Metric declare/update/remove, backfill, and per-line apply_all_metrics
now run on the profile actor (engine::profile) instead of inline in the
primary's ingest closure, per idea 0001's resolved decision that the
primary handles params only. The profile actor reuses EngineSnapshot as
its own snapshot type -- declare_metric/update_metric/apply_all_metrics
work unchanged -- and mirrors the primary's param declarations into its
own snapshot on every wake (and synchronously before handling a
DeclareMetric/UpdateMetric message, since there's no ordering guarantee
between the watch-wake and message branches in tokio::select!) so
metric label/value resolution keeps working without storing a second
copy of raw log text; backfill reads lines through its cloned primary
IndexReader instead.
EngineHandle gained a profile_reader field so MetricDeclarations/
BucketValues/Promql queries route to the profile actor's snapshot
instead of the primary's -- without it, metrics could be declared but
never queried, since the primary no longer computes or stores them.
DeclareMetric/UpdateMetric/RemoveMetricDeclaration messages are still
sent to the primary's mpsc channel and proxied to the profile actor by
handle.rs's run() loop; EngineHandle's write-side API is otherwise
unchanged (ticket 106 gives API handlers a direct route). Alert
evaluation still reads the primary's (now metric-less) snapshot
directly -- an accepted gap until ticket 105 moves it, per idea 0001's
resolved decision that alerts move with metric computation.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
Adds a second actor task with its own left_right-backed ProfileSnapshot
and a tokio::sync::watch the primary's IndexWriter bumps after every
publish, so the profile actor wakes and re-reads newly-visible primary
lines through its own cloned IndexReader instead of polling. No metric
computation, alert evaluation, or query routing moves yet -- this is
pure plumbing, proven with a test that ingests a line on the primary
and observes the profile actor's watermark advance in response.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
Raw storage, log-stream header derivation, and auto-param detection
were inlined directly in the ingest actor's write closure alongside
metric computation. Pulling them into their own method gives the
ingest arm two clearly separated, independently-callable stages —
param extraction (ingest_params) and metric computation
(apply_all_metrics) — with no behavior change, in preparation for
splitting these onto separate actors (ticket 102).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
DataValue::LogLiteral and IngestedLine now carry an Arc<str> instead
of an owned String, so the two clones the ingest path paid per line
(once to move text into the write closure, once to store it) become
cheap refcount bumps instead of full string copies. Also cuts the
per-string-content cost of left_right's full-EngineSnapshot deep clone
on each reconciliation sync.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
The PromQL layer only had a fixed-width histogram() (per-bucket
counts), with no way to ask for p50/p95/p99. Adds
histogram_quantile(p, metric[range]), interpolating over each
series's retained raw sample values (linear interpolation between
closest ranks) rather than building a separate pre-aggregated
bucketed-histogram representation — exact for the data actually kept.
An optional bracketed range narrows the computation to a trailing
window, mirroring rate()'s range selector.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
Metrics previously only declared a data type (String/Number/...), so
rate() could be applied to a point-in-time gauge reading and sum()/avg()
could be applied directly to a cumulative counter reading with no guard
against either. Adds a MetricSemantics enum (Gauge, default; Counter)
threaded through DeclareMetricRequest, the engine's MetricDeclaration
data value, profile persistence, the API/MCP declaration routes, and the
Metrics panel's declaration form, then uses it in evaluate_promql to
reject rate() on a declared gauge and bare sum()/avg() on a declared
counter (suggesting rate() first). The guard only applies to metrics
with a captured value_param — bare occurrence-count metrics are always
safe under both operations, so existing dashboards are unaffected.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
Formats the whole backend to rustfmt defaults and fixes all 23
outstanding clippy warnings (collapsible ifs, clone-to-slice-ref,
type-complexity, single-arm match, unnecessary is_none check), then
wires cargo fmt --check and -D warnings into make lint so both stay
enforced going forward.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
A log line whose text parses as JSON was rendering every nested node
fully expanded, dumping large payloads into the timeline instead of a
collapsed summary the user opts into expanding.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
declare_metric and update_metric duplicated the resolve-params +
backfill_metric + insert-MetricDeclaration sequence; extracted into
resolve_backfill_and_store_metric so the declare-vs-rename difference
is expressed only in caller-side logic. Also extracted
auto_declare_params's per-candidate insertion body into
try_insert_auto_param so the candidate-building and insertion steps
read as separate concerns.
