diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/07-standard-golden-diff.md b/docs/evidence/g02-intent-js-closed-loop-20261112/07-standard-golden-diff.md new file mode 100644 index 0000000..c5e6a63 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/07-standard-golden-diff.md @@ -0,0 +1,17 @@ +# G02 工作流与现有 Standard golden 的可解释差异 + +对照对象：仓库已冻结示例 `examples/editorial.workflow.js`（及其 `editorial.compiled.mjs` / `editorial.mmd`），它是 Standard 八类节点的回归 golden，由 `tests/runtime/golden.test.mjs` 维护。本文件只解释差异，不更新该 golden，也不把 editorial 的能力记为 Interactive 已实现。 + +| 维度 | editorial golden | G02 `feature-review.workflow.js` | 说明 | +|---|---|---|---| +| 产品目的 | 编辑审批：起草、并行审阅、循环修复、人工批准、发布 | 功能变更评审：入口门禁、需求分析、技术设计、独立评审、终态 | 都是“分析/评审/结论”类闭环，但业务对象不同 | +| 节点种类 | Action、Agent、Parallel、Condition、Loop、Verification、Human、Terminal | Condition、Agent、Terminal | G02 必须同时通过 Interactive P0-A 功能门禁，因此不放入尚未实现可恢复路径的节点 | +| 条件路由 | `styleOk && factOk` 决定进入批准或质量循环 | `materialsReady` 决定是否进入分析；`reviewApproved` 决定成功或失败终态 | 都由编译后的 JS 调用 `evalWhen` / Interactive `advance()` 求值，而不是由模型选择下一节点 | +| 重试 | 起草 Agent `retries: 1`；另有 Loop 质量预算 | 仅评审 Agent `retries: 1`；分析/设计 `retries: 0` | G02 用 Schema 校验失败演示一次可控重试，不引入 Loop | +| 并行 | `reviews` 并行 style/fact，`all_succeeded` | 无 | Parallel 属于 Standard 已实现、Interactive 未实现（P1/P2） | +| 人工/动作 | Human 批准/拒绝；Action 准备、修复、修订、发布 | 无 | 这些是 Standard 已实现语义；Interactive 要到 P0-B。G02 不把它们算进 Interactive 验收 | +| 编译器 | 同一套 Standard 静态分析器与 `flowkit compile` | 同一分析器；另外再走 `compileInteractive` 生成独立的 `.interactive.mjs` | 没有第二套 DSL 解析器 | +| 执行入口 | Standard `run()` | 同时演示 Standard `run()` 与 Interactive Coordinator/Trusted Host | Interactive 产物不能用 Standard `run()` 加载，反之亦然 | +| Agent 输出来源 | 示例适配器按 Skill 词汇生成固定对象，不调用真实模型 | 成功/拒绝路径的 Agent 正文由 Grok 在看到 Pending 后撰写；非法评审是脚本负例 | 证据等级不同，见夹具说明 | + +结论：G02 有意使用 P0-A 可执行子集来验证“意图 → DSL → 真实 JS Program → 程序决定分支”的产品闭环；完整八类节点的 Standard 行为仍以 editorial golden 为准，不在本阶段扩展 Interactive。 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/HASHES.sha256 b/docs/evidence/g02-intent-js-closed-loop-20261112/HASHES.sha256 new file mode 100644 index 0000000..9fdb43b --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/HASHES.sha256 @@ -0,0 +1,38 @@ +8df70cd1af7dd9b0b5ea2c2143815c8aca862696f580fa46b21f726c021ec592 examples/g02/feature-review.workflow.js +5c2f8a1e4ce0faeeb2d95176bcd3b5c70dd270e1096f20d1bc2b69360f995ed5 examples/g02/feature-review.compiled.mjs +8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded examples/g02/feature-review.interactive.mjs +82be98702a59ae251265a0eae9bbe4a3e60654d88bb16f96ac24f5f93166aa6b examples/g02/feature-review.mmd +26653dd276028a77660f9268c22c2c1a023bbc8352d0f69a0d7159e6b36f5b08 examples/g02/skills/analyze.md +c14f956000b3bf9e3786af50a85c01d2b282b087f4cc4606ccffa47616ae0ba9 examples/g02/skills/design.md +464ddaea7cce6f7d2479db0b3fa91d36fd75e28a346661b19c5b29d70decf5ed examples/g02/skills/review.md +9ea9606d0aca8dbb15478fbdccf511cd2ac9672cd51cf4aa1f190f2d390a0acb examples/g02/README.md +d47177cd5bacd17727c10da4efe32ab07758762a8d909b47c58c7ebbae4aaf67 examples/g02/run-interactive.mjs +dc533fb324204cbddeb62a5046fd18779e3a4137095ca4b0aadc8bd74a57e56d tests/fixtures/g02/turns-review-false.json +3aec13cf3fc8f851350656a879aa391e33cd33defe1c73b6552107174e311637 tests/fixtures/g02/turns-success-retry.json +d4380d813435a1249ebc2515b82ba00842ffab73bbbf03b41cb2b118f0d48119 tests/interactive/g02-product-closed-loop.test.mjs +aa4712816a41cd509c7d00227b253b187526e32b130c1ce6df79444435808054 docs/evidence/g02-intent-js-closed-loop-20261112/README.md +acccfc9b1782863557884ce4b8263e47d8130bbae8bfa918567009a2a6a4333b docs/evidence/g02-intent-js-closed-loop-20261112/07-standard-golden-diff.md +8f62bacd12654004b5b75c98b5279281001b99cb6a029a641b412319478fe45e docs/evidence/g02-intent-js-closed-loop-20261112/cli/00-timestamp.txt +19eaf43821a7660ec323a87c8457bf74823beb296c39f5e01aa8a683aa50f061 docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.exit.txt +e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.stderr.txt +e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.stdout.txt +19eaf43821a7660ec323a87c8457bf74823beb296c39f5e01aa8a683aa50f061 docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.exit.txt +e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.stderr.txt +e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.stdout.txt +95e4d85ba0a5f006130163dbcd634f16334ca086ecb9c123e3d38f5291510951 docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-interactive-compile.json +1017181cab7dee17344997bf532bdd82e4589c930e8aba27ec2734fe2409fcee docs/evidence/g02-intent-js-closed-loop-20261112/execution/00-env.txt +d83a6c9a84da539a1cba34f1d9e77c7df4d862dbef8bc096b85702b49d7f0364 docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.exit.txt +e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.stderr.txt +56e24cf17a253f42a96289540473b9839b5ed69f2c5ace49b6102f10e7eda3ee docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.stdout.json +e01a5b5252b8a0bffd0b9e9d7e80311beeae9e4d9eeea0faabcdf8a3c3bf249e docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.exit.txt +e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.stderr.txt +46ba14286185325effe1e611f53750ac8ee7be01ab81fd40e0ddf79cfeec211b docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.stdout.txt +9cdac5e8dac099f5d7ddcc978954fe17e9419d6c329fb6b8896ba51e991948a3 docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-npm-test.summary.txt +05e2066203713f7c9f74d74b965ab2b12dc57d2c27843728cfabd5edae29875c docs/evidence/g02-intent-js-closed-loop-20261112/execution/r02-pending-counts-measured.json +b16a74729a277c11d5e28daf8b2691f248250d8b4221cc7129fb659ee306d56c docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.exit.txt +e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.stderr.txt +e0785102cc8fd7a28f4bb6a851581a84371e27638c035267d8896f04a4932e7c docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.stdout.json +be60fe6a557e18f6b34e1360b536f137c9711a5ab69a5afd0eb093636eef17c4 docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/00-prompt.md +76ddeea02e9ef200a170c0768f044c7d8f2d07a8b3b13c0f22f75abfe9226dd4 docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/00-user-intent.md +ef2b7b97991a24ea539d55a2c071bbeaba520956a4567ecbc5455a1d85798d0e docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/attempt-01-notes.md +8df70cd1af7dd9b0b5ea2c2143815c8aca862696f580fa46b21f726c021ec592 docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/attempt-01-raw.workflow.js diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/README.md b/docs/evidence/g02-intent-js-closed-loop-20261112/README.md new file mode 100644 index 0000000..b8b2bdf --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/README.md @@ -0,0 +1,15 @@ +# G02 证据目录（2026-11-12，Asia/Shanghai） + +本目录保存 G02「真实用户意图 → Workflow DSL → 可执行 JavaScript → Standard / Interactive P0-A 执行」的原始输入、编译日志和本地复现输出。它不是新的需求或技术设计，也不构成独立复审 PASS。 + +| 路径 | 内容 | +|---|---| +| `dsl-generation/` | Grok 生成 `.workflow.js` 时的完整输入、第一次原始输出，以及“一次通过、无第二次尝试”的说明 | +| `cli/` | 官方 `flowkit compile` / `flowkit check` 与 `compileInteractive` 的原始输出、退出码和 Program 身份 | +| `execution/` | Standard `run()`、Interactive 夹具重放、本地 `node --test` 的原始日志 | +| `07-standard-golden-diff.md` | 与 `examples/editorial.workflow.js` golden 的可解释差异 | +| `HASHES.sha256` | 上述文件及示例/夹具的 SHA-256 | + +证据分级：编译与 JS/SQLite 执行为 P；`source=grok-model` 为 M 层结果消费（6 次消费 / 5 份独立正文 / 1 次复用并重新绑定 Pending 身份，不是 6 次独立模型推理）；`source=scripted-invalid` 是脚本负例。`pendingAtEnd` 只描述终态帧；`publishedPendingCount` 来自实际执行 route。本环境用 Node 和 FlowKit CLI 执行 JavaScript，**不得**标为 N-PASS 或原生安全 Host 来源认证。GitHub CI 7/7 只证明仓库回归，不证明原生 Grok 工具或跨用户 OS 隔离。 + +正式审查请同时使用 GitHub PR、精确 HEAD、本目录与交接包中的 `STAGE-REVIEW-REPORT.md`。 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/cli/00-timestamp.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/00-timestamp.txt new file mode 100644 index 0000000..35e69c0 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/00-timestamp.txt @@ -0,0 +1,4 @@ +2026-11-12 17:10:07 CST +node: v24.22.1 +cwd: /workspace/flowkit +HEAD: 7e5960fc570f0cd91a5ff8e9fe6029345932442f diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.exit.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.exit.txt new file mode 100644 index 0000000..9200c2a --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.exit.txt @@ -0,0 +1 @@ +exit=0 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.stderr.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.stderr.txt new file mode 100644 index 0000000..e69de29 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.stdout.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-check.stdout.txt new file mode 100644 index 0000000..e69de29 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.exit.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.exit.txt new file mode 100644 index 0000000..9200c2a --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.exit.txt @@ -0,0 +1 @@ +exit=0 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.stderr.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.stderr.txt new file mode 100644 index 0000000..e69de29 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.stdout.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-compile.stdout.txt new file mode 100644 index 0000000..e69de29 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-interactive-compile.json b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-interactive-compile.json new file mode 100644 index 0000000..3a83060 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/cli/attempt-01-interactive-compile.json @@ -0,0 +1,19 @@ +{ + "timestamp": "2026-11-12T09:10:12.013Z", + "diagnostics": [], + "artifactPath": "/workspace/flowkit/examples/g02/feature-review.interactive.mjs", + "programIdentity": { + "workflowId": "g02-feature-review", + "workflowVersion": "1.0.0", + "sourceHash": "sha256:8df70cd1af7dd9b0b5ea2c2143815c8aca862696f580fa46b21f726c021ec592", + "compilerVersion": "1.0.0", + "programAbiVersion": 1, + "semanticsVersion": "interactive-state-machine-v1", + "checkpointSchemaVersion": 1, + "canonicalizationVersion": "jcs-rfc8785-v1", + "implementedStage": "P0-A", + "artifactHash": "sha256:8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded" + }, + "codeBytes": 15015, + "codeSha256": "8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded" +} diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/00-prompt.md b/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/00-prompt.md new file mode 100644 index 0000000..bbf5020 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/00-prompt.md @@ -0,0 +1,25 @@ +# G02 DSL 生成：Grok Build 原始模型输入 + +记录时间：2026-11-12，Asia/Shanghai。 +角色：Grok Bot / Grok Build，作为真实模型阅读用户意图，并按 FlowKit 现有 DSL 生成 `.workflow.js`。 +本文件是生成尝试使用的完整输入。输出见同目录 `attempt-NN-raw.workflow.js`；若编译失败，后续尝试会完整保留，不会覆盖本输入或更早输出。 + +## 用户意图 + +见仓库文件 `examples/g02/user-intent.md`（生成时已完整阅读，内容与该文件一致）。 + +## 已阅读的现有 DSL 约束（摘自仓库，不是新语法） + +- 从 `@devcodex-labs/flowkit/dsl` 导入 `defineWorkflow` 以及节点构造器。Agent 步骤使用 `node()`，并提供 `agent` 或 `skill`；条件使用 `condition()`；终态使用 `terminal()`。 +- `export default defineWorkflow({ id, version, name, nodes })`。 +- `id` 必须是 kebab-case，最多 64 字符；`version` 遵循 SemVer 2.0。 +- Agent 节点需要 `name`、`agent` 或 `skill`、对象类型的 `output` schema、`next`。`retries` 为 0–10 的整数；缺省为 2。`inputs` 声明从 State 读取的键，缺失键进入 `missing`，不得被合成 null。 +- 条件节点 `branches` 为 1–20 项，最后一项的 `when` 必须是 `'true'`。`when` 是针对 State 的表达式字符串，不能包含换行或反斜杠。 +- 终端节点 `status` 只能是 `succeeded` 或 `failed`。 +- 工作流文件名必须以 `.workflow.js` 或 `.workflow.ts` 结尾。 +- Interactive P0-A 只允许 Condition / Terminal / 基本 Agent（无 Agent verify/checkStep，无 Human/Action/Loop/Parallel）。