Round 21 (three reported items).
1. A hashboard that is still mining near full hashrate but has some failed chips
now shows yellow (degraded), not red. VNish emits per-board chip failures as
Warning while the board is mining (Error only when the board is not mining),
and board_healths() renders a warning-only board as Degraded instead of Down.
Verified live on 10.4.0.193 (mining ~83%, dead chips -> yellow).
2. New "Power limit" column: the configured power-target setpoint in watts
(MinerRow.power_limit_w, from a Power tuning target). WhatsMiner shows its
"Power Limit" (e.g. 2500 W); the value was already scanned but had no dedicated
column. Numeric (sortable/filterable), Performance category, opt-in.
3. Chip + hashboard model coverage:
- VNish hashboard model scanned from factory-info board_model (A3HB70601,
BHB42603, ...) -- 147-unit fleet gains HB model. Both VNish backends.
- WhatsMiner v1/v2 PSU serial scanned from get_psu serial_no (was hardcoded
None). 48-unit fleet gains PSU serial.
- Chip type derived from the detected model for Antminer hardware (VNish/stock/
LuxOS/Braiins/UMC never report a chip string on the wire), in naming.rs:
S21Pro->BM1370, S19jPro->BM1362, S19KPro->BM1366, S19/T19->BM1398, etc.
Only high-confidence mappings; WhatsMiner and unknown models stay blank.
PSU Vmin/Vmax remain Braiins/ePIC-only (not on the wire for WhatsMiner/VNish/
Antminer -- confirmed live).
asic-rs patch 0033. 95 unit tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
- Hashrate cell: a readable 0 hashrate now shows "0.00T" instead of a dash. The
dash is reserved for an unreadable/absent hashrate. New MinerRow.hashrate_known
records whether the miner reported a value; combined_si/num_value key off it,
and scan/fixtures set it from data.hashrate.is_some().
- Paused-miner Health icons: the control board (brain) and PSU keep their true
health (green when fine) instead of turning blue; fans turn blue only when
actually stopped (0 RPM) and stay green while spinning; hashboards stay blue
while paused (idle is the expected state).
- VNish per-hashboard faults: DataField::Messages now also reads the summary
chains (chip_statuses red/orange + per-chain state) and emits one message per
faulted board, attributed to MinerComponent::HashBoard{idx}, so the Health
column reds the specific board and its tooltip carries the fault instead of it
landing on the brain. Applied to v1_2_0 and v1_3_0. patch 0032.
Verified live: VNish 10.7.1.25 -> "Hashboard 2: 1 degraded chip(s)" on board idx
1; healthy 10.7.1.122 -> no messages. 93 unit tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
A miner an operator has stopped (e.g. 10.4.1.122, VNish miner_state "stopped" ->
OperatingState::Stopped) was read as broken: red hashboards and counted under "not
hashing" instead of "Paused".
Two causes:
1. is_paused() only matched the literal "pause" substring, so "Stopped"/"Stopping"/
"Suspended" -- every other firmware's word for an intentional, resumable halt --
fell through to not-paused. Now matches pause/stop/suspend (Error stays broken).
2. encode_component_health applied the paused-blue treatment to hashboards only
(round 19 #19), so a paused miner's idle fans/power still rendered red. A stopped
miner's off components are expected, not faults: colour every reported component
blue when paused (an undetected/Unreachable one still stays grey -- "not present"
is not "paused").
Verified live against 10.4.1.122 on a headless build: the row now shows all-blue
health (power grey, honestly unreachable) and counts under the Paused KPI, where it
previously showed red boards + "not hashing".
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Round 19 #2. The health-icon hover tooltips never appeared because the GTK
query-tooltip handler built the per-icon "tip area" from gtk_tree_view_get_cell_area
(BIN-WINDOW coordinates) and passed it to gtk_tooltip_set_tip_area, which expects
WIDGET coordinates. The area was offset upward by the column-header height, so it
did not contain the pointer, and GTK suppresses any tooltip whose tip area excludes
the cursor -- so nothing showed. The Errors column worked because it sets no tip
area. Prior fixes (rounds 15/18/19) only tuned the icon slot math, which was already
correct; the coordinate-system mismatch was the actual cause.
