Commit Graph
251 Commits
Author SHA1 Message Date
cflmy eae4815fc7 docs(design): 函数能否解释为 GFM 表 — 11 组实验 + 史料考古(未改代码)
回应「代码函数(## 名 + 参数 + 体)从某种角度能否解释成 GFM 表,
这样的解释运行时有没有巨大风险、有没有实际收益、历史上出现过吗为什么放弃」。

核心答案:命题在 Marqdo 里**已经实现**,但实现的是「关于函数的表」(契约表),
不是「函数本身」。界:表能描述外延(输入→结果)与签名(元数据),
不能描述内涵(怎么算);而调用恰恰是内涵操作。
[k](集合) 是投影已有数据,f(x) 是计算新数据,只有 f 已是预先枚举的表函数才重合。

一、固定表格几何(实测)
- 1 列 N 行 → List;≥2 列 1 行 → Map 记录(实测 {"x":1,"y":2})
- ≥2 列 N 行 → **列向聚合** Map(实测 {"档位":[1,2],"税率":[...]})
- 行向记录必须显式标记(行/@/row),文档原话「不得与默认列向字典静默混用」
→ 「行 = 参数」在 Marqdo 里不是自然形态

二、命题已存在:契约表(tests/contracts/ok_contract.mq.md)
src/contract.rs:表头含 参数/返回/字段 + 必须有「类型」列 → 契约块。
契约表的首列标头正是 §3.5 那种「显式形态标记」的同类做法。

三、11 组实验(1.5.2 实跑)划定边界
- E1b 表列 n 但体不读 → > f 无参调用成功 → 表**不能声明**形参
- E3 叙述声明 n + 表 + 传"七" → 参数 n 期望 number,实际 text(契约校验)
- E4 无表 → 七1,不校验
- E5 + `n` 旧式形参行 + 表 → 契约生效(用户草图形态可用)
- E6 表写 2 行、体只读 1 个 → too many positional(**表 ≠ arity**)
- E7 单列参数表 → 七1,**契约静默失效**(1 列 = List,且 check 报"无契约问题")
- E8 重复行 → 取首行无诊断;E9 幽灵行 → 静默忽略
- E10 同一张表加 `tc` = 绑定 → 契约消失
→ 隐含规则:**未绑定的表 = 声明/元数据;绑定的表 = 数据**

四、若表成为定义(把静默漂移从"少个校验"升级为"少个参数")
漏类型列/表体漂移/重复行/调换行序都会变成绑定语义;
最致命的是行序携带语义——「把 b 行移到 a 上面」这个无害编辑会改变位置参数绑定。
而 [键](集合) 之所以安全是因为**键是显式的**,取元不依赖顺序。

五、链接式调用第三次论证(带史料)
collection-access-link.md §1「方括号 = 取出的名字,圆括号 = 从哪取」;
§4 非目标排除点号/在/箭头/(键)[集合];call-arguments.md 锁定 `]` 紧邻 `(` = 取元。
→ [f](args) 不是未决定,而是**已被明确分配给取元**。

六、史料:这条路是被反向走过的
- v0.2 要求手写 `+ 名` 形参表 → v0.3 列为**代价**(仪式感强、与正文脱节),
  P4 改为体推断,`+ 名` 标注「过渡兼容,新代码不推荐」
- objects.md §6 非目标:**用表格声明字段 schema**(表格仍表示数据集合)
- tables-maps-footnotes §3.3 行向记录「不默认」,§3.5 显式标记才开启
- collection-access-link §3 记录了取元语法的迁移与定稿

七、收益(诚实)
tool schema 最实:tools 已能写成记录行(tool_names 接受含 name/工具 的 Object),
但 agent_format_tools 只渲染 `- name`,说明与参数被丢弃——能力已在那儿,只差接上签名。

