SW-05 p3 fragment 组成数据化：窗口/投影/Thread 注入进 profile.context（ADR-0017） - #13
Merged
Conversation
There was a problem hiding this comment.
🟡 Changes recommended
There is a confirmed config semantics bug (known_fact_fields: [] cannot be expressed) and an unresolved/contradictory definition of the window ordering (“near-first” vs current join order) across code/tests/ADR.
Once you've addressed the issues Copilot identified, you can request another Copilot review.
Pull request overview
This PR makes p3 fragment cross-episode context composition configurable via profile.context.* (window size for prev_episode_summary, projection fields for known_facts, and an inject_threads switch), with defaults intended to preserve existing behavior, and adds schema/ADR/tests to formalize the change.
Changes:
- Add
ContextSettingstoprofiles/_schema.pyand setProfile.contextwith defaults + whitelist validation. - Update
pipeline.py(run + recompile) to build p3 fragments usingcontext.prev_summary_window,context.known_fact_fields, andcontext.inject_threads. - Add end-to-end stubbed tests covering default behavior, window>1 behavior, and threads injection forwarding.
File summaries
| File | Description |
|---|---|
| tests/test_p3_context_config.py | Adds regression + configuration tests for known_facts projection, prev_summary windowing, and threads injection. |
| src/nsc/passes/pipeline.py | Implements the configurable window/projection/injection logic and helper utilities (_window_join, _threads_view, etc.). |
| src/nsc/passes/p3_beatsheet.py | Forwards injected threads from fragment into the LLM inputs when present. |
| profiles/short_video_v1.yaml | Adds explicit default context block (intended zero behavior change). |
| profiles/short_drama_v1.yaml | Adds explicit default context block (intended zero behavior change). |
| profiles/_schema.py | Introduces ContextSettings with extra=forbid and known_fact_fields whitelist validation. |
| adr/0017-p3-context-profile-section.md | Documents the decision and intended semantics for the new profile.context section. |
Review details
- Files reviewed: 7/7 changed files
- Comments generated: 4
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Comment on lines
+650
to
+653
| def _known_fact_fields_of(profile: dict[str, Any]) -> tuple[str, ...]: | ||
| """profile.context.known_fact_fields → 投影字段（白名单交集，缺省 = 原五字段）。""" | ||
| wanted = profile.get("context", {}).get("known_fact_fields") or list(_KNOWN_FACT_FIELDS) | ||
| return tuple(k for k in _KNOWN_FACT_FIELDS if k in wanted) |
Owner
Author
There was a problem hiding this comment.
已修正（10227b0）：_known_fact_fields_of 改为 is None 判定——显式 known_fact_fields: [] 现在是合法的空投影（有意隐藏全部字段），只有键缺失才回退五字段缺省。补了回归断言（tests/test_p3_context_config.py::test_known_facts_fields_from_profile）。
Comment on lines
+670
to
+676
| def _window_join(summaries: list[str], window: int) -> str: | ||
| """prev_episode_summary 窗口（SW-05）：近端在前、远端追加在后，逐行拼接。 | ||
|
|
||
| window=1 与原行为逐字节一致（单元素直接返回）；window=0 即恒空串。 | ||
| """ | ||
| n = max(0, int(window)) | ||
| return "\n".join(summaries[-n:]) if n else "" |
Owner
Author
There was a problem hiding this comment.
已修正（10227b0）：确认预期语义是时间序（远端在前、近端在后），与 compress_history 的【前情】→【上一集】布局一致——实现本身正确，初稿 docstring 表述反了。已统一 docstring、ADR-0017、测试注释三处表述；断言不变（窗口扩展只在前端追加远端集）。
Comment on lines
+18
to
+22
| | 键 | 语义 | 缺省 | 消费点 | | ||
| |---|---|---|---| | ||
| | `context.prev_summary_window` | p3 `prev_episode_summary` 看近端 N 集（0=恒空） | 1 | `run_pipeline` p3 循环 + `recompile_episode`（`_window_join`：近端在前、远端逐行追加） | | ||
| | `context.known_fact_fields` | `known_facts` 投影字段（白名单子集） | `[id, content, episode_no, status, type]` | `_known_facts` | | ||
| | `context.inject_threads` | 是否把 p2 的 Thread 表注入 p3（投影 `{id,title,status,state}`） | false | `run_pipeline`/`recompile_episode` 组装 fragment，p3 转发进 LLM 输入 | |
Owner
Author
There was a problem hiding this comment.
