Runs on models without a registered tokenizer (e.g. 9router brand models)
ended with the generic "Agent couldn't generate a response" fallback even
though the real request used about 55% of the context window.
PruneStage counted history with TokenCounter, which falls back to a
chars/2 heuristic for unregistered models and overcounted about 1.8x.
Once over budget it ran memory flush (~35s, invisible in traces), then
mid-loop compaction, which cannot summarize a history made only of tool
call/result pairs. The callback reported the untouched history as
compacted, PruneStage still saw it over budget and returned AbortRun
before any LLM call, and FinalizeStage replaced the empty reply with the
fallback.
- PruneStage and ContextStage overhead count with the request guard's
BudgetCounter. PruneStage no longer controls loop flow; the final
request guard in ThinkStage decides.
- CompactMessages returns ErrNotCompacted when history is unchanged.
Callers stop counting it as a compaction and do not retry it in the
same run, while post-run summarization still sees the pressure.
- When the guard exhausts every reduction step, ThinkStage stops the run
with a localized chat.context_budget_exceeded notice instead of an
error, so the run's tool results are still persisted. The stop reason
marks the trace and agent span as error; team tasks, cron and
heartbeat treat it as a failure via RunOutcome.Failure().
- Memory flush and mid-loop compaction emit event spans.
- Web and desktop UIs treat an unset context_pruning as enabled (the
backend default since 7639a8c0), keep it unset when untouched, and can
re-enable pruning after it was turned off.
Forum topic routing (Telegram supergroup threads) relies on local_key
carried through team task events to pick the correct thread for
notifications. Four broadcast sites dropped local_key by relying on
WithContextInfo(ctx) alone — but the member agent's isolated context
does not carry the originating local_key.
Read task.Metadata[TaskMetaLocalKey] explicitly and pass via
WithLocalKey() to match the already-correct pattern in
team_tool_dispatch.go and team_tasks_blocker.go.
Covered sites:
- team_tasks_lifecycle.executeComplete → EventTeamTaskCompleted
- gateway_consumer_post_turn agent-error → EventTeamTaskFailed
- gateway_consumer_post_turn loop-kill → EventTeamTaskFailed
- gateway_consumer_post_turn auto-complete → EventTeamTaskCompleted
Two fixes:
1. Remove assistant prefill from team task reminders. The injected
[user]+[assistant]+[user] pattern caused LLMs to treat the canned
ack as "turn complete", returning NO_REPLY for every user message
in group sessions with active tasks. Reminders are now merged into
the user message as prefix tags.
2. Add PeerKind propagation to team notification routing. TaskTicker
and progress notifications were missing PeerKind on InboundMessage,
causing them to route to phantom DM sessions instead of the correct
group session. PeerKind is now carried through event payloads,
notify queue metadata, and all inbound message publications.
- loop.go: replace blocking range with select+ctx.Done() for parallel tool collection;
document finalization trade-off on early cancel
- loop_cancel_test.go: tests for context cancellation and normal completion paths
- gateway_consumer_post_turn.go: always auto-complete team task in default case,
fallback message when outcome.Result is nil
- gateway.go: call sched.Stop() with 5s drain before context cancel on shutdown
thotam