Reimplemented directly against current main (ticket 085) after the
original worktree agent's branch was based on a stale pre-consolidation
snapshot module layout that no longer matches main.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
The Logs and Timeline arms of handle_query_inner each duplicated the same
"compile filter, bail out on error" block; the Logs arm also inlined its
sort-helper-index walk. Pull both into named helpers (compile_optional_filter,
logs_sorted_by_param) so the Logs arm reads as "compile filter, then dispatch
to sorted or unsorted." Pure decomposition, no behavior change.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
run() bundled settings loading, engine spawn, active-profile resolve/load,
ingest spawn, and session registration inline. Extract each into a named
function (spawn_engine, load_active_profile, spawn_ingest,
register_session_and_log) so run() reads as a sequence of steps. No
behavior change.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
Extract fetch_metrics_body(client, url) -> Result<String, String> to
handle the request/error_for_status/text chain with `?`, centralizing
the three warning messages. run_metrics_poll now matches once on the
result instead of nesting three match arms deep on the happy path.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
They were scattered across filter/metrics/params/query/sort.rs by an
earlier file-splitting pass. Files can be split as much as needed, but
a single type's impl should never be split across files — so all 15
impl EngineSnapshot blocks now live in mod.rs, while each submodule
keeps its own types, free helper functions, and tests.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
Log reads (logs_since, the unsorted Logs query path) now seek the
IngestedLog range directly from since_sequence/before_sequence instead
of scanning from IndexKey::Start and filtering every prior entry out
in a closure. Metric-comparison filters and bucket_values now seek to
a single metric's Bucket key range instead of scanning every declared
metric's buckets. evaluate_promql sits on top of bucket_values, so its
per-poll dashboard evaluation is narrowed by the same fix without
further changes; caching prior evaluations across polls is left for a
follow-up if profiling ever shows it's needed (per ticket 081's own
scope, not built speculatively).
Resolves tickets 077, 078, 081.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TeyhRfMYjMditek6geDRoQ
Pure code-motion refactor, no logic changes. Follows the repo convention
of one struct/enum + its impl per module file:
- engine/handle.rs: the left-right write/read plumbing (EngineOp,
IndexWriter, IndexReadHandle, IndexReader, EngineHandle) and the actor
loop (spawn/run) that owns EngineSnapshot.
- engine/message.rs: EngineQuery/EngineMessage/EngineResponse.
- engine/snapshot/: EngineSnapshot itself, split by concern into
params.rs (declaration + auto-detect), metrics.rs (declare/apply/
backfill), filter.rs (search-bar filter compilation/eval), sort.rs
(sort-helper indexes), and query.rs (handle_query_inner + PromQL
evaluation), with mod.rs holding the struct and the accessors shared
across all of them.
- engine/tests.rs: the inline test module, kept as one file (per the
task's own escape hatch) since most tests exercise ingest -> declare
-> query across several of the new submodules at once rather than one
in isolation.
Visibility was narrowed per call site: pub(super) where only sibling
snapshot submodules need an item, pub(crate) where engine::handle or
engine::tests (outside the snapshot subtree) also need it.
cargo build --all-targets and cargo test are green (291 passed, 1
ignored, same as before the split).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
- src/api/: HTTP router + DTOs + the MCP server (the HTTP/MCP-facing
surface): mod.rs (formerly api.rs), dto.rs, mcp_server.rs.
- src/cli/: argument parsing and process entry (formerly inline in
main.rs) plus session registry: mod.rs, sessions.rs. main.rs is now
a thin entry point that just declares modules and calls cli::run().
- src/engine/: the actor + everything it depends on to parse/index/
query log lines: mod.rs (formerly engine.rs), alerts.rs, promql.rs,
rate.rs, ingest.rs, logfmt.rs, filter_parser.rs/.pest,
filter_datetime.rs, index_key.rs, data_value.rs, metric_primitive.rs,
timestamp.rs.
