fix(opencode): 샌드박스 시작 실패를 공급자 장애와 구분

.github/workflows/agent-review-runtime-quality-ci.yml

@@ -15,6 +15,8 @@ on:
       - "docs/product-technical-gap-baseline.md"
       - ".github/workflows/opencode-review-dispatch.yml"
       - "scripts/ci/ensure_rust_llvm19.sh"
+      - "scripts/ci/emit_opencode_failed_check_fallback_findings.sh"
+      - "tests/test_strix_fallback_attribution.py"
       - "tests/test_opencode_rust_coverage_toolchain_contract.py"
       - "tests/test_pr_review_autofix_nvidia_nim_contract.py"
       - "docs/doctoring/opencode-rust-coverage-runtime-boundary.md"
@@ -184,6 +186,8 @@ jobs:
                 ;;
               .github/workflows/opencode-review-dispatch.yml|\
               scripts/ci/ensure_rust_llvm19.sh|\
+              scripts/ci/emit_opencode_failed_check_fallback_findings.sh|\
+              tests/test_strix_fallback_attribution.py|\
               tests/test_opencode_rust_coverage_toolchain_contract.py|\
               docs/doctoring/opencode-rust-coverage-runtime-boundary.md)
                 opencode_suite=true
@@ -343,8 +347,8 @@ jobs:
         if: steps.affected_suites.outputs.opencode == 'true'
         run: |
           set -euo pipefail
-          python -m pytest -q tests/test_opencode_rust_coverage_toolchain_contract.py
-          python -m compileall -q tests/test_opencode_rust_coverage_toolchain_contract.py
+          python -m pytest -q tests/test_opencode_rust_coverage_toolchain_contract.py tests/test_strix_fallback_attribution.py
+          python -m compileall -q tests/test_opencode_rust_coverage_toolchain_contract.py tests/test_strix_fallback_attribution.py
 
       - name: Verify exact-head path policy and syntax
         if: steps.affected_suites.outputs.strix == 'true'

.github/workflows/opencode-review-dispatch.yml

@@ -5602,7 +5602,15 @@ jobs:
               local path=".github/workflows/strix.yml"
               local line="1"
 
-              if ! grep -Eq "LLM CONNECTION FAILED|RateLimitError|Too many requests|budget limit|Configured model and fallback models were unavailable|provider infrastructure|STRIX_PROVIDER_UNAVAILABLE" "$strix_evidence_file"; then
+              if grep -Eq '(^|[[:space:]])STRIX_PROVIDER_UNAVAILABLE:[[:space:]]+STRIX_SANDBOX_UNAVAILABLE:' "$strix_evidence_file"; then
+                printf 'Non-source-backed Strix sandbox bootstrap failure: this signal alone does not establish a gateway or provider-pool failure. Do not approve from this incomplete current-head scan. Inspect sandbox startup, repair it, and rerun the intended current-head scan; review any independently reported vulnerabilities. No application source edit is justified by this signal alone.\n' >&2
+              fi
+
+              if ! awk '
+                { gsub(/(^|[[:space:]])STRIX_PROVIDER_UNAVAILABLE:[[:space:]]+STRIX_SANDBOX_UNAVAILABLE:/, " ") }
+                /LLM CONNECTION FAILED|RateLimitError|Too many requests|budget limit|Configured model and fallback models were unavailable|provider infrastructure|STRIX_PROVIDER_UNAVAILABLE/ { found = 1 }
+                END { exit !found }
+              ' "$strix_evidence_file"; then
                 return 0
               fi
 

AGENTS.md

@@ -68,6 +68,15 @@ The materialization contract is also covered by [`docs/doctoring/exact-artifact-
 
 ## Verification discipline
 
+OpenCode failed-check diagnostics must distinguish a Strix sandbox bootstrap
+failure from gateway availability. The paired `STRIX_PROVIDER_UNAVAILABLE:
+STRIX_SANDBOX_UNAVAILABLE:` marker is not a source-backed vulnerability or proof
+of provider failure. Keep both the canonical helper and missing-helper workflow
+fallback aligned; exclude only that paired marker from provider classification,
+preserve independent provider signals and mapped/unmapped vulnerability reports,
+and keep the current-head scan incomplete. Run
+`python -m pytest -q -W error tests/test_strix_fallback_attribution.py`.
+
 Many agent sessions work this organization concurrently under the same standing
 brief. Silence is not evidence: "I have not touched X" describes one session's
 history, never the organization's actual state.