Convert the slot rectangle from bin-window to widget coordinates with
gtk_tree_view_convert_bin_window_to_widget_coords before gtk_tooltip_set_tip_area.
The slot hit-test (block_left/adv, in bin-window coords) is unchanged and still
matches DrawShape's uniform health-icon advance. wx-rs patch 0005.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Round 19 #10 telemetry extension. Live-verified against the corpus miners, this
remodels the board temperature thresholds to item #10's target/hot/dangerous scheme
and adds the LuxOS and Braiins backends. Supersedes the ePIC-only critical/shutdown
model shipped in v0.72.0.
Core (asic-rs-core):
- MinerData: temp_critical/temp_shutdown -> temp_target/temp_hot/temp_dangerous.
- parse_temperature_limits (defaulted on the existing GetFluidTemperature trait)
now returns (target, hot, dangerous); parse_data wires all three fields.
Backends (each reads an already-fetched or single command; live-verified):
- LuxOS: new tempctrl RPC; ChipTarget/ChipHot/ChipDangerous (live 10.4.1.60 = 83/93/100).
- Braiins v25_07 + v26_04: configuration/constraints cooling_constraints defaults
(live 10.4.1.83 = 60/80/90); WEB_CONSTRAINTS already fetched for tuning caps.
- ePIC remap: Shutdown Temp -> hot, Critical Temp -> dangerous, no target.
- VNish (settings restart_temp/param) handled in a follow-up (needs settings auth);
Antminer/WhatsMiner expose no thresholds -> (None,None,None).
GUI: the two temp columns (Critical/Shutdown) become three (Target/Hot/Dangerous),
right-aligned, Temperature category, auto-hidden when empty. MinerRow gains
temp_target_c/temp_hot_c/temp_dangerous_c; to_row reads them. ALL_COLUMNS 70 -> 71.
Fixtures captured from live: luxminer tempctrl.json, braiins configuration_constraints.json.
Fixture tests for epic, luxminer, braiins v26_04 assert the parsed thresholds. Patch 0029.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Round 19 #10 telemetry extension. Adds the board temperature safety thresholds
("board dangerous temp" from item #10).
Core (asic-rs-core):
- New DataField::TemperatureLimits; MinerData gains temp_critical and temp_shutdown
(Celsius).
- Rides the existing GetFluidTemperature trait via a defaulted parse_temperature_limits
method returning (critical, shutdown) -- no supertrait blast radius; parse_data
wires both fields in.
Backend:
- ePIC: summary /Misc Critical Temp + Shutdown Temp (already-fetched summary).
- Antminer/WhatsMiner expose no thresholds; LuxOS needs an unwired tempctrl command
(no fixture) and VNish/Braiins need live verification -> left for a live-verified
pass, all inherit (None, None).
GUI: two new columns (Critical temp / Shutdown temp), right-aligned, Temperature
category, auto-hidden when empty. MinerRow gains temp_critical_c / temp_shutdown_c;
to_row reads them. ALL_COLUMNS 68 -> 70.
Fixture test for epic asserts the parsed thresholds. Patch 0028.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Round 19 #10 / #8 telemetry extension. Adds the fan duty cycle as a miner-level
percentage -- the normalized fan-load indicator the GUI wanted alongside per-fan
RPMs. No firmware exposes a maximum fan RPM, so a true actual/maximum pair is not
possible; the duty cycle is the closest normalized signal.
Core (asic-rs-core):
- New DataField::FanSpeedPercent; MinerData gains fan_speed_percent (Option<f64>).
- Modeled as a top-level miner-level value (most firmwares report a single duty),
so FanData is untouched -- no Eq removal / literal churn across its 17 crates.
- Rides the existing GetFans trait via a defaulted parse_fan_speed_percent method
(no supertrait blast radius); parse_data wires it in.
Backends (read already-fetched commands, no extra round-trips):
- ePIC: summary /Fans/Fans Speed (single %).
- LuxOS: max of FANS[].Speed.
- Braiins v25_07 + v26_04: max of cooling fans[].target_speed_ratio x 100.
- Antminer (pwm via config) and VNish (fan_duty) have no data fixtures -> left for a
live-verified pass; WhatsMiner is firmware-locked -> None.