- src/config/: persisted configuration: mod.rs (formerly settings.rs),
profile.rs.
- src/consts.rs stays at the crate root, used from every module above.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
The engine actor and API handlers previously called left_right's
WriteHandle/ReadHandleFactory methods (append/publish/enter/handle)
directly. IndexWriter now exposes mutate()/peek() and IndexReader
exposes handle() (returning an IndexReadHandle with its own enter()),
so how the single-writer/multi-reader index is accessed is named and
scoped to these two types instead of being raw left_right calls
scattered through engine.rs/api.rs/profile.rs.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
Profile (legacy-shape migration, load_into_engine, load_legacy_store),
MetricDeclarationRecord/Dashboard seed constructors, PromqlQuery::parse,
LogStreamHeader (detect/split_row/column_pattern/message_pattern), and
api.rs's EngineHandle::ask/*Error::into_response/ParamDeclarationDto::reserved
all had a single obvious receiver and are now methods on it instead of
free functions taking that type as a parameter.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
Free functions that took the engine's index/sort-helper-index state as
a parameter now live on EngineSnapshot (or, for a handful of clean
single-receiver cases, on MetricPrimitive/CompareOp/CompiledFilter/
EngineQuery). Route handlers in api.rs and alert evaluation in
alerts.rs are updated to call the new methods; alerts.rs now threads
an &EngineSnapshot through instead of a raw &BTreeMap so the per-line
alert backfill path doesn't clone the whole index just to call a method.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
Centralizes engine tuning values, poll intervals, logfmt column
regex fragments, and app/settings identity strings that were
previously defined ad hoc across engine.rs, ingest.rs, logfmt.rs,
profile.rs, settings.rs, and main.rs.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
- 076: gate auto-declared log-stream column/message params
(`ParamDeclaration::log_stream_gated`) behind `is_log_stream_line` at
extraction time, since their fixed-offset regex has no content anchor
and previously extracted garbage from continuation lines and the
header row itself (the regex crate doesn't support lookaround, so
this couldn't be baked into the regex text).
- 074: end-to-end regression test confirms sorting a real `log stream`
capture by a log-stream column (e.g. PID) now actually reorders
results, since 076's fix removes the garbage values that were
corrupting the sort helper index; also add a "Building sort
index…" indicator in the timeline UI for a slow first-time sort.
- 072: attach the reserved `ottyl_fmt` pseudo-param to each
`LogEntryDto`'s params so it can be picked as a Log Preview column
(previously listed as pickable but always blank).
- 073: track a rolling 5s high-watermark of ingestion-to-engine-processing
lag and surface it in the nav stats popover next to approx_bytes_total.
- 071: group the regex(...) help button back with the filter bar's other
controls, and restructure the Filtering docs section into scannable
per-feature subsections (also fixes stale log_type/process/pid mention).
Tickets 070 (CLI scoping) and 075 (API unresponsiveness investigation)
are left in place per their own scope — both are explicitly
conversation/investigation tasks, not implementation tickets.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
Adds a sessions/<name>.json registry (068) so a running ottyl instance can
be discovered by name; an `ottyl mcp --session <name>` subcommand (067)
that starts a short-lived MCP server over stdio (via the official rmcp
SDK), proxying query_logs/declare_param/declare_metric/create_dashboard/
add_panel/get_dashboard tool calls onto the existing HTTP API; and a
SKILL.md (069), embedded into the binary and served at GET /SKILL.md, that
documents how an agent connects to and drives a running instance.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
- 059: log stream header/column detection is now generic
whitespace-columnar splitting (2+ spaces, single tab) instead of one
fixed six-column regex. Declared params come from whatever columns
the header actually names, plus a trailing unlabelled message field
that's never swallowed by the last named column.
- 060: timeline sort-by-property now walks an incrementally-maintained
per-sort-key helper index (BTreeMap keyed so "has the param, ordered
by value" sorts before "missing it, ordered by ottyl_seq") instead of
materializing and re-sorting the full filtered range on every query.
