01 — the work diary
What the service says
Live from /v1/report/summary and /v1/report/sessions on
. If this errors, it says so — the old dashboard used to
fabricate a plausible week whenever the API failed, which is the habit this rebuild
exists to break.
By ticket
By source
Agent time is tagged separately at the point it is earned, so it can be priced or excluded later without going back through the records.
Sessions
02 — in jira
One entry that grows while you work
The old extension posted a fresh worklog every ten minutes — about 48 a day, 240 per ticket-week. This posts once and then updates the same entry, so the ticket shows a timer rather than a wall of noise. Progress notes accumulate inside its comment instead of becoming a comment each.
Rendered from the ADF the real comment builder produced. The
[33t #…] marker is how a timed-out create is recognised on retry
instead of being logged twice — and how reconciliation knows to leave
hand-logged entries alone.
03 — on the ticket
The button
Injected on Jira issue pages. It reads the key from the URL, so it survives Jira renaming its own class names. Press it — these are the three states it has.
It never silently stops a timer you might be relying on. Switching is always something you chose.
04 — from a terminal
And without a browser
$ 33t start NAM-51
✗ NAM-49 has been running for 1h 12m.
33t start NAM-51 --switch to stop it and start this one
$ 33t start NAM-51 --switch
⏹ NAM-49 1h 12m logged
▶ NAM-51 0m
$ 33t note "absolute-total PUTs so a missed flush self-heals"
✓ noted on NAM-51
$ 33t status
▶ NAM-51 34m worklog projector
• absolute-total PUTs so a missed flush self-healsThe same API backs an MCP server, so an agent can run its own timer and append what it did. No tool it can call accepts a duration, a timestamp or a worklog id — it can describe work, but only elapsed wall clock bills for it.
05 — what it will not claim
The number that used to be a lie
Every old worklog read 0 keystrokes · 0 clicks · 0% active, because the
counter only saw three domains and none of them were an IDE. That percentage is
removed rather than fixed. What replaces it is evidence anyone can check: commits,
files touched, agent tool calls, idle minutes, and which client earned the time.
A sleeping laptop credits nothing, because its active counter did not move.
A client offline for thirty minutes recovers at most thirty minutes, whatever its journal claims.
A Jira outage costs a delay. Time is committed locally before Jira is ever told.
Edit an entry by hand and it stays edited — the tracker detaches rather than overwriting you.
The CLI bills declared time, not observed time. A terminal cannot see a keyboard, and the design doc says so rather than implying otherwise.