八、建议:表做声明不做定义,补三件事
① 消灭静默:契约表缺类型列/幽灵行/重复行/与 arity 不符 → check 应 warn(当前全静默)
② 反向生成:marqdo view 从签名生成契约表草案(单一真相仍是体,表是投影)
③ 签名可供机器消费:knowledge 能力条目带参数 schema → tool schema 自动生成
零语法改动。§9 给了「双写 + 一致性校验」的可试写法(E5 已验证可跑)。

未修改任何代码。
2026-10-22 12:46:13 +08:00
cflmy 3260494e5c docs(design): 评估「函数作为变量 / 函数即表」(未改代码)
回应「函数能不能是变量的一种,`函数名` = 参数|函数体 表,取元即调用」的提议。

核心结论:这个直觉在数学上成立,但一句话里混了四种不同的「函数」观,必须拆开:
- 外延观:函数是图 → 就是表,调用是查表(有限域才成立)
- 内涵观:函数是规则 → 调用是执行(定义域无穷,写不成表)
- 闭包观:函数 = 代码 + 环境
- 名字观:函数是命名空间一条(Marqdo 现状)
用户草图是「参数声明 + 函数体」的混合体,而「像表格取元」的直觉来自外延观。

现状实测(1.5.2,六组实验):
1. Value 枚举十种无函数;函数是 AST,挂 Module.functions
2. CallExpr.callee 是 String 不是 Expr → 语法上无法调用「计算出来的值」
3. 对象本来就是 Value::Map + _type 标签(对象是值,函数不是,这个不对称是关键)
4. 已存在 call_fn 按名字间接调用,但零参、限当前模块、只在运行期校验
5. 实验:`**f = > greet**` 报 missing argument(> 名字 是调用语法,无法引用函数本身)
6. 实验:表里存函数名 + [h](reg) + call_fn → 动态派发今天就能做

关键实测纠正(我最初写错、已改):
GFM 两列表绑定为「列映射」{"档位":[1,2],"税率":[...]},不是按首列键控的映射,
所以 [2](两列表) 报 missing map key 2。真正键→值查表需 Map(json.parse)
或用 table.index_of + table.at 显式索引(两条配方均已跑通)。

三条断裂点:表格有限 vs 函数常无限;函数体是代码不是数据(需 eval,静态检查/契约/
EKC 全失效);[k](x) 会被语义重载(与上一轮 [函数名](参数) 冲突同源)。

五条路径与代价对比(值模型改动/静态可索引/契约/EKC 指纹/序列化沙箱/成本),
并逐项分析对 Marqdo 特有机制的冲击:契约 T2.1/T2.2 依赖静态 target、
EKC 指纹 collect_calls 读静态 callee(src/knowledge/fingerprint.rs:76)、
value_to_json 对 Value 穷举(src/host/json.rs:98)、字节码 Op::Call(fid) 用编译期 id
(src/bytecode/compile.rs:709)。

建议分阶段:
- 立即:决策表(零改动),但补 table.lookup / table.by_column 库缺口
- 短期:扩展 call_fn(args=、跨模块、字面量静态校验)——不动值模型
- 中期:若要一等函数,走 ADR 并回答匿名函数如何进 EKC 索引
- 不建议:「函数即表」强主张(语义重载 + 削弱静态可索引性)

§9 给出折中:数据在表、逻辑在命名函数,兼顾表驱动直觉与可索引性。
未修改任何代码。
2026-10-22 12:35:40 +08:00
cflmy 42a1a46383 docs: 链接式调用 [函数名](参数) 评估 + 开发环境脚本 dev-env.sh
两件事:

1. doc/design/link-style-call-evaluation.md(Proposed · 待决策)
   评估用户提议的 Markdown 链接式调用 [函数名](参数…)。

   结论:语法优雅,但与既有特性「链接索引 [key](map)」形状完全相同,无法可靠消歧。
   实测证据:
   - [a](m) → 链接索引,取成员,返回 1(既有特性,v0.3 明确推广)
   - [greet](alice) → 被当链接索引处理,报 undefined variable alice
   两者区别只在「greet 是函数还是键名」,需要符号表信息,而语法判定发生在
   解析期(跨模块时未知)。若靠「已知函数表」消歧,会把清晰的
   missing map key 变成误导性的 unknown function。

