fix: one unprobeable .jsonl no longer blanks the whole usage dashboard - #79

大家好

Problem

sortUsageFilesByEarliestTimestamp maps readEarliestTimestamp over every
discovered .jsonl via mapWithConcurrency, with no per-file error handling.
readEarliestTimestamp throws TimestampProbeLimitError when a file has more
than 1 MiB of leading lines that carry no timestamp field (covered by the
existing "timestamp probing is bounded when no valid timestamp exists" test).

That rejection is unhandled, so Promise.all rejects, the error unwinds out of
loadUsageRecords, and the outer catch returns { records: [] } with
filesFailed = Math.max(filesFailed, 1). A single such file empties the entire
file-based dashboard on every refresh. Diagnostics report failed=1 with
bytes=0. The API quota path is separate and keeps working, so only the
usage/history panel goes blank.

The file type that triggers this in normal use is a workflow journal.jsonl:
its records contain type, key, result, and agentId, never a
timestamp, and a workflow run with a few hundred sub-agents produces a
journal.jsonl over 1 MiB.

Origin

This surfaced from Claude Code's Workflow feature (multi-agent orchestration).
Each run writes a journal at
~/.claude/projects/<project>/<session>/subagents/workflows/<workflow-id>/journal.jsonl,
one line per sub-agent lifecycle event: a started line and a result line
per agent, each holding type, key, result, and agentId and no
timestamp. The result values make some lines large. A single run that
fanned out to roughly two hundred sub-agents produced a 1.32 MB, 402-line
journal, over the 1 MiB probe limit. From then on every refresh returned empty
records (failed=1, bytes=0) and the dashboard stayed blank while the
status-bar quota kept updating, which presented as a rendering fault rather
than a load abort.

Reproduction

1. Create a .jsonl under ~/.claude/projects/<project>/ whose first 1 MiB
   contains no line with a timestamp field:

   node -e 'const fs=require("fs"),os=require("os"),p=require("path");const d=p.join(os.homedir(),".claude","projects","repro");fs.mkdirSync(d,{recursive:true});fs.writeFileSync(p.join(d,"journal.jsonl"),`{"type":"result","key":"k","agentId":"a"}\n`.repeat(40000))'
   

   Equivalently, run any workflow whose
   subagents/workflows/<id>/journal.jsonl grows past 1 MiB.

2. Open the extension or trigger a refresh.

3. The dashboard is empty. Show Diagnostic Logs reports failed=1 and bytes=0
   on every refresh, while the status-bar quota continues to update.

With the fix, the same setup loads the remaining files normally and the
dashboard populates; the oversized file is skipped for ordering.

Fix

Catch the per-file probe failure in sortUsageFilesByEarliestTimestamp and fall
back to timestampMs: 0, so the file sorts by discovery order — the same value
readEarliestTimestamp already returns for a small file that has no timestamp.
One unprobeable or transiently unreadable file no longer aborts the batch, and
the file is still passed to the main reader.

readEarliestTimestamp and its guard test are unchanged.

Test

Adds "a single unprobeable file does not abort ordering the rest" to
src/test/claudeUsageFiles.test.ts: a >1 MiB no-timestamp journal.jsonl
beside a normal file. It asserts readEarliestTimestamp still rejects on the
journal, and that sortUsageFilesByEarliestTimestamp resolves and returns both
files.

CHANGELOG.md is updated. package.json version is left unchanged per
CONTRIBUTING.

谢谢