Continue subagent transcripts by lifu963 · Pull Request #3586 · esengine/DeepSeek-Reasonix

lifu963 · 2026-06-08T12:41:27Z

Summary

Add persisted subagent transcript storage for task and skill subagents.
Support continuing and forking subagent transcripts with model/effort/workspace validation.
Adapt the review CLI to the session-based subagent runner.

Test plan

go test ./...

lifu963 · 2026-06-09T00:40:17Z

我对 reasonix 的 subagent 做了会话复用能力的支持！subagent 可以继续延续上一次的会话记录（不再是仅支持 fresh window），可以用在主 agent 需要与 subagent 多轮对话的场景！例如：主 agent 完成开发 -> subagent review -> fix -> 再次 review -> 再次 fix -> ... -> 直到 review 通过！
本次代码变更从方案提出到实施、测试我都严格 review 过；辛苦能看一下吗？谢谢你！

SivanCola

Thanks for the transcript continuation work. I found one blocking issue around failed foreground runs.

In the foreground task path, RunSubAgentWithSession errors return immediately without saving the mutated session or marking the transcript failed (internal/agent/task.go). Subagent skills have the same pattern in internal/boot/boot.go. Since Agent.Run has already appended the new user prompt and any tool results before returning errors, a failed continue_from can leave the on-disk transcript/meta at the previous completed state. A later continue_from then reloads stale history and may repeat side-effecting work.

Could you persist the updated session and mark the run failed, or otherwise make the in-place continuation non-reusable after a failed mutation, for both foreground task and subagent skill paths?

Validation I ran:

go test ./internal/agent ./internal/skill ./internal/boot ./internal/cli passed
go test ./... still fails in internal/control on TestRemoveMCPServerRemovesUnconnectedLazyPlaceholder, but that reproduces on current origin/main-v2, so I treated it as existing/environment-related.

lifu963 · 2026-06-09T08:47:06Z

@SivanCola
Thanks again for the review. I pushed updates addressing the requested changes.

What changed:

SubagentStore.SaveFailed now persists the mutated subagent session JSONL before marking the metadata as failed.
Foreground task continuations now call SaveFailed when RunSubAgentWithSession fails after the run has started.
runAs=subagent skill continuations in boot.go now use the same failure persistence path.
Run errors and persistence errors are preserved with errors.Join, so the original failure is still visible while persistence risk is not hidden.
Pre-run setup/compatibility failures, such as incompatible tool scope, are still rejected before execution and do not mark an otherwise completed ref as failed.

Validation:

go test ./... passes locally.
I also ran a live CLI end-to-end validation against DeepSeek, which exercises the same boot.Build / Controller / agent kernel used by the desktop app:
1. Asked the main agent to call task once with subagent prompt: Reply with exactly: subagent first turn ok.
  - It returned a subagent ref and the expected answer.
2. Asked the main agent to call task again with continue_from=<same ref> and prompt: Continue the same subagent transcript. Reply with exactly: subagent second turn ok.
  - The same ref was reused, and the persisted subagent JSONL contained both user turns.
3. Asked the main agent to continue the same ref with max_steps=1 and prompt: Before any final answer, call read_file on missing-file-for-failure-test.txt.
  - The subagent performed the read_file tool call, then failed/paused after the max-step limit.
  - The subagent metadata was marked failed, and the failed continuation prompt/tool result were persisted in the JSONL transcript.
4. Tried continue_from=<same ref> again.
  - It was rejected with failed and cannot be continued, confirming the ref is no longer reusable after a failed mutation.

