[eric-engberg]

What

When Claude Code's payload has no rate_limits reset time (API-key, Bedrock and Vertex logins, accounts without a five-hour window, or a failed usage fetch), Block Timer, Block Reset Timer and Block % find the current 5-hour block from the transcripts. This fixes four ways that fallback got the block wrong:

• Rows Claude Code writes for rejected requests no longer count as block activity.
• A gap in a resumed session is no longer taken as the block boundary while a transcript the scan skipped could fill it.
• Transcripts are found when the config directory's name has glob characters, e.g. ~/claude (work)/.claude.
• Block cache entries written while the system clock ran ahead are no longer trusted after it is set back.

Why (evidence)

1. Rejected requests. For a rejected request (a usage limit, an expired login, an outage), Claude Code writes {"type":"assistant","isApiErrorMessage":true,"message":{"model":"<synthetic>","stop_reason":"stop_sequence","usage":{"input_tokens":0,"output_tokens":0,...}}}. The scan only checked that input_tokens/output_tokens are numbers.
   • With only rejected requests in the last 5 hours, it showed a running block (e.g. "Block: 3hr 49m | Reset: 1hr 10m") that doesn't exist.
   • With rejections before real work, the block started at the first rejection's hour, so the reset was reported hours early.
   • Either wrong start was cached for 5 hours. The token/context metrics and Cache Timer already skip these rows.
2. Skipped transcript. The scan reads in widening passes (10h, 20h, 48h). Each pass reads only the transcripts last written after its cutoff, but keeps all of their rows, so a session left open since yesterday brings in rows from long before the cutoff.
   • Example: session A has one row 30h ago and then a row every 30 min from 9h ago; another project's session B has rows 12h–10.5h ago and was last written 10.5h ago.
   • main reads only A, takes A's 21h gap as the boundary, never runs the wider passes, and anchors the block 2 hours early: "Block: 4hr 54m | Reset: 5m".
   • With B's mtime moved to 9.5h ago, the same code gives the correct "2hr 57m | 2hr 2m". Only the mtime differs.
3. Glob characters. The pattern was <config dir>/projects/**/*.jsonl with the absolute directory written in. Under claude (work), cfg[1], brace{a}, star*x or q?mark it matched nothing, so Block Timer stayed at "0hr 0m" and Block Reset Timer at [Loading] for as long as that directory was in use.
4. Clock set back. A "no block" entry (written one minute ahead) was trusted whenever now < emptyUntil. After a 24h-fast clock was corrected, it hid the block for a day, even with fresh activity. A cached start in the future counted as valid (now <= start + 5h), freezing elapsed time at 0 and Reset at 5hr. The usage lock in usage-fetch.ts already bounds its deadline against a clock set back.

How

• getAllTimestampsFromFile skips rows with isApiErrorMessage === true, the same check the token/context metrics and Cache Timer use. Every main-chain <synthetic> row in a sample of real transcripts carried the flag and all-zero usage (session limit, spend limit, model limit, expired login, not logged in, sleep, unreachable API). This makes Block Timer differ from ccusage blocks, which counts these zero-token rows as entries; the scan's own comment says to treat only entries with real token usage as activity.
• In the gap search, a gap of 5h or more is a boundary only if the part after the pass's cutoff spans 5h by itself. A transcript the pass skipped was last written before the cutoff, so it can only add rows before it. Otherwise the next pass reads it first, and the last pass takes any gap, as before.
  • Cost: in that layout, one more pass on a cache miss, which happens about once per block. It reads the transcripts written in the last 20h instead of 10h. On a synthetic 291 MB set that is 82 → 183 ms.
  • Layouts where the first pass finds a real gap, or no gap, are unchanged.
• The glob pattern is the relative projects/**/*.jsonl, with the existing cwd: rootDir. tinyglobby escapes and normalizes the cwd itself, so Windows paths keep working.
• Left as is: a backslash in the config directory's name on macOS or Linux. tinyglobby turns every backslash in the cwd into /, so a directory named back\slash still finds no transcripts, as on main.
• getCachedBlockMetrics trusts emptyUntil only while it ends within one interval, and a cached start only once it has begun. Otherwise it rescans. (If a transcript itself holds rows stamped by the fast clock, the fresh scan still waits for them to age out; that's outside the cache.)
• Unchanged: the lookback windows, how blocks are chained, the cache format, and the sidechain rule.

Demo

Piped renders of Block Timer and Block Reset Timer over fake transcripts in a scratch HOME. The payload has no rate_limits, there's no login and nothing is fetched. Times are relative to when each take was recorded.

Rejected requests. Line 1: only rejected requests, so there is no block. Line 2: rejections, then real requests from just after midnight, so the block is 00:00–05:00.

Powerline

Before:

After:

Plain

Before:

After:

A transcript the first pass skips. A resumed session plus another project's session last written 10.5h ago, whose rows chain into the current block.

Powerline

Before:

After:

Plain

Before:

After:

Config directory with parentheses. The same transcript under ~/claude-work/.claude and ~/claude (work)/.claude.

Powerline

Before:

After:

Plain

Before:

After:

Testing

• 12 new tests: 5 config-dir names, 2 for rejected requests, 2 for the lookback gap, 2 for the cache bounds and 1 for the other rows the scan skips. 10 fail on main. The other two check that a gap no transcript fills is still a boundary, and that unparseable lines, rows without both token counts, sidechain rows and rows without a valid timestamp still don't count as activity; both pass on both.
• bun test: 3180 pass, 0 fail (main: 3168), Bun 1.4.2. bun run lint is clean.
• Node v26.10.0 (vitest) on the two changed test files: 36/36 (main 24/24).
• Built dist under Bun 1.4.2 and Node 26.10.0: the baseline line matches main; a Block Timer + Block Reset Timer line over each of the three scenarios agrees across runtimes and differs from main as intended; the TUI opens and exits cleanly.