GUI: new "Fan %" column (right-aligned, Cooling category, auto-hidden when empty).
MinerRow gains fan_speed_percent; to_row reads it. ALL_COLUMNS 67 -> 68.
Fixture tests: epic, luxminer, braiins v26_04 assert the parsed percentage. Patch 0027.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Round 19 #10 telemetry extension, tier 2b. Adds the cumulative share statistics
from item #10 end to end: difficulty-weighted accepted/rejected work, raw stale
share count, and blocks found.
Core (asic-rs-core):
- New ShareSummary sub-struct (data/share.rs) with a parse_cgminer_share_summary
helper reading a cgminer/btminer SUMMARY[0] object.
- New DataField::ShareSummary; MinerData gains share_summary.
- Rides the existing GetBestShare trait via a defaulted parse_share_summary method
(no new supertrait bound, no per-backend empty impl); parse_data wires it in.
Backends (read the already-fetched SUMMARY, so no extra round-trips; these are
miner-level aggregates -> top-level MinerData fields, not per-pool PoolData):
- Antminer v2020 + v2023_07, LuxOS, WhatsMiner V1 + V2 override parse_share_summary.
- Braiins (no cumulative difficulty), ePIC (current difficulty only), WhatsMiner V3
and VNish expose none and inherit None -- fixtures confirm.
GUI: four new columns (Diff acc / Diff rej / Stale / Blocks), right-aligned, in the
Pools category, auto-hidden when empty. MinerRow gains the four fields; to_row reads
them from share_summary. ALL_COLUMNS 63 -> 67.
Fixture tests: whatsminer v1 and antminer v2020 assert the parsed values. Patch 0026.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Round 19 #10 telemetry extension, tier 2. Adds the rolling average hashrate
windows the GUI asked for (1M/5M/15M/24H) end to end.
Core (asic-rs-core):
- New HashrateAverages sub-struct (data/hashrate.rs): avg_1m/5m/15m/24h, each an
Option<HashRate> carrying the same unit and algorithm as the current hashrate.
- New DataField::HashrateAverages; MinerData gains hashrate_averages.
- The windows ride the EXISTING GetHashrate trait via a new defaulted
parse_hashrate_averages method (default None), so no new supertrait bound and no
empty impl in every backend -- only the backends that expose the data override
it. parse_data wires it in.
Backends (all read a command the collector already fetches -> no extra round-trips;
each window is filtered to > 0 and the whole struct is dropped when every window is
absent, so a firmware that reports nothing keeps hashrate_averages = None):
- WhatsMiner V1: btminer summary MHS 1m/5m/15m (MegaHash); no 24h.
- WhatsMiner V3: get.miner.status summary hash-1min / hash-15min (TeraHash) only.
- LuxOS: summary GHS 1m/5m/15m/24h (GigaHash) -- full window set.
- Braiins v25_07 + v26_04: miner_stats real_hashrate last_1m/5m/15m/24h (GigaHash).
- Antminer (5s/30m/av only), WhatsMiner V2 (av only), ePIC/VNish (session average
only), sealminer: no standard 1/5/15 window -> inherit None. This is honest: the
fixtures confirm those firmwares do not expose the standard windows.
GUI: four new columns (HR 1m/5m/15m/24h), single TH/s value each, right-aligned,
blank when the window is not reported, auto-hidden when empty across the fleet.
MinerRow gains the four ths fields; to_row populates them from hashrate_averages.
Fixture tests: whatsminer v1 and braiins v26_04 gain assertions on the parsed
windows (value, unit, and the expected present/absent pattern). Patch 0025.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Round 19 #10 telemetry extension, tier 1 sub-batch B. Both ride existing
DataFields (Psu/Hashboards) on already-fetched commands (no extra round-trips)
and reuse the PsuInfo.serial / BoardData.chip_type+hashboard_model fields added
in T1a, so no core or GUI change -- the existing columns just populate for more
firmwares. Patch 0024; fixture tests pass.
- WhatsMiner V3: add a DataField::Psu location (get.device.info /msg/power) and a
parse_psu that reads model/type, sw/hwversion, and the serial (sn). V3 had no
PSU parse at all (default None) despite the data being fetched for PsuFans.