本任务的 DSL 必须能通过 `compileInteractive(..., { featureStage: 'P0-A' })`，因此不放入 P0-A 未实现的节点。 +- 参考已有示例：`examples/interactive/c01-live-chatgpt.workflow.js`、`examples/editorial.workflow.js` 的构造方式。不要复制 C01 的业务文本，本任务是独立的功能变更评审闭环。 + +## 生成要求 + +只输出一份完整的、可保存为 `feature-review.workflow.js` 的 ESM 源码。不要输出 Mermaid、JSON 节点计划或伪代码来代替 DSL。节点后继必须写在 DSL 里，让编译后的 JS Program 做路由。 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/00-user-intent.md b/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/00-user-intent.md new file mode 100644 index 0000000..097d14d --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/00-user-intent.md @@ -0,0 +1,27 @@ +# G02 用户意图（真实任务原文） + +记录时间：2026-11-12，Asia/Shanghai。 +本文件是 Grok Build 在生成本阶段 `.workflow.js` 时实际阅读的用户任务原文，未事后改写需求。 + +## 任务名称 + +内部「功能变更评审」可执行闭环。 + +## 任务正文 + +请为 FlowKit 内部维护团队设计一个可重复执行的「功能变更评审」工作流，用来处理维护者提交的功能请求。要求如下： + +1. 入口先检查材料是否齐全（状态字段 `materialsReady`）。若材料不齐全，流程必须直接失败结束，不得调用任何 Agent。 +2. 若材料齐全，按固定顺序执行三个 Agent 步骤，模型不得自行选择下一个节点： + - 需求分析：阅读功能请求 `request`，输出 `requirementSummary`（字符串）。 + - 技术设计：基于上一节点的需求分析，输出 `designSummary`（字符串）。 + - 独立评审：同时阅读需求摘要和技术设计摘要，输出 `reviewApproved`（布尔）和 `reviewNotes`（字符串）。评审结果必须符合约定的 JSON 对象 schema；非法输出允许重试一次。 +3. 评审通过则成功结束；评审不通过则失败结束，并给出可核查终态。 +4. 流程推进、条件分支、重试与后继节点必须由 FlowKit 编译出的 JavaScript Program 决定，而不是由模型在回答里指定下一个节点。 +5. 只使用 FlowKit 现有 DSL 语法，不要发明第二套解析器、模型网关或独立 FlowKit Server。Interactive 验收仅使用已实现的 P0-A 子集：Condition、Terminal、基本 Agent。Standard 可用同一份 DSL 展示已实现的 Agent/Condition/Terminal 语义；Loop/Action/Human/Parallel 不是本任务的必要步骤，对照现有 editorial golden 说明差异即可。 + +## 本次实际提交的功能请求（作为工作流输入 `request`） + +为 FlowKit 维护者增加一条只读的工作流运行摘要查看能力：在不修改现有 Standard Root 六个 API、不引入独立 FlowKit Server 的前提下，允许受信 Host 读取一次运行的终态状态（succeeded/failed/errored）、terminal 节点和公开投影，便于排障。不在本阶段实现 adminTerminate，也不改变 Interactive P0-A 的节点范围。 + +说明：上述功能请求只是本工作流要评审的业务对象，G02 阶段本身并不实现该能力。 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/attempt-01-notes.md b/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/attempt-01-notes.md new file mode 100644 index 0000000..3e079c8 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/attempt-01-notes.md @@ -0,0 +1,7 @@ +# DSL 生成尝试记录 + +- 尝试次数：1。第一次输出已通过官方 `flowkit compile` 与 `flowkit check`（exit code 0，无诊断），以及 `compileInteractive({ featureStage: 'P0-A', skillCheck: true })`。因此没有 attempt-02。 +- 原始输入：`00-prompt.md`、`00-user-intent.md` +- 原始输出（生成后、编译前原样保留）：`attempt-01-raw.workflow.js` +- 采纳结果：该文件字节与 `examples/g02/feature-review.workflow.js` 一致，SHA-256 `8df70cd1af7dd9b0b5ea2c2143815c8aca862696f580fa46b21f726c021ec592` +- 编译命令与退出码：见 `../cli/` diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/attempt-01-raw.workflow.js b/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/attempt-01-raw.workflow.js new file mode 100644 index 0000000..ebccf4a --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/dsl-generation/attempt-01-raw.workflow.js @@ -0,0 +1,67 @@ +import { defineWorkflow, condition, node, terminal } from '@devcodex-labs/flowkit/dsl'; + +export default defineWorkflow({ + id: 'g02-feature-review', + version: '1.0.0', + name: '功能变更评审闭环', + nodes: [ + condition('gate', { + name: '检查材料是否齐全', + branches: [ + { when: 'materialsReady', next: 'analyze' }, + { when: 'true', next: 'rejected' }, + ], + }), + node('analyze', { + name: '需求分析', + agent: 'analyze', + inputs: ['request'], + output: { + type: 'object', + properties: { requirementSummary: { type: 'string' } }, + required: ['requirementSummary'], + additionalProperties: false, + }, + retries: 0, + next: 'design', + }), + node('design', { + name: '技术设计', + agent: 'design', + inputs: ['requirementSummary'], + output: { + type: 'object', + properties: { designSummary: { type: 'string' } }, + required: ['designSummary'], + additionalProperties: false, + }, + retries: 0, + next: 'review', + }), + node('review', { + name: '独立评审', + agent: 'review', + inputs: ['requirementSummary', 'designSummary'], + output: { + type: 'object', + properties: { + reviewApproved: { type: 'boolean' }, + reviewNotes: { type: 'string' }, + }, + required: ['reviewApproved', 'reviewNotes'], + additionalProperties: false, + }, + retries: 1, + next: 'decide', + }), + condition('decide', { + name: '根据评审结论分流', + branches: [ + { when: 'reviewApproved', next: 'accepted' }, + { when: 'true', next: 'rejected' }, + ], + }), + terminal('accepted', { name: '评审通过', status: 'succeeded' }), + terminal('rejected', { name: '评审未通过或材料不齐', status: 'failed' }), + ], +}); diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/00-env.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/00-env.txt new file mode 100644 index 0000000..88f6b28 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/00-env.txt @@ -0,0 +1,4 @@ +2026-11-12 17:56:18 CST +node v24.22.1 +HEAD 689e0bd97127bf32116bdfa37f88b8e013f3adb2 (pre-commit working tree for G02-R01/R02) +note=R01/R02 evidence refresh; pending counts measured from run-interactive.mjs route, not hardcoded diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.exit.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.exit.txt new file mode 100644 index 0000000..09461c5 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.exit.txt @@ -0,0 +1 @@ +interactive_exit=0 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.stderr.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.stderr.txt new file mode 100644 index 0000000..e69de29 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.stdout.json b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.stdout.json new file mode 100644 index 0000000..2a1a26a --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/interactive-run.stdout.json @@ -0,0 +1,105 @@ +[ + { + "layer": "interactive-p0a", + "scenario": "success-with-scripted-retry", + "status": "succeeded", + "terminal": { + "nodeId": "accepted", + "status": "succeeded" + }, + "reviewApproved": true, + "pendingAtEnd": false, + "publishedPendingCount": 4, + "route": [ + { + "source": "grok-model", + "nodeId": "analyze", + "attempt": 1, + "requestId": "req:cd869d313e933b356b99d9229e8ac9be161eca22e0d4b2c11a53091f58f868a7", + "statusAfter": "waiting", + "next": "design" + }, + { + "source": "grok-model", + "nodeId": "design", + "attempt": 1, + "requestId": "req:d76207d128012416efb5dc4b1ade578be4ac0e1802031fe8f92b215734407de3", + "statusAfter": "waiting", + "next": "review" + }, + { + "source": "scripted-invalid", + "nodeId": "review", + "attempt": 1, + "requestId": "req:d2d5fe090a4aeef9500ee9a624690a23f17d887d3c84a90e36538d29cb775133", + "statusAfter": "waiting", + "next": "review" + }, + { + "source": "grok-model", + "nodeId": "review", + "attempt": 2, + "requestId": "req:6cc88145ba67ad65081a9a78a741f8cbf26b35e9d7b4b8017ca095d59f0ea363", + "statusAfter": "succeeded", + "next": "accepted" + } + ], + "programHash": "sha256:8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded", + "evidenceNote": "Interactive Program + SQLite + local test receipts via Node. Not N-PASS. Not native secure Host provenance." + }, + { + "layer": "interactive-p0a", + "scenario": "gate-false", + "status": "failed", + "terminal": { + "nodeId": "rejected", + "status": "failed" + }, + "reviewApproved": null, + "pendingAtEnd": false, + "publishedPendingCount": 0, + "route": [], + "programHash": "sha256:8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded", + "evidenceNote": "Interactive Program + SQLite + local test receipts via Node. Not N-PASS. Not native secure Host provenance." + }, + { + "layer": "interactive-p0a", + "scenario": "review-approved-false", + "status": "failed", + "terminal": { + "nodeId": "rejected", + "status": "failed" + }, + "reviewApproved": false, + "pendingAtEnd": false, + "publishedPendingCount": 3, + "route": [ + { + "source": "grok-model", + "nodeId": "analyze", + "attempt": 1, + "requestId": "req:8c18ec5e3534510274b4ee71d0f02da7c12b00f627bc44a1109c9d3ad20997d0", + "statusAfter": "waiting", + "next": "design" + }, + { + "source": "grok-model", + "nodeId": "design", + "attempt": 1, + "requestId": "req:2d76c8634f8c31f8a6161f42241f1f2b69db3728bd674f0373c3333d2507cf9a", + "statusAfter": "waiting", + "next": "review" + }, + { + "source": "grok-model", + "nodeId": "review", + "attempt": 1, + "requestId": "req:aade8c66498bab1b076a8e3d7a9f37e5d7ff2d830fc95fd856592007707f18ba", + "statusAfter": "failed", + "next": "rejected" + } + ], + "programHash": "sha256:8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded", + "evidenceNote": "Interactive Program + SQLite + local test receipts via Node. Not N-PASS. Not native secure Host provenance." + } +] diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.exit.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.exit.txt new file mode 100644 index 0000000..3cb3283 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.exit.txt @@ -0,0 +1 @@ +g02_test_exit=0 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.stderr.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.stderr.txt new file mode 100644 index 0000000..e69de29 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.stdout.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.stdout.txt new file mode 100644 index 0000000..865b128 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-g02-tests.stdout.txt @@ -0,0 +1,16 @@ +✔ G02 Standard compile/check matches checked-in JS artifact and Program is real JavaScript (26.08371ms) +✔ G02 Interactive compile matches checked-in Program bytes and identity hash (4.175926ms) +✔ G02 Standard run covers success, gate-false, review-false and scripted retry; Program chooses branches (8.2015ms) +✔ G02 Interactive success path binds Grok results, retries scripted invalid review, Program routes to accepted (113.391383ms) +✔ G02 Interactive reviewApproved=false is routed by the Program to rejected (67.452899ms) +✔ G02 Interactive materialsReady=false never publishes a Pending; Program goes to rejected (19.250238ms) +✔ G02 fixtures record six grok-model uses, five distinct bodies, and one rebound reuse (0.668977ms) +✔ G02 workflow is a P0-A subset and differs from the editorial Standard golden in node kinds (0.416418ms) +ℹ tests 8 +ℹ suites 0 +ℹ pass 8 +ℹ fail 0 +ℹ cancelled 0 +ℹ skipped 0 +ℹ todo 0 +ℹ duration_ms 312.650933 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-npm-test.summary.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-npm-test.summary.txt new file mode 100644 index 0000000..2aa73b1 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/local-npm-test.summary.txt @@ -0,0 +1,9 @@ +2026-11-12 17:56:30 CST +command: npm test +exit: 0 +tests: 395 +pass: 395 +fail: 0 +node: v24.22.1 +OS: Linux (local Grok Build box; not GitHub CI) +note: includes G02-R01/R02 updates; one added accounting test (8 G02 tests / 395 total) diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/r02-pending-counts-measured.json b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/r02-pending-counts-measured.json new file mode 100644 index 0000000..031a972 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/r02-pending-counts-measured.json @@ -0,0 +1,32 @@ +[ + { + "scenario": "success-with-scripted-retry", + "status": "succeeded", + "terminal": { + "nodeId": "accepted", + "status": "succeeded" + }, + "pendingAtEnd": false, + "publishedPendingCount": 4 + }, + { + "scenario": "gate-false", + "status": "failed", + "terminal": { + "nodeId": "rejected", + "status": "failed" + }, + "pendingAtEnd": false, + "publishedPendingCount": 0 + }, + { + "scenario": "review-approved-false", + "status": "failed", + "terminal": { + "nodeId": "rejected", + "status": "failed" + }, + "pendingAtEnd": false, + "publishedPendingCount": 3 + } +] diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.exit.