OpenCode Review에 구조화된 적대적 검증을 강제

.github/workflows/opencode-review.yml

@@ -3271,13 +3271,13 @@ jobs:
           OPENCODE_SMALL_CHANGE_FILE_THRESHOLD: "3"
           OPENCODE_MEDIUM_CHANGE_FILE_THRESHOLD: "20"
           OPENCODE_SMALL_CHANGE_RUN_TIMEOUT_SECONDS: "300"
-          OPENCODE_SMALL_CHANGE_TOTAL_BUDGET_SECONDS: "900"
-          OPENCODE_MEDIUM_CHANGE_RUN_TIMEOUT_SECONDS: "600"
-          OPENCODE_MEDIUM_CHANGE_TOTAL_BUDGET_SECONDS: "1500"
-          OPENCODE_LARGE_CHANGE_RUN_TIMEOUT_SECONDS: "900"
-          OPENCODE_LARGE_CHANGE_TOTAL_BUDGET_SECONDS: "1500"
-          OPENCODE_UNKNOWN_CHANGE_RUN_TIMEOUT_SECONDS: "600"
-          OPENCODE_UNKNOWN_CHANGE_TOTAL_BUDGET_SECONDS: "1500"
+          OPENCODE_SMALL_CHANGE_TOTAL_BUDGET_SECONDS: "600"
+          OPENCODE_MEDIUM_CHANGE_RUN_TIMEOUT_SECONDS: "420"
+          OPENCODE_MEDIUM_CHANGE_TOTAL_BUDGET_SECONDS: "900"
+          OPENCODE_LARGE_CHANGE_RUN_TIMEOUT_SECONDS: "420"
+          OPENCODE_LARGE_CHANGE_TOTAL_BUDGET_SECONDS: "1080"
+          OPENCODE_UNKNOWN_CHANGE_RUN_TIMEOUT_SECONDS: "420"
+          OPENCODE_UNKNOWN_CHANGE_TOTAL_BUDGET_SECONDS: "900"
           OPENCODE_DYNAMIC_MAX_CYCLES: "1"
           CENTRAL_REVIEW_PROCESS_FALLBACK_ELIGIBLE: ${{ steps.central_review_process_fallback_scope.outputs.eligible || 'false' }}
           CENTRAL_REVIEW_PROCESS_FALLBACK_SCOPE_LABEL: ${{ steps.central_review_process_fallback_scope.outputs.scope_label || 'unsupported' }}
@@ -3290,6 +3290,7 @@ jobs:
           OPENCODE_AGENT: ci-review-fallback
           OPENCODE_EVIDENCE_FILE: ${{ runner.temp }}/opencode-review-evidence.md
           OPENCODE_CHANGED_FILES_FILE: ${{ runner.temp }}/opencode-changed-files.txt
+          OPENCODE_REQUIRE_ADVERSARIAL_VALIDATION: "true"
           OPENCODE_OUTPUT_FILE: ${{ runner.temp }}/opencode-review-model-pool.md
           OPENCODE_REVIEW_WORKDIR: ${{ runner.temp }}/opencode-review-project
           OPENCODE_SOURCE_WORKDIR: ${{ runner.temp }}/opencode-pr-head
@@ -3397,6 +3398,8 @@ jobs:
           # re-rejects it (NO_CONCLUSION / exit 4), failing an otherwise valid
           # review instead of publishing it.
           OPENCODE_EVIDENCE_FILE: ${{ runner.temp }}/opencode-review-evidence.md
+          OPENCODE_CHANGED_FILES_FILE: ${{ runner.temp }}/opencode-changed-files.txt
+          OPENCODE_REQUIRE_ADVERSARIAL_VALIDATION: "true"
           # The publish gate re-runs source-backed validation against PR-head data.
           OPENCODE_SOURCE_WORKDIR: ${{ runner.temp }}/opencode-pr-head
           PR_BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.inputs.pr_base_sha }}
@@ -3666,6 +3669,13 @@ jobs:
           HEAD_REF: ${{ github.event.pull_request.head.ref || github.event.inputs.pr_head_ref || '' }}
           RUN_ID: ${{ github.run_id }}
           RUN_ATTEMPT: ${{ github.run_attempt }}
+          OPENCODE_MODEL_POOL_OUTPUT_FILE: ${{ runner.temp }}/opencode-review-model-pool.md
+          OPENCODE_EVIDENCE_FILE: ${{ runner.temp }}/opencode-review-evidence.md
+          OPENCODE_CHANGED_FILES_FILE: ${{ runner.temp }}/opencode-changed-files.txt
+          OPENCODE_SOURCE_WORKDIR: ${{ runner.temp }}/opencode-pr-head
+          OPENCODE_REQUIRE_ADVERSARIAL_VALIDATION: "true"
+          PR_BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.inputs.pr_base_sha }}
+          PR_HEAD_SHA: ${{ github.event.pull_request.head.sha || github.event.inputs.pr_head_sha }}
           CENTRAL_REVIEW_PROCESS_FALLBACK_CHANGED_COUNT: ${{ steps.central_review_process_fallback_scope.outputs.changed_count || '0' }}
           CENTRAL_REVIEW_PROCESS_FALLBACK_SCOPE_LABEL: ${{ steps.central_review_process_fallback_scope.outputs.scope_label || 'unsupported' }}
           APPROVAL_CHECK_WAIT_ATTEMPTS: "12"
@@ -3683,6 +3693,27 @@ jobs:
             exit 1
           fi
 