   给出四个受限方案并对比,推荐 L3:用「] 与 ( 之间是否有空格」消歧
   - [key](coll)  紧贴 → 链接索引,语义一字不动(零回归)
   - [callee] (args) 有空格 → 调用,参数显式成组,顺带解决 D8 空格函数名
   并附 8 条验证计划。未改任何代码。

2. scripts/dev-env.sh(新增,可 source 复用)
   本机测试环境缺失导致基线本来就有失败,本次装好并固化:
   - 约束1:$HOME 不可写 → zig/rust 缓存显式重定向
   - 约束2:/tmp 是 noexec tmpfs → TMPDIR 必须指向可写且可执行目录
     (这是 Go 链接报 "signal: aborted"、zig 报 AccessDenied 的真正原因)
   - 提供 rust 1.81 / go 1.25 / python 3.12 / gcc shim(zig, glibc 2.36)

   环境修复前后:
   - 无 python3:231 passed / 2 failed
   - 加 python3:233 / 0(lib)
   - 加 go + libweb:gold 284 / 0
   - 全量:554 passed / 0 failed

   libweb.so 构建要点:Go cgo 需要 gcc,用 zig cc 充当
   (/workspace/qdagent/.tools/zigcc/bin/gcc,过滤 Go 传入的
   -Qunused-arguments,该参数会让 zig 驱动 abort)。

变更文档同步追加 §8 记录环境安装过程与前后对比。
2026-10-22 11:59:14 +08:00
cflmy b6629e5450 fix(parse): D9 未闭合引号不再静默吞行(分类层 + 解析层双层修复)
缺陷:`**s = "abc`(引号未闭合)被静默当成叙述丢弃,退出码 0,
没有输出也没有报错,该行凭空消失。对「文档即程序」来说这是最危险的一类
错误——文档看起来完整、能跑、check 通过,实际少了一节。

根因在**分类层**而不是解析层:
parse_quoted_string 里 `unterminated " string` 的报错一直都在,
但 matching_bold_close 扫描时遇到未闭合的 '"' 后一直处于 in_str 状态,
再也找不到收尾 '**',于是 is_marqdo_star_line 返回 false,
该行被判为 LineKind::Comment 成为叙述,解析器根本看不到它。
解析层精心准备的报错不可达。

修复(两层,缺一不可):
1. src/lex/mod.rs
   - 新增 pub fn has_unterminated_string:统计未被 \ 转义的 '"',奇数即「开了没收」
   - is_marqdo_star_line:粗体起始且找不到收尾时,若引号未闭合则仍判 Code
     (判 Comment 会静默丢行;判 Code 才能让解析器说话)
2. src/parse/mod.rs
   - parse_simple_stmt 紧跟既有粗体分支之后,对「粗体起始 + 引号未闭合」
     给出精确报错,而不是笼统的 unrecognized statement

影响面(改前量化,避免凭直觉改核心解析):
- 全仓库 845 个 .mq.md、620 个含代码行文件
- 引号计数为奇数的代码行:**0**
- 结论:不改变任何现有文件的解析结果,只作用于畸形输入

测试(TDD,先写测试确认失败再修):
- 新增 4 个单测:未闭合→Code、闭合→Code、叙述含引号→Comment、
  叙述含孤立引号→Comment
- 修复前 unterminated_quote_inside_bold_is_code_not_comment 如期失败
- 修复后 lex 14/14;cargo test --lib 237 passed 0 failed(基线 233 + 新增 4)
- cargo test --test gold 276 passed 8 failed;git stash 验证基线同样 8 failed,
  失败全部是 ext_web_*(依赖 Go 插件 libweb,本机环境不具备),与本次无关

