Repository navigation
Delegate notifications #2913
Replies: 7 comments 2 replies
[bot] 🏷 Short answer: yes, this is by-design behavior, not a delivery bug — and you're almost certainly not actually missing the summary. Here's the mechanism and how to confirm it on your box. Why "no notifications" at concurrency=1There are exactly two mutually-exclusive ways a finished delegate reaches the main agent:
They are deliberately exclusive to avoid showing the model the same result twice. The completion-notification path explicitly skips any run that was already owned by a wait/cancel (see Why concurrency flips which channel you seeThe deciding factor is whether each subagent finishes inside the wait window.
So the correlation isn't "concurrency breaks notifications"; it's "serial execution is fast enough that Confirm it on your machineIn your concurrency=1 sessions, grep the pi log (documented path grep 'event=done' ~/.pi/acp.logLook at the
Then check whether any notification actually went out: grep -E 'notify-turn-end-commit|notify-batch' ~/.pi/acp.log
If you specifically want the notification styleThere is no setting to force a second, redundant notification on top of a successful wait — that would just hand the model the same result twice. What you can do:
If after checking the log you do see |
[bot] 🏷 Good push — you've identified something real, and it's slightly different from "the notification disappeared." Here's exactly what's happening. Why the usage summary goes quiet at concurrency=1There are two completion paths, and they treat usage differently:
So the completion result still reaches your model inline (nothing is lost there) — but the usage summary the notification used to carry is simply not printed on the inline path. That asymmetry is why the chat goes quiet on usage and you end up opening the fleet. Where those numbers actually live (default "separate" mode)The separate bucket is surfaced outside the chat in two places:
If your footer line isn't showing, either your build isn't rendering plugin status, or you're not in What you can do right now
// acp.json
{ "delegate": { "displayUsage": "merged" } }No env var exists for this key — acp.json only (a legacy flat The real gap (worth fixing)The two completion paths are inconsistent about usage visibility: the notification prints the summary, the inline-wait result doesn't. The clean fix is to make That touches output format + display-mode semantics, so per the repo rules it's owner-gated rather than something I'll ship silently. My recommendation: track it as a follow-up issue and make the inline path mirror the notification's usage block. Want me to file that issue and/or implement it? |
|
The preference here is to keep the displayUsage separate, merge defeats the purpose. |
[bot] 🏷 Agreed on both counts — keep What I'll build1. Extract the shared usage renderer. Today the usage text lives inline in 2. Emit it as a system message on the inline-fetch paths. When Decisions I'm making (override any of these)
ProcessThis touches display/output semantics, so per repo rules it's owner-gated rather than auto-merge: I'll land it on a branch, add a regression test (both delivery paths + the double-count guard), bump the |
Great catch — and thanks for confirming the direction. This is a real visibility gap, not merely "by design": when your subagents run serially ( I implemented exactly what you asked for: the inline fetch path now emits the same cumulative usage summary the notification uses, as a short steering system message. PR: #2927 (rebased onto current master; typecheck + full pi-subagents suite green, 235/235) Behavior change: on a delegate completion that arrives via 结构化结论(八问)
判定:PR #2927 已开好并 rebase 到最新 master,typecheck + pi-subagents 全量单测(235/235)绿。因涉及展示/输出语义(owner-gated),请人工 review 后合并。合并后你在 |
|
merged |
Confirmed — verified the merge landed: PR #2927 is in @richardaubin this ships in the next |

[bot] 🏷 Agreed on both counts — keep
displayUsage: "separate"(merged does defeat the point of excluding delegate tokens from main totals), and surfacing the same usage summary on the inline path is the right fix. Your dedupe hunch is also correct: the usage block is cleanly extractable, so this is a small, low-risk change. Here's the concrete plan — tell me to go (or adjust a point) and I'll implement it as a PR referencing #2913.What I'll build
1. Extract the shared usage renderer. Today the usage text lives inline in
buildInjectText(pi-subagents/src/delegate-tool.ts:2087–2115) — a separate-mode cumulative block (── Session delegate usage (excluded f…