已修正（10227b0）：ADR 表述改为"按时间序逐行拼接，远端在前、近端在后，与 compress_history 布局一致"，并在约束一节补记 review 的两处澄清（显式空投影合法、窗口时间序）。
Comment on lines
+113
to
+115
| # 窗口=2：第 3 集的窗口严格包含第 2 集的窗口（多了第 1 集摘要） | ||
| assert summaries[2].startswith(summaries[1]), "近端摘要在前，远端追加在后" | ||
| assert len(summaries[2]) > len(summaries[1]) |
Owner
Author
There was a problem hiding this comment.
已修正（10227b0）：注释改为"时间序拼接（远端在前、近端在后，同 compress_history 布局）——第 3 集的窗口以第 1 集摘要开头、第 2 集摘要收尾"。断言本身（startswith + 长度递增）在新旧表述下都成立于时间序实现，未改。
12 tasks
randypanding
added a commit
that referenced
this pull request
Aug 21, 2026
Copilot review（PR #13）： - _known_fact_fields_of 用 or 判断，显式 [] 与未配置混同 → 改为 is None 判定 - _window_join 实现是时间序（远端在前），初稿 docstring/ADR/测试注释写成 "近端在前" → 统一改为时间序表述（与 compress_history【前情】→【上一集】布局一致）
randypanding
enabled auto-merge (squash)
August 22, 2026 04:32
randypanding
force-pushed
the
sw/sw-05-p3-context-config
branch
3 times, most recently
from
August 22, 2026 04:59
f9d4f85 to
b4bd1a7
Compare
randypanding
force-pushed
the
sw/sw-05-p3-context-config
branch
from
August 22, 2026 05:00
b4bd1a7 to
85cff88
Compare
randypanding
added a commit
that referenced
this pull request
Aug 26, 2026
* SW-05 p3 fragment 组成数据化：窗口/投影/Thread 注入进 profile.context（ADR-0017）
- context.prev_summary_window（缺省 1=原行为）：run_pipeline 与 recompile_episode
统一走 _window_join（近端在前、远端逐行追加；window=1 逐字节同原实现）
- context.known_fact_fields（缺省原五字段，白名单校验）：_known_facts 投影可窄化
- context.inject_threads（缺省 false）：p2 Thread 表按 {id,title,status,state}
投影注入 p3 fragment，p3 转发进 LLM 输入
- profiles/_schema.py 新增 ContextSettings（extra=forbid + field_validator）
- 测试：tests/test_p3_context_config.py（5 例，先红后绿）
* SW-05 review 修正：known_fact_fields 显式空列表不再回退缺省；窗口序表述改为时间序
Copilot review（PR #13）：
- _known_fact_fields_of 用 or 判断，显式 [] 与未配置混同 → 改为 is None 判定
- _window_join 实现是时间序（远端在前），初稿 docstring/ADR/测试注释写成
"近端在前" → 统一改为时间序表述（与 compress_history【前情】→【上一集】布局一致）
* sync: 远端 main 快照(API tarball;解决 fetch 阻断)
* style: ruff format(union 解冲突文件)
* lab round10: p4 beat_to_scene 类型矫正(随机后端结构漂移)+lab_smoke_v1 三集评测切片+p3 指令 v2(round8/9:escalation/枚举/PENDING/钩子纪律)
* lab round10b: _to_int 宽容转换(字符串下标/嵌套值,TypeError 实证)
* lab round12: resolve_pending 降级语义(无 donor 解除契约不致命;随机后端 PENDING 悬空最高频死法实证)
* lab round12: resolve_pending 降级语义(无 donor 解除契约不致命)+修正最终返回输入;测试 8 绿
* lab round12b: p1 Prop sku_ref null 归一(NarrativeIR ValidationError 实证)
