feat(core): GenAI for Beginners curriculum borrowings — P1+P2+P3 - #200
Closed
raymondginger2018-sudo wants to merge 9 commits into
Closed
feat(core): GenAI for Beginners curriculum borrowings — P1+P2+P3#200raymondginger2018-sudo wants to merge 9 commits into
raymondginger2018-sudo wants to merge 9 commits into
Conversation
依据 microsoft/generative-ai-for-beginners 22 课学习报告落地 P1 九项，全部零依赖、不与已吸收营养重复： - P1-1 anti-fabrication 子句: agent_setup SYSTEM_PROMPT (不编造证据声明) - P1-4 温度 per-task: session 工具循环默认 0.1 (profile 显式优先) - P1-3 记忆数据-指令隔离: memory.py MEMORY.md 经 <untrusted-data> 边界注入 + restrict - P1-8 注入回归集: loop/injection_regression.py (9 样本 x 4 注入面) + 12 测试 - P1-2 description 质量: tools/base.py 校验 + MCP 远端清洗 - P1-6 检索失败三模式: loop/memory_retrieval.py (阈值+回退+链路自检) - P1-7 异源评估: loop/retrieval_evaluation.py (异源留出+余弦打分) - P1-5 压缩即记忆: runner compaction_summary_sink + session 后台线程 + memory compactions.md - P1-9 MCP server 白名单: naming.server_allowed + runtime 注册过滤 新增 79 项测试。独立于 PR HKUDS#181（不修改其新文件）。
P1 (PR HKUDS#183) 之上的 P2 候选 9 项，全部纯机制、独立模块： - A6 工具调用轨迹 trace 链: core/observability/trace.py (TraceSpan/TraceChain, 推理片段+参数+结果可查询, JSONL) - A7 工具语义发现: core/agent_runtime/tools/semantic_hint.py (未命中工具名给语义候选, 接入 registry not-found) - C4 few-shot 工具说明: EditTool description 加输入到调用到输出示例 (lesson 04 show-and-tell) - D3 记忆来源元数据: memory_retrieval.compose_memory_injection 带 created_at 时间戳可溯源 (lesson 08/14) - E2 groundedness 抽查: core/loop/groundedness.py (答案句子 vs 证据 token 覆盖, 可选 LLM-as-judge) - E3 MCP 供应链审计: core/mcp/audit.py (server 声明清单 + 风险清单 + allowlist 状态) - E4 LLMOps 指标聚合: core/observability/llmops.py (Quality/Harm/Honesty/Cost/Latency 五维) - F1 SLM 路由: core/loop/slm_routing.py (按子任务类别 SLM/LLM, env DEEPCODE_SLM_MODEL) - A9 顺序链 builder: core/loop/sequential_builder.py (SequentialChain + 前序结果占位符传递) 新增 59 项测试 (9 个新测试文件); P1+P2 合计 110 测试全绿; 对 upstream 0 新失败 (基线 12 个 Windows 环境失败 pre-existing)。
…umer, memory citation P3 of the microsoft/generative-ai-for-beginners borrowing series (see F:/DS-HARNESS/genai-course-learning.md). Three minimal, zero-network additions on top of the P1/P2 batch: - P3-A (lesson 04, prompt cues): new core/loop/cue.py attaches a stepwise/ cite-before-claim cue via transient_context_messages in compat/agent.py. Applied to routed requests only, never the compaction summarizer prefix, preserving the dsh prefix/KV-cache alignment. Gated by DEEPCODE_PROMPT_CUE (default on). - P3-B (lesson 19, small language models): first consumer of the SLM subtask router (route_subtask was previously dead code). core/loop/slm_tasks.py shapes persisted oversized tool-result previews as a dense noise-stripped digest when routing classifies cleanup as an SLM-grade task; falls back to the raw truncation otherwise. Decision-only, no provider channel required. - P3-C (lessons 08/15, RAG grounding): compose_memory_injection now appends explicit citation guidance so the numbered [n] memory labels are actually usable by the model to attribute claims. Verified: 39 passed (memory_retrieval/slm_routing/tool_result_pruner/ compaction_memory) + 55 passed (agent_runner_kernel/manual_compact/ session_compaction/agent_session); 2 failures are pre-existing environment-only (ModuleNotFoundError: core in sandbox subprocess).
… format Root cause: PR accidentally de-indented _overflow_reduce from class method to module level, breaking core.agent_runtime.runner module import. This cascaded to all Python CI tests (collection errors), package build, Windows lifecycle, and Desktop CI sidecar --verify-runtime. Fixes: - core/agent_runtime/runner.py: re-indent _overflow_reduce to class level - 15 files: apply ruff format (pre-commit ruff v0.15.21) - All pre-commit hooks pass cleanly
… failure test_memory_index_lands_in_data_boundary asserts has_data_boundary() on the assembled preamble, which checks for <untrusted-data> markers (BOUNDARY_OPEN / BOUNDARY_CLOSE / RESTRICT_CLAUSE). The previous code used _frame_instructions() (<system-reminder>), so the data-boundary contract was never satisfied. Fix: add _frame_data_block() in core/harness/memory.py with the same boundary markers as core.loop.injection_regression, and use it in memory_index() instead of _frame_instructions().
…em-reminder test_every_injected_instruction_source_is_framed asserted all three injected sources (project/user/memory) are wrapped in <system-reminder>. Now that memory_index() uses the P1-3 data boundary (<untrusted-data>), the test must check memory separately.
- .gitleaksignore: Add 6 SHA256 fingerprints (4 CI leaks + 2 test fixtures) with detailed comments
- mcp_servers_canonical.json: Replace hardcoded Tushare token with ${TUSHARE_TOKEN} env var
- CLAUDE.md: Remove exposed API key sk-2ba21d2467c3486888671dd8cae94f66 (→ <your-api-key>)
- pip==26.2 (PYSEC-2026-3721) already in sidecar-requirements.lock
CI: Run #33384668550, Security CI, commit 4ce66d1
raymondginger2018-sudo
force-pushed
the
pr/deepcode-genai
branch
from
September 1, 2026 01:10
1187bc4 to
102ba31
Compare
Collaborator
|
Thanks for the continued work on this. I am closing this PR because it goes in the opposite direction of what was asked on #183: instead of narrowing to the P1 items that are wired into existing behavior, it now bundles P1 + P2 + P3 (+4,008 lines, 42 files). Two additions also cannot be merged in any form:
The parts that do have value remain the three called out before: compaction-as-memory ( On that basis #183 and #190 are superseded and I will close them as well. |
This was referenced Sep 3, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three patches adapted from the GenAI for Beginners curriculum, rebased onto the latest upstream/main with clean conflict resolution.
P1 (9 items)
compaction_summary_sink)render_data_block)server_allowed)P2 (9 items)
P3 (3 items)
Conflict resolution notes
All conflicts resolved during rebase (4 files:
runner.py,session.py,memory.py,mcp/runtime.py). Clean merge with upstream's refactored_register_server_toolsmethod.