Performance round: measure on real data, then fix the top costs

The August performance issues (#225, #564, #617, #618, #639, #640) are all closed and their numbers are stale. This round starts from fresh measurements on today's real data.

Baseline (2026-10-05, live data, read-only)

Store file: coder-store.json is 6.0 MB and parses in 18 ms. Boot parse is no longer an issue.
Threads: 2,097, of which 958 are live (not archived), across 14 projects. Message shards: 2,097 files, 788 MB, the largest 3.6 MB.
Running Nightly (not idle: the orchestrating thread runs inside it): main 347 MB RSS, renderer 575 MB RSS, GPU 60 MB.

Phase 1: measure (no code changes)

PM1, main process, on a clone of the userData dir: the boot timeline (window, first list, interactive); IPC payload sizes and rates for hot channels at 2k threads; a CPU profile during boot and during simulated streaming; what holds the main-process heap; disk writes while streaming.
PM2, renderer, the real Electron app on the clone driven over CDP: first render; sidebar with ~958 live rows; switching to the largest threads; typing latency; long tasks while streaming; heap and DOM-node breakdown.

Phase 2: fix

Pick the top costs from the Phase 1 numbers. Each fix needs a before/after measurement taken the same way.

Open from before: #1398 (scope threads:summaries); PM1 will check whether #1424 already covered it.

Incident during PM2 (resolved, no data lost)

What happened: booting the app against a cp -Rc clone of userData is NOT isolated. The cloned worktrees/*/.git files point into the real repos' .git/worktrees/<id>. At 19:37, PM2's clone boot ran sweepOrphanWorktrees, which switched three real worktrees of archived threads to recovered/* branches and committed their uncommitted work there:

e055dfbb in ~/Desktop/Apps/saas/appfeedback
7f93b763 and 08472090 in ~/code/huskyscout

Restore (by PM2, at 19:39), verified independently:

All three worktrees are back on their original branches.
Their full contents (untracked included, compared via a temp index) are identical to the 19:37 snapshots.
The original branch tips have no ref changes today.
No recovered/* branches remain in any of the 14 project repos.
All 21 live worktree registrations are intact.
The live memory.db (which the clone's memory-server.json also pointed at) passes PRAGMA quick_check.

Prevention: the safe launch recipe (PM1 used it with zero leaks) is in agent memory as solenta-cloned-data-dir-launch. PM2 has been told to use only that recipe from here on. Any future perf or dogfood worker gets it in its prompt.

PM1 results (main process): `docs/perf/2026-10-05-main-process.md`

Streaming rewrites the whole 5.9 MB envelope on every flush: 19 MiB/s, ≈92% of main-process JS busy time while streaming.
The first threads:summaries call per session reads 463 MiB of transcripts synchronously, freezing the main process for 2.07 s when the Agents panel opens. P1: Scope threads:summaries so opening Agents does not read every transcript #1398 is only partly fixed.
The login-shell PATH capture blocks boot for 0.71 s.
Sync git in projects:list and git:listBranches: 140 ms on the boot path, and ~170 ms stalls on run starts.
files:resolve fan-out: 424 IPCs and 131 ms per thread open.
The message cache never shrinks: +4 MB heap per large thread visited.

Idle is fine: 50 ms of main JS per 30 s.

PM2 results (renderer): `docs/perf/2026-10-05-renderer.md`

Provenance is recomputed over the whole transcript on every push: 30% of renderer JS while streaming into a big thread, and ~48 ms per big switch.
The open transcript grows without bound while streaming: DOM 24.7k → 62.2k nodes, RSS 275 → 579 MB, and blocked time 220 → 337 ms/s.
A forced synchronous layout in useStickToBottom runs on every push: ~25–30 ms/s.
The whole streaming message is re-parsed as markdown on every push, so cost grows O(n²) with answer length.
threads:changed ships 2.68 MB (all 2,004 rows): a ~21 ms renderer hitch per status change. PM1 found the same.
A big-thread switch takes 120–220 ms.
Cold start: ~500 ms to FCP.
localStorage: the boot snapshot is at 2.7M of the 4M cap, with 1,667 stale paneLayout keys.

Fine: the sidebar at 966 rows, small-thread switches, idle typing, streaming into a thread that isn't open, and memory after switching (no leaks).

