json extracts the reviewer's text from the payload (falling back to raw stdout when it is not JSON); text uses stdout verbatim.

critics[].timeoutSeconds

• Type: number, at least 1
• Default: 900

Hard bound on one critic round, so a hung critic CLI cannot wedge the review loop.

critics[].enabled

• Type: boolean
• Default: true

Set false to keep the entry but refuse to run it.

Review rounds

The top-level reviews block — the review-round policy the skill follows before it escalates to a human. Moved here from the repository's harness config in issue #352: the operator who runs the rounds sets how many. Every key has a default, so the block may be omitted; the-loop critic policy prints the effective values, and a harness with no CLI installed follows the same defaults.

reviews:
  selfReviewCount: 3
  criticReviewCount: 3
  stopOnNoNewFindings: true
  escalateOnRepeatFinding: true

reviews.selfReviewCount

• Type: integer, at least 0
• Default: 3

Cap on self-review rounds. The loop stops early when a round yields zero new actionable findings.

reviews.criticReviewCount

• Type: integer, at least 0
• Default: 3

Cap on critic-review rounds, run with the critics[] above. Also stops early on zero new findings; a critic that could not run does not count toward it.

reviews.stopOnNoNewFindings

• Type: boolean
• Default: true

Stop the review loop as soon as a round surfaces no new actionable finding.

reviews.escalateOnRepeatFinding

• Type: boolean
• Default: true

Diminishing-returns guard: when two consecutive rounds surface the same finding, escalate to a human instead of looping.