文档:
- doc/design/parser-diagnostics-d8-d9.md  设计依据、根因链路、候选方案评判、影响面
- doc/design/change-d9-unterminated-string.md  变更记录(现象/根因/改法/测试/边界)
- doc/design/spaced-function-names-d8.md  D8 方案评估(待审阅,未改代码)

已知边界:单引号 '...' 未闭合仍是旧行为,留作后续评估
(单引号兼作英文撇号,误判风险高于双引号)。
2026-10-22 11:52:27 +08:00
cflmy aeb5a2a43d docs(tutorial): 新增《常见陷阱手册》,并补 D8–D10、按实测收敛 D3
把 qdagent 重构期间踩到的坑整理成**使用向**手册 doc/tutorial/15-常见陷阱.md
(417 行,全部条目在 1.5.2 逐条实测,含现象/原因/改法三段式与提交前自查清单)。

手册要点:
- §1 写文档:行首数字必报错(2026 年 / 42 条 / 3 个要点 都中招);
  行首 + 报错;未闭合引号静默吞行(最危险)
- §2 变量作用域:模块级变量在 # main 与各 ## 之间不可见;
  顶层不许裸语句;table.put 不原地修改
- §3 函数:函数名含空格时定义成功但调用必崩
- §4 字符串:引号内 ** / | / # 实为安全,只有引号不配对才出事
- §5 导入:相对路径以当前文件为基准;标准库必须带库名前缀
- §6 易记错的函数签名(txt.contains 的 sub、table.put 的 in/at/value …)
- §7 结构:# main 体不被 ## 截断;1./2./3. 是 if/elif/else 链
- §8 自查清单(marqdo check + run 双跑)

同步修正缺陷实录 research/qdagent-refactor-defects-2026-10.md:
- D3 收敛:实测证明 - / > / | / [ 在叙述区**安全**,只有行首数字与 +
  会被当语句;判据是「该行能否被解析成完整语句」
- 新增 D8 函数名含空格:定义成功、调用必崩,check 却通过(不对称)
- 新增 D9 未闭合引号静默吞行,退出码 0(静默类最危险)
- 新增 D10 table.put 非原地,累积计数静默失效
- 同步复现清单与变更记录

索引:tutorial/README 增第 15 章;doc/README 增手册与缺陷实录两行。
手册内所有代码示例均实测跑通(含参数传递修法、if 链、标准库前缀)。
2026-10-22 09:46:17 +08:00
cflmy f2d651e063 docs(research): qdagent 重构期间发现的运行时缺陷实录(D1–D7)
在 qdagent 实现「智能体输出必须是可运行 .mq.md」的机械门禁
(提取 → 结构检查 → 沙箱试跑)时,暴露 Marqdo 1.5.2 的一批
行为与文档不一致、诊断不可用、静默降级问题,逐条记录并给出建议:

- D1 高 sandbox.profile 仅回显不生效:readonly 与 default 等价,
  RunOptions 硬编码 allow_fs_write=true(安全语义静默失效)
- D2 高 全角冒号 调用:<name> 无法解析,但提示文案推荐全角;
  extract_tool_name 用 split sep=":",与 READ/DECISION 的全半角支持不一致
- D3 中 叙述行行首执行标记(数字/-/>/|/[/**/*)导致中文文档解析失败;
  对 LLM 生成文档是高频陷阱,建议增强诊断与 check 规则
- D4 中 sandbox.run source= 报错丢上下文:文件名显示 <memory>,
  未提示「源文本必须含 # main」
- D5 中 [key](map) 键缺失硬崩:sandbox.run 成功形状无 error 键,
  按防御式写法在成功路径崩溃;建议统一形状
- D6 低 未绑定 GFM 表不可迭代,报错不指向根因
- D7 低 沙箱内 subtask spawn path= 设计性不可用,未在沙箱文档明示