+          model_output_copy="$(mktemp)"
+          normalized_control_file="$(mktemp)"
+          if [ ! -s "${OPENCODE_MODEL_POOL_OUTPUT_FILE:-}" ]; then
+            echo "::error::CENTRAL_FAST_APPROVAL_NO_MODEL_OUTPUT: selected current-head model output is unavailable."
+            exit 1
+          fi
+          perl -pe 's/\x1b\[[0-9;?]*[A-Za-z]//g' "$OPENCODE_MODEL_POOL_OUTPUT_FILE" >"$model_output_copy"
+          if ! python3 scripts/ci/opencode_review_normalize_output.py \
+            "$HEAD_SHA" "$RUN_ID" "$RUN_ATTEMPT" "$model_output_copy"; then
+            echo "::error::CENTRAL_FAST_APPROVAL_ADVERSARIAL_INVALID: selected model output did not satisfy the structured adversarial contract."
+            exit 1
+          fi
+          gate_result="$(
+            bash scripts/ci/opencode_review_approve_gate.sh \
+              "$HEAD_SHA" "$RUN_ID" "$RUN_ATTEMPT" "$model_output_copy" "$normalized_control_file"
+          )"
+          if [ "$gate_result" != "APPROVE" ]; then
+            echo "::notice::Central fast approval skipped because the adversarially validated model verdict was ${gate_result:-unknown}, not APPROVE."
+            exit 0
+          fi
+
           api_url="https://api.github.com"
           api_timeout="${REVIEW_PUBLISH_GH_API_TIMEOUT_SECONDS:-15}"
           read_token="${CHECK_LOOKUP_GH_TOKEN:-$GH_TOKEN}"
@@ -3800,19 +3831,28 @@ jobs:
             exit 1
           fi
 
+          model_reason="$(jq -r '.reason' "$normalized_control_file")"
+          model_summary="$(jq -r '.summary' "$normalized_control_file")"
+          adversarial_evidence="$(jq -c '.adversarial_validation' "$normalized_control_file")"
           body="$(printf '%s\n' \
             "## Pull request overview" \
             "" \
-            "OpenCode reviewed the current-head central review-process repair evidence and found no blocking issues." \
+            "$model_summary" \
             "" \
             "## Findings" \
             "" \
             "No blocking findings." \
             "" \
+            "## Adversarial validation" \
+            "" \
+            '```json' \
+            "$adversarial_evidence" \
+            '```' \
+            "" \
             "## Evidence" \
             "" \
             "- Result: APPROVE" \
-            "- Reason: central review-process fast approval; model-pool output succeeded and current-head peer checks, code-scanning alerts, and review threads were clear." \
+            "- Reason: ${model_reason}" \
             "- Scope: \`${CENTRAL_REVIEW_PROCESS_FALLBACK_SCOPE_LABEL:-unknown}\`" \
             "- Changed files: \`${CENTRAL_REVIEW_PROCESS_FALLBACK_CHANGED_COUNT:-unknown}\`" \
             "- Head SHA: \`${HEAD_SHA}\`" \
@@ -3850,6 +3890,8 @@ jobs:
           OPENCODE_EVIDENCE_FILE: ${{ runner.temp }}/opencode-review-evidence.md
           OPENCODE_FAILED_CHECK_EVIDENCE_FILE: ${{ runner.temp }}/opencode-failed-check-evidence.md
           OPENCODE_FAILED_CHECK_DIAGNOSIS_FILE: ${{ runner.temp }}/opencode-failed-check-diagnosis.md
+          OPENCODE_CHANGED_FILES_FILE: ${{ runner.temp }}/opencode-changed-files.txt
+          OPENCODE_REQUIRE_ADVERSARIAL_VALIDATION: "true"
           COVERAGE_EVIDENCE_RESULT: ${{ needs.coverage-evidence.result || 'skipped' }}
           COVERAGE_EVIDENCE_SUMMARY: ${{ needs.coverage-evidence.outputs.coverage_summary || 'Coverage evidence job did not run or did not publish coverage evidence.' }}
           OPENCODE_REVIEW_WORKDIR: ${{ runner.temp }}/opencode-review-project
@@ -4654,9 +4696,11 @@ jobs:
             local summary
             local reason
             local findings
+            local adversarial_evidence
 
