[eric-engberg]

What

• Output, Input and Total Speed with a window set (1-120s) now read the speed the responses actually ran at, instead of up to hundreds of times more.
• On the Node build (npm/npx), long sessions no longer lose their transcript widgets. Previously, once one transcript passed about 125k usage rows, any speed widget turned Tokens to 0, speed to "—", Compaction Counter to a false ↻ 0 and Session Name to blank. When neither Claude Code nor the usage API gave a 5-hour reset time (so the transcripts are scanned for the block), Block Timer, Block Reset Timer and the block usage-percent fallback showed no block, with or without a speed widget.

Why (evidence)

Windowed speed. buildSpeedMetrics picks each request whose response lands inside the window and adds all of its tokens, but divides by only the part of its user-to-response interval inside the window. A response longer than the window, or a tool result arriving soon after one (it moves the window's end), leaves a thin slice of time against the whole request's tokens. For one 90s response at 50 t/s:

Transcript	Session avg	10s	30s	60s
tool result 9.9s after the response	50.0	45.0k	223.9	89.8
response is the last row	50.0	450.0	150.0	75.0
agentic loop: 20s responses, 3s tools	50.0	142.9	83.3	58.8

Long transcripts. mergeCollectedSpeedMetrics does requests.push(...part.requests) and the block scan does timestamps.push(...fileTimestamps). Both pass one argument per usage row, and the call throws "RangeError: Maximum call stack size exceeded" past about 125k arguments under Node 26 (works at 124k, throws at 126k), or about 640k under Bun 1.4 (works at 600k, throws at 700k).

• The speed merge runs inside the transcript scan's single try, whose catch returns the empty result, so everything already gathered is lost.
• The block scan's catch returns "no block". That result is cached for only 60s, so the failing full scan runs again every minute.

With a 140k-row transcript, the built dist under Node prints Total: 0 | Out: 0 | Out: — | ↻ 0 | Block: 0hr 0m. The same input under Bun, or under Node without the speed widget, shows the tokens (Total: 4.2M | Out: 2.8M …); without the speed widget, Node still shows no block. At 700k rows Bun fails the same way.

How

• Each request in a speed window keeps the share of its tokens that matches the share of its interval inside the window. A single request reads its own speed at any window size, and several requests average by the time each spent inside the window, so the window stays a time window. The subagent window test expected the old result (15 t/s for a 12 t/s request) and now expects 12 t/s.
• Both merges push one item at a time.
• fix: count each API response once in speed and token metrics #622, fix(jsonl): dedup cumulative token totals by message.id #725 and fix(speed): count each API call once in token speed widgets #735 already count each API response once; this keeps that and changes only how a windowed speed counts a response that started before the window. Three of fix: count each API response once in speed and token metrics #622's window tests expected a response's full tokens over its clipped time, so their windowed numbers now take the in-window share (for example 150 in / 90 out becomes 125 / 75 for a 6s response with 5s inside the window). Their session totals don't change.

Demo

Four Output Speed widgets (session average, 10s, 30s, 60s) on fake transcripts where every response runs at 50 t/s.

Powerline

Before:

After:

Plain

Before:

After:

A long session's 140k-row transcript through the built dist under Node (npx) and Bun (bunx). Each render starts from an empty cache.

Powerline

Before:

After:

Plain

Before:

After:

Testing

• On main as of [low] perf(claude-status): hold a lock while the status page fetch is in flight #635 and [medium] perf(git-review): key the PR cache through the cached runGit #639 (62a371b): merges cleanly, bun run lint is clean, and bun test gives 3706 pass, 0 fail. The built CLI under Bun 1.4.4 and Node 26.13.0 matches main in plain and Powerline mode with the baseline configs, and the TUI opens and exits cleanly. This branch's tests against main's code: 9 fail, 2 for the long transcripts and 7 for windowed counts (including the three fix: count each API response once in speed and token metrics #622 window tests above).
• New tests: a request that straddles the window (10s and 40s windows), a 700k-row transcript through getTranscriptAnalysis, and a 700k-row transcript through getBlockMetrics. Together with the updated subagent window test, all 4 fail on main under both Bun and Node and pass here. The large-transcript tests use 700k rows so they fail under Bun too; each adds about 0.3s under Bun and 1s under Node.
• bun test: 3171 pass, 0 fail. bun run lint: clean.
• Vitest under Node: 2962 passed; the 186 failures are the same tests that fail on main. Both changed test files pass 38/38 under Node and Bun.
• The built dist was checked under Bun 1.4.2 and Node 26.10, plain and Powerline, with the default payload and with a 140k-row transcript. Every runtime agrees on this branch; on main they disagree for the large transcript.