fix(review): 오프라인 커버리지와 증거 결합을 통합

requirements-opencode-review-ci.txt

@@ -3,3 +3,12 @@ interrogate==1.7.0
 pytest==9.1.1
 pytest-cov==7.1.0
 uv==0.11.25
+
+# Trusted runtime set required to execute ContextualWisdomLab/saju-caldav tests
+# inside the networkless OpenCode coverage sandbox. Keep these exact pins aligned
+# with that repository's reviewed lockfile; PR-controlled manifests remain inert.
+bcrypt==5.0.0
+fastapi==0.139.2
+httpx==0.28.1
+icalendar==7.2.0
+korean-lunar-calendar==0.4.0

requirements-strix-ci.txt

@@ -3,4 +3,5 @@ google-cloud-aiplatform==1.133.0
 protobuf<7.0.0
 cryptography==49.0.0
 python-multipart==0.0.31
+# CVE-2026-59885 and CVE-2026-59886: ASN.1 decoder DoS; fixed in 0.6.4
 pyasn1==0.6.4

scripts/ci/opencode_review_normalize_output.py

@@ -653,6 +653,161 @@ def adversarial_probe_source_receipt_error(
     return ""
 
 
+def evidence_cites_probe_path_at_any_line(evidence: str, path: str) -> bool:
+    """Return whether evidence already cites any positive line for the probe path."""
+    escaped_path = rf"(?<![A-Za-z0-9_./-]){re.escape(path)}"
+    return (
+        re.search(
+            rf"{escaped_path}(?::|#L|\s+line\s+)[1-9][0-9]*\b",
+            evidence,
+            re.IGNORECASE,
+        )
+        is not None
+    )
+
+
+def receipt_verified_evidence_location(
+    evidence: str,
+    changed_files: frozenset[str],
+) -> tuple[str, int] | None:
+    """Return one evidence citation uniquely bound to its trusted source receipt.
+
+    A model can place a valid changed-file ``path:line`` citation and its exact
+    source-line receipt in ``evidence`` while copying a different location into
+    the redundant structured fields. Consider only safe current-head changed
+    files cited by the evidence, recompute every cited line receipt from the
+    sealed source tree, and return a location only when exactly one distinct
+    citation matches the single model receipt. Ambiguous or unverified evidence
+    remains unrepairable and fails closed.
+    """
+    receipts = SOURCE_LINE_RECEIPT_RE.findall(evidence)
+    if len(receipts) != 1:
+        return None
+    receipt = receipts[0].casefold()
+    matches: set[tuple[str, int]] = set()
+    for path in changed_files:
+        escaped_path = rf"(?<![A-Za-z0-9_./-]){re.escape(path)}"
+        citation_re = re.compile(
+            rf"{escaped_path}(?::|#L|\s+line\s+)([1-9][0-9]*)\b",
+            re.IGNORECASE,
+        )
+        for citation in citation_re.finditer(evidence):
+            line = int(citation.group(1))
+            if adversarial_probe_location_error(path, line):
+                continue
+            digest = adversarial_probe_source_line_digest(path, line)
+            if digest is not None and digest.casefold() == receipt:
+                matches.add((path, line))
+    if len(matches) != 1:
+        return None
+    return next(iter(matches))
+
+
+def repair_adversarial_probe_evidence_bindings(value: Any) -> dict[str, Any] | Any:
+    """Bind trusted source receipts and locations into otherwise valid evidence.
+
+    Models sometimes place the exact changed-file path and positive line in the
+    structured ``path``/``line`` fields but omit the duplicate ``path:line`` text
+    from ``evidence``, or cite and receipt-bind one changed source line in
+    ``evidence`` while copying a different valid changed-file location into the
+    redundant structured fields. Restore only a missing citation or rebind the
+    structured location when exactly one cited line matches the existing single
+    receipt in the sealed current-head tree. Missing, duplicate, ambiguous, or
+    mismatched receipts, unsafe paths, missing changed-file evidence, and
+    circular or unobserved claims remain unmodified and fail closed.