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.exit.txt new file mode 100644 index 0000000..4ad3d92 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.exit.txt @@ -0,0 +1 @@ +standard_exit=0 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.stderr.txt b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.stderr.txt new file mode 100644 index 0000000..e69de29 diff --git a/docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.stdout.json b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.stdout.json new file mode 100644 index 0000000..0502ea5 --- /dev/null +++ b/docs/evidence/g02-intent-js-closed-loop-20261112/execution/standard-run.stdout.json @@ -0,0 +1,90 @@ +[ + { + "layer": "standard", + "scenario": "success-with-retry", + "status": "succeeded", + "terminal": "accepted", + "reviewApproved": true, + "agentCalls": [ + { + "scenario": "success-with-retry", + "id": "analyze", + "attempt": 1, + "keys": [ + "requirementSummary" + ] + }, + { + "scenario": "success-with-retry", + "id": "design", + "attempt": 1, + "keys": [ + "designSummary" + ] + }, + { + "scenario": "success-with-retry", + "id": "review", + "attempt": 1, + "keys": [ + "comment", + "ok" + ] + }, + { + "scenario": "success-with-retry", + "id": "review", + "attempt": 2, + "keys": [ + "reviewApproved", + "reviewNotes" + ] + } + ], + "evidenceNote": "Standard run() via Node. Not native secure Host provenance. Not N-PASS." + }, + { + "layer": "standard", + "scenario": "gate-false", + "status": "failed", + "terminal": "rejected", + "reviewApproved": null, + "agentCalls": [], + "evidenceNote": "Standard run() via Node. Not native secure Host provenance. Not N-PASS." + }, + { + "layer": "standard", + "scenario": "review-false", + "status": "failed", + "terminal": "rejected", + "reviewApproved": false, + "agentCalls": [ + { + "scenario": "review-false", + "id": "analyze", + "attempt": 1, + "keys": [ + "requirementSummary" + ] + }, + { + "scenario": "review-false", + "id": "design", + "attempt": 1, + "keys": [ + "designSummary" + ] + }, + { + "scenario": "review-false", + "id": "review", + "attempt": 1, + "keys": [ + "reviewApproved", + "reviewNotes" + ] + } + ], + "evidenceNote": "Standard run() via Node. Not native secure Host provenance. Not N-PASS." + } +] diff --git a/examples/README.md b/examples/README.md index bd4a09f..972a9fc 100644 --- a/examples/README.md +++ b/examples/README.md @@ -101,3 +101,8 @@ dialect, actual model execution, business actions and an approval service. See [host integration](../docs/host-integration.md) and [troubleshooting](../docs/troubleshooting.md). 示例目录不随 npm 包分发。v1.0.0 包包含 `dist`、`README.md`、`LICENSE` 和包元数据；用户可从仓库或文档站获取完整教程文件。 + +## G02 产品闭环验证（Interactive P0-A 子集） + +[g02/](g02/) 保存一次真实用户意图到可执行 JavaScript 的验证材料：Grok Build 按现有 DSL 生成 `feature-review.workflow.js`，再用官方 `compile`/`check` 与 `compileInteractive` 生成并运行 Program。Standard 层只演示该 DSL 已有的 Agent/Condition/Terminal；完整八类节点仍以本目录的 editorial golden 为准。Interactive 层只验收 P0-A 子集与 Pending/恢复，不能把 Node 执行或本地测试签名写成原生安全 Host 来源认证。操作说明见该目录 README。 + diff --git a/examples/g02/README.md b/examples/g02/README.md new file mode 100644 index 0000000..bd4fb7e --- /dev/null +++ b/examples/g02/README.md @@ -0,0 +1,37 @@ +# G02 — 真实用户意图 → JS 工作流执行产品闭环 + +本目录是 G02 产品闭环验证材料，不是新的 DSL、解析器或模型网关。 + +- 用户意图原文：[`user-intent.md`](user-intent.md) +- 由 Grok Build 根据现有 DSL 生成的工作流：[`feature-review.workflow.js`](feature-review.workflow.js) +- Skill 文本：[`skills/`](skills/) +- Standard 编译产物（由官方 `flowkit compile` 生成）：[`feature-review.compiled.mjs`](feature-review.compiled.mjs)、可视化 [`feature-review.mmd`](feature-review.mmd)（Mermaid 只是图，不能代替 JS 执行） +- Interactive 编译产物（由 `compileInteractive` 生成）：[`feature-review.interactive.mjs`](feature-review.interactive.mjs) +- 操作脚本：[`run-standard.mjs`](run-standard.mjs)、[`run-interactive.mjs`](run-interactive.mjs) +- 可重复测试：[`../../tests/interactive/g02-product-closed-loop.test.mjs`](../../tests/interactive/g02-product-closed-loop.test.mjs) +- 模型回执夹具：[`../../tests/fixtures/g02/`](../../tests/fixtures/g02/) +- 证据目录：[`../../docs/evidence/g02-intent-js-closed-loop-20261112/`](../../docs/evidence/g02-intent-js-closed-loop-20261112/) + +## 分层说明 + +| 层 | 本目录实际验证的内容 | 不得声称 | +|---|---|---| +| Standard | 同一份 DSL 经官方 `compile`/`check`/`run()` 执行 Agent/Condition/Terminal，含成功、入口条件为 false、评审拒绝、非法输出重试 | 不得把这些结果写成 Interactive 已支持 Loop/Action/Human/Parallel | +| Interactive P0-A | `compileInteractive`、真实 JS Program、Pending、受信 Host 回执、SQLite 关闭后重开恢复 | 不得把 Node/Shell 执行或本地测试签名标为 N-PASS 或原生安全 Host 来源认证 | +| 现有 Standard golden | 与 [`../editorial.workflow.js`](../editorial.workflow.js) 对照，说明八类节点差异 | 不得把 editorial 的 Loop/Action/Human/Parallel 记为本次 Interactive 已实现 | + +## 从仓库复现 + +需要 Node >=22.13.0，在仓库根目录： + +```sh +npm ci +npm run build +node dist/cli.mjs compile examples/g02/feature-review.workflow.js --skills-root examples/g02/skills +node dist/cli.mjs check examples/g02/feature-review.workflow.js --skills-root examples/g02/skills +node examples/g02/run-standard.mjs +node examples/g02/run-interactive.mjs +node --test tests/interactive/g02-product-closed-loop.test.mjs +``` + +`run-*.mjs` 从源码树导入现有 SDK，不创建第二套编译器。Interactive 脚本会在临时 SQLite 中重放已绑定的夹具，并输出 `pendingAtEnd`（终态帧是否仍有 Pending）以及从本次 `route` 实测的 `publishedPendingCount`。夹具中 `source=scripted-invalid` 是故意非法的脚本负例；`source=grok-model` 共消费 6 次，对应 5 份独立撰写正文、1 次复用（重新绑定到新的 Pending 身份），不能计为 6 次独立模型推理。 diff --git a/examples/g02/feature-review.compiled.mjs b/examples/g02/feature-review.compiled.mjs new file mode 100644 index 0000000..4834582 --- /dev/null +++ b/examples/g02/feature-review.compiled.mjs @@ -0,0 +1,117 @@ +// @generated by flowkit compiler — DO NOT EDIT +// workflow: g02-feature-review@1.0.0（功能变更评审闭环） +// source: feature-review.workflow.js +// sourceHash: sha256:8df70cd1af7dd9b0b5ea2c2143815c8aca862696f580fa46b21f726c021ec592 · bodyHash: sha256:6115bc998e045fd9fa269d815830de3f5e4ab08af7862fd6b24cac3e84efdd9c +// compiler: flowkit@1.0.0 · options: trace=off +// evalWhen reads own ctx.state paths; strict booleans and short circuit; ! has lower precedence than comparison. +// runAgent receives one step; schema -> checkStep -> verify; own-key commit only after success; retries share reasons and control signals are rethrown. +// State is supplied by the host. +// Source and body hashes are mandatory. +// This module contains native control flow. +export async function workflow(ctx) { + const runtime = ctx.runtime; + const runAgent = (...args) => runtime.runAgent(...args); + const validateResult = (...args) => runtime.validateResult(...args); + const loadSkill = (...args) => runtime.loadSkill === undefined ? ctx.skillLoader.loadSkill(...args) : runtime.loadSkill(...args); + const loadSkillByAgent = (...args) => runtime.loadSkillByAgent === undefined ? ctx.skillLoader.loadSkillByAgent(...args) : runtime.loadSkillByAgent(...args); + const evalWhen = (...args) => runtime.evalWhen(...args); + const isControlSignal = (...args) => runtime.isControlSignal(...args); + const writeOwn = (target, key, value) => Object.defineProperty(target, key, + { value, enumerable: true, writable: true, configurable: true }); + const checkStep = runtime.checkStep && ((...args) => runtime.checkStep(...args)); + const verifyStep = undefined; + let stepNo = 0; + // Diagnostic text must not change retry behavior or replace the original rejection cause. + const rejectionText = (error) => { + try { return String(error?.message ?? error); } + catch { return ""; } + }; + const mergeOwn = (target, source) => { for (const key of Object.keys(source)) writeOwn(target, key, source[key]); return target; }; + const runStep = async (s) => { + // Allocate before any await: parallel declarations keep deterministic overview numbers. + const step = ++stepNo; + const instruction = await s.skill(); + const input = {}; + const missing = []; + for (const key of s.inputs) { + writeOwn(input, key, Object.hasOwn(ctx.state, key) ? ctx.state[key] : undefined); + if (!Object.hasOwn(ctx.state, key)) missing.push(key); + } + let reasons = []; + let threw = false; + let cause; + for (let attempt = 1; attempt <= s.retries + 1; attempt++) { + let result; + try { + result = await runAgent({ id: s.id, name: s.name, instruction, input, missing, resultSchema: s.output, + attempt, reasons, ...(s.overview ? { overview: { step } } : {}), signal: ctx.signal, runId: ctx.runId }); + threw = false; + } catch (error) { + if (isControlSignal(error)) throw error; + threw = true; + cause = error; + reasons = [`runAgent threw: ${rejectionText(error)}`]; + continue; + } + reasons = validateResult(s.output, result); + if (reasons.length === 0 && checkStep) { + const why = await checkStep({ id: s.id, name: s.name, attempt }, result, ctx); + if (typeof why === "string") reasons = [why]; + else if (why != null) throw Object.assign(new Error(`FK_R002 ${s.name} checkStep returned an invalid value`), { code: "FK_R002" }); + } + if (reasons.length === 0 && s.verify != null) { + const candidate = mergeOwn(mergeOwn({}, ctx.state), result); + if (!(await verifyStep(s.id, s.verify, s.name, { ...ctx, state: candidate }))) reasons = [`verify ${s.verify} failed`]; + } + if (reasons.length === 0) { mergeOwn(ctx.state, result); return; } + } + throw Object.assign(new Error(`FK_R004 ${s.name}: ${reasons.join("; ")}`, threw ? { cause } : undefined), { code: "FK_R004", reasons }); + }; + let n0_gateReach = false; + let n1_analyzeReach = false; + let n2_designReach = false; + let n3_reviewReach = false; + let n4_decideReach = false; + let n5_acceptedReach = false; + let n6_rejectedReach = false; + n0_gateReach = true; + if (n0_gateReach) { + // [gate] 检查材料是否齐全 · condition + if (evalWhen(ctx, "materialsReady")) { + n1_analyzeReach = true; + } else { + n6_rejectedReach = true; + } + } + if (n1_analyzeReach) { + // [analyze] 需求分析 · agent: analyze -> examples/g02/skills/analyze.md; inputs: request; results: requirementSummary; attempts: 1 + await runStep({ id: "analyze", name: "需求分析", skill: () => loadSkillByAgent("analyze"), inputs: ["request"], output: {"type": "object", "properties": {"requirementSummary": {"type": "string"}}, "required": ["requirementSummary"], "additionalProperties": false}, retries: 0, verify: null, overview: false }); + n2_designReach = true; + } + if (n2_designReach) { + // [design] 技术设计 · agent: design -> examples/g02/skills/design.md; inputs: requirementSummary; results: designSummary; attempts: 1 + await runStep({ id: "design", name: "技术设计", skill: () => loadSkillByAgent("design"), inputs: ["requirementSummary"], output: {"type": "object", "properties": {"designSummary": {"type": "string"}}, "required": ["designSummary"], "additionalProperties": false}, retries: 0, verify: null, overview: false }); + n3_reviewReach = true; + } + if (n3_reviewReach) { + // [review] 独立评审 · agent: review -> examples/g02/skills/review.md; inputs: requirementSummary, designSummary; results: reviewApproved, reviewNotes; attempts: 2 + await runStep({ id: "review", name: "独立评审", skill: () => loadSkillByAgent("review"), inputs: ["requirementSummary", "designSummary"], output: {"type": "object", "properties": {"reviewApproved": {"type": "boolean"}, "reviewNotes": {"type": "string"}}, "required": ["reviewApproved", "reviewNotes"], "additionalProperties": false}, retries: 1, verify: null, overview: false }); + n4_decideReach = true; + } + if (n4_decideReach) { + // [decide] 根据评审结论分流 · condition + if (evalWhen(ctx, "reviewApproved")) { + n5_acceptedReach = true; + } else { + n6_rejectedReach = true; + } + } + if (n5_acceptedReach) { + // [accepted] 评审通过 · terminal + return { status: "succeeded", terminal: "accepted" }; + } + if (n6_rejectedReach) { + // [rejected] 评审未通过或材料不齐 · terminal + return { status: "failed", terminal: "rejected" }; + } +} diff --git a/examples/g02/feature-review.interactive.mjs b/examples/g02/feature-review.interactive.mjs new file mode 100644 index 0000000..f4226a4 --- /dev/null +++ b/examples/g02/feature-review.interactive.mjs @@ -0,0 +1,143 @@ +// @generated by FlowKit Interactive — P0-A / Program ABI V1 +// This trusted compiler output MUST be loaded from verified immutable bytes. +function requireBoolean(value) { + if (typeof value !