Cross-Phase Review
SkillMediaDispatch a structured cross-phase review of a phase artefact. MANDATORY for design-spec and test-spec; the gate validator checks for the resulting review record before passing. Recommended for requirements-spec.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the Cross-Phase Review skill
What this skill tells your AI
The instructions your AI receives, as published by stevegjones/ai-first-sdlc-practices in skills/phase-review/SKILL.md and read by ahel’s review.
Dispatch a structured review of a phase artefact against its cited prior artefacts.
Arguments
<phase>— one ofrequirements,design,test<feature-id>— short slug for the feature
Why this is mandatory for design and test
phase-gate checks structure: do citations resolve to declared IDs? phase-review checks meaning: do the citations actually capture the requirement / design intent? Programme Article 14 requires the review record for design-spec and test-spec — the gate validator looks for docs/specs/<feature-id>/reviews/<phase>-review-*.md before passing.
Steps
PHASE="$1"
FEATURE_ID="$2"
if [ -z "$PHASE" ] || [ -z "$FEATURE_ID" ]; then
echo "Usage: phase-review <phase> <feature-id>" >&2
exit 1
fi
case "$PHASE" in
requirements|design|test) ;;
*)
echo "phase must be one of: requirements, design, test" >&2
exit 1
;;
esac
FEATURE_DIR="docs/specs/$FEATURE_ID"
SPEC_PATH="$FEATURE_DIR/$PHASE-spec.md"
if [ ! -f "$SPEC_PATH" ]; then
echo "$SPEC_PATH does not exist; run phase-init $PHASE $FEATURE_ID first" >&2
exit 1
fi
REVIEWS_DIR="$FEATURE_DIR/reviews"
mkdir -p "$REVIEWS_DIR"
# Determine reviewer agent based on phase
case "$PHASE" in
requirements)
REVIEWER_AGENT="sdlc-team-common:solution-architect"
REVIEWER_LABEL="solution-architect"
;;
design)
REVIEWER_AGENT="sdlc-team-common:solution-architect"
REVIEWER_LABEL="solution-architect"
;;
test)
REVIEWER_AGENT="sdlc-team-fullstack:backend-architect"
REVIEWER_LABEL="backend-architect"
;;
esac
REVIEW_PATH="$REVIEWS_DIR/$PHASE-review-$REVIEWER_LABEL.md"
Graceful fallback when the configured agent is not available
If the $REVIEWER_AGENT listed above is not loaded in the current environment (e.g., the sdlc-team-common or sdlc-team-fullstack plugin was not installed), the dispatch will return an error like "subagent_type not found" or "agent not loaded".
In that case, fall back to the superpowers:requesting-code-review pattern: dispatch the general-purpose subagent with the same review prompt below, and prefix the review record's ## Reviewer section with a Fallback: line so auditors can see which agent actually performed the review.
Concretely:
- First attempt the configured agent (
$REVIEWER_AGENT). - If that returns an "agent not found" / "subagent_type not loaded" error, retry with
subagent_type=general-purposeand prepend this line to the review prompt:**Fallback dispatch:** the configured reviewer ($REVIEWER_AGENT) was not loaded; you are acting as a fallback reviewer. Note this in the review record's Reviewer section. - The resulting review record is still valid — Article 14's gate checks for the FILE, not for which agent produced it. The audit trail captures the fallback explicitly via the
Fallback:line.
Do NOT silently skip the review just because the configured agent is absent. Either dispatch the configured agent OR explicitly fall back; never produce no review record.
Now use the Agent tool to dispatch $REVIEWER_AGENT (substituting the value from the bash variable) with this prompt:
You are reviewing a $PHASE-spec for feature $FEATURE_ID.
Read the artefact at: $SPEC_PATH
Required prior artefacts to read:
- For design phase: docs/specs/$FEATURE_ID/requirements-spec.md (the requirements being satisfied)
- For test phase: both docs/specs/$FEATURE_ID/requirements-spec.md AND docs/specs/$FEATURE_ID/design-spec.md
Your review must answer:
1. **Coverage**: are all REQ-IDs from the prior phase(s) addressed by this artefact's elements?
2. **Soundness**: do the satisfies references make semantic sense — does each cited DES/TEST actually address what the cited REQ/DES says?
3. **Completeness**: are there obvious requirements / design elements missing that should have been declared?
4. **Out-of-scope cleanliness**: does the artefact stay within scope, or does it sneak in elements that don't trace back?
Return your review as a markdown file with these sections:
- ## Reviewer
- ## Coverage assessment
- ## Soundness assessment
- ## Completeness assessment
- ## Out-of-scope concerns
- ## Verdict: APPROVE / NEEDS-REWORK / NOT-APPROVED
Save the review to: $REVIEW_PATH (use the Write tool)
After the agent dispatch completes, verify the review file was created:
if [ -f "$REVIEW_PATH" ]; then
echo "Review record created at $REVIEW_PATH"
git add "$REVIEW_PATH"
echo "Stage the review record with the next commit."
else
echo "ERROR: $REVIEW_PATH was not created" >&2
exit 1
fi
Done
Report:
- Phase artefact reviewed
- Reviewer agent used (
solution-architectfor design;backend-architectfor test) - Review record path
- Verdict from the review (APPROVE / NEEDS-REWORK / NOT-APPROVED)
If verdict is NEEDS-REWORK or NOT-APPROVED, the user must address the review's findings before commit/push.
Model selection
This skill is structured comparison and reasoning — comparing artefacts against prior phases, surfacing semantic gaps. Prefer a model with strong reasoning over one optimised for throughput. Reviewer agents (solution-architect, backend-architect) are dispatched via the Agent tool; they should run on a high-quality model.
Signals
- GitHub stars
- 41
- Forks
- 6
- Last commit
- Aug 2026
Advanced
- Catalog kind
- skill
- Gateway key
phase-review- Source
- github.com/stevegjones/ai-first-sdlc-practices