Repository navigation
groupmind poller: carry reply_to and resolve reply targets into agent-facing lines - #70
Conversation
…e line A reply's meaning lives in its target. The client sent reply_to, the server stored it, the API returned it - and this adapter dropped it in normalize(), so every agent read replies as freestanding statements and guessed at referents (Aug 27: petrus's 'spec this' built the wrong spec). - fetch(): resolve targets from the same batch (fresh replies are near always within the poll window) - normalize(): payload carries reply_to + resolved reply_target - formatLine(): '⤷ in reply to <from>: "<snippet>"' - and an unresolved reply is still marked as a reply, never shown freestanding - 4 regression tests; suite 210/210
|
Blocking integration issue: |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ca1231f2af
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".
| reply_to: msg.reply_to || null, | ||
| reply_target: msg._replyTarget || null |
There was a problem hiding this comment.
Update the GroupMind adapter imported by the CLI
When users run ide-agent-kit platform watch, this reply metadata and formatting never execute: bin/cli.mjs imports src/team-relay/adapters/groupmind.mjs, which still has the old implementation, while a repo-wide import search finds this changed adapter used only by test/unified-poller.test.mjs. Consequently, the production poller continues presenting replies as freestanding messages even though the new tests pass; apply the change to the team-relay adapter used by the CLI (and test that module directly).
Useful? React with 👍 / 👎.
What
The GroupMind poller adapter dropped
reply_toinnormalize(), so every agent-facing surface showed replies as freestanding statements. Today that made "Great! Please spec this" (a reply to a specific message) get interpreted by guess, and the wrong spec was built (THI-26 / the AgentOS mixup).The client always sent
reply_to; the server stores and returns it. The drop was here.How
fetch(): resolves reply targets from the same poll batch (a fresh reply's target is nearly always within the window)normalize(): payload now carriesreply_to+ resolvedreply_target {from, body}formatLine(): renders⤷ in reply to <from>: "<snippet>"; an unresolved reply is still marked as a reply with its target id, never shown freestandingTests
4 new regression tests (resolved, unresolved, non-reply formats); full suite 210/210.
Part of THI-19's re-aim (agent-side reply reading). The remaining half - the Android app's vanishing reply preview - is a separate CodeWatch client fix.
🤖 Generated with Claude Code