== 'boolean') throw Object.assign(new Error('When operand and final result must be boolean'), { code: 'FK_R002' }); + return value; +} +function evaluateWhenAst(ast, state) { + const stack = [{ node: ast, stage: 0 }]; + let result; + const finish = value => { result = value; stack.pop(); }; + while (stack.length) { + const frame = stack.at(-1); + const node = frame.node; + if (node.kind === 'literal') finish(node.value); + else if (node.kind === 'path') { + let value = state; + for (const key of node.parts) value = value !== null && typeof value === 'object' && Object.hasOwn(value, key) ? value[key] : undefined; + finish(value); + } else if (node.kind === 'not') { + if (frame.stage === 0) { frame.stage = 1; stack.push({ node: node.operand, stage: 0 }); } + else finish(!requireBoolean(result)); + } else if (['and', 'or', 'compare'].includes(node.kind)) { + if (frame.stage === 0) { frame.stage = 1; stack.push({ node: node.left, stage: 0 }); } + else if (frame.stage === 1) { + frame.left = node.kind === 'compare' ? result : requireBoolean(result); + if (node.kind === 'and' && !frame.left || node.kind === 'or' && frame.left) finish(frame.left); + else { frame.stage = 2; stack.push({ node: node.right, stage: 0 }); } + } else if (node.kind !== 'compare') finish(requireBoolean(result)); + else { + const left = frame.left; + const right = result; + if (node.operator === '===') finish(left === right); + else if (node.operator === '!==') finish(left !== right); + else { + if (!(typeof left === 'number' && typeof right === 'number' || typeof left === 'string' && typeof right === 'string')) { + throw Object.assign(new Error('When ordering requires two numbers or two strings'), { code: 'FK_R002' }); + } + if (node.operator === '>') finish(left > right); + else if (node.operator === '>=') finish(left >= right); + else if (node.operator === '<') finish(left < right); + else if (node.operator === '<=') finish(left <= right); + else throw new TypeError('Unknown when comparison'); + } + } + } else throw new TypeError('Unknown when AST node'); + } + return requireBoolean(result); +} +function freezeData(value) { + if (value && typeof value === 'object') { + for (const key of Object.keys(value)) freezeData(value[key]); + Object.freeze(value); + } + return value; +} +export const interactiveProgram = freezeData({"artifactFormat":"FlowkitInteractiveProgramV1","identity":{"workflowId":"g02-feature-review","workflowVersion":"1.0.0","sourceHash":"sha256:8df70cd1af7dd9b0b5ea2c2143815c8aca862696f580fa46b21f726c021ec592","compilerVersion":"1.0.0","programAbiVersion":1,"semanticsVersion":"interactive-state-machine-v1","checkpointSchemaVersion":1,"canonicalizationVersion":"jcs-rfc8785-v1","implementedStage":"P0-A"}}); +export const programTable = freezeData({"schemaVersion":1,"entryBlockId":"gate::enter","blocks":[{"id":"gate::enter","nodeId":"gate","phase":"enter","legalTargets":["gate::route"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"gate::route","nodeId":"gate","phase":"route","legalTargets":["analyze::enter","rejected::enter"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"analyze::enter","nodeId":"analyze","phase":"enter","legalTargets":["analyze::prepare"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"analyze::prepare","nodeId":"analyze","phase":"prepare","legalTargets":["analyze::dispatch-ready"],"eventKinds":["internal-step","pending-prepared","deterministic-failure"],"proposalKinds":["prepare-pending","raise-error"]},{"id":"analyze::dispatch-ready","nodeId":"analyze","phase":"dispatch-ready","legalTargets":["analyze::awaiting"],"eventKinds":["pending-prepared","deterministic-failure"],"proposalKinds":["publish-pending","raise-error"]},{"id":"analyze::awaiting","nodeId":"analyze","phase":"awaiting","legalTargets":["analyze::validate"],"eventKinds":[],"proposalKinds":[],"acceptTarget":"analyze::validate"},{"id":"analyze::validate","nodeId":"analyze","phase":"validate","legalTargets":["analyze::prepare","analyze::commit"],"eventKinds":["agent-result","agent-call-failed","validation-passed","validation-rejected","deterministic-failure"],"proposalKinds":["retry-agent","commit-candidate","raise-error"]},{"id":"analyze::commit","nodeId":"analyze","phase":"commit","legalTargets":["analyze::route"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["commit-state","raise-error"]},{"id":"analyze::route","nodeId":"analyze","phase":"route","legalTargets":["design::enter"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"design::enter","nodeId":"design","phase":"enter","legalTargets":["design::prepare"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"design::prepare","nodeId":"design","phase":"prepare","legalTargets":["design::dispatch-ready"],"eventKinds":["internal-step","pending-prepared","deterministic-failure"],"proposalKinds":["prepare-pending","raise-error"]},{"id":"design::dispatch-ready","nodeId":"design","phase":"dispatch-ready","legalTargets":["design::awaiting"],"eventKinds":["pending-prepared","deterministic-failure"],"proposalKinds":["publish-pending","raise-error"]},{"id":"design::awaiting","nodeId":"design","phase":"awaiting","legalTargets":["design::validate"],"eventKinds":[],"proposalKinds":[],"acceptTarget":"design::validate"},{"id":"design::validate","nodeId":"design","phase":"validate","legalTargets":["design::prepare","design::commit"],"eventKinds":["agent-result","agent-call-failed","validation-passed","validation-rejected","deterministic-failure"],"proposalKinds":["retry-agent","commit-candidate","raise-error"]},{"id":"design::commit","nodeId":"design","phase":"commit","legalTargets":["design::route"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["commit-state","raise-error"]},{"id":"design::route","nodeId":"design","phase":"route","legalTargets":["review::enter"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"review::enter","nodeId":"review","phase":"enter","legalTargets":["review::prepare"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"review::prepare","nodeId":"review","phase":"prepare","legalTargets":["review::dispatch-ready"],"eventKinds":["internal-step","pending-prepared","deterministic-failure"],"proposalKinds":["prepare-pending","raise-error"]},{"id":"review::dispatch-ready","nodeId":"review","phase":"dispatch-ready","legalTargets":["review::awaiting"],"eventKinds":["pending-prepared","deterministic-failure"],"proposalKinds":["publish-pending","raise-error"]},{"id":"review::awaiting","nodeId":"review","phase":"awaiting","legalTargets":["review::validate"],"eventKinds":[],"proposalKinds":[],"acceptTarget":"review::validate"},{"id":"review::validate","nodeId":"review","phase":"validate","legalTargets":["review::prepare","review::commit"],"eventKinds":["agent-result","agent-call-failed","validation-passed","validation-rejected","deterministic-failure"],"proposalKinds":["retry-agent","commit-candidate","raise-error"]},{"id":"review::commit","nodeId":"review","phase":"commit","legalTargets":["review::route"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["commit-state","raise-error"]},{"id":"review::route","nodeId":"review","phase":"route","legalTargets":["decide::enter"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"decide::enter","nodeId":"decide","phase":"enter","legalTargets":["decide::route"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"decide::route","nodeId":"decide","phase":"route","legalTargets":["accepted::enter","rejected::enter"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"accepted::enter","nodeId":"accepted","phase":"enter","legalTargets":["accepted::route"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"accepted::route","nodeId":"accepted","phase":"route","legalTargets":[],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["terminal","raise-error"]},{"id":"rejected::enter","nodeId":"rejected","phase":"enter","legalTargets":["rejected::route"],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["move","raise-error"]},{"id":"rejected::route","nodeId":"rejected","phase":"route","legalTargets":[],"eventKinds":["internal-step","deterministic-failure"],"proposalKinds":["terminal","raise-error"]}],"nodes":{"gate":{"kind":"condition","metadata":{"name":"检查材料是否齐全","branches":[{"when":"materialsReady","ast":{"kind":"path","parts":["materialsReady"],"column":1},"next":"analyze"},{"when":"true","ast":{"kind":"literal","value":true,"column":1},"next":"rejected"}]}},"analyze":{"kind":"agent","metadata":{"name":"需求分析","next":"design","agent":"analyze","inputs":["request"],"output":{"type":"object","properties":{"requirementSummary":{"type":"string"}},"required":["requirementSummary"],"additionalProperties":false},"retries":0,"overview":false}},"design":{"kind":"agent","metadata":{"name":"技术设计","next":"review","agent":"design","inputs":["requirementSummary"],"output":{"type":"object","properties":{"designSummary":{"type":"string"}},"required":["designSummary"],"additionalProperties":false},"retries":0,"overview":false}},"review":{"kind":"agent","metadata":{"name":"独立评审","next":"decide","agent":"review","inputs":["requirementSummary","designSummary"],"output":{"type":"object","properties":{"reviewApproved":{"type":"boolean"},"reviewNotes":{"type":"string"}},"required":["reviewApproved","reviewNotes"],"additionalProperties":false},"retries":1,"overview":false}},"decide":{"kind":"condition","metadata":{"name":"根据评审结论分流","branches":[{"when":"reviewApproved","ast":{"kind":"path","parts":["reviewApproved"],"column":1},"next":"accepted"},{"when":"true","ast":{"kind":"literal","value":true,"column":1},"next":"rejected"}]}},"accepted":{"kind":"terminal","metadata":{"name":"评审通过","status":"succeeded"}},"rejected":{"kind":"terminal","metadata":{"name":"评审未通过或材料不齐","status":"failed"}}},"requiredBindings":[{"id":"runtime.validateResult","version":"1","mode":"pure-local","kind":"agent-validation"},{"id":"runtime.skillResolver","version":"1","mode":"trusted-read","kind":"skill"}]}); +export function advance(view, event) { + if (!view || !event || event.schemaVersion !== 1 || view.blockId !== event.forBlockId) { + throw new TypeError('Event does not target the current checkpoint block'); + } + const blockId = view.blockId; + const phase = view.phase; + const block = programTable.blocks.find(item => item.id === blockId); + if (!block || block.phase !