feat(engine): conversation-insertion primitives — keepInbox steer queue + steer events + deterministic cancel - #16
Merged
h3c-hexin merged 7 commits intoAug 25, 2026
Conversation
|
Thanks @qiuYliangM for taking the time to contribute. This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered. Please read |
qiuYliangM
force-pushed
the
feat/conversation-insertion-primitives
branch
from
August 17, 2026 18:07
d2972b9 to
2f376e4
Compare
…ue + steer events + deterministic cancel
Two insertion semantics needed base support that was only best-effort:
- idle insertion (steer): messages could be silently dropped at turn
boundaries and hosts had no delivery confirmation
- forced insertion (interrupt): cancel never released a deterministic
"slot free" signal, and shell processes could survive the tool-future
drop race
Changes:
- steer queue becomes session-scoped (keepInbox): leftover steers are no
longer drained at new-turn start; collected-but-uninjected steers live on
the Engine and are re-injected by the next turn step boundary, surviving
cancel (aligned with deepseek-harness cancel(cause, { keepInbox: true }))
- steer_keep_inbox switch on EngineHandle (set before cancel): interrupt
(true) parks unconsumed steers for the next turn; stop (false) drops them
and emits Event::SteerDropped per message so hosts can remove the queued
chip and tell the user — nothing hangs invisible
- new events Event::SteerCommitted / Event::SteerDropped carry an FNV-1a
content hash so hosts can correlate delivery without full text; committed
is emitted at all three injection points (step boundary, post-stream,
post-tool)
- cancel now kills running shell process groups explicitly at both the
parallel and sequential tool select sites (drop(tool_tasks) does not
kill OS processes); shell.rs foreground poll switched from 100ms polling
to a select on the cancel token (broadcast wake, no longer loses to the
turn-loop drop)
Verification: cargo check + codewhale-tui --lib tests (9888 passed;
9 pre-existing Windows-environment failures reproduced on baseline).
评审修复（对应 pinvou-agent#308 宿主评审）：
- 关联契约：SteerCommitted/SteerDropped 改携带入队时生成的 opaque
steer_id（SteerMessage { id, content }），替代 FNV-1a 内容哈希——
UTF-8 字节与 UTF-16 码元哈希对非 ASCII 内容必然不一致，事件关联全断。
- 撤回：EngineHandle::withdraw_steer 记入共享撤回集合，所有收集/注入点
过滤并恰好发一条 SteerDropped；撤回标记跨轮存活，SyncSession/Shutdown
清场时清除。宿主排队占位的取消由此在注入前真正生效。
- 停止语义：keep_inbox=false 在全部 Interrupted 出口统一 settle
（pending_steers + channel 残留逐条 SteerDropped），不再只清字段。
- 清场覆盖：Op::SyncSession 换会话前 drop_all_steers 发事件（修跨会话
注入），Op::Shutdown 同样处理；Drop for Engine 以 try_send 兜底。
- kill 收敛：cancel 只杀本轮前台（spawned_as_foreground 且无
owner_agent）shell 进程组，后台任务与子智能体 shell 不再被连带 kill。
测试：engine/tests.rs 新增 13 条（keepInbox park/drop、三处注入点、
SyncSession 清场、Drop 兜底、撤回五场景），shell/tests.rs 新增
kill_running_turn_foreground 范围回归。cargo test -p codewhale-tui --lib
--locked 无新增失败（基线 Windows 环境失败与负载 flake 除外）。
Signed-off-by: luzeyang (INT) <lu.zeyang@h3c.com>
Signed-off-by: hexin <372726039@qq.com>
h3c-hexin
force-pushed
the
feat/conversation-insertion-primitives
branch
from
August 24, 2026 08:55
a5ef9eb to
e109b9b
Compare
|
我已在 e109b9b 上完成一轮重构，烦请作者重点复核以下语义和边界：
测试做了精简：原方案测试净增约 852 行，现在约 407 行，保留 6 个关键生命周期契约及 Shell 范围测试。PR 总体由 +1432/-107 调整为 +1175/-142，净增 1325 降至 1033。 本地验证：
请重点确认两项产品语义是否符合预期：一是 Stop 丢弃未提交 steer、Interrupt 保留；二是取消父 turn 不终止后台任务和子 Agent Shell。 |
steer_lifecycle 测试精简中丢失的覆盖：宿主不经 Op::Shutdown 直接回收 引擎（evict/reclaim）时，Drop 必须为每条未消费 steer 尽力补发 SteerDropped（try_send），否则宿主排队占位悬挂。实现在 engine.rs:6354，本测试重新锁定该契约。 Signed-off-by: luzeyang (INT) <lu.zeyang@h3c.com>
Author
|
已复核 两点产品语义确认符合预期：
另外几点复核结论： 我补回了一条被精简掉的 |
|
/lgtm |
This was referenced Aug 25, 2026
This was referenced Aug 25, 2026
h3c-hexin
pushed a commit
to qiuYliangM/CodeWhale
that referenced
this pull request
Aug 27, 2026
Hosts that re-send a queued input through another path (interrupt-and-send) must know whether the engine copy can still be committed, otherwise the same message can be delivered twice: the withdrawal used to be fire-and-forget, and "no SteerCommitted received yet" cannot be distinguished from "already committed" on the async event path. SteerControlState::withdraw now reports from the authoritative unsettled ledger, and EngineHandle::withdraw_steer returns it: - Retired: the id was pending and is now marked withdrawn — guaranteed never to be injected, settles with exactly one SteerDropped; the host may safely re-send. - NotPending: the id already settled (committed or dropped) or was never seen — no-op with no event; the host must not re-send. The withdrawal-set invariants are unchanged: unknown ids do not grow the set (bounded), settled ids stay no-ops. Behavior test steer_lifecycle_withdrawal_is_bounded_and_prevents_commit now asserts all three outcomes. No-Issue: driven by pinvou-agent#308 review (interrupt-and-send resend race); upstream-neutral, same area as the conversation-insertion primitives in Pinvou#16. Signed-off-by: luzeyang (INT) <lu.zeyang@h3c.com>
h3c-hexin
pushed a commit
that referenced
this pull request
Aug 27, 2026
Hosts that re-send a queued input through another path (interrupt-and-send) must know whether the engine copy can still be committed, otherwise the same message can be delivered twice: the withdrawal used to be fire-and-forget, and "no SteerCommitted received yet" cannot be distinguished from "already committed" on the async event path. SteerControlState::withdraw now reports from the authoritative unsettled ledger, and EngineHandle::withdraw_steer returns it: - Retired: the id was pending and is now marked withdrawn — guaranteed never to be injected, settles with exactly one SteerDropped; the host may safely re-send. - NotPending: the id already settled (committed or dropped) or was never seen — no-op with no event; the host must not re-send. The withdrawal-set invariants are unchanged: unknown ids do not grow the set (bounded), settled ids stay no-ops. Behavior test steer_lifecycle_withdrawal_is_bounded_and_prevents_commit now asserts all three outcomes. No-Issue: driven by pinvou-agent#308 review (interrupt-and-send resend race); upstream-neutral, same area as the conversation-insertion primitives in #16. Signed-off-by: luzeyang (INT) <lu.zeyang@h3c.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
2 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.
背景
两种插入语义需要底座支持（原为尽力而为）：
改动
验证
cargo check --all-targets + codewhale-tui --lib 9888 测试通过（9 个 Windows 环境既有失败在基线复现）。
No-Issue: pinvou3 打断/插入语义的底座配套，由 pinvou-agent PR Hmbown#308 的同一需求驱动