* lab round12c: 通用 null→字段默认值归一(characters/locations/motifs 同型错误一例多杀)
* lab round13b: present_character_ids 空表全集兜底(scenes.N=[] ValidationError 实证)
* lab round14: p3/p4 结构机械修复——STR-014 承重节拍兜底/BM-002 植入间距重排/STR-010 主角补位(attempt4/5 同门连死实证,相位重试只复述诊断不改结构)
* lab round15: est_duration_s 等比缩放到集目标时长(DLG-006 六集全灭根因)+_retry_pass 传输容错(APIConnectionError 杀死整轮实证)
* lab round16: p5 对白目标区间与 DLG-006 对齐(旧 lo=0.8x 全顺从也死)+_expand_if_thin 欠量当场定点扩写(系统性欠量 26% 实证)
* lab round16b: 对白瞄准线=门禁线+3pp(毫厘之死实证 341/342/344 vs 344.25)+扩写循环至达标(上限2次,只留更厚稿)
* lab round17: p6 prompt 瘦身(_slim_scenes/_slim_profile/_slim_bible_for_episode 投影,实证 46631 字符撞护栏;id 全保留 anchor_map 契约不动)
* lab round18: 暗线步进钳制(INV-19 机械前置,实证 8 章全产物死于 final 门 current_stage 5/7 超界)
* lab round19: CMP-001 绝对化用语机械替换(门禁 fix 要求的确定性执行)+p1 必填 str 空串占位(characters.4.need 实证)
* lab round20: 南浪仔(海南文旅 IP)brief+brand 资产(用户定稿故事五幕:雨林→浪尖→旅程→暗线→陪伴)
* lab round20b: DLG-006 瞄准余量 +3pp→+6pp+扩写上限 2→3(ep8 差 12 字实证)
* lab round20c: canonical_name 改为'南浪仔'(BM-009 系统性失败实证:slogan 本身就违规,'南浪仔 NOLAN'全名在叙事中永不自然;NOLAN 单用降级为偶发可重试)
* lab round20d: ir_io.save 用 mode='json'(datetime 序列化,实证全绿产物死于 ir.json 导出)
* test 修正:Project 必填字段补齐(provenance_id/logline)
* docs: README 快速开始修正(实际 CLI 语法)+独立使用说明(端点配置/新品牌/profile/测试/战役分支)
* lab round23: 战役实测版 prompts 入库(p3 v3.2/p5 round13,全绿产物复现所需;B1 生成物例外,PR 内声明)+ADR-0015/16/17 状态 accepted
* merge 收尾:ruff 全绿(pairwise/values()/去过期 noqa)
* merge 收尾2:ruff format+check 全仓合规(CI lint & typecheck 门禁)
* merge 收尾3:pyright 合规(测试替身 cast PassContext)+format
---------
Co-authored-by: sync <t@t>
Co-authored-by: cnb <cnb@example.com>
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
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.
这个 PR 改的是哪一层？（必选其一）
spec/profiles/brands/cases/export/）→ 必须打标签asset-change并附 ADRsrc/tests/）→ 无需 ADRADR：adr/0017-p3-context-profile-section.md（status: proposed，含 profiles/_schema.py 变更，等确认）
（编号说明：ADR-0016 已被 sw/sw-07 分支占用，本卡顺延 0017）
工单
Closes # · 工单号：SW-05（上游依赖卡，Lab 仓 docs/WORK_ORDERS.md §上游依赖卡）
问题
p3 的跨集上下文组成硬编码在
pipeline.py：prev_episode_summary窗口恒 1、known_facts投影恒五字段、p2 的threads表永不注入。弱/强模型需要不同窗口与投影面。改动
context.prev_summary_window_window_join：近端在前、远端逐行追加；window=1 与原实现逐字节一致；recompile 同享context.known_fact_fieldscontext.inject_threads{id,title,status,state}注入 fragment，p3 转发进 LLM 输入（同 revivable_ideas 模式）profiles/_schema.py新增ContextSettings（extra=forbid）；两个在库 profile 显式写入缺省值，行为零变化验收命令（贴出你本地跑通的输出）
检查表
prompts/影响生成结果吗？