AI Agent 全栈构建参考：24 篇生产级 TypeScript 指南（ReAct / MCP / RAG / 多 Agent / 记忆 / 护栏 / 部署 / 产品化），坚持不重复造轮子，逐层对标业界成熟方案（Vercel AI SDK / LangGraph / LlamaIndex / Langfuse / LiteLLM），附 2026 技术生态现状与开源项目二次开发对标清单。Production-grade AI Agent reference for any coding agent, with build-vs-buy decision tree + benchmark vs industry open-source.

文件	最后提交记录	最后更新时间
.github	remove-ci-workflow	18 天前
.probe-runtime	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
assets	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
docs	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
references	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
tools	fix: split offline verify from runtime probe (no false green) - npm run verify = check + audit + audit:selftest (offline, no network needed) - npm run verify:full = verify + probe (requires .probe-runtime install) - probe prints explicit SKIP when scratch deps absent, and no longer advertises a nonexistent --install flag - rationale: a gate that silently skips must not be reported as passing	5 天前
.gitignore	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
CHANGELOG.md	fix: split offline verify from runtime probe (no false green) - npm run verify = check + audit + audit:selftest (offline, no network needed) - npm run verify:full = verify + probe (requires .probe-runtime install) - probe prints explicit SKIP when scratch deps absent, and no longer advertises a nonexistent --install flag - rationale: a gate that silently skips must not be reported as passing	5 天前
CONTRIBUTING.md	chore: add project governance files - VERSION (1.0.0) - semantic versioning - CHANGELOG.md - following Keep a Changelog format - CONTRIBUTING.md - contribution guidelines - SECURITY.md - vulnerability reporting policy - .github/ - issue templates (bug + feature) + PR template Project structure is now complete with proper open-source conventions.	29 天前
LICENSE	fix: comprehensive audit — resolve 170 issues across all 24 references P0 fixes (38): - 17-streaming: SSE event field hardcoded 'event' → use parameter - 19-tenancy: Redis INCRBYBY → INCRBY, remove fake billing data - 20-marketplace: semver regex → semver lib, fix await in sync fn - 21-fault-recovery: remove 'stream' from serialization filter - 22-guardrails: OpenAI Agents SDK API doesn't exist → custom impl - 23-artifacts: isDestructiveAction always false, fake ariaSnapshot - 24-product: tiktoken import path, model names don't exist - 01-foundation: LLMMessage missing id/createdAt, content_filter mapping - 02-tools-skills: tiktoken incorrect API usage - 03-memory: BM25 drops Chinese chars, wrong chunk offset - 04-sandbox: seccomp=default invalid, whitelist ineffective - 07-security: zero-width regex broken, seccomp no serialization - 08-eval: claude-sonnet-4 model doesn't exist P1 fixes (52): - Token estimation inconsistencies standardized - Session storage LRU eviction added - Guardrail Promise.all → Promise.allSettled - Rate limiting atomic operations (Lua INCR+check) - HTML sanitizer bypass hardened - Path traversal guard fixed for Windows - Langfuse cost reporting split correctly - Model routing interface gaps filled - Nginx SSE rate limit added - Docker multi-stage COPY fix - Reflection uses independent evaluator LLM - HITL timeout clearTimeout Cross-file: - SECURITY.md: private disclosure flow + SLA - .gitignore: fix typo, add .turbo/.next/.vercel - QUICK_START: ORM统一为Drizzle, Node 15+, 24 nodes - PRE_LAUNCH_CHECKLIST: 21→24 nodes - CHANGELOG: Gite→Gitee, license year fix	25 天前
README.en.md	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
README.md	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
README.zh.md	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
ROADMAP.md	docs(roadmap): record v1.5.1 and mirror sync status (GitCode+Gitee aligned at 2fe2417)	5 天前
SECURITY.md	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
SKILL.md	fix: split offline verify from runtime probe (no false green) - npm run verify = check + audit + audit:selftest (offline, no network needed) - npm run verify:full = verify + probe (requires .probe-runtime install) - probe prints explicit SKIP when scratch deps absent, and no longer advertises a nonexistent --install flag - rationale: a gate that silently skips must not be reported as passing	5 天前
VERSION	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
package-lock.json	feat: v1.5.0 runtime probe + roadmap + node 26 components; fix verified API/version defects Runtime-verified fixes (probe found real defects, not theory): - AI SDK version baseline was WRONG: ai ^4.0.0 + @ai-sdk/openai-compatible ^1.0.0 throws AI_UnsupportedModelVersionError at streamText line 1. Locked to ai 4.3.19 + openai 1.3.24 + anthropic 1.2.12 + compat 0.2.16 (all live-tested). Added 00-types §6.1 pairing matrix with real error text. - LangGraph versions were missing: added @langchain/langgraph 1.5.3 + @langchain/core 1.3.2 (core 0.3.x -> Package subpath './utils/uuid' error). - 28.6 event bridge had 2 wrong API assumptions: streamMode:'updates' chunks are {nodeName: delta} not {node,type,payload}; interrupt() payload appears in invoke().__interrupt__ (tasks>0 signals pending), not polled from getState().values. Both rewritten per measurement. - 28.4 added measured constraint: compile({checkpointer}) requires configurable.thread_id on every call. Added: - tools/probe-runtime.mjs (npm run probe, wired into verify): verifies version pairing, streamText API surface, LangGraph chunk shape, interrupt/resume semantics, thread_id requirement. 5 notes, 0 failures. - ROADMAP.md: single source of truth for路线 (done v1.0-v1.5, in-progress, P0-P3 backlog with acceptance criteria, explicit non-goals, release cadence). - 26.4.1 full session-list components + 26.4.2 detail Sheet (558 -> 949 lines). Fixed drift: - SKILL.md metadata version stuck at 1.1.0 since v1.1.0 (4 versions behind) -> 1.5.0 - SECURITY.md supported versions 1.0.x -> 1.5.x - PRE_LAUNCH appendix line-count table: 24 rows of estimates -> 28 rows measured by wc -l (22/23/24 were off by 4-5x) - CHANGELOG: added missing v1.1.0 section, cleared stale Unreleased, added version links - 26.4.1 scheduleDelete arg order + empty restore function gates: 414 blocks parse 0 fail, audit 0 block 0 warn, selftest pass, probe 0 fail, linter PASS	5 天前
package.json	fix: split offline verify from runtime probe (no false green) - npm run verify = check + audit + audit:selftest (offline, no network needed) - npm run verify:full = verify + probe (requires .probe-runtime install) - probe prints explicit SKIP when scratch deps absent, and no longer advertises a nonexistent --install flag - rationale: a gate that silently skips must not be reported as passing	5 天前