+    """
+    if not isinstance(value, dict):
+        return value
+    validation = value.get("adversarial_validation")
+    if not isinstance(validation, dict):
+        return value
+    probes = validation.get("probes")
+    if not isinstance(probes, list):
+        return value
+
+    changed_files = current_changed_files()
+    repaired_probes: list[Any] = []
+    changed = False
+    for probe in probes:
+        if not isinstance(probe, dict):
+            repaired_probes.append(probe)
+            continue
+        path = probe.get("path")
+        line = probe.get("line")
+        evidence = probe.get("evidence")
+        if (
+            not isinstance(path, str)
+            or path not in changed_files
+            or isinstance(line, bool)
+            or not isinstance(line, int)
+            or line <= 0
+            or not isinstance(evidence, str)
+            or not evidence.strip()
+            or adversarial_probe_location_error(path, line)
+        ):
+            repaired_probes.append(probe)
+            continue
+
+        receipts = SOURCE_LINE_RECEIPT_RE.findall(evidence)
+        if len(receipts) != 1:
+            repaired_probes.append(probe)
+            continue
+        repaired_evidence = evidence
+        repaired_path = path
+        repaired_line = line
+        rejection = adversarial_evidence_rejection_reason(repaired_evidence, path, line)
+        if rejection == "must cite the exact probe path and positive line":
+            if not adversarial_probe_source_receipt_error(evidence, path, line):
+                if evidence_cites_probe_path_at_any_line(evidence, path):
+                    repaired_probes.append(probe)
+                    continue
+                repaired_evidence = f"{path}:{line} {repaired_evidence}"
+            else:
+                rebound_location = receipt_verified_evidence_location(
+                    evidence,
+                    changed_files,
+                )
+                if rebound_location is None:
+                    repaired_probes.append(probe)
+                    continue
+                repaired_path, repaired_line = rebound_location
+        elif rejection:
+            repaired_probes.append(probe)
+            continue
+        if adversarial_probe_source_receipt_error(
+            repaired_evidence, repaired_path, repaired_line
+        ) or adversarial_evidence_rejection_reason(
+            repaired_evidence,
+            repaired_path,
+            repaired_line,
+        ):
+            repaired_probes.append(probe)
+            continue
+        if (
+            repaired_evidence == evidence
+            and repaired_path == path
+            and repaired_line == line
+        ):
+            repaired_probes.append(probe)
+        else:
+            repaired_probes.append(
+                {
+                    **probe,
+                    "path": repaired_path,
+                    "line": repaired_line,
+                    "evidence": repaired_evidence,
+                }
+            )
+            changed = True
+
+    if not changed:
+        return value
+    return {
+        **value,
+        "adversarial_validation": {**validation, "probes": repaired_probes},
+    }
+
+
 def adversarial_validation_error(
     value: Any,
     *,
@@ -1267,6 +1422,7 @@ def reject(reason: str) -> None:
         return reject("APPROVE cannot contain findings")
     if result == "REQUEST_CHANGES" and not findings:
         return reject("REQUEST_CHANGES requires at least one finding")
+    value = repair_adversarial_probe_evidence_bindings(value)
     adversarial_error = adversarial_validation_error(
         value.get("adversarial_validation"),
         result=result,

