Menu
Feature-centric deep incremental security review harness for Git-backed code changes, with autonomous, agents, and hybrid execution modes.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Optional graph-context enrichment for feature-centric incremental security review using CodeGraph or code-review-graph when available.
Pre-report finding verification contract. Re-checks whether each vulnerability has a real external Source, reachable Sink, bypassable or missing sanitization, and practical exploitability before final reporting.
Sink chain deep tracing methodology with graded code output templates (Critical full-chain / High-Medium key-nodes), reverse tracking rules, and per-hop Read verification.
Agent contract templates for R1 and R2+ rounds, including output format, token budget management, truncation defense, and auto-injection prompt templates.
Jalor internal Spring framework audit extension for endpoint operation authorization and service audit logging coverage.
Two-layer checklist architecture with D1-D10 security coverage matrix and language-specific semantic prompts for gap verification after free audit.
| name | audit-diff-harness |
| description | Feature-centric deep incremental security review harness for Git-backed code changes, with autonomous, agents, and hybrid execution modes. |
功能中心的 Git 增量深度安全检视协议。核心目标: 审计一个由多个 commit / PR / patch 实现的功能，而不是孤立审查几行 diff。
增量审计只有 deep。
默认由 AI 自主漏洞挖掘，D1-D10 只作为审后覆盖矩阵。
Agent 是可选放大镜，不是默认探索边界。
功能语义是审计边界，Git diff 是证据入口。
功能文档是上下文，不是漏洞成立证据。
最终 finding 必须同时绑定:
优先要求用户给出功能名或功能说明:
/git-audit --feature "报表导出" --range main..HEAD
/git-audit --feature "SSO 登录改造" --commits a1b2c3,d4e5f6
/git-audit --feature "支付回调接入" --patch patches/payment.diff --docs specs/payment.md
自然语言也可接受:
审计这次“批量导入客户”功能，代码在 PR 21、22、25，需求文档在 docs/import-customer.md
若未提供功能名，调度器必须从 branch 名、commit message、PR 标题、diff 文件名和变更模块推断 [FEATURE_PROFILE]，并把 feature_name_inferred 写入 [CONTEXT_GAPS]。
支持以下 Git-backed 范围:
| 输入 | 含义 |
|---|---|
--worktree | staged + unstaged + untracked，本地工作区整体改动 |
--staged | 只审暂存区 |
--unstaged | 未暂存改动 + untracked |
--commit <sha> | 审单个 commit，对比 <sha>^..<sha> |
--commits a,b,c | 审一组 commit；用于一个功能分散在多个提交 |
--range <base>..<head> | 审 revision range |
--base <ref> --head <ref> | 显式 base/head |
--merge-base <base> <head> | 使用 merge-base 作为 base |
--patch <file.diff> | 审离线 patch |
--prs 1,2,3 | 可选解析器；若无法解析，要求用户提供 branch/range/patch |
| 输入 | 用途 |
|---|---|
--docs <paths> | 功能需求、设计、接口、权限矩阵、测试计划 |
--focus D1,D3,D9 | 用户指定重点，但不能跳过明显相关风险 |
| `--engine autonomous | agents |
--include/--exclude | 限制或排除路径 |
{target}/audit-context.md | 项目级技术栈、暴露面、内部框架语义 |
.codegraph/ / .code-review-graph/ | 可选图上下文，补充 impact radius / affected flows |
autonomous (default)主审计 AI 自主完成漏洞挖掘:
feature intent
-> expected security properties
-> trust boundaries
-> implementation diff
-> supporting code
-> source/control/sink proof
-> exploitability and severity
D1-D10 在该模式下只用于审后 coverage self-check 和报告分类。
agents结构化并行审计。调度器根据 [FEATURE_SECURITY_PROFILE] 启动相关专业 Agent:
| 风险信号 | Agent |
|---|---|
| injection / query / template / expression | audit-d1-injection |
| auth / permission / ownership / workflow | audit-d2d3d9-control |
| deserialization / RCE / script engine | audit-d4-rce |
| file / archive / download / outbound URL | audit-d5d6-file-ssrf |
| crypto / config / dependency / CI | audit-d7d8d10-config |
每个 Agent 必须只审 [FEATURE_PROFILE] + [DIFF_SCOPE] + [RELEVANT_SUPPORTING_FILES]，不得扩散为全仓库审计。
hybrid自主主线 + 定向 Agent。主审计 AI 先自由建模和发现风险，再把高风险簇派给对应 Agent 深挖或复核。
调度器在进入漏洞挖掘前必须生成:
[FEATURE_PROFILE]
name: {provided|inferred feature name}
intent: {what capability is added or changed}
actors: {anonymous|user|admin|service|third_party|unknown}
assets: {data, files, credentials, money, tenant data, control plane}
entrypoints: {routes, jobs, consumers, CLI, callbacks, config toggles}
trust_boundaries: {internet->app, user->tenant data, service callback->internal, ...}