附「确认正常的项」以避免误报,与逐条复现清单。
2026-10-22 09:03:07 +08:00
cflmy 6ed70fd8b8 docs(design): defer G7 and G8 from the v3 target
Keep the cross-document scope and non-Python/Node adapter designs in the document for future work, but remove them from the current implementation target and priorities.
2026-10-18 14:51:12 +08:00
cflmy 115d1f6c0b feat(view): explicit persist and nested-plan boundary
Add per-cell Persist backed by lib/writeback and /api/session/cell/persist. Reject nested foreign.session plans with a top-level diagnostic and tests.
2026-10-18 14:47:54 +08:00
cflmy 514caa55ae docs(design): mark G9 snapshot capture implemented
view capture and view output --snapshot are implemented with source-hash validation; remaining gaps are G3/G7/G8/G10.
2026-10-18 14:46:20 +08:00
cflmy 74908ebf04 feat(view): capture live snapshots for static export
Add view capture to record per-cell outputs and session summary with source hashes. view output --snapshot consumes the capture without re-executing and rejects source drift. Includes snapshot parsing tests.
2026-10-18 14:45:26 +08:00
cflmy 0f852593bc docs(design): record foreign v3 P0/P1 implementation status
Mark stdin/per-cell inputs/interrupt and plan/snapshot APIs as implemented; keep G3/G7/G8/G9/G10 as pending design items.
2026-10-18 14:44:29 +08:00
cflmy a7266185b2 docs(design): propose foreign v3 capability plan
Document the remaining foreign runtime gaps: persistent raw stdin, per-cell input, dynamic plan boundaries, runtime plan updates, interrupt, state export/replay, cross-document scope, non-Python/Node adapters, live-to-static snapshots and explicit writeback policy.
2026-10-18 14:31:44 +08:00
cflmy fb28eee62f feat(foreign): own view sessions and clear outputs
foreign.session returns a handle from ordinary kernel/cell tables; view consumes only the 会话 binding. Sessions live in host foreign_doc. Without 会话, view cells run stateless one-shot. Add /api/session/clear and a Clear outputs button for every view page; clearing keeps persistent kernel state.
2026-10-18 12:42:15 +08:00
cflmy f3791d11b4 refactor(view): evaluate cell plans through the interpreter
Delete the source-line table scanner. 内核表/会话表 are now evaluated as ordinary Marqdo bindings (with simple assignment dependencies) via Interpreter::capture_declarations. Every view page uses the same cell toolbar; run/watch semantics are unchanged; no new CLI or Markdown marker.
2026-10-18 12:17:42 +08:00
cflmy 7fef690167 refactor(view): unify cells into the view surface
Cell controls now appear only when a document declares 内核表/会话表. Remove Notebook branding and Jupyter In/Out prompts; ordinary view pages stay classic. There is no new CLI command or second interactive mode; run remains canonical execution and view is the only interactive surface.
2026-10-18 12:03:17 +08:00
cflmy 1415aee08f docs(notebook): record D1–D3 implementation status 2026-10-18 01:45:41 +08:00
cflmy 0f2c12b636 feat(notebook): link statement debugger from view
Expose /debug and /api/debug/* on the live view server; add a notebook Debug button so Marqdo statement stepping is reachable from the same page/session as cell stepping.
2026-10-18 01:44:05 +08:00
cflmy 9873df7099 feat(notebook): reload, journal and state panel polish
D3 (partial): reload re-parses the document while preserving kernels, marks changed cells dirty; Journal and Reload view actions; filtered Node variable names for the state panel; per-session timeout parsing before spawn.
2026-10-18 01:41:36 +08:00
cflmy 3001ed7ba5 feat(notebook): document session with shared kernels and view controls
D1/D2: 内核表/会话表 parsing, DocumentSession cell index, named/default kernels, shared state + isolation, run_cell/step/run_all/restart/assign, state bridge, journal, POST /api/session/*, and a live-view notebook bar with status badges, per-cell kernel selector, step/run-all and push/pull state form.
2026-10-18 01:39:36 +08:00
cflmy c21a28c3bc docs(foreign): design kernel runtime and document session
Record the Jupyter-beyond discussion, decisions, and the document-native notebook syntax (内核表/会话表). Split the design into F (foreign kernel runtime) and D (document session) layers, with roadmaps, state bridge, stepping, journal, and security rules.
2026-10-18 00:43:31 +08:00
cflmyandCursor 8353ba38fe Add Streamable HTTP MCP and harden web /v1 invoke auth.
agent_mcp_serve gains http transport; invoke injects Bearer/X-Api-Key and maps
unauthorized to 401; openai/agent field access prefers json.get over footnotes.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-16 17:33:24 +08:00
cflmyandCursor 1908bf1a53 fix(ai): soft LLM transport errors and strip ext/ai json glue
Host/plugin HTTP failures return structured finish|status=error instead of
aborting the VM; agent/llm/openai/智能体 result maps use GFM tables and
table.put. SKILL gains positive copy-paste examples so agents stop mirroring
legacy json.set chains.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-16 14:27:45 +08:00
cflmyandCursor d489f7c088 fix(ext/ai): cold agent.run EFI — self.preflight and resolve llm_free
Miss resolve omitted llm_free and run called bare preflight, which crashed
qdagent on empty OKF. Align miss fields, use self.preflight + json.get, add
regression gold and research note.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-16 12:54:00 +08:00
cflmyandCursor 0f72173d08 feat(ext/ai): land Agent/LLM optimize waves O1–O5
Wire opt-in sandbox trial into agent, require sandbox proof before OKF
promote, add step_route/EFI/cost surfaces, style lint + quarantine, and
offline policy optimize script—keeping GFM tables over json glue.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-16 11:47:28 +08:00
cflmyandCursor a707758d47 docs: research and plan Agent/LLM optimization with OKF and sandbox
Capture the 2026-10 landscape (skills, Skill Lift, routing, GEPA) against Marqdo’s code-as-docs assets and list feasible optimize directions without editing doc/next.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-16 11:13:48 +08:00