- 061: helper indexes are swept for staleness on ingest and dropped
after 5 minutes unused; a purged (or never-built) index transparently
rebuilds via the same lazy one-time scan on the next sorted query.
- Also invalidate a param's helper index on manual re-declaration,
removal, or ClearParams, so a redefined regex can't leave a sort
helper index serving values derived under the old declaration.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1ZJ3SrayzLzVndhVq9BmS
A rate() series over a metric whose occurrences are confined to a short
burst within a much wider queried range (e.g. sum by (code)
(rate(KeyPressed[5m])) against a ~10h log where all key presses happen
in the first 10s) rendered as broken: the line simply stopped at the
last real occurrence, and the chart's x-axis auto-scaled to the tiny
span of actual data points rather than the queried range, silently
narrowing a 10-hour view down to ~10 seconds with no indication anything
was off.
Backend (src/engine.rs): windowed_rate_series now appends a synthetic
boundary point at start_ms/end_ms (computed via the existing
rate::rate_at against the original occurrence points, so a boundary
inside a still-warm window gets its true partial value rather than a
hardcoded 0) whenever the real series doesn't already reach that edge.
Wired into both the bare rate() and sum()/avg(rate()) evaluate_promql
branches, which already had start_ms/end_ms in scope.
Frontend (PanelChart.tsx/Dashboards.tsx): the line chart's XAxis now
uses an explicit numeric domain pinned to the panel's queried range
(passed down from Dashboards.tsx's PanelPreview) instead of recharts'
default auto-scale-to-data-extent, so sparse data can no longer make
the axis look narrower than what was actually queried.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
GET /api/params now includes the three non-deletable ottyl_* reserved
entries (ticket 055), so the "clear removes everything" assertion needs
to count only real (non-reserved) declarations.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
Resolved conflicts with ticket 051's unknown-metric guard in evaluate_promql
(reordered so ottyl_fmt's categorical rejection and the reserved-name
skip both run before 051's declaration/bucket-data check, since
ottyl_seq/ottyl_ts are synthesized from the log index and never have a
MetricDeclaration or bucket entries) and merged both tickets' additions
to the reserved-params Docs.tsx paragraph.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
auto_declare_params (src/engine.rs) previously never overwrote or pruned a
declaration, so a burst of unusual key=value-shaped lines could permanently
inflate the params UI forever. A real macOS `log stream` dump
(huig_dump.txt, copied into src/fixtures/log_stream_sample.txt) discovered
220 distinct auto-declared params from under 700 lines, mostly one-off
WiFi/radio stat counters (20+ pairs on a single line), plus base64-blob
false positives from key()'s greedy alnum matching landing right before
`==` padding.
Fix, combining two bounds per the ticket's options:
- src/logfmt.rs: extract_pairs/pair now rejects a candidate pair whose
bareword value starts with a bare '=' (looks_like_base64_padding_artifact)
— the exact shape produced when a base64/hash blob's padding sits right
after a run of alnum chars that key() consumed as if it were a key.
- src/engine.rs: auto_declare_params now enforces a per-line cap
(MAX_NEW_AUTO_PARAMS_PER_LINE = 8) on new params from one line, and a
total ceiling (MAX_AUTO_DECLARED_PARAMS = 64) on auto-derived params
overall — the per-line cap alone doesn't bound unbounded key *variety*
across many lines, so both are needed.
Also adds structured extraction for log stream's fixed-column shape
(Timestamp Thread Type Activity PID TTL, then `process: message`):
detect_format gains a new "log_stream" ottyl_fmt value (checked first,
since it's the most specific of the four), and
auto_declare_log_stream_params declares log_type/process/pid as standalone
queryable params the first time such a line is seen — derived positionally
from the column layout rather than =-scanning, so there's no ambiguity or
false-positive risk. This is a separate always-attempted extraction
alongside the existing key=value pass, not a replacement for it: a log
stream line's message tail can still contain real key=value pairs (e.g.
the WiFi stat lines) that auto_declare_params keeps discovering
independently. This also fully addresses ticket 054, confirming huig_dump.txt
is space-padded fixed-width columns, not TSV, and that no separate TSV
format is needed.