             summary="$(jq -r '.summary // ""' "$control_json")"
             reason="$(jq -r '.reason // ""' "$control_json")"
+            adversarial_evidence="$(jq -c '.adversarial_validation' "$control_json")"
             findings="$(
               # shellcheck disable=SC2016
               jq -r '
@@ -4684,6 +4728,8 @@ jobs:
               printf '%s\n\n' "$findings"
               printf '## Summary\n\n'
               printf '%s\n\n' "$summary"
+              printf '## Adversarial validation\n\n'
+              printf '```json\n%s\n```\n\n' "$adversarial_evidence"
               printf -- '- Result: REQUEST_CHANGES\n'
               printf -- '- Reason: %s\n\n' "$reason"
               printf -- "- Head SHA: \`%s\`\n" "$HEAD_SHA"
@@ -6049,7 +6095,7 @@ jobs:
                 ' >"$output_file"
           }
 
-          approve_current_head_after_model_unavailable() {
+          publish_blockers_after_model_unavailable() {
             local pending_wait_status body
 
             if [ "${COVERAGE_EVIDENCE_RESULT:-skipped}" != "success" ]; then
@@ -6136,30 +6182,8 @@ jobs:
               return 0
             fi
 
-            body="$(printf '%s\n' \
-              "## Pull request overview" \
-              "" \
-              "OpenCode model providers were unavailable for this same-head run, but deterministic current-head evidence is clean: coverage evidence passed, peer GitHub Checks are complete, medium-or-higher code-scanning alerts are clear, mergeability is clean, and reviewer threads are resolved or outdated." \
-              "" \
-              "## Findings" \
-              "" \
-              "No blocking findings." \
-              "" \
-              "## Evidence" \
-              "" \
-              "- Result: APPROVE" \
-              "- Reason: current-head model-unavailable evidence fallback; coverage, docstring, peer GitHub Checks, code-scanning alerts, mergeability, and review threads were clear for current head." \
-              "- Scope: \`${CENTRAL_REVIEW_PROCESS_FALLBACK_SCOPE_LABEL:-unknown}\`" \
-              "- Changed files: \`${CENTRAL_REVIEW_PROCESS_FALLBACK_CHANGED_COUNT:-unknown}\`" \
-              "- Model-pool outcome: \`${OPENCODE_MODEL_POOL_OUTCOME:-unknown}\`" \
-              "- Head SHA: \`${HEAD_SHA}\`" \
-              "- Workflow run: ${RUN_ID}" \
-              "- Workflow attempt: ${RUN_ATTEMPT}" \
-              "" \
-              "This fallback does not suppress failed checks, medium-or-higher code-scanning alerts, merge conflicts, unresolved reviewer threads, or failed coverage evidence; any of those conditions still publish REQUEST_CHANGES or leave the approval state unchanged."
-            )"
-            create_pull_review "APPROVE" "$body"
-            return 0
+            printf '::notice::MODEL_OUTPUT_UNAVAILABLE: current-head checks and alerts are clean, but no APPROVE review will be published without mandatory structured adversarial probes for head %s.\n' "$HEAD_SHA"
+            return 1
           }
 