- ePIC: parse_hashboards now reads capabilities "Chip Type" (e.g. "BM1366") and
"Model Subtype" (e.g. "BHB56801"), applied to every board (capabilities reports
them miner-wide). The capabilities command is already fetched for Hashboards.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Extend vendored asic-rs to scan more per-firmware telemetry (round 19 #10
follow-up, tier 1 sub-batch A). All new values ride on existing DataFields
(Psu/Hashboards/BestShare) and commands already fetched, so no new core
DataField plumbing and no extra network round-trips. Patch 0023.
Vendored (asic-rs):
- PsuInfo gains `serial`. Braiins v26_04 parse_psu corrected to read model_name
+ fw_version (it used the numeric type code as the model and dropped the fw
version) and to capture serial_number. ePIC / WhatsMiner v1/v2 report no serial.
- BoardData gains `chip_type` (e.g. "BM1370") + `hashboard_model` (e.g.
"A3HB70601"); Braiins miner/hw/hashboards populates both.
- best_share (all-time) wired for Antminer v2020 and WhatsMiner v1/v2 from the
already-fetched summary "Best Share" (those backends had the empty default).
- refresh-checksums --all; existing braiins/antminer/whatsminer fixture tests pass.
GUI:
- MinerRow gains chip_type, hashboard_model, psu_serial; populated in to_row.
- New columns: Chip, HB model, PSU serial. Best share now populates for Antminer
and WhatsMiner (column already existed).
Feasibility analysis (5-agent, all firmwares) recorded in memory; further tiers
(fan %, rolling hashrate averages, difficulty/shares, temp thresholds, WhatsMiner
V3 / VNish PSU serial) follow. Confirmed genuinely-unavailable: fan max RPM,
WhatsMiner/LuxOS real logs, chip type on Antminer/LuxOS.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Health column (operator-reported): give every component-health icon a UNIFORM
fixed-width slot in the SegmentRenderer (GetSize + DrawShape), so power/control/
fan/board icons line up in fixed columns across rows regardless of glyph aspect,
and left-align the icon columns (Health/Access/Light) so the block anchors at a
consistent left edge. The per-icon tooltip hit-test uses the same uniform slot.
Code-review-harness fixes:
- High: isolate per-miner telemetry with catch_unwind so one miner's get_data
panic can't abort the whole scan (and strand the fast-path partial rows).
- Medium: fill the zebra stripe across the full cell for right-aligned colored
columns (was anchored at the content's left edge, leaving the left unstriped).
- Medium: lower the scan-concurrency ceiling 512 -> 256 (128 is fully reliable,
192 already dropped responders on the live fleet).
- Medium: board voltage/frequency sort key skips placeholder-zero boards
(avg_positive), matching how the cell renders them as "-".
- Low: saturating_add for accepted/rejected share sums; moved an orphaned doc
comment; corrected a stale comment.
CI: fix the README-badge push-back -- `git push -c http.extraheader` is invalid
(`-c` must precede the subcommand), so the badge/coverage/bench-baseline commit
never actually reached main; now `git -c http.extraheader=... push`. This also
lets the bench-delta baseline persist so the delta can compute.
Demo: the offline-demo paused miner now has detected (idle) hashboards so the
paused-blue Health icons (round 19 #19) are exercised.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Table/columns:
- Fix the "columns shrink to tiny" bug on language change / add / hide (#3, #24):
the rebuild-time capped autosize read column widths before GTK's measurement
landed. Columns keep their default widths; autosize is on-demand only.
- Right-align every column except IP (#21): new alignment param on the bridge
append calls; the SegmentRenderer and the tooltip hit-test honour it.
- "Auto-size all columns" header menu item; Ctrl+A selects all rows (#4).
- Remove the Boards (dots) column (#7).
- Split the slash-joined Fan column into per-fan columns Fan 1-4 (#8).
- Efficiency: unit moved into the value and made algorithm-aware (J/TH for
SHA256 families, J/kSol for Equihash), fixing the blank/garbage figure on
non-SHA256 miners (#9, #23).
- Pool account column before Worker by default (#22).
Scan:
- Fast scan (#18): emit a lightweight identity row the instant a miner is
identified, then upsert the full telemetry row when get_data completes.
