fix(recall): judge native Stop replies from session recall evidence - #1374

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

… sends the reply (#1396)

Claude Code sends last_assistant_message in Stop along with transcript_path.
Since #1374 the reply took precedence, so a Claude session was judged from the
recall log instead of its transcript, and a doc the transcript shows but no
settled run of the session records was no longer judged.

Use the transcript whenever it shows a recall. Fall back to the reply and the
recall log only when it shows none: Codex sends its own rollout as
transcript_path, which is not a Claude transcript.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

…tor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

* feat(tools): state each tool's run-time and AI support in the descriptor (#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of #1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (#1193, #1224, #1245, #1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

* docs: generate the tool support page from the descriptor (#1221)

docs/guide/tool-support.md and its Chinese page are rendered from TOOLS
and diffed by a snapshot test, as the command reference is. Headers and
the legend are translated; cell text stays English on both pages.
AGENTS.md records that exception and the review rule for descriptor
updates.

* docs(readme): mark usage and sessions the descriptor records, and JoyCode's inlined rules (#1221)

Usage: skill use is recorded for OpenCode, Pi, Oh My Pi and DeepSeek Harness
since #1245 (`features.stats`). Sessions: `teamai session` reads OpenCode
(`TOOLS.opencode.sessions`). Rules: JoyCode's user rules are inlined into
rules.txt, so they stay on, as for every other tool marked ✓¹.

* docs(readme): add culture, workflows and scorers to the agent matrix, and check it against the descriptor (#1221)

Learnings, codebase and teamwiki were the same for every agent, so Team
Context is one column. That leaves room for culture (team culture and
shared instructions need an instruction-blocks channel, which Qoder,
Qoder CN, Kiro, Trae and Trae CN do not have), workflows (every agent
drives one from its session) and scorers (the agents whose sessions
teamai can judge: a session adapter in `TOOLS[tool].sessions`). Team
Execution is ordered skills, rules, culture, agents, hooks, mcp,
workflows, env, models, and the layer bullets follow it.

readme-matrix.test.ts keeps every README's cells equal and checks the
cells the code owns: culture, models, usage, scorers and ✓¹.

* docs: regenerate the tool support page for Kiro's hook file (#1221)

- 让 opt-in judge 接受 Codex、Muse 与 ZCode Stop 事件中的 last_assistant_message，不再强制要求 Claude 格式 transcript。
- 从同一会话已确认并关联的 recall 日志提取候选路径及投票键，继续保留作用域 eligibility、受信知识根校验与每会话投票账本去重。
- 新增三种工具的原生回复、重复 Stop、继承用户范围只读和伪造召回标记回归，更新中英说明及技能排查指南。
- 329 个评分与会话归属单测、构建、类型检查、lint 和隔离模型替身的真实 CLI recall/Stop/投票 E2E 通过；修复前三个原生回复用例失败。

… sends the reply (Tencent#1396)

Claude Code sends last_assistant_message in Stop along with transcript_path.
Since Tencent#1374 the reply took precedence, so a Claude session was judged from the
recall log instead of its transcript, and a doc the transcript shows but no
settled run of the session records was no longer judged.

Use the transcript whenever it shows a recall. Fall back to the reply and the
recall log only when it shows none: Codex sends its own rollout as
transcript_path, which is not a Claude transcript.

… (Tencent#1406)

* feat(tools): state each tool's run-time and AI support in the descriptor (Tencent#1221)

The env, workflows, recall, stats, AI backend, votes judge and HTTP agent
hook rows of Tencent#1221's matrix become a required `features` facet, one
`SupportCell` per row: the mark and the evidence letter, as the matrix
writes them. A cell a later fix changed (Tencent#1193, Tencent#1224, Tencent#1245, Tencent#1374) says
so in its note and keeps teamai's-code-only evidence until it is rerun.
Data only: nothing reads it but the support page.

* docs: generate the tool support page from the descriptor (Tencent#1221)

docs/guide/tool-support.md and its Chinese page are rendered from TOOLS
and diffed by a snapshot test, as the command reference is. Headers and
the legend are translated; cell text stays English on both pages.
AGENTS.md records that exception and the review rule for descriptor
updates.

* docs(readme): mark usage and sessions the descriptor records, and JoyCode's inlined rules (Tencent#1221)

Usage: skill use is recorded for OpenCode, Pi, Oh My Pi and DeepSeek Harness
since Tencent#1245 (`features.stats`). Sessions: `teamai session` reads OpenCode
(`TOOLS.opencode.sessions`). Rules: JoyCode's user rules are inlined into
rules.txt, so they stay on, as for every other tool marked ✓¹.

* docs(readme): add culture, workflows and scorers to the agent matrix, and check it against the descriptor (Tencent#1221)

Learnings, codebase and teamwiki were the same for every agent, so Team
Context is one column. That leaves room for culture (team culture and
shared instructions need an instruction-blocks channel, which Qoder,
Qoder CN, Kiro, Trae and Trae CN do not have), workflows (every agent
drives one from its session) and scorers (the agents whose sessions
teamai can judge: a session adapter in `TOOLS[tool].sessions`). Team
Execution is ordered skills, rules, culture, agents, hooks, mcp,
workflows, env, models, and the layer bullets follow it.

readme-matrix.test.ts keeps every README's cells equal and checks the
cells the code owns: culture, models, usage, scorers and ✓¹.

* docs: regenerate the tool support page for Kiro's hook file (Tencent#1221)