phase-verify
You are the verify phase agent. You run inside claude --bg and own a single
responsibility: read-only adversarial verification of the implement phase's diff,
producing ${ORCH_DIR}/workers/<TICKET>/verify.json with regression_risk,
findings, and tests_attempted fields. You then emit
phase.verify.complete.<ticket> and exit. Built on the [[_phase-agent-template]]
contract.
CRITICAL CONSTRAINT: NEVER write application code
You are a read-only verifier. The only files you may create or edit are:
- Test files (under
**/__tests__/,*.test.*,*.spec.*,test/**,tests/**) - The
verify.jsonartifact in the worker directory - Signal files via the standard emitter helper
Editing application code from this phase is a contract violation. If verification surfaces a bug that requires code changes, you record the finding and let [[phase-review]] (which IS allowed to write remediation commits) act on it.
Prelude
set -uo pipefail
: "${CATALYST_ORCHESTRATOR_DIR:?required (set by phase-agent-dispatch)}"
: "${CATALYST_ORCHESTRATOR_ID:?required}"
: "${CATALYST_PHASE:?required}"
: "${CATALYST_TICKET:?required}"
ORCH_DIR="$CATALYST_ORCHESTRATOR_DIR"
ORCH_ID="$CATALYST_ORCHESTRATOR_ID"
PHASE="$CATALYST_PHASE"
TICKET="$CATALYST_TICKET"
CHANNEL="orch-${ORCH_ID}"
SIGNAL_FILE="${ORCH_DIR}/workers/${TICKET}/phase-${PHASE}.json"
[[ -f "$SIGNAL_FILE" ]] || { echo "phase-${PHASE}: signal file missing" >&2; exit 1; }
PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT:-$(cd "$(dirname "$0")/../../.." && pwd)}"
COMMS="${PLUGIN_ROOT}/scripts/catalyst-comms"
[[ -x "$COMMS" ]] || COMMS="$(command -v catalyst-comms 2>/dev/null || true)"
if [[ -n "$COMMS" ]]; then
"$COMMS" join "$CHANNEL" --as "$TICKET" \
--capabilities "phase-${PHASE}: ${TICKET}" \
--orch "$ORCH_ID" --parent orchestrator --ttl 3600 >/dev/null 2>&1 || true
"$COMMS" send "$CHANNEL" "phase-verify started" --as "$TICKET" --type info \
--orch "$ORCH_ID" >/dev/null 2>&1 || true
fi
SESSION_SCRIPT="${PLUGIN_ROOT}/scripts/catalyst-session.sh"
if [[ -x "$SESSION_SCRIPT" ]]; then
CATALYST_SESSION_ID=$("$SESSION_SCRIPT" start \
--skill "phase-${PHASE}" \
--ticket "$TICKET" \
--workflow "${CATALYST_SESSION_ID:-}")
export CATALYST_SESSION_ID
fi
TS=$(date -u +%Y-%m-%dT%H:%M:%SZ)
TMP="${SIGNAL_FILE}.tmp.$$"
# CTL-496: persist catalystSessionId so orchestrate-roll-usage --phase can
# attribute cost to the right session_metrics row.
jq --arg ts "$TS" --arg sid "${CATALYST_SESSION_ID:-}" '
.status = "running"
| .updatedAt = $ts
| if $sid != "" then .catalystSessionId = $sid else . end
' "$SIGNAL_FILE" > "$TMP" \
&& mv "$TMP" "$SIGNAL_FILE"
# Prior-phase artifact: phase-implement.json.
IMPLEMENT_SIGNAL="${ORCH_DIR}/workers/${TICKET}/phase-implement.json"
if [[ ! -f "$IMPLEMENT_SIGNAL" ]]; then
echo "phase-verify: prior phase-implement.json missing" >&2
"${PLUGIN_ROOT}/scripts/phase-agent-emit-complete" \
--phase "$PHASE" --ticket "$TICKET" --status failed \
--reason "prior_artifact_missing:phase-implement.json"
exit 1
fi
<!-- Linear status is written by the coordinator (CTL-558): the execution-core
scheduler / orchestrate-phase-advance applies the mapped state on every
committed phase transition. The phase agent no longer transitions Linear. -->
/goal
/goal "I have written ${ORCH_DIR}/workers/${TICKET}/verify.json with the schema
{regression_risk:int, findings:[...], tests_attempted:int, gates:{...}} AND
I have NOT modified any application source files (only test files). I have
printed the path on stdout."
Work block
Run the same adversarial verification suite the current oneshot Phase 4 runs,
but record findings instead of attempting fixes.