- User-adjustable scan concurrency with a bounded max (#25).
- Surface more scanned values as columns (#10): current tuning target, accepted
/ rejected shares + reject %, and per-hashboard hashrate / voltage / frequency.
Logs (#11-16): fetch the miner log through asic-rs's per-backend read_logs()
instead of the narrow curl paths, fixing Braiins / LuxOS / Z15 / BixBit / stock
Antminer; WhatsMiner keeps the fault-code fallback (no API log). Removed the
summary tidbits from the log viewer (#13).
Misc: fix per-icon health tooltips (advance math matched to the renderer) (#2);
paused miners show blue hashboard icons, healthy stays green (#19); state-aware
Pause/Resume in the right-click menu (#20); export opens a column-picker dialog
(#6) and the default filename carries a UTC timestamp (#5); the scan status bar
says "Scan Paused" / "Scan Stopped" and no longer overwrites Stop with paused
(#17); the bench badge shows the weighted-average microbench delta vs the
previous release (#1).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Add a best-effort Coverage step to release.yml: it adds the llvm-tools-preview
rustup component, fetches cargo-llvm-cov as a prebuilt binary (the vendored
.cargo/config blocks `cargo install` from crates.io), runs ci/coverage.sh to
merge unit-test and headless self-test line coverage, and feeds the TOTAL
line-coverage percent to the badge refresh so badges/coverage.svg tracks the
real number on every release. The step is gated on a release and marked
continue-on-error, so coverage tooling can never block a release.
coverage.sh now runs its instrumented builds --offline by default (the
dependency tree is fully vendored), overridable via CARGO_LLVM_COV_FLAGS.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
The self-test builds the real wxWidgets table and renders every fixture through
the real widgets, exercising ~27% of ui.rs (up from ~4%) -- but it runs as a
separate binary, so a plain `cargo llvm-cov` (which only runs the #[test]s) never
counted it, understating coverage. New ci/coverage.sh runs both the unit tests and
the self-test under one instrumentation pass and merges them, excluding the
vendored asic-rs / patched wx-rs from the report.
Our-source line coverage is now ~46% (model 81%, naming 95%, fixtures 100%,
credentials 92%); the remainder is click-driven dialogs and live-miner network
I/O, which unit tests can't reach without hardware or heavy mocks. Badge, report,
and README dev section updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Collapse the dual MIT OR Apache-2.0 licensing to MIT only (operator's choice):
set the workspace `license` field to "MIT" (all crates inherit it), replace
LICENSE-MIT + LICENSE-APACHE with a single LICENSE file, and update the README
license section and the license badge. deny.toml keeps its permissive allow-list
(that governs dependency licenses, which remain a mix of MIT/Apache-2.0/BSD/etc.);
only its project-license comment changed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
[profile.release] opt-level="s" + fat LTO + one codegen unit + strip cuts the
Linux binary from ~38M to ~19M (50%). panic stays "unwind" deliberately: the scan
runs many miners on tokio tasks and the wx dispatch relies on unwinding, so one
miner's panic must never abort the whole app. Verified: stripped binary, SELFTEST
PASS, 99 tests, clippy clean.
Adds code-review-report.md: the ship verdict and a full accounting of the five
recursive review rounds (all confirmed critical/high fixed; remaining items are
documented perf/dev-CLI tradeoffs). Badges refreshed (tests 99, coverage ~35%).
This completes the round-18 marathon: all 12 queued items plus Round C G3 (LuxOS
named presets) and G7 (per-board V/F read side), CI hardening (verify/release
split, clippy gate, audit, badges), README + docs, and the recursive multi-agent
code review.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Fifth (final) review round; this clears its functional findings:
- HIGH config actions / pool writes could run concurrently with a live scan
(start_action / apply_pools guarded flash + restore but not scanning); added
the self.scanning guard, and begin_scan now defers while an action/restore runs,
closing the race symmetrically.
- HIGH blanking a per-firmware credential row saved a live ("","") override that
then broke that firmware's auth fleet-wide; a fully-blank row now clears the
override (use the shipped default).
- HIGH a ChangePassword action never updated the toolkit's own saved credentials,
so after a fleet-wide change it locked itself out; on full success it now updates
the saved global password (and on a partial change logs a reconcile note rather
than guessing).