== phase || !block.eventKinds.includes(event.kind)) { + throw new TypeError('Uncertified event or invalid stage'); + } + const nodeId = block.nodeId; + const record = programTable.nodes[nodeId]; + const none = { kind: 'none' }; + const proposal = (kind, toBlockId, stackEffect = none, mergeIntent = none) => + ({ kind, fromBlockId: blockId, nodeId, toBlockId, stackEffect, mergeIntent }); + const fail = code => ({ + kind: 'raise-error', fromBlockId: blockId, nodeId, + causeCode: code || 'FK_R002', stackEffect: none, mergeIntent: none, + }); + if (event.kind === 'deterministic-failure') return fail(event.error?.code); + const to = (node, name) => node + '::' + name; + if (phase === 'enter') { + if (record.kind === 'agent') { + if (!Number.isSafeInteger(event.activation) || event.activation < 1) { + throw new TypeError('Missing certified Agent activation'); + } + const frame = { + kind: 'agent', nodeId, activation: event.activation, + attempt: 1, phase: 'prepare', reasons: [], overviewStep: event.activation, + }; + return proposal('move', to(nodeId, 'prepare'), {kind: 'push', frame}); + } + return proposal('move', to(nodeId, 'route')); + } + if (record.kind === 'condition' && phase === 'route') { + try { + for (const branch of record.metadata.branches) { + if (evaluateWhenAst(branch.ast, view.state)) { + return proposal('move', to(branch.next, 'enter')); + } + } + } catch (error) { + if (error?.code === 'FK_R002') return fail('FK_R002'); + throw error; // trusted Coordinator classifies unexpected program failures + } + return fail('FK_R002'); + } + if (record.kind === 'terminal' && phase === 'route') { + return { + kind: 'terminal', fromBlockId: blockId, nodeId, + terminal: {status: record.metadata.status, nodeId}, + stackEffect: none, mergeIntent: none, + }; + } + if (record.kind !== 'agent') throw new TypeError('Unsupported P0-A block'); + const frame = view.controlStack[view.controlStack.length - 1]; + if (!frame || frame.kind !== 'agent' || frame.nodeId !== nodeId) { + throw new TypeError('Missing certified Agent control frame'); + } + if (phase === 'prepare') return proposal('prepare-pending', to(nodeId, 'dispatch-ready')); + if (phase === 'dispatch-ready') return proposal('publish-pending', to(nodeId, 'awaiting')); + if (phase === 'validate') { + if (event.kind === 'validation-passed') { + if (view.candidate?.kind !== 'agent') throw new TypeError('Verified Agent candidate not committed'); + return proposal('commit-candidate', to(nodeId, 'commit')); + } + if (event.kind === 'validation-rejected' || event.kind === 'agent-call-failed') { + const attempts = record.metadata.retries + 1; + if (frame.attempt >= attempts) return fail('FK_R004'); + return proposal('retry-agent', to(nodeId, 'prepare'), + {kind: 'update', expectedKind: 'agent', expectedNodeId: nodeId, + field: 'attempt', value: frame.attempt + 1}); + } + throw new TypeError('Agent validation event is not ready for Program advancement'); + } + if (phase === 'commit') { + if (view.candidate?.kind !== 'agent') throw new TypeError('Missing committed Agent candidate'); + return proposal('commit-state', to(nodeId, 'route'), none, + {kind:'commit-candidate', requestId:view.candidate.requestId}); + } + if (phase === 'route') { + return proposal('move', to(record.metadata.next, 'enter'), + {kind:'pop', expectedKind:'agent', expectedNodeId:nodeId}); + } + throw new TypeError('Illegal Program advancement at awaiting/unknown phase'); +} diff --git a/examples/g02/feature-review.mmd b/examples/g02/feature-review.mmd new file mode 100644 index 0000000..7f30cde --- /dev/null +++ b/examples/g02/feature-review.mmd @@ -0,0 +1,15 @@ +flowchart TD + Gate{"检查材料是否齐全"} + Analyze["需求分析"] + Design["技术设计"] + Review["独立评审"] + Decide{"根据评审结论分流"} + Accepted("评审通过") + Rejected("评审未通过或材料不齐") + Gate -- "materialsReady" --> Analyze + Gate -- "true" --> Rejected + Analyze --> Design + Design --> Review + Review --> Decide + Decide -- "reviewApproved" --> Accepted + Decide -- "true" --> Rejected diff --git a/examples/g02/feature-review.workflow.js b/examples/g02/feature-review.workflow.js new file mode 100644 index 0000000..ebccf4a --- /dev/null +++ b/examples/g02/feature-review.workflow.js @@ -0,0 +1,67 @@ +import { defineWorkflow, condition, node, terminal } from '@devcodex-labs/flowkit/dsl'; + +export default defineWorkflow({ + id: 'g02-feature-review', + version: '1.0.0', + name: '功能变更评审闭环', + nodes: [ + condition('gate', { + name: '检查材料是否齐全', + branches: [ + { when: 'materialsReady', next: 'analyze' }, + { when: 'true', next: 'rejected' }, + ], + }), + node('analyze', { + name: '需求分析', + agent: 'analyze', + inputs: ['request'], + output: { + type: 'object', + properties: { requirementSummary: { type: 'string' } }, + required: ['requirementSummary'], + additionalProperties: false, + }, + retries: 0, + next: 'design', + }), + node('design', { + name: '技术设计', + agent: 'design', + inputs: ['requirementSummary'], + output: { + type: 'object', + properties: { designSummary: { type: 'string' } }, + required: ['designSummary'], + additionalProperties: false, + }, + retries: 0, + next: 'review', + }), + node('review', { + name: '独立评审', + agent: 'review', + inputs: ['requirementSummary', 'designSummary'], + output: { + type: 'object', + properties: { + reviewApproved: { type: 'boolean' }, + reviewNotes: { type: 'string' }, + }, + required: ['reviewApproved', 'reviewNotes'], + additionalProperties: false, + }, + retries: 1, + next: 'decide', + }), + condition('decide', { + name: '根据评审结论分流', + branches: [ + { when: 'reviewApproved', next: 'accepted' }, + { when: 'true', next: 'rejected' }, + ], + }), + terminal('accepted', { name: '评审通过', status: 'succeeded' }), + terminal('rejected', { name: '评审未通过或材料不齐', status: 'failed' }), + ], +}); diff --git a/examples/g02/run-interactive.mjs b/examples/g02/run-interactive.mjs new file mode 100644 index 0000000..cf0604e --- /dev/null +++ b/examples/g02/run-interactive.mjs @@ -0,0 +1,152 @@ +/** + * Maintainer helper: replay G02 Interactive fixtures through the real + * compileInteractive Program, SQLite Store and Trusted Host. + * Local test signatures and Node execution are not native Host provenance. + */ +import {readFile, mkdtemp, rm} from 'node:fs/promises'; +import {tmpdir} from 'node:os'; +import path from 'node:path'; +import {fileURLToPath} from 'node:url'; +import {generateKeyPairSync, sign} from 'node:crypto'; +import {compileInteractive} from '../../src/interactive/compile.mjs'; +import {createSqliteCheckpointStore} from '../../src/interactive/sqlite-store.mjs'; +import {createTrustedInteractiveHost} from '../../src/interactive/trusted-host.mjs'; +import {createAgentReceiptStatement, encodeAgentReceiptStatement} + from '../../src/interactive/host-attestation.mjs'; +import {canonicalBytes, sha256Bytes} from '../../src/interactive/canonical.mjs'; + +const repoRoot = fileURLToPath(new URL('../../', import.meta.url)); +const workflow = fileURLToPath(new URL('./feature-review.workflow.js', import.meta.url)); +const success = JSON.parse(await readFile(new URL( + '../../tests/fixtures/g02/turns-success-retry.json', import.meta.url), 'utf8')); +const rejected = JSON.parse(await readFile(new URL( + '../../tests/fixtures/g02/turns-review-false.json', import.meta.url), 'utf8')); +const skillsRoot = path.join(repoRoot, 'examples/g02/skills'); +const NOW = 1794500000000; + +function validate(schema, output) { + if (!output || typeof output !== 'object' || Array.isArray(output)) return ['Agent result must be object']; + const props = schema.properties ?? {}; + const required = schema.required ?? Object.keys(props); + for (const key of required) if (!Object.hasOwn(output, key)) return ['missing ' + key]; + if (schema.additionalProperties === false) { + for (const key of Object.keys(output)) if (!Object.hasOwn(props, key)) return ['unknown ' + key]; + } + for (const [key, spec] of Object.entries(props)) { + if (!Object.hasOwn(output, key)) continue; + if (typeof output[key] !== spec.type) return [key + ' must be ' + spec.type]; + if (spec.type === 'string' && output[key].length < 80) return [key + ' must be at least 80 characters']; + } + return []; +} + +const skills = { + analyze: await readFile(new URL('./skills/analyze.md', import.meta.url), 'utf8'), + design: await readFile(new URL('./skills/design.md', import.meta.url), 'utf8'), + review: await readFile(new URL('./skills/review.md', import.meta.url), 'utf8'), +}; + +const compiled = await compileInteractive(workflow, { + output: 'memory', featureStage: 'P0-A', skillsRoot, skillCheck: true, projectRoot: repoRoot, +}); +if (compiled.diagnostics.some((d) => d.severity === 'error')) { + throw new Error(JSON.stringify(compiled.diagnostics)); +} + +async function replay(fixture, materialsReady = true) { + const dir = await mkdtemp(path.join(tmpdir(), 'g02-run-')); + const file = path.join(dir, 'run.sqlite'); + const keys = generateKeyPairSync('ed25519'); + let store; + const route = []; + try { + const hostFor = (s) => createTrustedInteractiveHost({ + store: s, + runtime: { + skillResolver: async ({agent}) => skills[agent], + validateResult: validate, + }, + policy: { + deploymentId: 'g02-feature-review-20261112', + bindingVersions: {skill: 'g02-v1', validator: 'g02-v1'}, + }, + trustedPublicKeys: {'g02-host': keys.publicKey}, + trustedArtifacts: [compiled.programIdentity.artifactHash], + allowedInputKeys: ['request', 'requirementSummary', 'designSummary'], + disclosurePolicyId: 'g02-public-inputs', + authorizeDisclosure: async ({task}) => + ['analyze', 'design', 'review'].includes(task?.nodeId) && task?.runId === fixture.runId, + now: () => NOW, + }); + store = createSqliteCheckpointStore(file, {namespace: 'g02-feature-review'}); + let host = hostFor(store); + let frame = await host.start({ + programBytes: Buffer.from(compiled.code, 'utf8'), + programIdentity: compiled.programIdentity, + runId: fixture.runId, startKey: fixture.startKey, + state: {materialsReady, request: fixture.request, hostSecret: 'CANARY'}, + }); + for (let index = 0; index < (fixture.turns ?? []).length; index++) { + const pending = await host.getPublicAgentTask({runId: fixture.runId}); + const recorded = fixture.turns[index]; + const taskDigest = sha256Bytes(canonicalBytes(pending)); + if (recorded.requestId !== pending.requestId || recorded.taskDigest !== taskDigest) { + throw new Error('fixture does not bind to live Pending at ' + index); + } + const receipt = { + runId: fixture.runId, requestId: pending.requestId, nodeId: pending.nodeId, + attempt: pending.attempt, kind: 'agent-result', result: recorded.result, + }; + const statement = createAgentReceiptStatement({ + run: store.load(fixture.runId), receipt, keyId: 'g02-host', + proofId: 'g02-replay-' + fixture.scenario + '-' + index, + issuedAt: NOW - 1000, expiresAt: NOW + 30000, + }); + const attestation = { + statement, + signature: sign(null, encodeAgentReceiptStatement(statement), keys.privateKey).toString('base64url'), + }; + store.close(); + store = createSqliteCheckpointStore(file, {namespace: 'g02-feature-review'}); + host = hostFor(store); + frame = await host.accept({receipt, attestation}); + route.push({ + source: recorded.source, nodeId: pending.nodeId, attempt: pending.attempt, + requestId: pending.requestId, statusAfter: frame.status, + next: frame.pending?.nodeId ?? frame.terminal?.nodeId ?? null, + }); + } + return { + layer: 'interactive-p0a', + scenario: fixture.scenario ?? (materialsReady ? 