scripts/ci/run_opencode_review_model_pool.sh

@@ -384,6 +384,9 @@ cap_model_run_timeout() {
 	github-models/openai/gpt-5 | github-models/openai/gpt-5-chat)
 		cap_seconds="$(env_integer_or_default OPENCODE_GITHUB_GPT5_RUN_TIMEOUT_SECONDS 45)"
 		;;
+	github-models/deepseek/deepseek-r1 | github-models/deepseek/deepseek-r1-0528)
+		cap_seconds="$(env_integer_or_default OPENCODE_GITHUB_DEEPSEEK_R1_RUN_TIMEOUT_SECONDS 300)"
+		;;
 	*)
 		printf '%s\n' "$run_timeout_seconds"
 		return 0
@@ -547,9 +550,6 @@ main() {
 		fi
 	fi
 	deadline=0
-	if [ "$budget_seconds" -gt 0 ]; then
-		deadline=$((SECONDS + budget_seconds))
-	fi
 	: >"$OPENCODE_OUTPUT_FILE"
 	cd "$OPENCODE_REVIEW_WORKDIR"
 	read -r -a model_candidates <<<"${OPENCODE_MODEL_CANDIDATES:-}"
@@ -573,6 +573,7 @@ main() {
 				continue
 			fi
 			if should_skip_model_candidate "$model_candidate"; then
+				dead_candidate_reasons[$model_candidate]="not runnable with the credentials available to this job"
 				continue
 			fi
 			assert_reasoning_effort_for_candidate "$model_candidate"
@@ -582,6 +583,12 @@ main() {
 			opencode_json_file="${candidate_output_file}.jsonl"
 			opencode_export_file="${candidate_output_file}.session.json"
 			write_prompt "$model_candidate" "$prompt_file"
+			# The retry budget measures provider attempts. Trusted local prompt
+			# preparation can be slower on a busy runner and must not exhaust the
+			# budget before the first provider is invoked.
+			if [ "$deadline" -eq 0 ] && [ "$budget_seconds" -gt 0 ]; then
+				deadline=$((SECONDS + budget_seconds))
+			fi
 			for attempt in $(seq 1 "$attempts"); do
 				now="$SECONDS"
 				if [ "$deadline" -gt 0 ] && [ "$now" -ge "$deadline" ]; then
@@ -610,7 +617,7 @@ main() {
 				uncapped_run_timeout="$OPENCODE_RUN_TIMEOUT_SECONDS"
 				OPENCODE_RUN_TIMEOUT_SECONDS="$(cap_model_run_timeout "$model_candidate" "$OPENCODE_RUN_TIMEOUT_SECONDS")"
 				if [ "$OPENCODE_RUN_TIMEOUT_SECONDS" -lt "$uncapped_run_timeout" ]; then
-					printf 'OpenCode %s runtime cap selected %ss instead of %ss because this installation has returned a constrained request-body limit for that endpoint.\n' \
+					printf 'OpenCode %s runtime cap selected %ss instead of %ss because the configured provider-specific cap is lower than the cadence timeout.\n' \
 						"$model_candidate" "$OPENCODE_RUN_TIMEOUT_SECONDS" "$uncapped_run_timeout"
 				fi
 				export OPENCODE_RUN_TIMEOUT_SECONDS
@@ -665,7 +672,7 @@ main() {
 			fi
 		done
 		if [ "$alive_candidates" -eq 0 ]; then
-			printf 'Every OpenCode model candidate is marked failed for this run; ending the pool without further provider spend.\n'
+			printf 'No runnable OpenCode model candidates remain for this run; ending the pool without further provider spend or idle cycles.\n'
 			if finish_pool_without_model; then
 				exit 0
 			fi