- HIGH/perf fill_rows re-rated: a scan's streaming rows triggered a whole-table
refill on every fast tick. New rows_dirty latch flushes the refill at most once
per second during a scan (immediately once it ends).
- MED an indeterminate access-probe result no longer blanks a miner's already-known
locked/unlocked state.
- LOW compute-version.sh header doc corrected to match the broadened rule;
release.yml sets WX_RS_WXCONFIG once at job level instead of per step.
Accepted (documented) perf items that remain: per-render health recompute and
per-column cell-string allocation (small, fleet-scale), the linear upsert scan,
and synchronous scan-range parsing (bounded by the range caps).
99 tests + clippy-clean + selftest green. This was review round 5 (the harness's
iteration cap); every confirmed critical/high across all five rounds is fixed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Fourth review round found further issues, two being gaps in earlier fixes; this
clears the actionable ones:
- HIGH board_healths() painted a genuinely-undetected board RED when a fault text
also named it (e.g. "chain 2 not found" -> Board(2)); an undetected board now
stays grey, matching the documented grey-vs-red invariant. Regression-tested.
- HIGH the Change-password prompt used an unmasked field (the Credentials dialog
was masked in wave 1 but this prompt was missed) -> masked for PromptKind::Password.
- MED the WhatsMiner access probe aborted all remaining password candidates on one
transient error -> transient errors now skip to the next candidate and never
report a confident "locked" when the answer was really indeterminate.
- MED config actions / refresh / pools / flash ignored the operator's configured
scan timeout (used asic-rs's 10s default) -> configured_factory now applies it,
so a slow-link action reaches the same miners a longer-timeout scan found.
- MED the auto-rescan loop set a bogus "scanning..." status even when the scan was
deferred (mid-flash) -> begin_scan reports whether it started.
- LOW is_aml_140mm matched every S21, not just the S21 Pro its doc names -> fixed.
- Offline fixtures now reuse scan::infer_pic instead of a diverging copy.
Accepted (documented) from this round: miner_probe's argv secrets (diagnostic CLI),
the rebuild_table/update_headers sort-marker duplication, the LuxOS substring
success check (command-specific reply), restore/flash target snapshot (dialogs are
modal), non-cancellable restore/probe batches, Braiins host-key (LAN reflash), and
the perf items (per-tick refill, health recompute, linear upsert, cell formatting).
99 tests + clippy-clean + selftest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Third review round (over the wave 3/4 fixes) surfaced further findings, mostly in
new code paths; this clears them:
- HIGH the fully-built "Boards" column (Column::Hashboards: renderer, sort/filter
logic, tests, a self-test check) was missing from ALL_COLUMNS, so it was
unreachable in the UI and its self-test check was dead. Added to ALL_COLUMNS
(opt-in; Health still shows composite board health by default) and the self-test
now makes it visible -- the dot-sentinel check is live again.
- HIGH apply_pools() could start a pool write during a restore-to-stock (it sets
self.action directly, bypassing start_action's guard) -> same restore guard.
- HIGH superseding an in-flight action/pools job left the old miners' Health cells
frozen on the "…ing" label forever (the replaced handle's drain never runs) ->
clear the superseded action's op-status.
- HIGH clean_antminer() byte-sliced a miner-supplied model string without a
char-boundary check (panic DoS) -> use get(..9).
- MED access-probe errors were silently discarded -> surfaced in the indicator.
- MED the pool-password file was briefly world-readable before chmod -> created
0600 from the start (OpenOptions mode).
- MED a miner-controlled log response was buffered unbounded (OOM DoS) -> capped.
- MED the language-menu id range was hardcoded (and 790-794 collided with the
pool-preset / fw-preset dialog ids!) -> derived from Lang::all() and moved the
dialog ids to a clear block.
- LOW restore_vnish now sends the unlock body via stdin (password off argv), and
the LuxOS restore treats a reboot-dropped/empty reply as success.
Remaining round-3 findings are the previously-accepted perf tradeoffs (bulk-serial
writes, per-tick refill, per-render health recompute, linear upsert), documented
in code and in code-review-report.md.