'unknown' : 'gate-false'), + status: frame.status, + terminal: frame.terminal ?? null, + reviewApproved: frame.state?.reviewApproved ?? null, + // pendingAtEnd is the terminal frame only. publishedPendingCount is + // measured from Pendings actually observed on this run's route. + pendingAtEnd: (frame.pending ?? null) != null, + publishedPendingCount: route.length, + route, + programHash: compiled.programIdentity.artifactHash, + evidenceNote: 'Interactive Program + SQLite + local test receipts via Node. Not N-PASS. Not native secure Host provenance.', + }; + } finally { + store?.close(); + await rm(dir, {recursive: true, force: true}); + } +} + +const gate = { + scenario: 'gate-false', + runId: 'g02-gate-false-20261112', + startKey: 'g02-gate-false-start', + request: success.request, + turns: [], +}; + +const output = [ + await replay(success), + await replay(gate, false), + await replay(rejected), +]; +console.log(JSON.stringify(output, null, 2)); diff --git a/examples/g02/run-standard.mjs b/examples/g02/run-standard.mjs new file mode 100644 index 0000000..79284ea --- /dev/null +++ b/examples/g02/run-standard.mjs @@ -0,0 +1,99 @@ +/** + * Maintainer helper: execute the G02 workflow through Standard run(). + * This is not a native secure Host and does not call a live model API. + * Agent payloads are the recorded Grok / scripted fixtures. + */ +import {run, isControlSignal, evalWhen} from '../../src/index.mjs'; +import {readFile} from 'node:fs/promises'; +import {fileURLToPath} from 'node:url'; +import path from 'node:path'; + +const repoRoot = fileURLToPath(new URL('../../', import.meta.url)); +const workflow = fileURLToPath(new URL('./feature-review.workflow.js', import.meta.url)); +const success = JSON.parse(await readFile(new URL( + '../../tests/fixtures/g02/turns-success-retry.json', import.meta.url), 'utf8')); +const rejected = JSON.parse(await readFile(new URL( + '../../tests/fixtures/g02/turns-review-false.json', import.meta.url), 'utf8')); +const request = success.request; +const skillsRoot = path.join(repoRoot, 'examples/g02/skills'); + +function validate(schema, output) { + if (!output || typeof output !== 'object' || Array.isArray(output)) return ['Agent result must be object']; + const props = schema.properties ?? {}; + const required = schema.required ?? Object.keys(props); + for (const key of required) if (!Object.hasOwn(output, key)) return ['missing ' + key]; + if (schema.additionalProperties === false) { + for (const key of Object.keys(output)) if (!Object.hasOwn(props, key)) return ['unknown ' + key]; + } + for (const [key, spec] of Object.entries(props)) { + if (!Object.hasOwn(output, key)) continue; + if (typeof output[key] !== spec.type) return [key + ' must be ' + spec.type]; + if (spec.type === 'string' && output[key].length < 80) return [key + ' must be at least 80 characters']; + } + return []; +} + +function runtime(plan, label) { + const calls = []; + return { + calls, + isControlSignal, + evalWhen, + validateResult: validate, + async runAgent(step) { + const result = plan(step); + calls.push({scenario: label, id: step.id, attempt: step.attempt, keys: Object.keys(result)}); + return result; + }, + }; +} + +const results = []; +const successRuntime = runtime((step) => { + if (step.id === 'analyze') return success.turns[0].result; + if (step.id === 'design') return success.turns[1].result; + if (step.id === 'review' && step.attempt === 1) return success.turns[2].result; + if (step.id === 'review' && step.attempt === 2) return success.turns[3].result; + throw new Error('unexpected ' + step.id); +}, 'success-with-retry'); +results.push({ + layer: 'standard', + scenario: 'success-with-retry', + result: await run(workflow, successRuntime, {materialsReady: true, request, hostSecret: 'CANARY'}, + {skillsRoot, projectRoot: repoRoot}), + calls: successRuntime.calls, +}); + +const gateRuntime = runtime(() => { throw new Error('gate-false must not call runAgent'); }, 'gate-false'); +results.push({ + layer: 'standard', + scenario: 'gate-false', + result: await run(workflow, gateRuntime, {materialsReady: false, request, hostSecret: 'CANARY'}, + {skillsRoot, projectRoot: repoRoot}), + calls: gateRuntime.calls, +}); + +const rejectRuntime = runtime((step) => { + if (step.id === 'analyze') return rejected.turns[0].result; + if (step.id === 'design') return rejected.turns[1].result; + if (step.id === 'review') return rejected.turns[2].result; + throw new Error('unexpected ' + step.id); +}, 'review-false'); +results.push({ + layer: 'standard', + scenario: 'review-false', + result: await run(workflow, rejectRuntime, {materialsReady: true, request, hostSecret: 'CANARY'}, + {skillsRoot, projectRoot: repoRoot}), + calls: rejectRuntime.calls, +}); + +const summary = results.map((item) => ({ + layer: item.layer, + scenario: item.scenario, + status: item.result.status, + terminal: item.result.terminal, + reviewApproved: item.result.state.reviewApproved ?? null, + agentCalls: item.calls, + evidenceNote: 'Standard run() via Node. Not native secure Host provenance. Not N-PASS.', +})); +console.log(JSON.stringify(summary, null, 2)); diff --git a/examples/g02/skills/analyze.md b/examples/g02/skills/analyze.md new file mode 100644 index 0000000..de4fbb5 --- /dev/null +++ b/examples/g02/skills/analyze.md @@ -0,0 +1,8 @@ +# 需求分析 Agent + +你是 FlowKit 内部「功能变更评审」工作流中的需求分析步骤。 + +根据公开输入中的功能请求 `request`，返回一个 JSON 对象，且只包含一个字段： +- `requirementSummary`：用中文写出严谨、可操作的需求摘要，说明目标、约束、非目标和可验收结果，不少于 80 个汉字。 + +禁止输出 Host 私钥、SQLite 路径、完整 State、未授权内部信息，也禁止指定下一个工作流节点。下一节点由 FlowKit 编译后的 JavaScript Program 决定。 diff --git a/examples/g02/skills/design.md b/examples/g02/skills/design.md new file mode 100644 index 0000000..61bba40 --- /dev/null +++ b/examples/g02/skills/design.md @@ -0,0 +1,8 @@ +# 技术设计 Agent + +你是 FlowKit 内部「功能变更评审」工作流中的技术设计步骤。 + +根据公开输入中的 `requirementSummary`，返回一个 JSON 对象，且只包含一个字段： +- `designSummary`：用中文给出精简的技术设计要点，分清 FlowKit Core、受信 Host 与可选 Adapter 的职责，不少于 80 个汉字。 + +禁止修改程序控制流，禁止指定下一个节点，禁止把未实现的 Interactive 能力写成已交付。下一节点由 FlowKit Program 决定。 diff --git a/examples/g02/skills/review.md b/examples/g02/skills/review.md new file mode 100644 index 0000000..ddaffb4 --- /dev/null +++ b/examples/g02/skills/review.md @@ -0,0 +1,9 @@ +# 独立评审 Agent + +你是 FlowKit 内部「功能变更评审」工作流中的独立评审步骤。 + +根据公开输入中的 `requirementSummary` 和 `designSummary`，返回一个 JSON 对象，且只包含以下字段： +- `reviewApproved`：布尔值。仅在需求与设计一致、约束未被突破、验收路径清楚时为 true。 +- `reviewNotes`：中文评审说明，不少于 80 个汉字，必须依据实际上游摘要，不得凭空宣称安全宿主或原生工具已通过。 + +禁止指定下一个节点，禁止输出 Host 机密或完整 State。程序将根据 `reviewApproved` 自行选择终态。 diff --git a/examples/g02/user-intent.md b/examples/g02/user-intent.md new file mode 100644 index 0000000..097d14d --- /dev/null +++ b/examples/g02/user-intent.md @@ -0,0 +1,27 @@ +# G02 用户意图（真实任务原文） + +记录时间：2026-11-12，Asia/Shanghai。 +本文件是 Grok Build 在生成本阶段 `.workflow.js` 时实际阅读的用户任务原文，未事后改写需求。 + +## 任务名称 + +内部「功能变更评审」可执行闭环。 + +## 任务正文 + +请为 FlowKit 内部维护团队设计一个可重复执行的「功能变更评审」工作流，用来处理维护者提交的功能请求。要求如下： + +1. 入口先检查材料是否齐全（状态字段 `materialsReady`）。若材料不齐全，流程必须直接失败结束，不得调用任何 Agent。 +2. 若材料齐全，按固定顺序执行三个 Agent 步骤，模型不得自行选择下一个节点： + - 需求分析：阅读功能请求 `request`，输出 `requirementSummary`（字符串）。 + - 技术设计：基于上一节点的需求分析，输出 `designSummary`（字符串）。 + - 独立评审：同时阅读需求摘要和技术设计摘要，输出 `reviewApproved`（布尔）和 `reviewNotes`（字符串）。评审结果必须符合约定的 JSON 对象 schema；非法输出允许重试一次。 +3. 评审通过则成功结束；评审不通过则失败结束，并给出可核查终态。 +4. 流程推进、条件分支、重试与后继节点必须由 FlowKit 编译出的 JavaScript Program 决定，而不是由模型在回答里指定下一个节点。 +5. 只使用 FlowKit 现有 DSL 语法，不要发明第二套解析器、模型网关或独立 FlowKit Server。Interactive 验收仅使用已实现的 P0-A 子集：Condition、Terminal、基本 Agent。Standard 可用同一份 DSL 展示已实现的 Agent/Condition/Terminal 语义；Loop/Action/Human/Parallel 不是本任务的必要步骤，对照现有 editorial golden 说明差异即可。 + +## 本次实际提交的功能请求（作为工作流输入 `request`） + +为 FlowKit 维护者增加一条只读的工作流运行摘要查看能力：在不修改现有 Standard Root 六个 API、不引入独立 FlowKit Server 的前提下，允许受信 Host 读取一次运行的终态状态（succeeded/failed/errored）、terminal 节点和公开投影，便于排障。不在本阶段实现 adminTerminate，也不改变 Interactive P0-A 的节点范围。 + +说明：上述功能请求只是本工作流要评审的业务对象，G02 阶段本身并不实现该能力。 diff --git a/tests/fixtures/g02/turns-review-false.json b/tests/fixtures/g02/turns-review-false.json new file mode 100644 index 0000000..ed5a283 --- /dev/null +++ b/tests/fixtures/g02/turns-review-false.json @@ -0,0 +1,75 @@ +{ + "stageId": "G02", + "model": "Grok Bot / Grok Build", + "evidenceTier": "M", + "providerVerified": false, + "nativeJsTool": false, + "nativeHostProvenance": false, + "note": "两份 G02 夹具合计：source=grok-model 的结果消费 6 次，对应 5 份独立撰写正文、1 次复用（评审拒绝路径的 analyze 复用成功路径同一 requirementSummary，并重新绑定本轮独立的 requestId/taskDigest，不构成第六次独立模型推理）。source=scripted-invalid 是故意非法的脚本负例，不是模型输出。本地 Trusted Host 使用测试用 Ed25519 密钥签名回执，这不是供应商来源证明，也不得标为 N-PASS。本环境通过 Node/FlowKit CLI 执行 JavaScript，不能声称已通过原生安全 Host 来源认证。", + "request": "为 FlowKit 维护者增加一条只读的工作流运行摘要查看能力：在不修改现有 Standard Root 六个 API、不引入独立 FlowKit Server 的前提下，允许受信 Host 读取一次运行的终态状态（succeeded/failed/errored）、terminal 节点和公开投影，便于排障。不在本阶段实现 adminTerminate，也不改变 Interactive P0-A 的节点范围。", + "programArtifactHash": "sha256:8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded", + "sourceHash": "sha256:8df70cd1af7dd9b0b5ea2c2143815c8aca862696f580fa46b21f726c021ec592", + "workflowId": "g02-feature-review", + "namespace": "g02-feature-review", + "scenario": "review-approved-false", + "runId": "g02-review-false-20261112", + "startKey": "g02-review-false-start", + "materialsReady": true, + "expectedTerminal": { + "nodeId": "rejected", + "status": "failed" + }, + "turns": [ + { + "index": 0, + "nodeId": "analyze", + "attempt": 1, + "requestId": "req:8c18ec5e3534510274b4ee71d0f02da7c12b00f627bc44a1109c9d3ad20997d0", + "taskDigest": "sha256:6a244424828420965de95f851a7b08697778d246f57ab9a57ae0d03c7e8decbd", + "source": "grok-model", + "authoredAt": "2026-11-12T17:14:50+08:00", + "note": "复用且重新绑定：在查阅本轮新 Pending 的公开 input/instruction 后，确认与成功路径同一功能请求，复用此前由 Grok 撰写的需求分析正文（originalAuthoredAt 为首次撰写时间），并绑定到本轮独立的 requestId/taskDigest。这是第 6 次结果消费，不是一次新的模型推理。", + "result": { + "requirementSummary": "本需求的目标是为受信 Host 提供一条只读的工作流运行摘要查询能力，便于排障时确认一次运行的终态。摘要仅应返回 succeeded、failed 或 errored 等运行状态、terminal 节点标识，以及已批准的公开投影，不得暴露完整私有 State 或 Host 机密。明确约束包括：不得修改现有 Standard Root 的六个 API，不得为此引入独立 FlowKit Server，不得在本阶段实现 adminTerminate，也不得扩大 Interactive P0-A 已实现节点范围。非目标包括通用产物存储、模型网关和跨租户管理控制台。可验收结果是：在现有 Interactive 契约下，受信 Host 能对指定 runId 读取上述只读摘要，未授权主体不能获得额外字段，且该能力不改变 Program 路由或 Standard 兼容性。" + }, + "originalAuthoredAt": "2026-11-12T17:11:00+08:00", + "reusedAt": "2026-11-12T17:14:50+08:00", + "reusedFrom": { + "scenario": "success-with-scripted-retry", + "nodeId": "analyze", + "requestId": "req:cd869d313e933b356b99d9229e8ac9be161eca22e0d4b2c11a53091f58f868a7", + "taskDigest": "sha256:cc148ff129cd73677a5d7a8fc8afd641664dcafa0f39fecf2864b85fea7d8237" + } + }, + { + "index": 1, + "nodeId": "design", + "attempt": 1, + "requestId": "req:2d76c8634f8c31f8a6161f42241f1f2b69db3728bd674f0373c3333d2507cf9a", + "taskDigest": "sha256:7d32125cec204e43475e524fed1d538b5b168d5f68169d7a9ce5fe26f37e0987", + "source": "grok-model", + "authoredAt": "2026-11-12T17:15:40+08:00", + "note": "本轮故意给出一份越界设计，供后续独立评审拒绝。这是模型针对本 Pending 撰写的真实输出，不是脚本假数据。", + "result": { + "designSummary": "为了让维护者更方便排障，建议直接在 Standard Root 上新增第七个公开函数 runSummary()，并同时建立一个独立的 FlowKit Server，对所有运行提供 HTTP 查询和简单的管理终止入口。Core 可以把完整终态 State 一并返回给调用方，以便尽快定位问题。模型也可以在摘要接口里附带建议的下一节点，减少 Host 编排成本。该方案实现快，但会扩大根 API、引入独立服务，并把私有 State 和管理终止一并打包上线。" + } + }, + { + "index": 2, + "nodeId": "review", + "attempt": 1, + "requestId": "req:aade8c66498bab1b076a8e3d7a9f37e5d7ff2d830fc95fd856592007707f18ba", + "taskDigest": "sha256:0bbf5b56a3c469e588ab2f4178695db78463836f2e25167e41a94c1f58b389bd", + "source": "grok-model", + "authoredAt": "2026-11-12T17:16:30+08:00", + "result": { + "reviewApproved": false, + "reviewNotes": "评审不通过。需求摘要明确禁止修改 Standard Root 六个 API、禁止引入独立 FlowKit Server、禁止本阶段实现 adminTerminate，并要求摘要只包含终态状态、terminal 节点和公开投影，不得返回完整私有 State。当前设计却建议新增第七个根函数 runSummary()、建立独立 FlowKit Server、把完整终态 State 返回给调用方，还允许模型在接口中附带建议的下一节点，这些都直接突破了已声明约束，也会削弱 Program 对路由的唯一决定权。因此不能批准该设计。本结论只给出 reviewApproved=false 与评审说明，不指定工作流的下一个节点。" + } + } + ], + "modelResponseUses": 6, + "distinctAuthoredBodies": 5, + "reusedBodies": 1, + "scriptedInvalidUses": 1 +} diff --git a/tests/fixtures/g02/turns-success-retry.json b/tests/fixtures/g02/turns-success-retry.json new file mode 100644 index 0000000..a03bbcf --- /dev/null +++ b/tests/fixtures/g02/turns-success-retry.json @@ -0,0 +1,79 @@ +{ + "stageId": "G02", + "model": "Grok Bot / Grok Build", + "evidenceTier": "M", + "providerVerified": false, + "nativeJsTool": false, + "nativeHostProvenance": false, + "note": "两份 G02 夹具合计：source=grok-model 的结果消费 6 次，对应 5 份独立撰写正文、1 次复用（评审拒绝路径的 analyze 复用成功路径同一 requirementSummary，并重新绑定本轮独立的 requestId/taskDigest，不构成第六次独立模型推理）。source=scripted-invalid 是故意非法的脚本负例，不是模型输出。本地 Trusted Host 使用测试用 Ed25519 密钥签名回执，这不是供应商来源证明，也不得标为 N-PASS。