🤖 agent-builder

生产级 AI Agent 构建从零参考 — 对齐 OpenAI / Anthropic 官方水准

English Docs

这是什么

一套帮助开发者和 AI Agent 从零构建企业级 AI Agent 的完整参考指南，以通用 Agent 可读格式组织。覆盖从 monorepo 脚手架到 Playground UI 的全栈 Agent 开发链路 —— 不重复造轮子：每个节点给出"手写实现 → 成熟方案"的映射与集成代码（Vercel AI SDK、LangGraph、LlamaIndex、E2B、OpenAI Agents SDK、Langfuse、LiteLLM 等），并附工程决策理由（🔗 工程逻辑）、AI 避坑注释（🤖 AI 避坑），经两轮以上交叉审计。

v1.1 起支持执行器模式：只给一段需求，AI 按零介入协议（不提问、只假设）自主完成规格补全 → 纵向切片 → 逐片实现 → 逐片回检 → 全量回归，交付可运行产品。

安装

# Claude Code / 通用 Agent
git clone https://gitcode.com/badhope/agent-builder.git ~/.claude/skills/agent-builder
# ZCode
git clone https://gitcode.com/badhope/agent-builder.git ~/.zcode/skills/agent-builder
# 验证
cd ~/.claude/skills/agent-builder && npm install && npm run verify
# verify = 代码块解析 + 自述审计 + 审计自测 + 运行期版本探测