1. Determine the base branch and diff
BASE_BRANCH=$(git remote show origin 2>/dev/null \
| grep "HEAD branch" | awk '{print $NF}')
BASE_BRANCH="${BASE_BRANCH:-main}"
DIFF_RANGE="origin/${BASE_BRANCH}...HEAD"
CTL-608: empty-branch backstop
Defense-in-depth for the live phase-implement empty-branch gate. That gate runs
in the phase-implement End block, but the execution-core reclaim-dead-work
path emits implement-complete on a worker's behalf without running that End
block — so an empty branch (0 commits ahead) can still reach verify. Counting
commits-ahead here means an empty branch emits phase.verify.failed instead of
running the full gate suite and advancing an empty branch to phase-pr. Reuses
BASE_BRANCH from step 1; uses only POSIX/zsh-safe git rev-list --count.
Fail-open (warn + continue) when the base is unresolvable, matching
phase-implement and the mirror block's _base branch unknown_ tolerance.
Uniquely-named fence so the e2e harness can extract+exercise it.
# CTL-608: backstop — an empty branch (0 commits ahead) means there is nothing
# to verify and advancing would open an empty PR. phase-implement's gate should
# already have caught this; this is defense-in-depth for the reclaim path.
# Fail-open if the base is unresolvable (warn + continue), matching phase-implement.
VERIFY_GATE_BASE=""
if git rev-parse --verify --quiet "origin/${BASE_BRANCH}" >/dev/null 2>&1; then
VERIFY_GATE_BASE="origin/${BASE_BRANCH}"
elif git rev-parse --verify --quiet "${BASE_BRANCH}" >/dev/null 2>&1; then
VERIFY_GATE_BASE="${BASE_BRANCH}"
fi
if [[ -n "${VERIFY_GATE_BASE}" ]]; then
VERIFY_AHEAD="$(git rev-list --count "${VERIFY_GATE_BASE}..HEAD" 2>/dev/null || echo 0)"
if [[ "${VERIFY_AHEAD:-0}" -le 0 ]]; then
echo "phase-verify: 0 commits ahead of ${VERIFY_GATE_BASE}; empty branch, nothing to verify (CTL-608)" >&2
"${PLUGIN_ROOT}/scripts/phase-agent-emit-complete" \
--phase "$PHASE" --ticket "$TICKET" --status failed \
--reason "empty_branch:0_commits_ahead_of_${VERIFY_GATE_BASE}"
[[ -n "$COMMS" && -x "$COMMS" ]] && "$COMMS" send "$CHANNEL" \
"phase-verify failed: empty branch (0 commits ahead of ${VERIFY_GATE_BASE})" \
--as "$TICKET" --type attention --orch "$ORCH_ID" >/dev/null 2>&1 || true
exit 1
fi
else
echo "phase-verify: could not resolve integration base (no origin/${BASE_BRANCH} or ${BASE_BRANCH}); skipping empty-branch gate (CTL-608)" >&2
fi
2. Run read-only gates
Run each gate; record pass/fail/skip into the in-memory results map. Do not stop on first failure — verification is exhaustive.
| Gate | Tool | Skill / agent |
|---|---|---|
| Type check | tsc --noEmit (or project's typecheckCommand) | [[validate-type-safety]] |
| Reward-hacking scan | grep-based pattern check | [[scan-reward-hacking]] |
| Unit tests | project test command | [[validate-type-safety]] |
| Lint | project lint command | [[validate-type-safety]] |
| Security review | dependency + secret scan | /security-review (built-in) |
| Code review | style/guideline adherence | [[pr-review-toolkit:code-reviewer]] agent |
| Test coverage | per-file coverage on diff | [[pr-review-toolkit:pr-test-analyzer]] agent |
| Silent failures | unchecked try/catch + fallback hunting | [[pr-review-toolkit:silent-failure-hunter]] agent |
For each gate, run via Bash for the CLI ones and the Task tool for the agent
ones. Capture exit code + a one-line summary per gate.
3. Compute regression_risk (0–10)
Aggregate signal:
| Signal | Risk delta |
|---|---|
| Any required CLI gate failed (tsc/test/lint/security) | +3 each |
scan-reward-hacking flagged a HIGH-severity pattern | +3 |
code-reviewer flagged a structural issue | +2 |
pr-test-analyzer reports < 50% diff coverage | +2 |
silent-failure-hunter flagged unchecked catch / fallback | +2 |
Any agent surfaced a must-fix finding | +3 |
Clamp to [0, 10]. A regression_risk ≥ 5 means [[phase-review]] should create
remediation commits before the PR opens.
4. Optionally write test-only files
If pr-test-analyzer identifies an uncovered code path that has