Review Decision

Run an independent challenge (decision kernel stage D10) against a drafted decision record and produce a review verdict — without modifying the record. Delegates the analysis to @decision-critic, which checks for framing errors, missing options, violated constraints, fragile assumptions, automation bias, and runs a premortem.

User invocation:

/review-decision

where ` is the decision record identifier (e.g., ADR-0001`) or the record's number for the relevant type.

This command is read-only by default: it produces a review artifact/verdict in its response and does NOT write, edit, stage, or commit any file. The caller decides whether and how to act on the findings (e.g., feed them back to @decision-advisor via /write-decision update, or escalate to a human reviewer for R3).

• id='$1' or allArguments='$ARGUMENTS': the decision record ID or number (e.g., ADR-0001, PDR-0003, or 0001 which defaults to ADR when the type is genuinely unspecified).

$ARGUMENTS

1. Parse the ID into TYPE + zeroPad4 (e.g., ADR-0001 → type ADR, number 0001). If only digits are given, default the type to ADR only when the type is genuinely unspecified.
2. Locate the record file by scanning doc/decisions/--*.md.
3. If not found, return NOT_FOUND with guidance to run /plan-decision + /write-decision first, and stop.

1. Resolve and read the target decision record read-only (read/grep/glob only).
2. Hand the record's problem, evidence, constraints, and options to @decision-critic. Where practical, withhold the record's recommendation/decision initially so the critic forms its own view, then compare against the record's stated conclusion.
3. @decision-critic performs the independent challenge (see its prompt): framing, option space, constraint compliance, assumption/sensitivity check, premortem, stakeholder harm, confidence assessment, automation-bias check.
4. Capture the critic's verdict — exactly one of PASS / PASSWITHRISKS / REWORK — and its structured findings.
5. Modify nothing. Do not write/edit/stage/commit. Output the review verdict and findings only.

State at the top of the review which independence level applies (per @decision-critic's honesty framing):

• Single-model configuration: the critic is a first-pass check, NOT independent assurance (same-model/same-prompt lineage is not independent evidence).
• Multi-model configuration: the critic runs on a different model family, providing genuine independence (recommended for R3, not mandated).

Regardless of the verdict, R3 decisions ALWAYS require a human reviewer. A PASS from the critic does not authorize an R3 decision.

Return a structured review artifact (in the response only — no file writes):

• Status: REVIEWED | NOT_FOUND | NEEDS_INPUT
• Record: - and path reviewed
• Verdict: PASS | PASSWITHRISKS | REWORK
• Independence note: single-model (first-pass check) | multi-model (independent) — plus the R3-requires-human-review reminder
• Findings: framing / option-space / constraint-compliance / assumption-sensitivity / premortem / stakeholder-harm / confidence / automation-bias
• Accepted risks (for PASSWITHRISKS): explicit list the human decider must accept
• Required changes (for REWORK): explicit, actionable list
• Next step: e.g., "Return findings to @decision-advisor to address" / "Escalate to human reviewer (R3)" / "Decision is sound; proceed to human authorization"

• Read-only: do NOT use write/edit; do NOT stage/commit; do NOT modify the decision record or any file.
• Do NOT transition the record's status (no auto-Accept, no status change).
• Do NOT use the network.
• The verdict is advisory; it never authorizes a decision on its own. R3 always needs a human.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: juliusz-cwiakalski
• Source: juliusz-cwiakalski/agentic-delivery-os
• License: MIT

Install and usage instructions live in the source repository linked above.