I also merged latest origin/main-v2 into this PR branch and resolved the boot conflict. The PR should no longer be blocked by merge conflicts.

lifu963 · 2026-06-09T14:44:47Z

@SivanCola @esengine

改说中文好了，哈哈。

更新一下本轮补充内容：

这次在 subagent 会话复用能力之外，补齐了对应的 lifecycle 管理，避免主会话删除后遗留 subagent transcript。

具体做了两部分：

1）为 subagent transcript 写入可靠的父会话归属

新建的 task / subagent skill transcript 现在会写入 ParentSession。
continue_from 只允许继续当前父会话 owned 的 subagent，避免两个父会话共同修改同一个 subagent transcript
fork_from 主要用于主会话 fork/branch 场景：新会话从旧会话 fork 出来时，历史里可能还引用旧的 subagent ref。如果新会话想复用这段 subagent 历史，就用 fork_from 复制出一个新 ref，并让它归属于新会话。这样后续删除、恢复、永久删除，都能按各自父会话清晰管理。

2）主会话 trash / restore / purge / delete 时级联处理 subagent artifacts

Desktop 中把父 session 放入回收站时，会把它拥有的 sa_*.jsonl 和 sa_*.meta.json 一起移入同一个 trash item。
Desktop restore 时，会把这些 subagent artifacts 一起恢复回 sessions/subagents/。
Desktop permanent delete / purge 时，会一起永久删除。
ACP session/delete 和 HTTP /delete-session 也会删除该父 session 拥有的 subagent artifacts。

测试方面，我做了三层验证：

单元/集成测试：
- go test ./internal/agent ./internal/boot ./internal/control ./internal/acp ./internal/serve
- cd desktop && go test .
真实 DeepSeek API E2E：
- fresh task 能创建持久化 subagent，并写入正确 ParentSession。
- 同一父会话下 continue_from 可以继续追加 transcript。
- 不同父会话下 continue_from 会被拒绝，并提示使用 fork_from。
- fork_from 可以复制旧 subagent transcript，生成归属于当前父会话的新 ref。
- HTTP /delete-session 删除父会话时，只删除该父会话 owned 的 subagent，不会误删 fork 后属于其他父会话的 subagent。
真实 macOS Desktop GUI 验收：
- 编译并启动 Wails app。
- 在真实 app 中调用 task 生成 subagent。
- 将父 session 移入回收站后，确认父 .jsonl / .meta / checkpoint 和 owned subagent .jsonl / .meta.json 一起进入 .trash/<parent>.jsonl/。
- 从回收站恢复后，确认父 session 和 subagent artifacts 一起回到 active session 目录。
- 永久删除后，确认父 session trash item 和 owned subagent artifacts 都被彻底删除。

另外，我也检查了当前分支与最新 origin/main-v2 的合并情况：没有阻塞性 merge conflict。本轮更新已经推送到 PR 分支。

Co-authored-by: Cursor <cursoragent@cursor.com>

Ensure persisted subagent transcripts record the active parent session and prevent in-place continuation from a different parent, keeping fork_from as the explicit copy path. Co-authored-by: Cursor <cursoragent@cursor.com>

Move owned subagent artifacts with desktop trash/restore and delete them from hard-delete paths so parent session cleanup no longer leaves reusable child transcripts behind. Co-authored-by: Cursor <cursoragent@cursor.com>

Persisted sub-agent transcripts made a parent session mandatory, so `reasonix run` (headless, which never mints a session path) failed every task/subagent-skill call with "subagent transcript parent session is required". Fall back to a non-persisted run when no parent session is active — restoring pre-transcript behaviour — and reserve continue_from/ fork_from for sessions that actually have a persisted owner. Adds a headless-run regression guard plus task-level ephemeral coverage.

esengine · 2026-06-09T16:21:59Z

/e2e

github-actions · 2026-06-09T16:22:27Z

🤖 e2e bot failed — see the run log.

esengine · 2026-06-09T16:30:17Z

/e2e

github-actions · 2026-06-09T16:32:04Z

🤖 Reasonix e2e benchmark

Accuracy: 4/5 (80%) · Cache hit: 78% · Tokens: 254,258 (prompt 251,307 / completion 2,951) · Compactions: 3 · Cost: ¥ 0.0647

Task	Result	Steps	Prompt	Completion	Cache hit	Compact	Cost
`compaction`	❌ fail	8	128,553	1,108	58%	3	¥ 0.0571
`fix-add-bug`	✅ pass	4	37,637	325	99%	0	¥ 0.0019
`fizzbuzz`	✅ pass	3	28,308	522	99%	0	¥ 0.0019
`palindrome`	✅ pass	3	28,226	498	99%	0	¥ 0.0018
`subagent-delegation`	✅ pass	3	28,583	498	99%	0	¥ 0.0020

