feat(memory+diagnostics): surface compaction events + memory stats#36
Merged
Conversation
Compaction telemetry was already collected (CompactionCounters, 10 atomics
behind /v1/diagnostics/memory) but invisible to both users and operators:
the agent silently called memory.compact() each iteration with no event,
the Diagnostics page only listed worktrees/runs, and the SlidingWindowMemory
backend (the default `JARVIS_MEMORY_MODE=window`) had no stats provider at
all. This makes the existing telemetry actionable end-to-end.
Backend (harness-core / harness-memory)
- New memory_event task-local channel mirroring plan/progress; carries
CompactionInfo { source, turns_kept, turns_dropped, summary_chars,
working_context_chars, model_input_tokens_est }. Emits are silent
no-ops outside the agent loop, keeping memory unit tests trivial.
- AgentEvent::MemoryCompacted variant; Agent::run_stream scopes the
channel around each build_request call and yields drained events
before the next LLM iteration.
- SummarizingMemory::compact now classifies its outcome into one of
CompactionSource {NoOp, WindowDropped, CacheMemory, CacheStore,
FreshLlm, PtlRoundOne, PtlRoundTwo, SummaryUnavailable} and emits
one CompactionInfo per call. Priority: PTL > summary source >
SummaryUnavailable > NoOp.
- SlidingWindowMemory gains a minimal SlidingCounters
{ compactions_total, window_dropped } so the default deployment
also serves /v1/diagnostics/memory instead of 503.
- jarvis-cli renders an inline one-liner for non-NoOp events.
Frontend (jarvis-web)
- Settings -> System -> Diagnostics now has a Memory panel polling
/v1/diagnostics/memory every 10s; renders backend / compactions /
cache hit rate / LLM failure rate / circuit opens-skips / PTL fallback
for the summarizing backend, and a trimmed view (backend +
compactions + window-dropped) for the sliding backend so no
misleading zeros show up.
- New memorySlice (per-conversation ring buffer, cap 20) +
memoryFrames handler. `reset` clears markers across conversations
to mirror planFrames.
- New CompactionMarker component renders an inline dashed-border card
in the chat transcript at the right message boundary; NoOp emits
are filtered out (counters still record them). SummaryUnavailable
and PTL fallbacks render with a warn tone (amber).
- Full zh-CN + en i18n: diagnosticsMemoryHeading, MemoryUnavailable,
and 7 compactionSourceXxx keys.
Quality gates
- cargo clippy --workspace --all-targets --exclude jarvis-desktop
-D warnings: clean
- cargo test --workspace --exclude jarvis-desktop: passed
- harness-memory: 47/47 (5 new event tests + 1 sliding counter test)
- harness-core: agent.rs run_stream_emits_memory_compacted +
4 memory_event::tests
- pnpm typecheck: clean; vitest 346/346 (9 new across memorySlice /
memoryFrames)
- Browser preview: zh-CN labels resolve; populated panel renders
6 rows for summarizing (80% cache hit rate over the sample
payload); sliding shape renders 2 rows; warn-tone marker renders
amber dashed border + tinted pill.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
AgentEvent::MemoryCompactedevent lets the chat surface show an inline marker when the memory backend compacts earlier turns; renders the source (cache hit / fresh summary / fallback prune / summary unavailable) and a few quick metrics.Settings → 系统 → 诊断 → Memorypanel polls/v1/diagnostics/memoryevery 10s and surfaces the 10 CompactionCounters that were already collected but had no UI consumer.SlidingWindowMemorynow ships a minimal stats provider so the defaultJARVIS_MEMORY_MODE=windowdeployment serves the endpoint instead of 503.Why
Commit 53500a0 landed
SummarizingMemorythree-tier cache +CompactionCounters(10 atomics) +MemoryStatsProvidertrait +GET /v1/diagnostics/memory. Audit found everything was tracked but invisible:agent.rssilently calledmemory.compact()each iteration with no event, so a user watching their chat history get shorter had no idea why.mode=windowdeployment had no stats provider at all → 503.This makes the existing telemetry actionable end-to-end without changing any compaction logic, cache strategy, or trait shape (
Memory::compactsignature unchanged).Test plan
cargo clippy --workspace --all-targets --exclude jarvis-desktop -- -D warnings— cleancargo test --workspace --exclude jarvis-desktop— passed🤖 Generated with Claude Code