CHANGELOG.md

@@ -52,6 +52,10 @@
 - Raised `hourly-review-repair.yml`'s discovery ceiling from 50 to 200 while rotating deterministic 50-PR deep-inspection windows by hourly run number. The scheduler hydrates only the selected window and stops immediately after its single dispatch, preserving access to newer PRs without quadrupling expensive review/check/comment work. See `docs/doctoring/hourly-review-repair-single-file-consolidation.md`'s 2026-09-03 follow-up.
 
 ## [Unreleased]
+- Failed security-scan diagnostics now identify sandbox startup failures without
+  blaming the model gateway or proposing an unrelated source edit. Independently
+  reported provider errors and vulnerabilities remain visible; the failed scan
+  still cannot support approval.
 - Include merge-scheduler entrypoint, core, and regression-test changes in
   the existing runtime-quality workflow's trigger and suite selector. Scheduler
   workflow edits retain queue checks and also select the full review-repair

scripts/ci/emit_opencode_failed_check_fallback_findings.sh

@@ -650,7 +650,18 @@ emit_strix_provider_failure_finding() {
 	local path=".github/workflows/strix.yml"
 	local line="1"
 
-	if ! grep -Eq "LLM CONNECTION FAILED|RateLimitError|Too many requests|HTTPStatusError|401 Unauthorized|api\\.deepseek\\.com|Authentication Fails|DeepseekException|budget limit|Configured model and fallback models were unavailable|provider infrastructure|Below-threshold findings detected|Unable to map Strix findings" "$strix_evidence_file"; then
+	if grep -Eq '(^|[[:space:]])STRIX_PROVIDER_UNAVAILABLE:[[:space:]]+STRIX_SANDBOX_UNAVAILABLE:' "$strix_evidence_file"; then
+		printf 'Non-source-backed Strix sandbox bootstrap failure: this signal alone does not establish a gateway or provider-pool failure. Do not approve from this incomplete current-head scan. Inspect sandbox startup, repair it, and rerun the intended current-head scan; review any independently reported vulnerabilities. No application source edit is justified by this signal alone.\n' >&2
+		if [ -s "$unmapped_strix_reports_file" ]; then
+			awk -F '\t' '{ printf "Unmapped Strix report: %s reported \"%s\" (%s; %s). Obtain a repository Code Location before proposing a source edit.\n", $1, $2, $3, $4 }' "$unmapped_strix_reports_file" >&2
+		fi
+	fi
+
+	if ! awk '
+		{ gsub(/(^|[[:space:]])STRIX_PROVIDER_UNAVAILABLE:[[:space:]]+STRIX_SANDBOX_UNAVAILABLE:/, " ") }
+		/LLM CONNECTION FAILED|RateLimitError|Too many requests|HTTPStatusError|401 Unauthorized|api\.deepseek\.com|Authentication Fails|DeepseekException|budget limit|Configured model and fallback models were unavailable|provider infrastructure|Below-threshold findings detected|Unable to map Strix findings/ { found = 1 }
+		END { exit !found }
+	' "$strix_evidence_file"; then
 		return 0
 	fi
 