_{Real provider run. Cache-hit % is cached prompt tokens / total prompt tokens.}

agent: PR head (cf4a3d2) · triggered by @esengine

lifu963 · 2026-06-09T17:42:29Z

🤖 Reasonix e2e benchmark

Accuracy: 4/5 (80%) · Cache hit: 78% · Tokens: 254,258 (prompt 251,307 / completion 2,951) · Compactions: 3 · Cost: ¥ 0.0647

Task Result Steps Prompt Completion Cache hit Compact Cost
compaction ❌ fail 8 128,553 1,108 58% 3 ¥ 0.0571
fix-add-bug ✅ pass 4 37,637 325 99% 0 ¥ 0.0019
fizzbuzz ✅ pass 3 28,308 522 99% 0 ¥ 0.0019
palindrome ✅ pass 3 28,226 498 99% 0 ¥ 0.0018
subagent-delegation ✅ pass 3 28,583 498 99% 0 ¥ 0.0020
Real provider run. Cache-hit % is cached prompt tokens / total prompt tokens.

agent: PR head (cf4a3d2) · triggered by @esengine

我看了这次 compaction e2e 失败的 run log。
这次失败看起来不是本 PR 的改动导致的。

日志里 compaction 任务没有调用 task / subagent；它正确走完了章节线索，也得到了正确答案内容：Aldermoor-Verrin；但最后尝试写入 story/answer.txt（其实应该写入 answer.txt），所以这是输出路径放错导致的 fail。

本 PR 没有改 internal/agent/compact.go、e2e harness、compaction 任务定义或 write_file；真正和本 PR 相关的 subagent-delegation e2e 是通过的。

因此，我倾向于认为这条 compaction fail 不是本 PR 引入的问题，而是任务提示在输出路径上不够明确，导致模型把正确答案写到了错误目录。

* feat(agent): add subagent transcript store Co-authored-by: Cursor <cursoragent@cursor.com> * feat(agent): continue task subagent transcripts Co-authored-by: Cursor <cursoragent@cursor.com> * feat(skill): continue subagent skill transcripts Co-authored-by: Cursor <cursoragent@cursor.com> * refactor(agent): require subagent transcript store Co-authored-by: Cursor <cursoragent@cursor.com> * fix(boot): record effective subagent identity Co-authored-by: Cursor <cursoragent@cursor.com> * fix(agent): release source lock after fork copy Co-authored-by: Cursor <cursoragent@cursor.com> * fix(cli): run review with explicit subagent session Co-authored-by: Cursor <cursoragent@cursor.com> * fix(agent): persist failed subagent transcripts Co-authored-by: Cursor <cursoragent@cursor.com> * fix(agent): mark failed foreground continuations Co-authored-by: Cursor <cursoragent@cursor.com> * fix(agent): bind subagents to parent sessions Ensure persisted subagent transcripts record the active parent session and prevent in-place continuation from a different parent, keeping fork_from as the explicit copy path. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(session): cascade subagent lifecycle cleanup Move owned subagent artifacts with desktop trash/restore and delete them from hard-delete paths so parent session cleanup no longer leaves reusable child transcripts behind. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(agent): run sub-agents ephemerally without a parent session Persisted sub-agent transcripts made a parent session mandatory, so `reasonix run` (headless, which never mints a session path) failed every task/subagent-skill call with "subagent transcript parent session is required". Fall back to a non-persisted run when no parent session is active — restoring pre-transcript behaviour — and reserve continue_from/ fork_from for sessions that actually have a persisted owner. Adds a headless-run regression guard plus task-level ephemeral coverage. --------- Co-authored-by: lifu963 <lifu963@lifu963deMacBook-Air-2.local> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: esengine <359807859@qq.com>

lifu963 requested review from SivanCola and esengine as code owners June 8, 2026 12:41

