Quality phase02 I2: transcript-authoritative context (Native #41, Backend #115/#118, Squad #49) #286

Merged
weichen merged 17 commits from feature/agent-quality-phase02 into main 2026-10-09 17:36:03 +00:00
Owner

摘要

P2 上下文保真 · 增量 I2 的 Dev 组合 PR(P2 收口)。组件均已合入:
Agent-Native #41(思考捕获/回放)、Backend #115(转录为权威上下文)、
Backend #118 + Agent-Squad #49(嵌套指针对齐)。I1 已随 Dev #280 合入。

I2 内容:首轮整段组装一次性播种内核会话转录;后续轮只传当前用户消息,完整
工具结果/参数/附件原文由转录承载;结构分歧(排除/变体/重新生成,指纹判定)
→ 重播种(等价映射,单测覆盖);动态上下文(尾注/环境/新命中模块)经 Reminders
逐轮刷新;接近窗口触发 9 段压缩;MiniMax 思考随转录持久并回放。

本 PR 另含:架构棘轮补 runmgr→cmdbus 边(Backend #114 遗漏的 Dev 侧更新,
main 此前组合即红;本 PR 修复)。

真实模型三方对照(同 golden 集,plan-followup 3 轮)

指标 P1 后 I1 后 I2 后
工具调用 78 42 29
工具失败 13 10 7
检查失败 7 1 2
未核实数字 13 0 4(单轮方差)

pm-tool 检查失败 4→0→0;终稿声称缺章节 0→3→0。读原件归零属设计使然
(原件全文已在转录上下文)。三份报告:.task/phase02-eval/。

Evidence

  • make test(组合,同步后最终 head):PASS(evidence 行,SELF)
  • 组件 PR CI 全绿(GATE);真实模型对照 evidence 行 + 报告入库

剩余(P4/P5 另立卡)

关键词评分死循环、composeFinal 漂移、验证器解析——裁定 4 处理。

## 摘要 P2 上下文保真 · 增量 I2 的 Dev 组合 PR(P2 收口)。组件均已合入: **Agent-Native #41**(思考捕获/回放)、**Backend #115**(转录为权威上下文)、 **Backend #118 + Agent-Squad #49**(嵌套指针对齐)。I1 已随 Dev #280 合入。 I2 内容:首轮整段组装一次性播种内核会话转录;后续轮只传当前用户消息,完整 工具结果/参数/附件原文由转录承载;结构分歧(排除/变体/重新生成,指纹判定) → 重播种(等价映射,单测覆盖);动态上下文(尾注/环境/新命中模块)经 Reminders 逐轮刷新;接近窗口触发 9 段压缩;MiniMax 思考随转录持久并回放。 本 PR 另含:架构棘轮补 runmgr→cmdbus 边(Backend #114 遗漏的 Dev 侧更新, main 此前组合即红;本 PR 修复)。 ## 真实模型三方对照(同 golden 集,plan-followup 3 轮) | 指标 | P1 后 | I1 后 | **I2 后** | |---|---|---|---| | 工具调用 | 78 | 42 | **29** | | 工具失败 | 13 | 10 | **7** | | 检查失败 | 7 | 1 | 2 | | 未核实数字 | 13 | 0 | 4(单轮方差) | pm-tool 检查失败 4→0→**0**;终稿声称缺章节 0→3→**0**。读原件归零属设计使然 (原件全文已在转录上下文)。三份报告:`.task/phase02-eval/`。 ## Evidence - make test(组合,同步后最终 head):PASS(evidence 行,SELF) - 组件 PR CI 全绿(GATE);真实模型对照 evidence 行 + 报告入库 ## 剩余(P4/P5 另立卡) 关键词评分死循环、composeFinal 漂移、验证器解析——裁定 4 处理。
weichen added 15 commits 2026-10-09 17:12:56 +00:00
Why: I2 switches the direct path to the kernel transcript as the authoritative
context; the backend head carries that switch plus the nested kernel pin.

Evidence: SELF - gitlinks equal the pushed branch heads; combined make test
recorded on the task card after this commit.

Impact: Pointer only.
Why: Backend #115 merged the transcript-authoritative switch; the Dev
combination must test the merged head (with main's concurrent EXT-14).

Evidence: SELF - gitlink equals backend origin/main head.

Impact: Pointer only.
Why: include the current integration baseline before verification.

Evidence: fetched origin/main=6c763cc1cf4b3e980835de0ad864220f97f3c93c.

Impact: integrate upstream changes without rewriting task history.