tests/test_pr_review_autofix_nvidia_nim_contract.py

@@ -17,7 +17,7 @@
 DOCTORING_RECORD = Path("docs/doctoring/hourly-nvidia-nim-autofix.md")
 CHANGELOG = Path("CHANGELOG.md")
 REVIEW_DISPATCH_WORKFLOW = Path(".github/workflows/opencode-review-dispatch.yml")
-REVIEW_DISPATCH_BLOB_SHA = "26e8555967171a5f3974602ac05700c27bddebf1"
+REVIEW_DISPATCH_BLOB_SHA = "a3c7c4fa7b389549fb58a2aefa5f6cbb0592f760"
 
 
 def _workflow_text(path: Path) -> str:

tests/test_strix_fallback_attribution.py

@@ -0,0 +1,135 @@
+"""Execute both failed-check emitters without GitHub or model access."""
+
+import os
+import subprocess
+import textwrap
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SANDBOX_SIGNAL = (
+    "STRIX_PROVIDER_UNAVAILABLE: STRIX_SANDBOX_UNAVAILABLE: bootstrap failed"
+)
+
+
+@pytest.mark.parametrize("changed_path", [
+    "scripts/ci/emit_opencode_failed_check_fallback_findings.sh",
+    "tests/test_strix_fallback_attribution.py",
+])
+def test_either_owned_path_selects_the_existing_opencode_contract_job(changed_path):
+    """Helper-only and test-only edits must execute this check without another job."""
+    workflow_text = (REPO_ROOT / ".github/workflows/agent-review-runtime-quality-ci.yml").read_text(
+        encoding="utf-8"
+    )
+    trigger_text = workflow_text.split("on:\n", 1)[1].split("\nconcurrency:\n", 1)[0]
+    assert f'      - "{changed_path}"' in trigger_text
+    selector_text = workflow_text.split('            case "$changed_path" in\n', 1)[1].split(
+        "            esac", 1
+    )[0]
+    result = subprocess.run([
+        "bash", "-c",
+        'set -euo pipefail\nopencode_suite=false\nchanged_path="$1"\ncase "$changed_path" in\n'
+        + selector_text + 'esac\nprintf "%s" "$opencode_suite"',
+        "selector-test", changed_path,
+    ], capture_output=True, text=True, check=True, env={"PATH": os.defpath})
+    assert result.stdout == "true"
+    step_text = workflow_text.split("- name: Verify OpenCode Rust coverage toolchain contract", 1)[1].split(
+        "\n      - name:", 1
+    )[0]
+    assert "tests/test_strix_fallback_attribution.py" in step_text
+
+
+def run_emitter(tmp_path: Path, route_name: str, evidence_text: str) -> subprocess.CompletedProcess:
+    """Run the real helper or its actual missing-helper workflow fallback."""
+    evidence_file = tmp_path / "failed-checks.md"
+    evidence_file.write_text(evidence_text, encoding="utf-8")
+    fixture_repo = tmp_path / "repository"
+    fixture_repo.mkdir()
+    (fixture_repo / "README.md").write_text("Fixture source line.\n", encoding="utf-8")
+    helper_file = REPO_ROOT / "scripts/ci/emit_opencode_failed_check_fallback_findings.sh"
+    command_args = ["bash", str(helper_file), str(evidence_file), str(fixture_repo)]
+    if route_name == "inline":
+        workflow_text = (REPO_ROOT / ".github/workflows/opencode-review-dispatch.yml").read_text(
+            encoding="utf-8"
+        )
+        start_index = workflow_text.index("          emit_line_specific_fallback_findings() {")
+        end_index = workflow_text.index("\n          }", start_index) + len("\n          }")
+        function_text = textwrap.dedent(workflow_text[start_index:end_index])
+        command_args = [
+            "bash", "-c",
+            'set -euo pipefail\n' + function_text + '\nemit_line_specific_fallback_findings "$1"',
+            "fallback-test", str(evidence_file),
+        ]
+    return subprocess.run(
+        command_args, cwd=fixture_repo, capture_output=True, text=True, check=False,