涵盖内容

模块	核心能力
执行器	需求十二维补全 · 假设台账 · 纵向切片 · 交付状态机 · 降级矩阵 · 交付报告
基础	Turborepo · ReAct 循环 · SSE 流式 · Next.js · Electron 桌面端
工具与技能	Token 预算 · ToolRegistry · MCP Client · Skills 加载
记忆与 RAG	三层记忆（短期/长期/情景）· 混合检索 · Qdrant/Milvus/Chroma
多 Agent	Supervisor+Worker · LangGraph 集成 · 框架对比（LangGraph/CrewAI/AutoGen）
安全	Prompt Injection 防御 · Guardrails 校验 · 工具风险分类
生产部署	Docker + K8s · CI/CD · OpenTelemetry · Webhook · Playground UI
产品细节	收纳决策表 · 输入区与消息动作 · 会话列表 · 四态 · 键盘 · 响应式
回检	需求追踪矩阵 · 接线五跳 · 假实现检测 · 证据台账 · 三轮回检
对标参考	技术生态现状 · 开源项目对标与二次开发 · 调研报告

快速开始

在你的 AI 编程助手中直接说：

"帮我搭一个 Agent"
"Agent 的安全怎么保障"
"怎样让 Agent 自我反思"
"Agent 输出如何渲染成 Artifact"
"怎么让多个 Agent 协作"
"聊天产品需要哪些 UI 细节"
"只要能跑起来就行，别问我问题"

Agent 会自动加载本参考并按需读取对应参考节点。

对标标准

OpenAI Agents SDK — Guardrails 三层校验 + Runner 架构 + Session 持久化 + Tracing 内建
Anthropic Claude — Artifacts 协议 + Computer Use + Harness 架构 + CLAUDE.md/Auto Memory/Skills/Hooks
MCP (Model Context Protocol) — 最新规范下的 Client + Server 双向开发
Langfuse / OpenTelemetry — 全链路可追踪、可度量、可告警

技术栈

层	技术
前端	Next.js + React + Tailwind + shadcn/ui
后端	Node.js + Fastify + tRPC
LLM	OpenAI / Anthropic / Gemini（多 Provider 路由）
向量库	Qdrant / Milvus / Chroma（混合检索）
桌面端	Electron（contextIsolation + sandbox）
沙箱	Docker（cgroups + seccomp + cap-drop + 只读 rootfs）
部署	Docker Compose + Kubernetes
监控	OpenTelemetry + Langfuse

28 参考节点

#	节点	覆盖内容
00	types	跨层类型契约（LanguageModel / UIMessage / AgentEvent）
01	foundation	Turborepo + ReAct 循环 + SSE + Next.js + Electron
02	tools-skills	Token 预算 + ToolRegistry + MCP + Skills
03	memory-rag	三层记忆 + 混合检索
04	sandbox	Docker 沙箱 + 七层安全边界
05	multi-agent	Supervisor+Worker + LangGraph 集成
06	hitl-reflection	四级 HITL + LLM-as-Judge 自反思
07	security	Prompt Injection 防御 + Guardrails
08	eval-observability	Eval 框架 + Langfuse 全链路
09	deploy	Docker Compose + Electron + K8s + CI/CD
10	prompt-protocol	五层 Prompt + 版本管理 + A/B
11	function-calling	三 Provider 适配 + 并行派发
12	vector-db-practical	Qdrant + Milvus + Chroma
13	structured-output	Zod Schema + 容错 JSON 解析
14	mcp-server-dev	MCP Server 开发 (Stdio + SSE)
15	knowledge-base	多格式 RAG + 增量索引
16	multimodal-agent	多模态 + TTS/STT
17	streaming-advanced	SSE 优化 + 智能重连
18	framework-comparison	LangGraph/CrewAI/AutoGen 对比
19	multi-tenancy	多租户 + RBAC + 审计
20	agent-marketplace	市场 + 安全扫描 + 收益分成
21	fault-recovery	状态快照 + 幂等工具
22	guardrails-tracing	Guardrails + Tracing + 成本追踪
23	artifacts-computer-use	Artifacts + Computer Use + CLI
24	product-layer	Playground + 多模型路由 + Webhook
25	autonomous-delivery	零介入交付：需求补全 + 切片 + 状态机 + 降级
26	product-detail-catalog	产品细节目录：收纳 + 输入区 + 消息动作 + 四态
27	implementation-audit	实现回检：矩阵 + 五跳 + 假实现检测
28	framework-integration-playbook	框架融合流水线：七阶段 + 契约对齐 + 事件桥