Refs: task-sync
Why: bind the task validation checkpoint to the latest integration baseline.

Evidence: SELF recursive component checks; sync_base=6c763cc1cf4b3e980835de0ad864220f97f3c93c; candidate=2b46d055a09ae285c9b42cbd6dd97dd21ca448b0.

Impact: task metadata only.

Refs: .task/feature-agent-quality-phase02.md
Why: Squad #49 aligned its nested agent-native pin with the Dev top-level
pointer; the Dev worktree records the merged head.

Evidence: SELF - gitlink equals Agent-Squad main head.

Impact: Pointer only.
Why: Backend #115 (transcript-authoritative switch) and #118 (squad pin
alignment) merged; the Dev combination tests the merged head.

Evidence: SELF - gitlink equals backend origin/main head; make test recorded
on the task card after this commit.

Impact: Pointer only.
Why: Backend #114 (EXT-14, commands-as-tools) added the runmgr -> cmdbus
production import but its Dev-side ratchet update went missing, so make test
fails on any Dev head combining it (backend CI does not run the Dev
architecture check). Recorded here with the merge that surfaces it.

Evidence: SELF - lint-boundaries passes after the entry; the edge exists in
backend main (internal/runmgr/runmgr.go imports cmdbus).

Impact: Ratchet baseline only; unblocks the Dev combination for all in-flight
work.
Why: include the current integration baseline before verification.

Evidence: fetched origin/main=13e03ab15d89d13e0b1cc6848d76776c723a6f96.

Impact: integrate upstream changes without rewriting task history.

Refs: task-sync
Why: bind the task validation checkpoint to the latest integration baseline.

Evidence: SELF recursive component checks; sync_base=13e03ab15d89d13e0b1cc6848d76776c723a6f96; candidate=6caac63ec0b979df9ff692df2fe3b8f88d32f8ad.

Impact: task metadata only.

Refs: .task/feature-agent-quality-phase02.md
weichen added 8 commits 2026-10-09 17:12:57 +00:00
Why: Squad #49 aligned its nested agent-native pin with the Dev top-level
pointer; the Dev worktree records the merged head.

Evidence: SELF - gitlink equals Agent-Squad main head.

Impact: Pointer only.
Why: Backend #115 (transcript-authoritative switch) and #118 (squad pin
alignment) merged; the Dev combination tests the merged head.

Evidence: SELF - gitlink equals backend origin/main head; make test recorded
on the task card after this commit.

Impact: Pointer only.
Why: Backend #114 (EXT-14, commands-as-tools) added the runmgr -> cmdbus
production import but its Dev-side ratchet update went missing, so make test
fails on any Dev head combining it (backend CI does not run the Dev
architecture check). Recorded here with the merge that surfaces it.

Evidence: SELF - lint-boundaries passes after the entry; the edge exists in
backend main (internal/runmgr/runmgr.go imports cmdbus).

Impact: Ratchet baseline only; unblocks the Dev combination for all in-flight
work.
Why: include the current integration baseline before verification.

Evidence: fetched origin/main=13e03ab15d89d13e0b1cc6848d76776c723a6f96.

Impact: integrate upstream changes without rewriting task history.

Refs: task-sync
Why: bind the task validation checkpoint to the latest integration baseline.

Evidence: SELF recursive component checks; sync_base=13e03ab15d89d13e0b1cc6848d76776c723a6f96; candidate=6caac63ec0b979df9ff692df2fe3b8f88d32f8ad.

Impact: task metadata only.

Refs: .task/feature-agent-quality-phase02.md
weichen added 2 commits 2026-10-09 17:31:44 +00:00
Why: include the current integration baseline before verification.

Evidence: fetched origin/main=7d52861f15f2b25dd5121301ac1e92b5a5e16f8d.

Impact: integrate upstream changes without rewriting task history.

Refs: task-sync
Record the sync checkpoint
ci / test (pull_request) Successful in 2m13s
fe0333e44f
Why: bind the task validation checkpoint to the latest integration baseline.

Evidence: SELF recursive component checks; sync_base=7d52861f15f2b25dd5121301ac1e92b5a5e16f8d; candidate=1164b96838753528ae55eb8dff87729da05f9815.

Impact: task metadata only.

Refs: .task/feature-agent-quality-phase02.md
weichen merged commit 911bec70a4 into main 2026-10-09 17:36:03 +00:00
Sign in to join this conversation.
No Reviewers
No labels
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: weichen/Agent-System-Dev#286