+        env={"PATH": os.defpath, "GITHUB_WORKSPACE": str(fixture_repo)},
+    )
+
+
+@pytest.mark.parametrize("route_name", ["helper", "inline"])
+def test_sandbox_bootstrap_does_not_invent_a_provider_or_source_finding(tmp_path, route_name):
+    """A sandbox-only failed scan remains incomplete without a fabricated finding."""
+    result = run_emitter(tmp_path, route_name, f"## Failed check: Strix Security Scan/strix\n{SANDBOX_SIGNAL}\n")
+    assert result.returncode == 1
+    assert "Strix sandbox bootstrap failure" in result.stderr
+    assert "Do not approve" in result.stderr
+    assert "No PR review was posted" in result.stderr
+    assert "### 1." not in result.stdout
+    assert "gateway or its discovered provider pool was unavailable" not in result.stdout
+
+
+@pytest.mark.parametrize("route_name", ["helper", "inline"])
+@pytest.mark.parametrize("evidence_text", [
+    "## Failed check: Strix Security Scan/strix\nSTRIX_SANDBOX_UNAVAILABLE: bootstrap failed\n",
+    f"## Failed check: Strix Security Scan/strix\nUnknown failure\n## Failed check: Unit tests\n{SANDBOX_SIGNAL}\n",
+])
+def test_sandbox_signal_requires_the_paired_marker_in_the_strix_check(tmp_path, route_name, evidence_text):
+    """Unrelated checks and a bare substring are not authoritative sandbox evidence."""
+    result = run_emitter(tmp_path, route_name, evidence_text)
+    assert result.returncode == 1
+    assert "Strix sandbox bootstrap failure" not in result.stderr
+    assert "### 1." not in result.stdout
+
+
+@pytest.mark.parametrize("route_name", ["helper", "inline"])
+@pytest.mark.parametrize("sandbox_line", ["", SANDBOX_SIGNAL + "\n", SANDBOX_SIGNAL + "; "])
+def test_provider_evidence_keeps_its_existing_classification(tmp_path, route_name, sandbox_line):
+    """Adding a sandbox distinction must not remove independent provider evidence."""
+    result = run_emitter(tmp_path, route_name, f"## Failed check: Strix Security Scan/strix\n{sandbox_line}RateLimitError\n")
+    assert result.returncode == 0
+    assert "### 1. HIGH" in result.stdout
+    assert "provider" in result.stdout
+    assert ("Strix sandbox bootstrap failure" in result.stderr) == bool(sandbox_line)
+
+
+@pytest.mark.parametrize("source_path", ["README.md", "missing_source.py"])
+def test_sandbox_signal_preserves_a_separately_reported_vulnerability(tmp_path, source_path):
+    """Emitted findings are not a passing scan, even when the helper exits zero."""
+    evidence_text = f"""## Failed check: Strix Security Scan/strix
+{SANDBOX_SIGNAL}
+
+### Strix vulnerability report window 1
+Model openai/example-model Vulnerabilities 1
+│ Vulnerability Report
+│ Title: Fixture vulnerability
+│ Severity: HIGH
+│ Endpoint: /example
+│ Code Locations
+│ Location 1: {source_path}:1
+"""
+    result = run_emitter(tmp_path, "helper", evidence_text)
+    if source_path == "README.md":
+        assert result.returncode == 0
+        assert "HIGH README.md:1 - Strix report from openai/example-model: Fixture vulnerability" in result.stdout
+    else:
+        assert result.returncode == 1
+        assert "Unmapped Strix report" in result.stderr
+        assert "openai/example-model" in result.stderr
+        assert "Fixture vulnerability" in result.stderr
+        assert "HIGH" in result.stderr
+        assert "did not include a mappable Code Location" in result.stderr
+        assert "### 1." not in result.stdout
+    assert "### 2." not in result.stdout
+    assert "Strix sandbox bootstrap failure" in result.stderr
+    assert "no Strix Vulnerability Report window was produced" not in result.stdout