Phase 2: fixes running (six parallel workers, disjoint files)

F-store: envelope-dirty tracking plus a 2 s envelope throttle, and an LRU message cache (PM1 Solenta X launch trailer #1, Add Settle all action to the collapsed Settled tail bar #6)
F-summaries: a lastActivity snippet on the thread row, so there are no transcript reads (PM1 Add settings toggle to isolate new threads in a git worktree by default #2 / P1: Scope threads:summaries so opening Agents does not read every transcript #1398)
F-boot: cached login-shell PATH with async recapture, and async git in projects:list / listBranches (PM1 Grok provider runs fail immediately after first assistant message #3, fix: grok headless runs die on first gated tool; map asking modes to auto and surface run errors in thread_status #4)
F-transcript: bounded window while streaming, content-visibility, no forced layout per push, idle cache write (PM2 Add settings toggle to isolate new threads in a git worktree by default #2, Grok provider runs fail immediately after first assistant message #3, Add Settle all action to the collapsed Settled tail bar #6)
F-provenance: incremental provenance (PM2 Solenta X launch trailer #1)
F-markdown: incremental parse of the streaming message (PM2 fix: grok headless runs die on first gated tool; map asking modes to auto and surface run errors in thread_status #4)

Next batch, after these land:

threads:changed row patches (PM1 fix: keep the sidebar's New-thread dropdown inside the app window #7, PM2 fix: assistant text between tool calls uses the full timeline column width #5)
files:resolve batching (PM1 fix: assistant text between tool calls uses the full timeline column width #5)
cold-start lazy views (PM2 fix: keep the sidebar's New-thread dropdown inside the app window #7)
localStorage hygiene (PM2 Thread runner kills agent child processes on turn settle; permission requests fail while settling #8)

No fix worker launches the app on cloned data. They measure with Store-on-a-copy benchmarks, unit benchmarks or the dev renderer.

Merge pull request #1476 from currentbits/coder/perf-pm1-measure-main…

Merge pull request #1477 from currentbits/coder/perf-pm2-measure-rend…

Merge pull request #1478 from currentbits/coder/perf-f-provenance-inc…

Progress:

Merged: perf(provenance): compute transcript provenance incrementally (#1475) #1478, which computes renderer provenance incrementally (PM2 Solenta X launch trailer #1).
CI pending:
- perf(boot): cached login PATH, async recapture; no sync git in projects:list / listBranches (#1475) #1479: login PATH is now captured async and some git calls are async (PM1 Grok provider runs fail immediately after first assistant message #3 and fix: grok headless runs die on first gated tool; map asking modes to auto and surface run errors in thread_status #4).
- perf(summaries): read lastActivity off the thread row, never a shard (#1475) #1480: threads:summaries no longer reads transcripts (PM1 Add settings toggle to isolate new threads in a git worktree by default #2). On cloned real data the first call went from 7.4 s to 4.7 ms, and the output is byte-identical across all 2,009 rows.
In progress: F-store (envelope rewrites, PM1 Solenta X launch trailer #1), F-transcript (window, layout, switch; PM2 Add settings toggle to isolate new threads in a git worktree by default #2, Grok provider runs fail immediately after first assistant message #3, Add Settle all action to the collapsed Settled tail bar #6), and F-markdown (incremental parse, PM2 fix: grok headless runs die on first gated tool; map asking modes to auto and surface run errors in thread_status #4).

Merge pull request #1479 from currentbits/coder/perf-f-boot-path-capt…

Merge pull request #1480 from currentbits/coder/perf-f-summaries-no-t…

Merge pull request #1481 from currentbits/coder/perf-f-store-stop-env…

Merge pull request #1482 from currentbits/coder/perf-f-transcript-win…

Batch 1 status: #1478, #1479, #1480, #1481 and #1482 are merged. #1483 is in CI, re-running a Windows-only failure.

Batch 2 started. Four workers, each owning separate files:

G-rowpatch: threads:changed sends row patches instead of the 2.68 MB full list (PM1 fix: keep the sidebar's New-thread dropdown inside the app window #7 / PM2 fix: assistant text between tool calls uses the full timeline column width #5).
G-resolve: batch files:resolve; today one thread open makes 424 IPCs costing 131 ms (PM1 fix: assistant text between tool calls uses the full timeline column width #5).
G-boot-renderer: React.lazy for rarely opened views, plus localStorage hygiene: drop archived threads from the boot snapshot and prune stale paneLayout keys (PM2 fix: keep the sidebar's New-thread dropdown inside the app window #7, Thread runner kills agent child processes on turn settle; permission requests fail while settling #8).
G-envelope: move hypotheses, suggestions and usageThreadsByDay out of the envelope, and load work logs lazily (PM1 #1c, Thread runner kills agent child processes on turn settle; permission requests fail while settling #8).