99 tests + clippy-clean + selftest green (Boards-column check now live).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Clears the rest of the second review round's confirmed findings:
- HIGH the Boards numeric filter compared the composite sort score (online*100+
total) instead of the board count, so a filter like "Boards <= 3" matched
nothing on healthy 3-board miners. num_value now returns the detected count;
regression-tested with per-board health present (the path the old test missed).
- HIGH the WhatsMiner access probe ignored the operator's saved scan password
(tried only admin/super), mis-reporting writable miners as locked. It now tries
the caller's password first.
- HIGH firmware flash and config actions were not blocked during a restore-to-
stock; start_action and both flash guards now refuse while a restore runs.
- MED log tail() could panic slicing at a non-UTF8 char boundary -> nudge to a
boundary first (regression-tested).
- MED board_index_from_text grabbed an unrelated number far past the keyword
(e.g. a temperature) -> only accept a number immediately after the keyword.
- MED a restore worker-thread spawn failure wedged the restore subsystem forever
-> emit Done on spawn failure and treat a disconnected channel as done.
- MED LuxOS restore reported failure when the miner rebooted mid-reply -> an empty
reply or dropped connection after a successful logon is the expected reboot.
- MED two tests raced on the process-global $HOME -> a shared serialization lock.
- LOW CSV export now neutralizes spreadsheet formula-injection prefixes on
miner-supplied fields.
- Made miner_probe's global-only credential resolution explicit (it can't share
the GUI's per-firmware configured_factory across the bin boundary).
99 tests (+3 regression tests) + clippy-clean + selftest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Second review round (over the wave 1/2 fixes) surfaced further, mostly new-path
findings; this clears the substantive ones:
- HIGH log_scan.rs passed miner admin credentials on the curl command line (same
class as the access-probe leak fixed in wave 1). Its curl() now takes optional
stdin: the Antminer digest login goes through a curl config on stdin (-K -),
the VNish unlock password through the POST body on stdin (-d @-), and the
timeout path reaps the child so it no longer leaks a zombie.
- MEDIUM JSON injection/breakage: the VNish unlock body and the LuxOS
uninstall command spliced a password / miner-supplied SessionID into JSON with
format!. Both now build the JSON with serde_json, which escapes correctly.
- MEDIUM column scan cost: fill_rows computed table_columns() twice (my wave-2
built_columns check added one), and with auto-hide-empty on that scans every
cell of every row each time. It is now computed once and reused.
Accepted, documented tradeoffs from this round: miner_probe's setpools takes the
pool password as a CLI arg (inherent to a diagnostic CLI; it is the low-sensitivity
mining worker password), MinerRow::upsert_into's linear scan is fine at the tool's
fleet scale (hundreds), and bulk config-action writes stay serial deliberately.
96 tests + clippy-clean + selftest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6
Continuing the recursive review's confirmed findings:
- HIGH scan telemetry was serial: the consumer awaited each responder's to_row
(get_data + presets, several REST calls) inline in the identification stream
loop, so telemetry ran one miner at a time. Now buffer_unordered runs it
concurrently up to SCAN_CONCURRENCY_LIMIT; identification (which buffers
internally) is unaffected, and telemetry only hits live responders.
- HIGH table columns diverged from the built widget with auto-hide-empty on:
(1) drag-reorder was silently rejected because column_reordered fed
table_columns() (empty-filtered) to apply_visible_reorder, which compared it to
display_columns() (unfiltered) -- apply_visible_reorder now reorders the shown
subset in place; (2) a column flipping non-empty mid-scan left the widget stale,
so fill_rows now rebuilds when the shown set no longer matches built_columns.
Both covered by a new unit test.
- HIGH viewing a miner's log blocked the wx UI thread for up to ~20s. The fetch
now runs on a background thread; the viewer opens from the tick drain when the
result lands (tick returns the dialog for dispatch to show modally).
- LOW start_pools_action's fake branch now honours the cancel flag like its
siblings.
- Documented the intentional tradeoffs the review flagged: scan-restart uses
cooperative cancellation (a hard join would freeze the UI), and the aftermarket
re-probe stays per-scan for freshness (firmware can change between scans).
96 tests (+subset-reorder) + clippy-clean + selftest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Djv3oncM4LxTS5tb61Ako6