cflmyandCursor 63bcb882a6 feat(ext/ai): add CoW sandbox for Marqdo-only agent trials
Give agents a capability-capped run path that stages workspace edits until commit, so generated .mq.md can be verified without polluting the real tree.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-16 11:11:54 +08:00
cflmyandCursor 58ec2172d6 fix(view): mid-to-mid annot leaders and prose chip density
Unify frame chip ids with marks, reorder chips under anchors, and draw
SVG leaders from underline mid-bottom to chip mid-top so crowded bands
stay readable.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-15 20:36:12 +08:00
cflmyandCursor f47a1e143a feat(writeback): image form ![marqdo-out](…) with view parity
Auto-persist image paths/URLs as Markdown images (validated extension);
keep HTML comment for text. View and static export share styled source
markers and output-card <img> rendering.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-15 15:45:52 +08:00
cflmyandCursor 9cf98a98b5 feat: block formula meta bind and control-flow prose
Support $$ % marqdo: var= binding, arm/loop narrative (fix comment-paragraph
swallow of branch arms), foreach ZH/EN sugar, and condition-line narrative.
Welcome pages and gold samples updated; view skips meta formula fence bodies.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-15 15:28:46 +08:00
cflmyandCursor 7d49c358c1 fix(view): nested-call welcome + clean formula/code cards
Welcome pages put greet calls in # main so CallFrames/pickers appear;
document surface skips fence body lines and Bind chips under formula/code cards.
Adds view-static-interactive design and entry-vs-method regression tests.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-15 14:13:42 +08:00
cflmyandCursor 4e2419b598 feat(view): document-surface Structure with frame-aware annotations
Browse Structure keeps narration and surfaces post-stmt values, branch
arms, and call/loop pickers; static `view output` shares the same render
path so exported HTML matches live view.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-15 13:12:53 +08:00
cflmyandCursor 56ab52356f feat(knowledge): land Engineering Knowledge Compiler (EKC) from 008
Give agents Engineering First Information before codegen: L0–L4 catalog,
reuse/preflight (REUSE→ADAPT→CREATE), graph, and CLI surfaces so large
repos stop reinventing existing capabilities.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-28 16:44:39 +08:00
cflmyandCursor ff252fd469 feat(web): land Artifact serve/endpoint path and retire compose golds
Wire data_source into Document render, register Endpoint handlers via
run_artifact, restore marqdo-blog as an Artifact site, and replace old
compose_* web golds with ADR 0007 fixtures and docs.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-28 11:36:45 +08:00
cflmyandCursor 67bac1bf02 feat(web): Document/Endpoint/Resource artifact rewrite (ADR 0007)
Replace the monolithic compose_*/configure author surface with file-level
web|endpoint artifacts, split ext/{web,data,security,net}, and lock
no-degrade rules so the model cannot regress to God-object APIs.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-28 11:06:38 +08:00
cflmyandCursor 0ed57b6d54 feat(agent): ext APIs, goldens, and docs for MCP/resume/small-llm routing
Wire mcp_connect/list/call, resume_*, and demote Jev in author docs.
Offline goldens cover MCP client, resume, and small-llm/jev-alias route.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 20:26:55 +08:00
cflmyandCursor 2d412e3bba feat(llm): LLMResult.tool_calls from OpenAI-shaped responses
Parse choices[0].message.tool_calls into the ask Result; empty list when
absent. Offline gold covers tool_calls_from.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 20:07:59 +08:00
cflmyandCursor 46d78c6e15 feat(core): Phase 2 Metadata scope lift (no body ${})
Inject bound frontmatter keys into the entry env as ordinary variables.
Reject body ${env…}; bytecode entry prologue seeds the same locals.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 20:05:25 +08:00
cflmyandCursor 81a14f9b38 feat(llm): sys.module_source + entry prompt / meta offline paths
Expose entry source via sys.module_source; ask/stream fall back to
prompt_body of the entry document. Add meta-binding and entry-prompt golds.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 19:52:57 +08:00
cflmyandCursor 88dc6d5a54 feat(core): Artifact Metadata Binding (${env|arg|sys|secret})
Load-time resolve of frontmatter bindings into Module.metadata; Secret
masking; sys.meta / --bind; llm ctor prefers meta over env_get.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 19:51:24 +08:00
cflmyandCursor 4a2d89406b fix(agent): preserve llm_free resources under plan force
Stop plan from clobbering an existing resources workbook with an explore
skeleton. Add llm.prompt helpers and agent-plan-preserve gold.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 19:34:05 +08:00
cflmyandCursor d8d435fead feat(ai-stack): metrics, named handles, prompt docs, backends, zero-LLM
Finish ai-stack steps 1–5: model_turn usage, fast/reasoning, type:prompt
load, openai/ollama split, and agent-zero-llm gold. Fix compiled-skill
Marqdo syntax so llm_free await no longer falls through to plan+HTTP.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 19:30:25 +08:00
cflmyandCursor 1f8e61f3bf feat(llm): semantic ask/stream/collect surface (005)
Lift ext/ai/llm from OpenAI HTTP client to an intelligence primitive;
keep complete as compat. Add unified ai-stack roadmap linking agent + llm.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 13:35:38 +08:00
cflmyandCursor a9ee858f53 feat(agent): Skill Compilation + Adaptive Routing (v2/P4)
Episode → maybe_learn → llm_free skills, plus abstract router backends
(marqdo/small-llm/jev?/llm/auto) via plugins/agent; no core host changes.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-24 13:28:50 +08:00
cflmyandCursor cba51238b6 doc/next/002: 智能体框架「认知编译」构想(Learning → Compilation)
用户维护文档归档:每次执行沉淀 mq.md;>3 次调用自动总结优化出记忆文档;
完全学会后降能耗运行(少 LLM 思考 / Jev 决策模型 / 纯分支控制)。