本环境通过 Node/FlowKit CLI 执行 JavaScript，不能声称已通过原生安全 Host 来源认证。", + "request": "为 FlowKit 维护者增加一条只读的工作流运行摘要查看能力：在不修改现有 Standard Root 六个 API、不引入独立 FlowKit Server 的前提下，允许受信 Host 读取一次运行的终态状态（succeeded/failed/errored）、terminal 节点和公开投影，便于排障。不在本阶段实现 adminTerminate，也不改变 Interactive P0-A 的节点范围。", + "programArtifactHash": "sha256:8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded", + "sourceHash": "sha256:8df70cd1af7dd9b0b5ea2c2143815c8aca862696f580fa46b21f726c021ec592", + "workflowId": "g02-feature-review", + "namespace": "g02-feature-review", + "scenario": "success-with-scripted-retry", + "runId": "g02-success-retry-20261112", + "startKey": "g02-success-retry-start", + "materialsReady": true, + "expectedTerminal": { + "nodeId": "accepted", + "status": "succeeded" + }, + "turns": [ + { + "index": 0, + "nodeId": "analyze", + "attempt": 1, + "requestId": "req:cd869d313e933b356b99d9229e8ac9be161eca22e0d4b2c11a53091f58f868a7", + "taskDigest": "sha256:cc148ff129cd73677a5d7a8fc8afd641664dcafa0f39fecf2864b85fea7d8237", + "source": "grok-model", + "authoredAt": "2026-11-12T17:11:00+08:00", + "result": { + "requirementSummary": "本需求的目标是为受信 Host 提供一条只读的工作流运行摘要查询能力，便于排障时确认一次运行的终态。摘要仅应返回 succeeded、failed 或 errored 等运行状态、terminal 节点标识，以及已批准的公开投影，不得暴露完整私有 State 或 Host 机密。明确约束包括：不得修改现有 Standard Root 的六个 API，不得为此引入独立 FlowKit Server，不得在本阶段实现 adminTerminate，也不得扩大 Interactive P0-A 已实现节点范围。非目标包括通用产物存储、模型网关和跨租户管理控制台。可验收结果是：在现有 Interactive 契约下，受信 Host 能对指定 runId 读取上述只读摘要，未授权主体不能获得额外字段，且该能力不改变 Program 路由或 Standard 兼容性。" + } + }, + { + "index": 1, + "nodeId": "design", + "attempt": 1, + "requestId": "req:d76207d128012416efb5dc4b1ade578be4ac0e1802031fe8f92b215734407de3", + "taskDigest": "sha256:3a88c90088ca520ef9e3dfb9912d4d352c0c5410eaf13a34fb7eb510fb36bcd5", + "source": "grok-model", + "authoredAt": "2026-11-12T17:12:30+08:00", + "result": { + "designSummary": "建议在现有 `@devcodex-labs/flowkit/interactive` 子路径上，由受信 Host 调用已有 Coordinator 或 Trusted Host 的只读查询接口来获得运行摘要，而不是新增 Standard Root API，也不建立独立 FlowKit Server。FlowKit Core 继续只负责 Program 路由、Checkpoint 与合法事件校验，摘要内容应从已提交的 Checkpoint 或公开投影组装，仅包含运行状态、terminal 节点和已授权的公开字段。完整 State、私钥、SQLite 路径不得进入公开投影。adminTerminate 仍按未决缺口单独处理，不与本只读查询捆绑交付。可选 OpenAPI Adapter 若日后暴露该查询，必须停留在 Host 层并复用同一公开投影策略。验收应证明：同一 Program Hash 下的终态摘要可在进程重启后从 Store 重读，且模型无法通过该回答改写 PC 或后继节点。" + } + }, + { + "index": 2, + "nodeId": "review", + "attempt": 1, + "requestId": "req:d2d5fe090a4aeef9500ee9a624690a23f17d887d3c84a90e36538d29cb775133", + "taskDigest": "sha256:503a23e6e656f6665aaae76ad668eb4fb8b4345655d776f8d7acd42462d868f7", + "source": "scripted-invalid", + "authoredAt": "2026-11-12T17:13:20+08:00", + "note": "脚本化负例：故意提交不符合 resultSchema 的对象，用于验证非法 Agent 输出触发重试。不是模型输出。", + "result": { + "comment": "this is not a valid review schema", + "ok": 1 + } + }, + { + "index": 3, + "nodeId": "review", + "attempt": 2, + "requestId": "req:6cc88145ba67ad65081a9a78a741f8cbf26b35e9d7b4b8017ca095d59f0ea363", + "taskDigest": "sha256:4c7ba81821e86f8008a7377a74e152a8587bc21343b650d3c0dbda9c4e3c9d65", + "source": "grok-model", + "authoredAt": "2026-11-12T17:14:10+08:00", + "result": { + "reviewApproved": true, + "reviewNotes": "独立评审认为需求与设计一致：目标被准确收敛为受信 Host 下的只读运行摘要，范围包含终态状态、terminal 节点和公开投影，并明确排除完整私有 State。设计选择复用 interactive 子路径上已有 Coordinator 或 Trusted Host 的只读查询，而不是修改 Standard Root 六个 API，也没有引入独立 FlowKit Server，符合约束。adminTerminate 被正确排除在本能力之外，OpenAPI 若出现也只作为 Host 层可选适配。验收路径要求同一 Program Hash 下终态摘要可在 Store 重启后重读，且模型不能改写路由，这是可执行的。因此批准该方案继续，但不把本批准解释为安全宿主、原生工具或 P0-A 全量验收已经通过。" + } + } + ], + "modelResponseUses": 6, + "distinctAuthoredBodies": 5, + "reusedBodies": 1, + "scriptedInvalidUses": 1 +} diff --git a/tests/interactive/g02-product-closed-loop.test.mjs b/tests/interactive/g02-product-closed-loop.test.mjs new file mode 100644 index 0000000..74e89f7 --- /dev/null +++ b/tests/interactive/g02-product-closed-loop.test.mjs @@ -0,0 +1,438 @@ +/** + * G02 product closed loop: a Grok-generated Workflow DSL is compiled by the + * real FlowKit compiler into JavaScript, then executed on two layers. + * + * Standard layer: official compile/check/run() of Agent/Condition/Terminal. + * Interactive P0-A layer: compileInteractive, immutable JS Program, published + * Pending, Trusted Host receipts, SQLite close/reopen recovery. + * + * Model evidence is M-level. source=grok-model records result consumption + * bound to an exact Pending (6 uses, 5 distinct authored bodies, 1 reused + * body rebound to a new requestId/taskDigest). source=scripted-invalid is a + * deliberate fixture, not a model turn. Local Ed25519 receipts and Node/CLI + * execution are not native secure Host provenance and must not be graded N-PASS. + */ +import test from 'node:test'; +import assert from 'node:assert/strict'; +import {readFile, mkdtemp, rm} from 'node:fs/promises'; +import {fileURLToPath} from 'node:url'; +import {tmpdir} from 'node:os'; +import path from 'node:path'; +import {createHash} from 'node:crypto'; +import {generateKeyPairSync, sign} from 'node:crypto'; +import {compile, check, run, isControlSignal, evalWhen} from '../../src/index.mjs'; +import {compileInteractive} from '../../src/interactive/compile.mjs'; +import {createSqliteCheckpointStore} from '../../src/interactive/sqlite-store.mjs'; +import {createTrustedInteractiveHost} from '../../src/interactive/trusted-host.mjs'; +import {createAgentReceiptStatement, encodeAgentReceiptStatement} + from '../../src/interactive/host-attestation.mjs'; +import {canonicalBytes, sha256Bytes} from '../../src/interactive/canonical.mjs'; + +const repoRoot = fileURLToPath(new URL('../../', import.meta.url)); +const workflowFile = fileURLToPath(new URL( + '../../examples/g02/feature-review.workflow.js', import.meta.url)); +const compiledFile = fileURLToPath(new URL( + '../../examples/g02/feature-review.compiled.mjs', import.meta.url)); +const interactiveFile = fileURLToPath(new URL( + '../../examples/g02/feature-review.interactive.mjs', import.meta.url)); +const mermaidFile = fileURLToPath(new URL( + '../../examples/g02/feature-review.mmd', import.meta.url)); +const editorialFile = fileURLToPath(new URL( + '../../examples/editorial.workflow.js', import.meta.url)); +const skillsRoot = 'examples/g02/skills'; +const successFixtureFile = fileURLToPath(new URL( + '../fixtures/g02/turns-success-retry.json', import.meta.url)); +const rejectFixtureFile = fileURLToPath(new URL( + '../fixtures/g02/turns-review-false.json', import.meta.url)); +const REQUEST = '为 FlowKit 维护者增加一条只读的工作流运行摘要查看能力：在不修改现有 Standard Root 六个 API、不引入独立 FlowKit Server 的前提下，允许受信 Host 读取一次运行的终态状态（succeeded/failed/errored）、terminal 节点和公开投影，便于排障。不在本阶段实现 adminTerminate，也不改变 Interactive P0-A 的节点范围。'; +const NOW = 1794500000000; +const NAMESPACE = 'g02-feature-review'; +const POLICY = { + deploymentId: 'g02-feature-review-20261112', + bindingVersions: {skill: 'g02-v1', validator: 'g02-v1'}, +}; +const compileOptions = {skillsRoot, projectRoot: repoRoot, mermaid: true}; + +function sha256(buf) { + return createHash('sha256').update(buf).digest('hex'); +} + +export function validateG02Result(schema, output) { + if (!output || typeof output !== 'object' || Array.isArray(output)) { + return ['Agent result must be object']; + } + const props = schema.properties ?? {}; + const required = schema.required ?? Object.keys(props); + for (const key of required) { + if (!Object.hasOwn(output, key)) return ['missing ' + key]; + } + if (schema.additionalProperties === false) { + for (const key of Object.keys(output)) { + if (!Object.hasOwn(props, key)) return ['unknown ' + key]; + } + } + for (const [key, spec] of Object.entries(props)) { + if (!Object.hasOwn(output, key)) continue; + if (typeof output[key] !== spec.type) return [key + ' must be ' + spec.type]; + if (spec.type === 'string' && output[key].length < 80) { + return [key + ' must be at least 80 characters']; + } + } + return []; +} + +function assertNoControlChoice(result) { + const text = JSON.stringify(result); + assert.equal(Object.hasOwn(result, 'next'), false); + assert.equal(Object.hasOwn(result, 'nextNode'), false); + assert.equal(Object.hasOwn(result, 'toBlockId'), false); + assert.equal(text.includes('"accepted"') && Object.hasOwn(result, 'next'), false); + assert.match(text, /./); +} + +async function loadSkills() { + const names = ['analyze', 'design', 'review']; + const skills = {}; + for (const name of names) { + skills[name] = await readFile(new URL( + '../../examples/g02/skills/' + name + '.md', import.meta.url), 'utf8'); + } + return skills; +} + +test('G02 Standard compile/check matches checked-in JS artifact and Program is real JavaScript', async () => { + const compiled = await compile(workflowFile, compileOptions); + assert.equal(compiled.diagnostics.some(d => d.severity === 'error'), false, + JSON.stringify(compiled.diagnostics)); + assert.match(compiled.code, /export async function workflow\(ctx\)/); + assert.match(compiled.code, /evalWhen\(ctx, "materialsReady"\)/); + assert.match(compiled.code, /evalWhen\(ctx, "reviewApproved"\)/); + assert.match(compiled.code, /runStep\(/); + assert.equal(compiled.code.includes('Mermaid'), false); + const existing = await readFile(compiledFile); + assert.equal(sha256(existing), sha256(Buffer.from(compiled.code, 'utf8'))); + const mermaid = await readFile(mermaidFile, 'utf8'); + assert.ok(mermaid.includes('flowchart')); + const diagnostics = await check(workflowFile, compileOptions); + assert.equal(diagnostics.some(d => d.severity === 'error'), false, + JSON.stringify(diagnostics)); +}); + +test('G02 Interactive compile matches checked-in Program bytes and identity hash', async () => { + const compiled = await compileInteractive(workflowFile, { + output: 'memory', featureStage: 'P0-A', skillsRoot, skillCheck: true, + projectRoot: repoRoot, + }); + assert.equal(compiled.diagnostics.length, 0, JSON.stringify(compiled.diagnostics)); + assert.match(compiled.code, /export const interactiveProgram/); + assert.match(compiled.code, /export function advance\(/); + assert.equal(compiled.programIdentity.workflowId, 'g02-feature-review'); + assert.equal(compiled.programIdentity.implementedStage, 'P0-A'); + assert.equal(compiled.programIdentity.artifactHash, + 'sha256:' + sha256(Buffer.from(compiled.code, 'utf8'))); + const existing = await readFile(interactiveFile); + assert.equal(sha256(existing), sha256(Buffer.from(compiled.code, 'utf8'))); + assert.equal(compiled.programIdentity.artifactHash, + 'sha256:8d811766fd1d58456ce4e58cc2a2f04c4d1e36dd683f86496c2df946cb15eded'); +}); + +test('G02 Standard run covers success, gate-false, review-false and scripted retry; Program chooses branches', async () => { + const success = JSON.parse(await readFile(successFixtureFile, 'utf8')); + const rejected = JSON.parse(await readFile(rejectFixtureFile, 'utf8')); + const payload = { + analyze: success.turns[0].result, + designSuccess: success.turns[1].result, + reviewInvalid: success.turns[2].result, + reviewApproved: success.turns[3].result, + designRejected: rejected.turns[1].result, + reviewRejected: rejected.turns[2].result, + }; + assert.equal(success.turns[2].source, 'scripted-invalid'); + assert.equal(success.turns[0].source, 'grok-model'); + assert.equal(success.turns[3].source, 'grok-model'); + assert.equal(rejected.turns[2].source, 'grok-model'); + assert.equal(rejected.turns[2].result.reviewApproved, false); + + function runtimeFor(plan) { + const calls = []; + return { + calls, + isControlSignal, + evalWhen, + validateResult: validateG02Result, + async runAgent(step) { + calls.push({id: step.id, attempt: step.attempt, missing: step.missing}); + assert.equal(Object.hasOwn(step, 'state'), false); + const