Claude quota: 4th outage in 9 days, 2nd weekly-limit — capacity arithmetic

[worktrunk-bot]

The bot hit a Claude weekly limit again on 2026-08-11, the fourth quota outage in nine days and the second weekly-flavour one. The 2026-08-05 sweep armed a watch for exactly this — "if a second weekly exhaustion lands within ~14 days, the cascade cost stops being an accepted architecture cost and becomes a capacity finding worth surfacing with the arithmetic" (#3691) — so here is the arithmetic. Nothing is broken; this is a capacity call, and it's yours to make.

The outages

Issue	Started	Flavour	Duration	Cost
#3715	2026-08-03 00:57Z	session	8h 20m	drained, no work lost
#3737	2026-08-04 08:39Z	weekly	15h 10m	3 stranded tend-review runs re-run by hand
#3780	2026-08-08 22:23Z	session	1h 32m	you re-ran 31281595492 by hand; without it #3779 merges unreviewed
#3800	2026-08-11 14:57Z	weekly	≤43m	none — self-healed

Yesterday's was the mild one: two runs died, and both recovered on their own. The tend-review for 581e1129f was picked up 8h later by 31545932359, whose review closes both prior findings by name. It's the frequency that's the signal, not this instance — 08-04's weekly window cost you three manual re-runs.

The spend behind it

Per-sweep API-list totals recorded in #3691 (each figure is a 24h window ending ~08:30Z, not a calendar day):

08-01	08-02	08-03	08-04	08-05	08-06	08-07	08-08	08-09	08-10	08-11	08-12
$51.81	$86.04	$49.74	$122.46	$38.21	$95.10	$105.40	$82.84	$92.24	$166.96	$60.68	$46.95

The seven windows ending at yesterday's outage total $641.43. These are API list prices — a large multiple of the effective subscription rate, so treat them as a consumption proxy rather than a bill, but the ratio between days and between workflows holds.

Today's $46.95 is the lowest in the log and splits as tend-mention $19.22 (14 runs) / tend-review $14.47 (8) / tend-nightly $8.78 (1) / tend-notifications $4.48 (5). Note the last one: 47 notification ticks ran, only 5 started a session, so the half-hourly cron is not the driver. 72% of spend is the mention+review loop on pull requests.

The cheapest lever is already merged, just not shipped

Three upstream fixes that cut no-op or duplicated sessions all merged on the evening of 2026-08-11, and all three are stranded behind the same release pin — every workflow here pins max-sixty/tend/claude@0.1.14, released 2026-08-06:

• max-sixty/tend#849 — skips the synthetic review container the bot's own inline reply creates. That class fired 8 times in the 08-11 window alone, each a full session that read the thread and correctly did nothing.
• max-sixty/tend#870 — stops nightly and tend-mention both actioning the same review. One observed instance burned $2.14 before the collision surfaced.
• max-sixty/tend#851 — the outage-drain recipe, which is what makes a stranded tend-review recover without you re-running it by hand.

So a tend release plus a pin bump reclaims measurable spend and closes the manual-rerun cost in the table above, with no change to what the bot does. Until then #3781 holds the drain guidance locally — I corrected its removal marker today, since "#851 merged" reads as the drop condition but the released action doesn't carry it yet.

Beyond that the levers are yours to weigh, and I'd rather name them than pick: the five open bot PRs each keep drawing re-reviews while they sit, so merge cadence and review spend are coupled; and the self-review depth on bot PRs (nine runs / $16.17 on #3799) is productive per round but is the largest single knob.

[worktrunk-bot]

Triaged as a capacity report — nothing to reproduce, and no worktrunk-side change to make yet. Two things the body doesn't have:

The upstream half is already filed, with disjoint evidence. max-sixty/tend#954 opened five minutes before this issue asking for the same release, but argues entirely from tend's own runs and names three different stranded fixes: tend#903 (review sessions cancelled mid-flight by a sibling push), tend#818 (outage annotation that names no cause, so a quota exhaustion surfaces only as exit=1), and tend#825 (a credential-environments false positive). Add this issue's tend#849 / tend#870 / tend#851 and six distinct fixes are stranded behind 0.1.14, which neither issue shows on its own.

That third one is worth a look on this side: #3729 tracks a credential-environments FAIL on this repo, refreshed nightly. tend#825 fixes that check's false positive when an all-tags ruleset has no bypass actors (tend#824) — I haven't confirmed worktrunk's ruleset matches that shape, so I can't say the release clears #3729, only that it's the first thing to re-check after the bump rather than re-diagnosing it another night.

The pin bump is still blocked, and it's a one-line change when it isn't. gh release list -R max-sixty/tend still returns 0.1.14 (2026-08-06) as latest — I re-checked at triage. All eight .github/workflows/tend-*.yaml here pin max-sixty/tend/claude@0.1.14 with no hand-maintained drift, so the moment a tag lands this is one sweep across those eight files, and #3781's removal marker becomes droppable at the same time.

The one thing I'd move now: the adopter-side arithmetic in this issue ($641.43 across seven windows, 72% in the mention+review loop) is precisely the evidence tend#954 lacks, since it makes the case from tend's own repo where no adopter spend is visible. Scope rules stop me commenting on another repo's thread uninvited. Two ways to unblock, whichever you prefer — I add it to tend#954 on your say-so, or we widen the running-tend exception for max-sixty/tend, which today covers filing new issues but not commenting on existing threads.

Triage checks

• Classification: capacity report (not a bug/feature/question) — no reproduction attempted, no fix PR opened.
• Duplicate scan: no other open worktrunk issue covers the quota/spend topic; #3691 is the tracker this reports into, not a duplicate.
• gh pr view <n> -R max-sixty/tend confirms 849 / 870 / 851 all merged 2026-08-11 23:03–23:48Z, i.e. after 0.1.14 was published 2026-08-06T00:36Z — so all three are genuinely on main and genuinely not on the fleet.
• tend#954's own claims were independently checked by tend-agent in-thread (72 commits ahead, generator/pyproject.toml still 0.1.14); I did not re-verify them here.
• Not verified: whether #3800's run annotation shows the bare exit=1 form tend#818 fixes — its issue body is the failure table only, and I didn't pull the run annotation.
• Pin inventory: rg -o 'max-sixty/tend/[a-z-]+@[0-9.]+' .github/workflows/ returns eight hits, all claude@0.1.14.