Throughput (src/fixtures/log_stream_sample.txt x50, release build):
before (caps disabled) ~369k lines/sec, 206 auto-declared params;
after (caps enabled) ~1.20M lines/sec, 64 auto-declared params — about
3.2x faster, confirming the ever-growing declaration set was the
bottleneck rather than per-line regex/nom scanning cost.
Removes tickets/050 and tickets/054 now that both are implemented.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
evaluate_promql previously resolved every metric name through bucket_values,
which only returns data for a real MetricDeclaration — so a dashboard panel
querying ottyl_seq/ottyl_ts/ottyl_fmt (already supported in the filter bar
and alert conditions) silently returned an empty series. ottyl_seq/ottyl_ts
are now synthesized directly from the ingested-log index instead of requiring
a declaration; ottyl_fmt is categorical (json/logfmt/none) with no sensible
numeric reading, so it's rejected up front with a clear error pointing back
at the filter bar rather than returning nothing.
GET /api/params now also lists the three reserved names (marked `reserved:
true`, distinct from `auto_derived`) purely for discoverability — merged in
at the API layer only, not the engine's internal ParamDeclarations query, so
profile-switch removal and metric-param-promotion logic are unaffected.
declare_param also now rejects declaring a param under one of these names.
Docs.tsx and the Params panel are updated to reflect the new dashboard
support and the reserved/non-editable listing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
Wires the engine's existing RemoveMetricDeclaration message (previously
only used internally by delete_profile's cleanup) up to a new DELETE
route, persists the removal into the active profile so it survives a
restart, and adds a delete affordance in the Metrics UI.
Decisions:
- Bucket-value history under the deleted metric's name is purged, not
left orphaned (engine.rs's new remove_metric_declaration reuses
update_metric's clear_metric_data helper). This is what lets a
dashboard panel or PromQL query still referencing the metric by name
fail with a readable "unknown metric" error instead of silently
returning stale data.
- evaluate_promql now errors when a metric has neither a declaration nor
any bucket data, since a metric ingested via --metrics-url scraping
never gets a MetricDeclaration but must stay queryable. Verified with
a seeded fixture (promql_query_against_deleted_metric_fails_with_readable_error
in src/api.rs).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
DELETE /api/params clears the entire standalone-param set — manual and
auto-derived alike — then immediately re-runs auto_declare_params against
every IndexKey::IngestedLog entry still held in the engine's index (oldest
to newest), so params come back fresh from currently-retained log text
without requiring new lines to be ingested. This is a deliberate full
reset, unlike reload_engine_declarations's profile-switch cleanup, which
preserves auto-derived params since those are engine-discovered facts
rather than profile state; here the immediate re-extraction step makes
clearing them safe.
Settings::clear_param_declarations empties the active profile's persisted
params so they don't reappear on restart. The webapp's Params section gets
a single "Clear all" button (with a confirm dialog and a status line
reporting removed/re-derived/logs_considered) rather than a per-param
delete loop; logs_considered surfaces that lines already evicted by
max_logs can't regain their params.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
Extends ticket 048's light_green pattern (light-inverted surface/border/text
scale + light's neutral danger colours, same base and accent hues as the
dark-surfaced counterpart) to the remaining five colour schemes across all
six files that must stay in sync: index.css, themes.ts, api.ts, settings.rs,
useSettingsPanel.ts, and NavMenu.tsx.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
Every built-in theme now exposes three accents (--c-accent/-2/-3) instead
of one, so PanelChart's series palette draws distinct colours per theme
via color-mix() rather than a fixed, theme-independent hex list. Existing
presets keep their current base surface/border/text/accent look — the two
new accent vars are purely additive. Adds a new "Light Green" scheme
(light-surfaced, same green/magenta hue family as "Green") selectable
alongside the existing built-ins, both in ColorScheme (frontend + backend)
and the settings picker.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XFUNZwCvz3kcod3352ZR2o