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-23 13:19:56 +08:00
cflmyandCursor 5fffb800f9 P4(b) 收口:mimo-v2.6-pro 两跑均端点 infra 不可评估,判据以 deepseek-flash 为准
- perf-validation-report §3 记录重跑终局(12/12→11/12 infra),(b) 判据判定不改
- three-problems-plan Phase 4 落地记录同步

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-23 13:10:28 +08:00
cflmyandCursor 1a729eb9ec P4 验证执行:(a)锚点消融✅/(c)契约消融✅/(d)防漂移✅/(e)守卫演练✅;(b)盲测 deepseek-flash 完成未达标、mimo-v2.6-pro 重跑中
- perf-validation.md:五实验协议/判据/材料(2026-09-23 自设计)
- perf-validation-report-2026-09-23.md:汇总报告(判据逐项对照 + 诚实结论)
- tests/blind_experiment.rs + tests/blind/:盲测 harness(查询通道/修复回路/usage 计量)+ 冻结材料(6 任务书 + docs-bundle)
- tests/contract_ablation.rs:四边界×三情形消融(拦截 100%/误报 0%/零回归)
- tests/drift_drill.rs:9 漂移必报 + 2 负对照
- src/mlsp.rs:syntax 卡片自持化(semantics/rule/sample/doc_quote)
- three-problems.md §5.1 判据矩阵逐行判定(问题 1✅ 2✅ 3❌);three-problems-plan.md Phase 4 落地记录
- 全量干净回归零失败

