The solver only ran behind ./run.sh, had one media query in the whole
stylesheet, and kept the inventory in a 46vh scroll box that showed maybe a
third of the 43 dice. Setting it up on the PC did nothing for the phone.
Deployment
dice.kuhy.duckdns.org, behind the existing gitea-caddy edge. DuckDNS already
wildcards *.kuhy.duckdns.org (checked against 1.1.1.1 and 8.8.8.8), so no
dashboard step. deploy_build.sh builds to dist.staging and rsyncs into dist/
with --delete-after --delay-updates; plain --delete removes stale assets
before writing the new ones, which 404s a client mid-load. A post-commit hook
starts a MemorySwapMax=0 systemd user unit, so a rebuild can neither slow a
commit nor thrash zram.
Three traps found by driving the live URL in a browser rather than curling it:
a site address of "127.0.0.1:8089" reads as a HOSTNAME, so Caddy enabled
auto-HTTPS and started an HTTP->HTTPS redirect listener that bound *:80 next
to the edge -- SO_REUSEPORT would then split inbound HTTP and intermittently
break ACME for every domain here. The same address also made a Host matcher,
so the edge's forwarded Host matched nothing and served an empty 200 while the
local health check passed. And watching dist/'s own inode could never detect a
detached bind mount: rsync into an existing directory never changes it. Each
now has a guard that asserts positively rather than failing open.
Dice list
All 43 on screen at every width, no inner scrollbar. Column count and size
tier are chosen together from a container query on the panel -- not the
viewport, which mispredicts by 2x at the point the columns split. Two
invariants govern it: the list fits above the fold, and the name column stays
wide enough to tell Painter's die B, G and R apart. The left column holds the
grid alone, so it is sized to fill the height it owns: one --die-row value
drives the type, buttons and bar strip, giving a 52px row and 44px buttons on
a 1296px-tall screen and 20px rows on a phone.
Three safe-design rules are broken inside .die-list deliberately: sub-16px
body text, square icon buttons, sub-8px intra-cell gaps. 43 items with a name,
a chart and a counter in one viewport is arithmetically incompatible with 16px.
Import / export
A file, the clipboard, or a link carrying the inventory in the URL fragment.
One validator behind all four entry points including localStorage; it keeps
what it can and reports what it dropped, so a die added in a later game patch
costs you that entry, not the inventory. A link never silently overwrites what
is saved here -- it offers Keep or Discard, and only a link that actually
carries dice is applied at all, because one encoding a count of zero parsed
cleanly, blanked the screen, disabled Clear and left Keep writing the emptiness
over the visitor's own loadout.
session_autopsy
parse.py (551) and report.py (504) were over the repo's 500-line limit, so
test_file_length has been failing on main since 5ffd035. Extracted TurnTracker
and the two candidate-cell helpers. Report output is byte-identical on real
data apart from its timestamp, checked against HEAD in a worktree.
Verified: 288 TS tests and 911 Python tests, 100% coverage both, lint and
pre-commit clean, 8 viewports measured on the live HTTPS site, solver confirmed
through the real module worker.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Cache read is charged per API round-trip on the whole conversation so far, so a
long session's cost is driven by TURN COUNT, not message length. One measured
session spent 267M cache-read against 856k output — a ratio of 290:1. The report
priced tokens but never counted the turns spending them.
Across the 263 stored transcripts this now shows 33,508 tool turns at an 8%
batching rate, of which 4,437 (13%) decided nothing: polling a counter, running
a linter, checking whether a push landed. At 606k average context re-read per
turn that is ~2.7B cache-read tokens on turns a script could have taken.
- turns.py classifies a command as poll / lint_test / vcs_check / substantive,
deliberately conservative: state-changing commands (git commit, push) count as
substantive even though they are scriptable, so the reported waste is a floor.
A turn is only mechanical when EVERY command in it is.
- Batching is grouped by requestId, not by transcript line: one API response can
span several lines, and counting per line reports every turn as single-tool
(0% batching where the real rate is 12%).
- records.py gains TurnEfficiency (schema v3, additive — old records read as
zeros) with derived rates written into to_dict so readers need not recompute.
- report.py gains a "Turn efficiency" section pricing each mechanical kind.
- _TurnTracker owns both the totals and the open turn, after pylint flagged
_Accumulator over its attribute budget; _Accumulator.turns is now a property,
so every existing caller is unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Per user verdicts: the ⚙ statusline glyph told them nothing (removed on
the ~/.claude side; the plain-words SessionStart line is the channel) and
explain-diff is not wanted at all (skill deleted there, verdict 'dropped').
The count now means 'significant AND not yet handled' — resolving a
candidate visibly shrinks it (7 -> 4), and handled table rows stop
re-suggesting a review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The ⚙ statusline badge counted all 182 candidates — meaningless. It now
counts candidates ≥100k est weekly tokens (7 today): a number small enough
to mean 'worth your time'. token-savers.zsh gives the compiled scripts
shell entry points (finish-gate / phone-deploy / autopsy / yay-clean) so
the deterministic halves cost zero LLM tokens. Adds the first explanation
doc produced by the newly compiled explain-diff scaffold.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The transcript analyzer behind /compile-candidate lived only in the working
tree. Untracked meant unlinted — pre-commit only inspects files handed to it —
so nothing had ever checked it, and it had been abandoned mid-refactor with
every parse raising AttributeError: '_Accumulator' object has no attribute
'first_timestamp'.
The accumulator had been moved onto the grouped records dataclasses, but three
increment sites and parse_session still read flat attributes that no longer
existed. parse_session now hands over acc.meta and acc.counts directly; only
subagent_count comes from elsewhere, since no transcript line carries it.
Two modules broke the 500-line limit, so they were split along seams that were
already there. signatures.py takes the text-signature helpers — traces.py was
importing them from parse.py, making parse an accidental hub, and now neither
imports the other. stats.py takes the pure number and time helpers. parse.py
521 -> 458, detectors.py 559 -> 467.
The structural lint findings were refactored, not configured away.
_counter_candidates took eight arguments; the five naming which counter field to
mine and at what thresholds became a frozen _CounterSpec, so both call sites are
one line. _extract_skill carried in_span and a line budget as loop variables
across thirteen branches; a _SpanBuffer owns that state now, leaving the walk at
four. record() took twelve flat kwargs and now takes one override per record
block, with session_meta() so a partial meta override keeps the standard
timestamp — a bare SessionMeta(slug=...) would have nulled started_at and
quietly changed the per-week arithmetic. All 36 call sites rewritten by an AST
pass. tests/factories.py moved into conftest.py, where every other package here
keeps shared builders and where name-tests-test does not object.
965 tests pass at 100% statement and branch coverage, every hook is green, and
no suppression was added anywhere. `candidates` and `traces skill-finish` both
run against the real corpus.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