test(markdown): 1.5k fuzz mixes so the file fits Windows CI's 20 s li…

Merge pull request #1484 from currentbits/coder/perf-g-resolve-batch-…

Merge pull request #1483 from currentbits/coder/perf-f-markdown-incre…

Merge pull request #1488 from currentbits/coder/perf-h-mainheap-byte-…

Round 3 status:

perf(store): cap hydrated transcripts by bytes as well as count (#1475) #1488 merged: the hydrated-transcript LRU now has a byte budget; heap growth from opening the 10 largest threads drops 41%.
perf(transcript): tail-first thread switch + parsed-markdown cache (#1475) #1489 in CI: tail-first switch plus parsed-markdown cache. Headless big-thread switch 55 → 11 ms.
fix(sidebar): stop auto-animate leaking detached rows via IntersectionObserver (#1475) #1490 in CI: auto-animate IntersectionObserver leak fixed; detached sidebar rows were kept alive on every status change.

Deep dives started:

The transcript window sometimes stops trimming. In 2 of ~10 dev soaks DOM climbed as if perf(transcript): bound the tail window, stop forced layout per push, idle cache write (#1475) #1482 weren't there; suspected loss of stick-to-bottom state. Reproduce, root-cause, add a regression test.
Renderer heap growth in Electron: +12 MB over the soak in Electron, against ~1.5 MB in the browser. Heap snapshots on the PM3 zero-leak recipe, to name the retainers.

Merge pull request #1489 from currentbits/coder/perf-h-switch-big-thr…

Merge pull request #1490 from currentbits/coder/perf-h-heap-renderer-…

Merge pull request #1491 from currentbits/coder/deep-dive-transcript-…

Deep dives done.

fix(thread): a trim's scrollTop clamp no longer unsticks the transcript (#1475) #1491, merged. A trim's own scrollTop clamp unstuck the transcript, so the window never trimmed again. Pin loss went from 9/24 to 0/24 runs.
fix(transcript): a past-the-end window resets to the char-budgeted tail; Electron heap report (#1475) #1492, in CI. Electron heap root cause: the same pin loss. A 30-stream Electron soak went from 44 → 154 MB heap and 150 → 482 MB vmmap footprint once the pin was lost; with fix(thread): a trim's scrollTop clamp no longer unsticks the transcript (#1475) #1491 it stays flat at 39–46 MB heap and ~145 MB footprint.
- Also fixed: a past-the-end window after a retention-cap trim mounted 120 entries regardless of size (48k nodes).
- The rest of PM3's +12 MB is bounded: a one-off full first thread:updated, the parse cache reaching its budget, and V8 code space.

Left (bounded, small):

stale transcript closures keep a switched-away thread's copy (~4 MB);
the first thread:updated after launch is a full 4 MB transcript;
the index shift for scrolled-up readers on a retention trim.

Merge pull request #1492 from currentbits/coder/deep-dive-electron-re…

Done: measurements and fixes landed for 0.24.0 (e45a09a before/after, 8e77940 row patches, 8ad23c8 provenance, 5255a69 byte-capped hydration, e0f8561 tail-first switch). Closing in the 2026-10-09 board cleanup (#1532).

Build devCoder on first use, not at import (#1439) (#1584)

Long threads load in pages and render lighter, safer file restores, a…

Navigation Menu

Repository navigation

Performance round: measure on real data, then fix the top costs #1475

Baseline (2026-10-05, live data, read-only)

Phase 1: measure (no code changes)

Phase 2: fix

Incident during PM2 (resolved, no data lost)

PM1 results (main process): `docs/perf/2026-10-05-main-process.md`

PM2 results (renderer): `docs/perf/2026-10-05-renderer.md`

Phase 2: fixes running (six parallel workers, disjoint files)

6 remaining items

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Performance round: measure on real data, then fix the top costs #1475

Description

Baseline (2026-10-05, live data, read-only)

Phase 1: measure (no code changes)

Phase 2: fix

Activity

currentbits commented on Oct 5, 2026

Incident during PM2 (resolved, no data lost)

PM1 results (main process): docs/perf/2026-10-05-main-process.md

currentbits commented on Oct 5, 2026

PM2 results (renderer): docs/perf/2026-10-05-renderer.md

Phase 2: fixes running (six parallel workers, disjoint files)

currentbits commented on Oct 5, 2026

currentbits commented on Oct 5, 2026

6 remaining items

currentbits commented on Oct 6, 2026

currentbits commented on Oct 6, 2026

currentbits commented on Oct 9, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

PM1 results (main process): `docs/perf/2026-10-05-main-process.md`

PM2 results (renderer): `docs/perf/2026-10-05-renderer.md`