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-23 05:06:30 +08:00
cflmyandCursor 00c763300e test(repair): mimo-v2.6-flash cross-check + measurement-side hardening
mimo-v2.6-flash 终测:anchor 臂一轮 20/20=100%、plain 90%(最终 95%);
三轮双臂一轮修复率全部 >=85%(目标 >=80%),护栏 0 越界。

'测试不稳定'定性为三类测量端问题并全部加固(与 Marqdo 语义无关):
- 网关空体抖动 -> curl --retry + 5s 退避(加固后 0 次失败)
- 推理模型 reasoning_content 吃光 token 预算 -> max_tokens 2000->8000
- 模型输出格式畸形 -> parse_edits 容错(数组/单对象/NDJSON 多对象流/
  字符串化元素/字符串型行号/围栏/数字后多余引号的确定性修补 unquote_line_nums)

报告按模型分文件(repair-experiment-<date>-<model>.md),cflmy/mimo 双报告并存;
计划记录更新为三轮对照表。干净环境 cargo test 444/444 全绿(零回归);
ext_llm_*/ext_agent_*_live 为环境开关型实况测试,与本期无关。

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-23 01:27:01 +08:00
cflmyandCursor 224b3d4a66 test(repair): §5.2(a) 模型在环修复实验 —— 锚点消融:anchor 100%(20/20) vs plain 85%,≥80% 目标达成
- tests/repair_experiment.rs:真实消费 validate → repair_targets → 模型行编辑 → repair_apply(越界必拒)
  → 复验 的完整闭环(≤2 轮;容错重试;cflmy @ llm.cflmy.cn,.env 读取,#[ignore] 不进常规回归)
- 双臂消融:anchor(诊断带 doc_anchor/doc_quote/suggestion)vs plain(剥掉)——
  anchor 一轮修复 20/20=100%;plain 17/20=85%(3 例失败均为网络空体抖动)
- tests/common/mod.rs:golden 20 例与诊断/应用共享件抽出(repair_loop + experiment 同源)
- parse_edits 容错(数组/单对象/字符串化元素/围栏);越界必拒/abstain 由 repair_loop 硬断言覆盖
- 报告:doc/roadmap/repair-experiment-2026-09-22.md;.gitignore 补 .env(密钥不入库)

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-22 21:10:03 +08:00
cflmyandCursor 81d26a265a docs(roadmap): 落地记录 —— T3.4 修复回路工程项完成(护栏 + golden 20 例);T3.3 编号归位
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-22 20:14:32 +08:00
cflmyandCursor 18d89d6cf2 docs(roadmap): 落地记录更新 —— T3.4 完成;T3.3 golden 实验待做
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-22 19:47:19 +08:00