critic

The seam by which the harness doing the work hands it to a different harness for a critic round, and reads back what it said.

the-loop critic list [--root .] [--format table|json]
the-loop critic run <name> (--prompt TEXT | --prompt-file PATH)
                    [--root .] [--cwd DIR] [--work-item ID] [--spec-dir PATH]
                    [--timeout SECONDS] [--output-file PATH]

A review is worth more when it comes from a model that did not write the code. This command is how that happens without either harness knowing about the other.

Where critics come from

critics[] in your CLI config — not any repository's harness config, since issue #352: a critic is a harness and a model installed on the machine that runs the round. The same file says how many rounds to run — its reviews block, printed by policy — and, with critics[], with what. Each entry is either:

• a harness the-loop has an adapter for — claude, cursor — where the invocation is derived; or
• an explicit command (argv[0]) plus args, with element-wise placeholders: {prompt}, {promptFile}, {model}, {workItem}, {specDir}, {cwd}.

env overlays the inherited environment, and outputFormat, timeoutSeconds, cwd and enabled complete the entry.

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.

list

Every configured critic with its executable and whether that executable is available.

Flag	Default	Meaning
--root	.	Project root — the default working directory for a round.
--format	table	table or json.

policy

The review-round policy the skill follows — the CLI config's reviews block with every key defaulted, so the output is the same shape whether or not the block exists.

Flag	Default	Meaning
--root	.	Project root (kept for the route's shape; the policy is the operator's, not the repository's).
--format	text	text (one key: value per line) or json.

{
  "selfReviewCount": 3,
  "criticReviewCount": 3,
  "stopOnNoNewFindings": true,
  "escalateOnRepeatFinding": true
}

A harness with no CLI installed follows these same defaults; the skill's reviewing reference is the procedure they drive.

run

Run one named critic round.

Flag	Default	Meaning
(positional)	—	The critics[] entry to run, by name.
--prompt	—	The review prompt, inline.
--prompt-file	—	File holding the prompt. Preferred — review prompts are long.
--root	.	Project root.
--cwd	the critic's cwd, else --root	Directory to run in.
--work-item	""	Work item id, e.g. issue-117 — fills {workItem}.
--spec-dir	<specDir>/<work item>	Fills {specDir}.
--timeout	the critic's timeoutSeconds	Override for this round.
--output-file	—	Also write the JSON envelope to this path.

--prompt and --prompt-file are mutually exclusive, and one is required.

One critic per invocation

There is deliberately no run-all mode. A critic round is a decision about which second opinion you want, not a batch job.

Safety

The critic is spawned as an argv list, never through a shell, under its timeout. So untrusted review material can be data but never a command — a prompt containing ; rm -rf / is a string the critic reads, not something the shell expands.

Output

stdout is exactly one JSON envelope; diagnostics go to the log stream, so stdout stays parseable:

{
  "critic": "cursor-gpt",
  "harness": "cursor",
  "model": "gpt-5.5",
  "attribution": "…",
  "ok": true,
  "exitCode": 0,
  "durationSeconds": 42.1,
  "output": "…",
  "error": null,
  "usage": {}
}

Exit codes

Code	Meaning
0	The round ran
1	The round failed — absent CLI, non-zero exit, timeout. The envelope is still printed.
2	Misconfigured — nothing was spawned

The 1-with-an-envelope behaviour is deliberate: a failed round is still a fact the calling harness needs to record, not an absence of output.

Scope

Repo-scoped, like check and scenarios: it runs in the project it is invoked in, reads only the operator's CLI config, and is no part of the daemon (decision-032).

The review loop itself — round counts, convergence, posting findings — stays with the harness following the reviewing reference. run runs one round; policy says how many there may be.

See also

• review-loop — the capability doc.
• decision-043 — why the critic is a spawned argv.