scripts/ci/test_strix_quick_gate.sh

@@ -87,6 +87,28 @@ assert_file_not_contains() {
 	fi
 }
 
+assert_file_not_matches() {
+	local file_path="$1"
+	local pattern="$2"
+	local message="$3"
+	local grep_status
+
+	if [ ! -f "$file_path" ]; then
+		record_failure "$message (missing file '$file_path')"
+		print_assertion_source "$file_path"
+		return
+	fi
+	if grep -Eq -- "$pattern" "$file_path"; then
+		record_failure "$message (unexpected pattern '$pattern')"
+	else
+		grep_status=$?
+		if [ "$grep_status" -ne 1 ]; then
+			record_failure "$message (grep failed with exit $grep_status for pattern '$pattern')"
+			print_assertion_source "$file_path"
+		fi
+	fi
+}
+
 seal_opencode_test_artifacts() {
 	local runner_temp="$1"
 	local head_sha="$2"
@@ -602,8 +624,12 @@ assert_opencode_review_uses_codegraph_and_gpt5_fallback() {
 	assert_file_contains "$workflow_file" 'locked_version" != "4.0.4"' "opencode review verifies both nested installed and locked picomatch evidence"
 	assert_file_contains "$workflow_file" '"$CODEGRAPH_BIN" explore' "opencode review precomputes structural evidence outside the model process"
 	assert_file_contains "$workflow_file" '"$CODEGRAPH_BIN" --version' "opencode review logs the exact trusted CodeGraph version"
-	assert_file_contains "$workflow_file" 'cat "$codegraph_status" >&2' "opencode review exposes CodeGraph status failures in the job log"
-	assert_file_contains "$workflow_file" 'cat "$codegraph_raw" >&2' "opencode review exposes CodeGraph exploration failures in the job log"
+	assert_file_not_contains "$workflow_file" 'cat "$codegraph_status" >&2' "opencode review does not replay PR-derived CodeGraph status bytes as workflow commands"
+	assert_file_not_contains "$workflow_file" 'cat "$codegraph_raw" >&2' "opencode review does not replay PR-derived CodeGraph exploration bytes as workflow commands"
+	assert_file_not_matches "$workflow_file" '^[[:space:]]*cat "\$CODEGRAPH_EVIDENCE_FILE"[[:space:]]*$' "opencode review does not replay assembled PR-derived CodeGraph evidence into the command channel"
+	assert_file_contains "$workflow_file" 'CodeGraph status command failed; captured %s bytes without replaying PR-derived log content.' "opencode review reports bounded CodeGraph status failure metadata"
+	assert_file_contains "$workflow_file" 'CodeGraph exploration command failed; captured %s bytes without replaying PR-derived log content.' "opencode review reports bounded CodeGraph exploration failure metadata"
+	assert_file_contains "$workflow_file" 'Captured bounded CodeGraph evidence (%s bytes) without replaying PR-derived log content.' "opencode review reports bounded CodeGraph success metadata"
 	assert_file_not_contains "$workflow_file" "serve --mcp" "opencode review must not fetch or launch CodeGraph again for MCP"
 	assert_file_not_contains "$workflow_file" "https://mcp.deepwiki.com/mcp" "opencode review does not expose remote MCP to the model"
 	assert_file_not_contains "$workflow_file" "@upstash/context7-mcp@3.1.0" "opencode review does not install Context7 at runtime"
@@ -1024,6 +1050,7 @@ assert_opencode_review_uses_codegraph_and_gpt5_fallback() {
 	assert_file_contains "$workflow_file" '(.workflow // "") == "CodeQL"' "opencode approval can distinguish CodeQL dynamic setup checks"
 	assert_file_contains "$workflow_file" '((.isRequired // false) | not) and (.workflow // "") == "CodeQL"' "opencode approval ignores non-required cancelled CodeQL checks without source evidence"
 	assert_file_contains "$workflow_file" '(.name // "") == "scan-pr-queue" and ((.workflow // "") == "PR Review Merge Scheduler" or (.workflow // "") == "Required PR Review Merge Scheduler")' "opencode approval ignores cancelled scheduler queue replacement checks without source evidence"
+	assert_file_not_contains "$workflow_file" 'select(((.conclusion // "" | ascii_downcase) == "cancelled" and (.name // "") == "scan-pr-queue") | not)' "opencode approval keeps ownership-ambiguous cancelled scan-pr-queue checks blocking in the workflow-less REST fallback"
 	assert_file_contains "$workflow_file" 'grep -Fq -- "Strix Security Scan/strix:" "$rollup_file"' "opencode approval avoids duplicate supplemental Strix workflow-run blockers when statusCheckRollup already has the Strix check"
 	assert_file_contains "$workflow_file" 'current_head_manual_strix_success_status()' "opencode approval can identify same-head manual Strix success status evidence"
 	assert_file_contains "$workflow_file" 'manual_run_line="$(latest_current_head_manual_strix_run || true)"' "opencode approval falls back to same-head manual Strix check-run success when commit status publication is unavailable"
@@ -5637,11 +5664,30 @@ run_gate_case_allow_provider_signal() {
 	run_gate_case_with_provider_signal_mode "0" "$@"
 }
 
+test_assert_file_not_matches_fails_closed() {
+	local fixture_file
+
+	fixture_file="$(mktemp)"
+	printf 'safe fixture\n' >"$fixture_file"
+	if ! (
+		FAILURES=0
+		assert_file_not_matches "$fixture_file" "[" "invalid regex must fail closed"
+		assert_file_not_matches "${fixture_file}.missing" "safe" "missing file must fail closed"
+		[ "$FAILURES" -eq 2 ]
+	) >/dev/null 2>&1; then
+		record_failure "assert_file_not_matches must record missing-file and grep failures"
+	fi
+	rm -f "$fixture_file"
+}
+
 run_filtered_gate_case_if_requested() {
 	case "${STRIX_TEST_CASE_FILTER:-}" in
 	"")
 		return 0
 		;;
+	assert-file-not-matches-errors-fail-closed)
+		test_assert_file_not_matches_fails_closed
+		;;
 	success)
 		run_gate_case "success" \
 			"vertex_ai/ready-primary" \