SivanCola requested changes Jun 9, 2026

View reviewed changes

lifu963 requested a review from SivanCola June 9, 2026 08:47

github-actions Bot added the desktop Wails desktop app (desktop/**) label Jun 9, 2026

esengine mentioned this pull request Jun 9, 2026

test(e2e): add subagent-delegation task forcing a task sub-agent #3725

Merged

lifu963 and others added 12 commits June 9, 2026 09:12

feat(agent): add subagent transcript store

1325884

Co-authored-by: Cursor <cursoragent@cursor.com>

feat(agent): continue task subagent transcripts

304728e

Co-authored-by: Cursor <cursoragent@cursor.com>

feat(skill): continue subagent skill transcripts

984effc

Co-authored-by: Cursor <cursoragent@cursor.com>

refactor(agent): require subagent transcript store

84f8502

Co-authored-by: Cursor <cursoragent@cursor.com>

fix(boot): record effective subagent identity

03f285f

Co-authored-by: Cursor <cursoragent@cursor.com>

fix(agent): release source lock after fork copy

f14432a

Co-authored-by: Cursor <cursoragent@cursor.com>

fix(cli): run review with explicit subagent session

abc1128

Co-authored-by: Cursor <cursoragent@cursor.com>

fix(agent): persist failed subagent transcripts

5f08be1

Co-authored-by: Cursor <cursoragent@cursor.com>

fix(agent): mark failed foreground continuations

c3d53d6

Co-authored-by: Cursor <cursoragent@cursor.com>

fix(agent): bind subagents to parent sessions

4ee5639

Ensure persisted subagent transcripts record the active parent session and prevent in-place continuation from a different parent, keeping fork_from as the explicit copy path. Co-authored-by: Cursor <cursoragent@cursor.com>

fix(session): cascade subagent lifecycle cleanup

4185c29

Move owned subagent artifacts with desktop trash/restore and delete them from hard-delete paths so parent session cleanup no longer leaves reusable child transcripts behind. Co-authored-by: Cursor <cursoragent@cursor.com>

esengine force-pushed the agent/subagent-transcripts-mainv2 branch from 825fa6d to cf4a3d2 Compare June 9, 2026 16:16

esengine mentioned this pull request Jun 9, 2026

ci(e2e-bot): build with go.mod's Go version, not pinned 1.22 #3731

Merged

esengine merged commit 5939500 into esengine:main-v2 Jun 9, 2026
13 checks passed

lifu963 deleted the agent/subagent-transcripts-mainv2 branch June 13, 2026 07:50

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Continue subagent transcripts#3586

Continue subagent transcripts#3586
esengine merged 12 commits into
esengine:main-v2from
lifu963:agent/subagent-transcripts-mainv2

lifu963 commented Jun 8, 2026

Uh oh!

lifu963 commented Jun 9, 2026

Uh oh!

SivanCola left a comment

Uh oh!

lifu963 commented Jun 9, 2026

Uh oh!

lifu963 commented Jun 9, 2026 •

edited

Loading

Uh oh!

esengine commented Jun 9, 2026

Uh oh!

github-actions Bot commented Jun 9, 2026

Uh oh!

esengine commented Jun 9, 2026

Uh oh!

github-actions Bot commented Jun 9, 2026

Uh oh!

lifu963 commented Jun 9, 2026

🤖 Reasonix e2e benchmark

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

lifu963 commented Jun 8, 2026

Summary

Test plan

Uh oh!

lifu963 commented Jun 9, 2026

Uh oh!

SivanCola left a comment

Choose a reason for hiding this comment

Uh oh!

lifu963 commented Jun 9, 2026

Uh oh!

lifu963 commented Jun 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

esengine commented Jun 9, 2026

Uh oh!

github-actions Bot commented Jun 9, 2026

Uh oh!

esengine commented Jun 9, 2026

Uh oh!

github-actions Bot commented Jun 9, 2026

🤖 Reasonix e2e benchmark

Uh oh!

lifu963 commented Jun 9, 2026

🤖 Reasonix e2e benchmark

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

lifu963 commented Jun 9, 2026 •

edited

Loading