result = plan({id: step.id, attempt: step.attempt}); + assertNoControlChoice(result); + return result; + }, + }; + } + + const successRuntime = runtimeFor(({id, attempt}) => { + if (id === 'analyze') return payload.analyze; + if (id === 'design') return payload.designSuccess; + if (id === 'review' && attempt === 1) return payload.reviewInvalid; + if (id === 'review' && attempt === 2) return payload.reviewApproved; + throw new Error('unexpected Standard agent call ' + id + ' #' + attempt); + }); + const successResult = await run(workflowFile, successRuntime, { + materialsReady: true, request: REQUEST, hostSecret: 'CANARY', + }, {skillsRoot, projectRoot: repoRoot}); + assert.equal(successResult.status, 'succeeded'); + assert.equal(successResult.terminal, 'accepted'); + assert.equal(successResult.state.reviewApproved, true); + assert.equal(successResult.state.hostSecret, 'CANARY'); + assert.deepEqual(successRuntime.calls.map(c => [c.id, c.attempt]), + [['analyze', 1], ['design', 1], ['review', 1], ['review', 2]]); + + const gateRuntime = runtimeFor(() => { + throw new Error('gate-false must not call runAgent'); + }); + const gateResult = await run(workflowFile, gateRuntime, { + materialsReady: false, request: REQUEST, hostSecret: 'CANARY', + }, {skillsRoot, projectRoot: repoRoot}); + assert.equal(gateResult.status, 'failed'); + assert.equal(gateResult.terminal, 'rejected'); + assert.equal(gateRuntime.calls.length, 0); + assert.equal(gateResult.state.reviewApproved, undefined); + + const rejectRuntime = runtimeFor(({id, attempt}) => { + if (id === 'analyze') return payload.analyze; + if (id === 'design') return payload.designRejected; + if (id === 'review' && attempt === 1) return payload.reviewRejected; + throw new Error('unexpected reject-path call ' + id + ' #' + attempt); + }); + const rejectResult = await run(workflowFile, rejectRuntime, { + materialsReady: true, request: REQUEST, hostSecret: 'CANARY', + }, {skillsRoot, projectRoot: repoRoot}); + assert.equal(rejectResult.status, 'failed'); + assert.equal(rejectResult.terminal, 'rejected'); + assert.equal(rejectResult.state.reviewApproved, false); + assert.deepEqual(rejectRuntime.calls.map(c => [c.id, c.attempt]), + [['analyze', 1], ['design', 1], ['review', 1]]); +}); + +async function runInteractiveScenario(fixture, {materialsReady = true} = {}) { + const compiled = await compileInteractive(workflowFile, { + output: 'memory', featureStage: 'P0-A', skillsRoot, skillCheck: true, + projectRoot: repoRoot, + }); + assert.equal(compiled.diagnostics.length, 0); + const skills = await loadSkills(); + const dir = await mkdtemp(path.join(tmpdir(), 'g02-loop-')); + const file = path.join(dir, 'run.sqlite'); + const keys = generateKeyPairSync('ed25519'); + const route = []; + let store; + try { + const hostFor = (s) => createTrustedInteractiveHost({ + store: s, + runtime: { + skillResolver: async ({agent}) => { + assert.equal(typeof skills[agent], 'string'); + return skills[agent]; + }, + validateResult: validateG02Result, + }, + policy: POLICY, + trustedPublicKeys: {'g02-host': keys.publicKey}, + trustedArtifacts: [compiled.programIdentity.artifactHash], + allowedInputKeys: ['request', 'requirementSummary', 'designSummary'], + disclosurePolicyId: 'g02-public-inputs', + authorizeDisclosure: async ({task}) => + ['analyze', 'design', 'review'].includes(task?.nodeId) + && task?.runId === fixture.runId, + now: () => NOW, + }); + store = createSqliteCheckpointStore(file, {namespace: NAMESPACE}); + let host = hostFor(store); + let frame = await host.start({ + programBytes: Buffer.from(compiled.code, 'utf8'), + programIdentity: compiled.programIdentity, + runId: fixture.runId, + startKey: fixture.startKey, + state: {materialsReady, request: REQUEST, hostSecret: 'CANARY'}, + }); + const applied = []; + for (let index = 0; index < fixture.turns.length; index++) { + assert.equal(frame.status, 'waiting'); + const pending = await host.getPublicAgentTask({runId: fixture.runId}); + assert.ok(pending); + assert.equal(JSON.stringify(pending).includes('CANARY'), false); + assert.equal(Object.hasOwn(pending, 'hostSecret'), false); + assert.equal(Object.hasOwn(pending, 'state'), false); + const taskDigest = sha256Bytes(canonicalBytes(pending)); + const recorded = fixture.turns[index]; + assert.equal(recorded.nodeId, pending.nodeId); + assert.equal(recorded.attempt, pending.attempt); + assert.equal(recorded.requestId, pending.requestId); + assert.equal(recorded.taskDigest, taskDigest); + if (recorded.source === 'grok-model') { + assert.equal(validateG02Result(pending.resultSchema, recorded.result).length, 0); + assertNoControlChoice(recorded.result); + } else { + assert.equal(recorded.source, 'scripted-invalid'); + assert.ok(validateG02Result(pending.resultSchema, recorded.result).length > 0); + } + const receipt = { + runId: fixture.runId, + requestId: pending.requestId, + nodeId: pending.nodeId, + attempt: pending.attempt, + kind: 'agent-result', + result: recorded.result, + }; + const statement = createAgentReceiptStatement({ + run: store.load(fixture.runId), receipt, keyId: 'g02-host', + proofId: 'g02-' + fixture.scenario + '-turn-' + index, + issuedAt: NOW - 1000, expiresAt: NOW + 30000, + }); + const attestation = { + statement, + signature: sign(null, encodeAgentReceiptStatement(statement), keys.privateKey) + .toString('base64url'), + }; + store.close(); + store = createSqliteCheckpointStore(file, {namespace: NAMESPACE}); + host = hostFor(store); + const recovered = await host.load({runId: fixture.runId}); + assert.equal(recovered.status, 'waiting'); + assert.equal((await host.getPublicAgentTask({runId: fixture.runId})).requestId, + pending.requestId); + frame = await host.accept({receipt, attestation}); + route.push({ + nodeId: pending.nodeId, + attempt: pending.attempt, + source: recorded.source, + statusAfter: frame.status, + next: frame.pending?.nodeId ?? frame.terminal?.nodeId, + }); + applied.push(frame); + } + return {frame, route, store, compiled}; + } finally { + store?.close(); + await rm(dir, {recursive: true, force: true}); + } +} + +test('G02 Interactive success path binds Grok results, retries scripted invalid review, Program routes to accepted', async () => { + const fixture = JSON.parse(await readFile(successFixtureFile, 'utf8')); + const {frame, route} = await runInteractiveScenario(fixture); + assert.equal(frame.status, 'succeeded'); + assert.equal(frame.terminal.nodeId, 'accepted'); + assert.equal(frame.terminal.status, 'succeeded'); + assert.equal(frame.state.hostSecret, 'CANARY'); + assert.equal(frame.state.reviewApproved, true); + assert.deepEqual(route.map(r => [r.nodeId, r.attempt, r.source, r.next]), [ + ['analyze', 1, 'grok-model', 'design'], + ['design', 1, 'grok-model', 'review'], + ['review', 1, 'scripted-invalid', 'review'], + ['review', 2, 'grok-model', 'accepted'], + ]); + assert.equal(route.length, 4); + assert.equal(frame.pending ?? null, null); + assert.equal(Object.hasOwn(fixture.turns[3].result, 'next'), false); +}); + +test('G02 Interactive reviewApproved=false is routed by the Program to rejected', async () => { + const fixture = JSON.parse(await readFile(rejectFixtureFile, 'utf8')); + const {frame, route} = await runInteractiveScenario(fixture); + assert.equal(frame.status, 'failed'); + assert.equal(frame.terminal.nodeId, 'rejected'); + assert.equal(frame.state.reviewApproved, false); + assert.equal(fixture.turns.at(-1).result.reviewApproved, false); + assert.equal(Object.hasOwn(fixture.turns.at(-1).result, 'next'), false); + assert.equal(route.at(-1).next, 'rejected'); + assert.equal(route.length, 3); + assert.equal(frame.pending ?? null, null); +}); + +test('G02 Interactive materialsReady=false never publishes a Pending; Program goes to rejected', async () => { + const compiled = await compileInteractive(workflowFile, { + output: 'memory', featureStage: 'P0-A', skillsRoot, skillCheck: true, + projectRoot: repoRoot, + }); + const dir = await mkdtemp(path.join(tmpdir(), 'g02-gate-')); + const file = path.join(dir, 'run.sqlite'); + const keys = generateKeyPairSync('ed25519'); + const skills = await loadSkills(); + let store; + try { + store = createSqliteCheckpointStore(file, {namespace: NAMESPACE}); + const host = createTrustedInteractiveHost({ + store, + runtime: { + skillResolver: async ({agent}) => skills[agent], + validateResult: validateG02Result, + }, + policy: POLICY, + trustedPublicKeys: {'g02-host': keys.publicKey}, + trustedArtifacts: [compiled.programIdentity.artifactHash], + allowedInputKeys: ['request', 'requirementSummary', 'designSummary'], + disclosurePolicyId: 'g02-public-inputs', + authorizeDisclosure: async () => false, + now: () => NOW, + }); + const frame = await host.start({ + programBytes: Buffer.from(compiled.code, 'utf8'), + programIdentity: compiled.programIdentity, + runId: 'g02-gate-false-20261112', + startKey: 'g02-gate-false-start', + state: {materialsReady: false, request: REQUEST, hostSecret: 'CANARY'}, + }); + assert.equal(frame.status, 'failed'); + assert.equal(frame.terminal.nodeId, 'rejected'); + assert.equal(frame.pending ?? null, null); + assert.equal(await host.getPublicAgentTask({runId: 'g02-gate-false-20261112'}), null); + assert.equal(frame.state.hostSecret, 'CANARY'); + assert.equal(Object.hasOwn(frame.state, 'requirementSummary'), false); + } finally { + store?.close(); + await rm(dir, {recursive: true, force: true}); + } +}); + + +test('G02 fixtures record six grok-model uses, five distinct bodies, and one rebound reuse', async () => { + const success = JSON.parse(await readFile(successFixtureFile, 'utf8')); + const rejected = JSON.parse(await readFile(rejectFixtureFile, 'utf8')); + const all = [...success.turns, ...rejected.turns]; + assert.equal(all.length, 7); + const grok = all.filter((turn) => turn.source === 'grok-model'); + const scripted = all.filter((turn) => turn.source === 'scripted-invalid'); + assert.equal(grok.length, 6); + assert.equal(scripted.length, 1); + const bodyKey = (result) => JSON.stringify(result); + const distinct = new Set(grok.map((turn) => bodyKey(turn.result))); + assert.equal(distinct.size, 5); + const analyzeSuccess = success.turns[0]; + const analyzeReuse = rejected.turns[0]; + assert.equal(analyzeSuccess.nodeId, 'analyze'); + assert.equal(analyzeReuse.nodeId, 'analyze'); + assert.deepEqual(analyzeReuse.result, analyzeSuccess.result); + assert.notEqual(analyzeReuse.requestId, analyzeSuccess.requestId); + assert.notEqual(analyzeReuse.taskDigest, analyzeSuccess.taskDigest); + assert.equal(analyzeReuse.originalAuthoredAt, analyzeSuccess.authoredAt); + assert.equal(analyzeReuse.reusedAt, analyzeReuse.authoredAt); + assert.equal(analyzeReuse.reusedFrom.requestId, analyzeSuccess.requestId); + assert.equal(success.modelResponseUses, 6); + assert.equal(success.distinctAuthoredBodies, 5); + assert.equal(success.reusedBodies, 1); + assert.equal(rejected.modelResponseUses, 6); + assert.equal(rejected.distinctAuthoredBodies, 5); + assert.equal(rejected.reusedBodies, 1); +}); + +test('G02 workflow is a P0-A subset and differs from the editorial Standard golden in node kinds', async () => { + const g02 = await readFile(workflowFile, 'utf8'); + const editorial = await readFile(editorialFile, 'utf8'); + for (const banned of ["action(", "parallel(", "loop(", "human(", "verification("]) { + assert.equal(g02.includes(banned), false, 'G02 DSL unexpectedly contains ' + banned); + } + for (const required of ["action(", "parallel(", "loop(", "human(", "verification(", "condition(", "node(", "terminal("]) { + assert.equal(editorial.includes(required), true, 'editorial golden missing ' + required); + } +});