                     request_changes_for_merge_conflict_if_present() {
@@ -6262,7 +6286,7 @@ jobs:
           # diagnostics. Do not read model output or change PR review state unless
           # the pool explicitly emitted a valid current-head control block.
           if [ "$opencode_review_outcome" != "success" ]; then
-            if approve_current_head_after_model_unavailable; then
+            if publish_blockers_after_model_unavailable; then
               echo "::endgroup::"
               exit 0
             fi
@@ -6482,6 +6506,7 @@ jobs:
               fi
               summary="$(jq -r '.summary' "$control_json")"
               reason="$(jq -r '.reason' "$control_json")"
+              adversarial_evidence="$(jq -c '.adversarial_validation' "$control_json")"
               body="$(printf '%s\n' \
                 "## Pull request overview" \
                 "" \
@@ -6495,6 +6520,12 @@ jobs:
                 "" \
                 "$summary" \
                 "" \
+                "## Adversarial validation" \
+                "" \
+                '```json' \
+                "$adversarial_evidence" \
+                '```' \
+                "" \
                 "- Result: APPROVE" \
                 "- Reason: ${reason}" \
                 "- Head SHA: \`${HEAD_SHA}\`" \

ci-review-prompt.md

@@ -135,6 +135,18 @@ false-positive dismissal for each plausible peer finding, request changes with
 your own line-specific finding and verification direction.
 When another review bot reports a plausible current-head static-analysis, linter, compiler, or accessibility defect, verify the claim independently before approving. For JSX/TSX and component templates, duplicate props such as repeated `aria-label`, repeated event handlers, or assignments overwritten later in the same element/object are blocking when they can mask the intended accessible name, event behavior, data binding, or runtime value. Do not approve by merely citing the peer bot; inspect the changed hunk or run the relevant parser/linter/typecheck in a scratch workspace, then either publish your own source-backed finding or explain the source-backed false-positive dismissal.
 
+Perform an explicit adversarial phase before every verdict. Assume the patch is
+wrong, derive concrete failure hypotheses for each materially changed surface,
+and attempt to trigger them with malformed or boundary inputs, authorization or
+tenant crossover, stale or concurrent state, dependency/runtime mismatch,
+error/rollback behavior, numerical extremes, or mobile/accessibility behavior
+as applicable. A green check or absence of a known bug is not a probe. Record
+the exact changed path, positive line, counterexample, executed or source-backed
+evidence, and whether the hypothesis was falsified or confirmed in the
+`adversarial_validation` control field. APPROVE needs two falsified probes for
+material code/workflow/config/package/test changes and one for non-code changes;
+REQUEST_CHANGES needs a confirmed probe anchored to a published finding.
+
 Review the diff first, then inspect surrounding code only when needed to
 understand impact. Evaluate correctness, API compatibility, security/privacy,
 data integrity, concurrency, error handling, observability, performance,
@@ -224,4 +236,4 @@ The final OpenCode output must still satisfy the existing
 `opencode-review-control-v1` JSON contract required by the approval gate. Use
 the reviewer rubric above for analysis and human-readable review quality, but
 return the sentinel and control block exactly as requested by the workflow
-prompt.
+prompt, including the mandatory structured `adversarial_validation` evidence.

code-reviewer-prompt.md

@@ -136,6 +136,16 @@ verify that `prefers-reduced-motion: reduce` users are not forced through smooth
 scrolling or animated motion.
 Treat peer review bot comments as adversarial seeds, not authority. If a peer bot flags a plausible current-head static-analysis, compiler, linter, or accessibility issue, independently verify it from the source hunk, parser/linter/typecheck output, runtime/library documentation, or a scratch repro before approving. In JSX/TSX and component templates, duplicate props such as repeated `aria-label`, repeated handlers, or assignments overwritten later in the same element/object are material defects when they can mask the intended accessible name, event behavior, data binding, or runtime value; report your own source-backed finding instead of merely quoting the peer bot.
 
+Run a dedicated adversarial phase before the verdict. Assume the proposed patch
+is wrong and build concrete counterexamples for each material changed surface:
+malformed or boundary inputs, authorization or tenant crossover, stale or
+concurrent state, dependency/runtime mismatch, error and rollback behavior,
+numerical extremes, or mobile and accessibility behavior as applicable. Trace
+or execute each probe and record the exact changed path, positive line,
+hypothesis, attack/counterexample, evidence, and falsified/confirmed outcome in
+the workflow's structured `adversarial_validation` control field. Green checks
+alone and absence of a known failure are not adversarial evidence.
+
 Implementation completeness is mandatory. Inspect changed runtime code and
 connected call sites for placeholder bodies such as `pass`, `...`,
 `NotImplementedError`, TODO-only branches, fake or constant returns, and

scripts/ci/opencode_review_approve_gate.sh

@@ -374,6 +374,7 @@ normalized = {
     "result": control["result"],
     "reason": control["reason"],
     "summary": control["summary"],
+    "adversarial_validation": control.get("adversarial_validation"),
     "findings": control.get("findings") or [],
 }
 Path(sys.argv[2]).write_text(