mirror of
https://github.com/tiennm99/goclaw.git
synced 2026-10-03 09:13:30 +00:00
thotam
3f90057c24
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.