fix(acp): keep agent DM replies top-level - #5661
Conversation
In a 1:1 DM, agents currently anchor every reply to the previous message, so each exchange nests one thread level deeper and the human has to expand a growing reply chain to follow a plain conversation (block#2748). This changes the reply-destination policy for DMs only: ordinary agent responses in a DM are published as top-level channel messages (no reply anchor). Channel and forum behavior is unchanged, and agent-to-agent subthreads inside DMs remain threaded. Signed-off-by: Ika Minami <ika@infiniteidol.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5ea2c32676
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let reply_anchor = if is_dm { | ||
| thread_tags | ||
| .root_event_id | ||
| .is_some() | ||
| .then(|| last_event.event.id.to_hex()) | ||
| None |
There was a problem hiding this comment.
When a threaded DM event comes from a profile marked is_agent and mentions no human, this unconditional DM branch bypasses turn_is_human_facing, sets no reply destination, and adds an instruction requiring a top-level send. Consequently agent-to-agent task subthreads in DMs are flattened too, contradicting the retained agent-only threading rule in base_prompt.md; apply the flat-DM policy only to human-facing turns and keep a threaded destination for agent-only ones.
Useful? React with 👍 / 👎.
2 participants
Addresses #2748 with a deliberately different design than #2766 and #2778 — filing as an explicit alternative, not a duplicate.
The fork in #2748
Two ways to stop DM conversations nesting one level deeper per exchange:
Why flat
A DM with an agent is a conversation, not a review thread. Root-anchoring still renders every agent response as "a reply" the client tucks under an expandable thread — the human keeps clicking into threads to read what is functionally a linear chat. Flat matches how every mainstream DM surface behaves and needs no client-side rendering special-cases.
Scope is DM-only and human-facing-only:
Implementation: reply-destination policy in
queue.rs+ a matching instruction in the ACP base prompt.Testing
cargo test -p buzz-acp --lib— 753 passedcargo clippy -p buzz-acp --all-targets/cargo fmt --check— cleanIf maintainers prefer the root-anchored design, happy to close in favor of #2766/#2778 — the value is in picking one; the nesting bug has six open PRs and none landed.