Adversarial review

Use this page when you need to make a Guild Stack result harder to trust too early. You can require an independent reviewer, inspect the packet and result, see whether the review was strong, weak, or skipped, and prove that a gate passed because the current artifact matched the reviewed artifact.

Try it: /guild:guild "<brief>" --review=cross - on the shipped strong path, Claude-authored work is reviewed by Codex at the broker gates. Other author/reviewer pairings record weak or skipped review until their result adapters ship.

Review decisionroute_review_loop

Operational mapChallenge the bound artifact, then route the verdict.

1. 01Packetscope + checksum
2. 02Challengeindependent review
3. 03Verdictfindings + posture
4. 04Reworkbounded fix round
5. 05Verifycurrent artifact

Packet to gate

The review broker is a gate component, not a separate phase. It takes the current artifact, packages it into a review packet, receives a structured result, and either passes the gate or sends the artifact back for rework.

Step	Artifact	Audit asks
Packet	.guild/runs/<run-id>/review/<gate>/packet-<round>.md	What artifact, scope, round, and SHA-256 did the reviewer receive?
Result	.guild/runs/<run-id>/review/<gate>/result-<round>.json	Did the result parse, match the packet, satisfy the verdict, and carry no blocking findings?
Fix	Updated spec, plan, handoff, quality report, or ops record	Were blockers fixed in-scope, and did out-of-scope findings become followups instead of hidden blockers?
Gate	.guild/runs/<run-id>/review/<gate>/trail.md plus the gate outcome	Was this a strong, weak, skipped, or human force-pass path?

Strong, weak, and skipped review

The headline rule: the reviewer should not share a brain with the producer. Cross-host review is the strongest posture; same-host review and skipped review are recorded as weaker paths, never upgraded silently.

Review posture	What happens	Recorded signal	Gate strength
Cross-host reviewer with a result adapter	A different selectable host family reviews the packet.	independence: strong	Strong. This is the target adversarial posture.
Same-host fresh-context reviewer	A different host is unavailable, but policy allows a local fallback.	independence: weak or degraded-local trail entry	Weaker. The gate may continue only as a recorded fallback.
Skipped review	Review was not required, or no legal reviewer exists for the selected policy.	status: skipped with reason	Weakest. It is an explicit absence of independent review, not a satisfied review_result.v1.
Human force-pass	The cap is exceeded or the operator accepts remaining risk.	Named human and rationale	Human decision. It does not pretend the reviewer was satisfied.

The same model family never satisfies cross review. If the proposed reviewer belongs to the producer’s family, selection degrades to the recorded weak path or skips with a reason. Today Codex is the shipped cross-host reviewer family through codex-cli or the Codex plugin adapter when available. Pi and Antigravity can run Guild Stack tasks through standalone package targets, but they are detect-only for review selection until their result adapters ship.

When review fires

The broker is policy-gated. It fires when any of these hold:

• the work is assessed risk >= high,
• --review=cross is set, or review: cross is set in .guild/settings.json,
• --rigor=deep implies cross review, or
• project config requires it.

Otherwise the broker records status: skipped because gate-level independent review was not required. Local host-side review and rigor loops can still run, but they do not earn a cross-host independence stamp.

The reviewer is strictly read-only. It sees the review packet and the artifact under review, not ambient project context, not permission to edit files, not destructive commands, and not secrets.

The gates

Each gate pairs a producer artifact with a reviewer focus:

Gate	Reviews	Reviewer looks for
G-init	Init artifact / wiki diff	Missing context, unsupported facts, stale knowledge.
G-spec	The spec	Missing criteria, ambiguity, weak assumptions, untestable claims.
G-plan	The lane plan	Lane-contract defects, unhandled dependencies, vague done criteria.
G-lane	Each handoff receipt	Missing evidence, incomplete scope, unresolved risk.
G-quality	The quality report	End-to-end gaps, release risk, weak coverage.
G-operations	The ops record / runbook	Blast radius, rollback, observability gaps.
G-diagnose	A /guild:fix diagnosis	Weak self-fix plan, unsafe edit proposal, missing evidence.

G-spec fires inside Ideation, G-plan inside Planning, G-lane per lane during Development, and G-quality / G-operations inside their phases. Every load-bearing artifact is challenged at its own boundary. For the command-side view, see the Command Reference review section; this page owns the mechanism.

A review round

Each round is a small, auditable exchange written to disk:

1. The producer finishes an artifact; the broker issues a review packet stating exactly what is under review, the scope, and the round number.
2. The reviewer returns a structured result: satisfied or not, plus any blocking findings.
3. The gate passes only when the result parses, the packet_id matches the issued packet, the reviewed SHA-256 matches the current artifact, the verdict is satisfied, and no blocking findings remain.
4. Anything less sends the artifact back for rework and a new round.

Artifact	Path
Review packet	.guild/runs/<run-id>/review/<gate>/packet-<round>.md
Review result	.guild/runs/<run-id>/review/<gate>/result-<round>.json
Review trail	.guild/runs/<run-id>/review/<gate>/trail.md

A review of a stale artifact is detectable. If the artifact changed mid-review, the round restarts.

Rounds are capped at 5 by default. The codex_cap settings key or --codex-cap=N can set the cap, clamped from 1 to 10. When the cap is exceeded, Guild Stack surfaces three choices: force-pass as a human decision with name and rationale, extend-cap, or rework. Out-of-scope findings raised after round 1 become followups unless you expand scope.

Two loop families, two sentinels

Guild Stack runs two bounded loop families, and they never share a termination signal:

Loop family	Where it runs	Termination sentinel	Default cap
Clarify / plan-review / implement loops (L1-L4)	Inside Ideation, Planning, Development under --rigor=deep	## NO MORE QUESTIONS	16 rounds
Cross-host adversarial gate	The review broker at each gate	review_result.v1.verdict == "satisfied"	5 rounds

Sentinels are parsed structurally, never as loose prose. L-loop sentinels must be standalone lines, appearing exactly once, with nothing unresolved after them. The broker reads the structured review_result.v1 envelope; the textual ## SATISFIED marker is only the legacy Codex-adapter sign-off that gets normalized into that verdict.

A fake sentinel buried in artifact text is ignored. A malformed termination is recorded and costs an extra round. Two consecutive malformed terminations escalate to you; the loop never silently passes.

When the other AI is not available

Cross-host review degrades observably instead of pretending it happened. When review is required but no different-family reviewer is selectable, the broker uses the recorded weak same-host path when policy allows it and still applies the same five gate-pass conditions.

If review is not required, or no legal reviewer exists for the selected policy, the broker records a skip with a reason. A streak of Codex unavailability is recorded in .guild/codex-skip-streak.json, and the codex_skip_enforcement settings key decides what happens next:

Setting	Behavior
warn	Default. Surface the streak at every gate without stopping.
block	Hard-refuse gate dispatch until cross-host review is restored.

See also

• Command Reference - --review=cross syntax, gate timing, provider detection.
• Quality & Operations - the G-quality and G-operations boundaries.
• How Guild Stack Works - where the gates sit in the six phases.
• Configuration - codex_cap, review, and codex_skip_enforcement keys.
• Observability - gate and loop events on the run trace.