security_properties: {authn, authz, ownership, validation, signing, replay, isolation, ...}
docs: {provided paths, missing, conflicts}
context_confidence: high|medium|low
[DIFF_SCOPE]
mode: worktree|staged|unstaged|commit|commits|range|base_head|merge_base|patch|prs
repo_root: {absolute path}
base: {ref or n/a}
head: {ref or n/a}
commits: {list or n/a}
patches: {list or n/a}
changed_files: {count and representative list}
changed_hunks: {file:line ranges}
renames_deletes: {if any}
excluded: {path and reason}
[FEATURE_SECURITY_PROFILE]
new_entrypoints: {...}
modified_controls: {...}
modified_sinks: {...}
modified_data_models: {...}
modified_file_network_ops: {...}
modified_crypto_config: {...}
dependency_ci_changes: {...}
risk_dimensions: {D1-D10 with reasons}
[RELEVANT_SUPPORTING_FILES]
{path}: reason={caller|callee|auth_chain|route_registry|sanitizer|sink_helper|config_profile|dependency_context}
[GRAPH_CONTEXT_PROFILE]
providers:
codegraph: available|missing|not_indexed|error
code-review-graph: available|missing|not_indexed|error
graph_confidence: none|low|medium|high
risk_score: {optional provider risk score}
changed_symbols: {name,file,line,provider}
impacted_files: {path,reason,provider}
affected_flows: {name,criticality,files}
test_gaps: {symbol,file,line}
supporting_files_added: {path,reason,provider}
limitations: {stale graph, missing graph, provider errors}
[CONTEXT_GAPS]
{missing feature docs, unknown exposure, unresolved PR fetch, conflicting doc/code assumptions}
必须使用确定性 worklist:
audit_generate_diff_worklist(...)
该工具由 .opencode/plugin/git-diff-harness.js 独立暴露。若 OpenCode plugin tool 不可用，才 fallback 到:
python3 references/core/git_diff_worklist.py ...
默认输出:
{target_project}/audit-output/git-diff-scans/{scan_id}/artifacts/02_worklist/diff_worklist.csv
{target_project}/audit-output/git-diff-scans/{scan_id}/artifacts/02_worklist/deep_review_input.csv
{target_project}/audit-output/git-diff-scans/{scan_id}/artifacts/02_worklist/work_ledger.md
{target_project}/audit-output/git-diff-scans/{scan_id}/artifacts/06_graph_context/graph_context.json
{target_project}/audit-output/git-diff-scans/{scan_id}/artifacts/06_graph_context/graph_context.md
每个 source-like changed file 必须有覆盖结论:
| disposition | 含义 |
|---|---|
covered | 完整审阅了 changed hunks 和必要上下文 |
not_applicable | 非安全相关或非源码，但说明理由 |
suppressed | 用户排除或 generated/vendor/test，说明证据 |
deferred | 因缺少 PR/patch/docs/依赖或预算无法闭合，说明 blocker |
只有搜索命中或只看 hunk 不算 covered。必须记录功能语义、变更行为、安全属性和候选结论。
在 worklist 生成后、漏洞挖掘前，调度器应加载 audit-graph-context 并运行:
audit_generate_graph_context(...)
该工具由 .opencode/plugin/git-diff-harness.js 独立暴露。若 OpenCode plugin tool 不可用，才 fallback 到:
python3 references/core/graph_context_adapter.py ...
Graph context 的作用:
[GRAPH_REVIEW_QUEUE]，改变漏洞挖掘的第一阅读顺序[RELEVANT_SUPPORTING_FILES]hybrid 模式决定是否派专项 AgentGraph context 的限制:
graph_review_queue 非空，必须先按队列读代码，再回到普通 diff 文件顺序必须输出:
[GRAPH_CONTEXT_PROFILE]
[GRAPH_REVIEW_QUEUE]
[GRAPH_SUPPORTING_FILES]
支持的 graph-derived supporting file reasons:
graph_caller
graph_callee
graph_impact
graph_flow
graph_test_gap
graph_route
graph_priority
最终 finding 必须回答:
不得报告:
当 engine=agents|hybrid 时，注入给每个子 Agent:
[FEATURE_PROFILE] 功能意图、角色、资产、入口、信任边界、预期安全属性
[DIFF_SCOPE] Git 范围、changed files、changed hunks
[FEATURE_DOC_CONTEXT] 已读取的功能文档摘要、冲突和缺口
[RELEVANT_SUPPORTING_FILES] 允许读取的 supporting files 和理由
[GRAPH_CONTEXT_PROFILE] 可选图上下文、provider 状态、impact/flow/test-gap 信号
[GRAPH_SUPPORTING_FILES] graph-derived supporting files，必须实际读取后才可作为证据
[INCREMENTAL_RULES]
- deep-only
- stay feature-scoped
- final candidate must bind to changed line/hunk/control or diff-affected old sink
- D1-D10 is coverage taxonomy, not exploration limit
子 Agent 输出中的 CANDIDATE_LEDGER 仍写入现有审计 DB；不得创建 per-agent JSONL 候选账本。
最终报告使用 references/core/git_diff_report_template.md。报告必须保存到:
{target_project}/audit-output/git-diff-scans/{scan_id}/report.md
报告必须包含: