#10 transport (obdcore/transport.py):
- TcpTransport.read raises IOError on a real socket error or peer-close instead
of swallowing it as a timeout, so a dead WiFi link surfaces (via the #8 poll
handler) as 'connection lost' rather than a frozen dashboard.
- TcpTransport.reset_input_buffer drains at most 64 chunks — never spins forever.
- BleTransport closes the client + stops the event-loop thread on connect
timeout (no leak), caps the notification buffer at 64 KiB, and close() is
robust when only partially initialised.
#11 controller (gui/controller.py, obdcore/store.py):
- connect() closes the transport and nulls the link if init()/connect() raises,
so a failed/retried connect doesn't orphan sockets/threads.
- stop_record() unhooks store.recorder BEFORE closing it, and CsvRecorder now
has a 'closed' guard so a poll-thread write racing close() is a no-op instead
of an I/O-on-closed-file crash.
Closes#10Closes#11
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
- parse_dtcs CAN branch is now message-aware: each ECU reply '<svc> <count>
<pairs>' has its header stripped per-message, instead of flattening all lines
and stripping svc+count once. With multiple ECUs the old code ate the second
header as a DTC pair -> phantom codes. Critically, it does NOT blind-scan for
svc (0x43 is a legal DTC first byte: C03xx) — a numbered ISO-TP continuation
is distinguished by its 'N:' frame-index prefix, not by value.
- _line_bytes strips hex frame indices A:-F: (ISO-TP index cycles 0-F), not just
0-9, so consecutive frames past the 10th aren't dropped.
- read_freeze_frame sends the correct '020200' (svc 02, PID 02, frame 00) and
skips SID+PID+frame (+3), fixing the off-by-one that mis-read the freeze DTC.
- tests/test_dtc_parse.py: single-frame, multi-ECU (no phantom), numbered
multiframe with a real C03xx continuation, hex index, non-CAN legacy.
Closes#9
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
- A transport exception in the poll loop killed the thread silently, leaving the
GUI on a frozen 'Connected' dashboard and blocking run_oneoff callers for the
full timeout. _loop now catches it -> stops, fails pending one-offs with the
real error, and calls an on_error callback. Controller wires on_error to flag
the connection dead; the GUI detects it in _tick and tears down with a
'Connection lost' dialog.
- A run_oneoff that timed out left its job queued, so it executed LATER on the
shared link -- a ghost/duplicate vehicle command. Jobs now carry
cancelled/started flags under a lock; on timeout a not-yet-started job is
cancelled (skipped by _drain_oneoffs), and a started one reports 'still
running -- do NOT retry'. stop() also frees stranded submitters.
- tests/test_scheduler.py: cancel-on-timeout, freed-on-death, loop-survives.
Closes#8
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
Untrusted profiles could bypass the confirmation and responses were mis-parsed:
- effective_risk(action): risk is now DERIVED from the actual service IDs the
steps send — any write/actuator/reset/transfer SID (2F/31/11/14/2E/27/34-37/…)
forces 'danger'; unknown SID / non-default session / security block force
'caution'. A profile can only RAISE risk, never label a reflash 'safe'. GUI
gates the confirmation (and the risk badge) on this derived value.
- Response checks use CONTIGUOUS subsequence matching + a hard '7F <sid>'
negative-response guard, so an NRC data byte (e.g. 0x7E) can't false-pass as a
positive response; applied to session/security/step checks.
- 0x78 (responsePending) is treated as in-progress, not terminal failure.
- controller.run_action restores slow ELM timing for the run (0x78 window).
- Tests: risk-cannot-be-downgraded, NRC false-positive rejected, pending handled.
Closes#7
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
The AST sandbox whitelisted ** and << with no magnitude bound, so a hostile
profile formula (9**9**9, 1<<10**9) computed a multi-hundred-MB integer on the
scheduler thread -> CPU pin + OOM. The scheduler except clause never catches a
runaway/OOM (not a raised exception), and a derived PID with empty deps fires
every tick on connect.
- _apply() guards each BinOp: shift amount <= 256, exponent <= 64, and any int
result bit_length > 512 raises FormulaError (caught by the scheduler -> sample
dropped, thread survives).
- compile-time caps: expr length <= 500, AST depth <= 60; parse also catches
RecursionError.
- test_formula_dos_bounded: giant-int payloads rejected in <0.5s; legit bit ops
and scaling still work.
Closes#6
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
- obdcore/transport.py: pluggable byte transports -- SerialTransport,
TcpTransport (WiFi ELM327, stdlib socket), BleTransport (experimental, via
optional 'bleak'; background asyncio loop buffering notifications). ble_scan().
- ElmLink refactored onto a transport with .serial()/.tcp()/.ble() factories
(close/cmd now go through self.io); no behavior change for serial.
- Controller.connect(conn={kind:serial|wifi|ble,...}); GUI connection bar gains
a transport selector (Serial/USB/BT-SPP | WiFi host:port | Bluetooth LE + Scan).
- Classic-Bluetooth needs no new code (pairs as a serial port); WiFi needs no
extra deps; BLE is opt-in (bleak not bundled, so CI binaries keep building).
- tests/test_transport.py: drives ElmLink over a fake ELM TCP server end-to-end
(connect, RPM, readiness, VIN). All suites pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
Compiled by the generic-dtc-db workflow (P0/U0/C0/B0 standard codes, system
tags + no-start flags). Lives in profiles/_data/generic-dtcs.json (bundled with
profiles/, not listed as a vehicle profile). DtcDatabase.get now falls back:
profile code -> generic code -> unknown, so any standard code resolves to a
description while vehicle profiles still override (e.g. P0148 keeps the 6.0 text).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
.gitea/workflows/release.yml builds the PySide6 GUI with PyInstaller on each
self-hosted runner and, on a v* tag, publishes a Gitea Release with every
platform binary attached:
- windows-latest -> OBDash-windows.exe (onefile, --windowed)
- self-hosted-mac -> OBDash-macos.zip (.app bundle, ditto-zipped)
- docker (linux) -> OBDash-linux-x86_64 (onefile + patchelf/GL libs)
- arm64 (Pi) -> OBDash-linux-aarch64
Uses softprops/action-gh-release with the auto GITEA_TOKEN; manual dispatch
build-only (no release).
profiles/ is bundled via --add-data; obdcore.profile.profiles_dir() now
resolves sys._MEIPASS when frozen so the binary finds its vehicle profiles.
Validated locally on Linux: PyInstaller onefile builds (88MB), launches
offscreen, and loads bundled profiles (6 found). Build artifacts gitignored.
Added .claude/gitea-ship.json.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
Gauges:
- Optional per-metric warning zones (warn_hi/redline_hi/warn_lo/redline_lo) in
the profile schema; gauges draw colored redline/warn bands and color the
needle + readout by zone. Default neutral when unset (no false redline).
- Removed the value progress-arc fill (it dominated the dial / looked wrong) ->
clean tach face: bezel, ticks, numeric scale, needle, redline band, readout.
- Auto-derivation rejected: bad direction/threshold vary per metric, so zones
are config (with a sensible neutral default).
Units:
- New Units menu: Temperature C / F. Converts gauges, graph, table, and PID
browser (values, scale, zones, unit labels) at display time; data stays C.
Ford 6.0 profile: zones for ICP (red<500), FICM (red<40/amber40-48/green48+),
ECT/EOT (high redline), RPM (redline 3800), boost, battery; tightened FICM
(38-52) and battery (9-15) ranges so redline bands land sensibly.
Docs: profiles/PROFILE_SPEC.md -- canonical, AI-agent-ready profile spec
(schema, formula language, zones, confidence, rules); README points to it.
Validated headless: zones parse/classify, F conversion (112C->233.6F, zones
converted), gauges render; obdcore + diagnostics tests pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
The polling thread owns the ELM327, so reading/clearing trouble codes from
the GUI thread would race PID reads and corrupt the stream. Add a one-off
command path that serializes ad-hoc link work onto the polling thread.
obdcore/scheduler.py:
- PollScheduler.run_oneoff(fn, timeout) enqueues a callable (queue.Queue +
threading.Event) and blocks for its result, re-raising the callable's
exception. tick() drains queued one-offs at its very top, so they run on
the same thread that does PID reads -- never concurrently. When the
scheduler thread isn't running, the job is drained inline on the caller
(still serialized vs tick(), safe because nothing else touches the link).
gui/controller.py:
- Controller.read_dtcs() -> {"stored","pending","permanent"} (modes 03/07/0A,
svc 0x43/0x47/0x4A) and clear_dtcs() -> bool. Both route through the
scheduler one-off when a scheduler exists, else call the link directly.
gui/main.py:
- Diagnostics menu (Read Codes / Clear Codes...) and a right-side QDockWidget
listing codes grouped Stored/Pending/Permanent. Each row is code +
description + system from DtcDatabase; no_start codes are flagged bold red.
- Clear is guarded by a confirmation warning (erases codes + freeze frame;
honest "the code comes right back" / permanent-codes-won't-clear tone from
run_clear in obd_reader.py). On confirm: clear, then re-read immediately and
show whatever returned, reporting active faults that came straight back.
tests/test_diagnostics.py:
- one-off returns its value, re-raises exceptions, is drained before a tick's
PID reads, and runs on a live background thread while polling continues.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
Vehicle data is now DATA, not code. PIDs/scaling/DTCs/presets live in
profiles/*.json; the app loads them at runtime, so it works across vehicles
and others can contribute profiles (open source).
Core:
- obdcore/formula.py: safe AST evaluator for scaling formulas (A/B/... byte
vars, Torque/FORScan convention). Only arithmetic/bitwise + min/max/abs/
round/int/float; names/attrs/arbitrary calls rejected at load -> a community
profile CANNOT execute code.
- obdcore/profile.py: load/save/list profiles; compiles each formula into a
decode callable. registry.py now profile-backed (PidRegistry/DtcDatabase
take a Profile); hardcoded Ford table removed.
- store.py: clear()/snapshot()/export_csv() for capture management.
Profiles:
- profiles/ford-6.0-powerstroke.json (27 PIDs, verified formulas, DTCs)
- profiles/generic-obd2.json (standard SAE Mode-01 base, any vehicle)
- profiles/README.md (schema + formula language + contributing)
GUI:
- Menu bar: File (new/record/export/replay capture, quit), Profile (switch/
load/import/reload/edit-JSON/export, live profile list), View (Graph/Table
views, gauges P2, toggle PID dock, normalize, light/dark theme), Help
(about/confidence legend/profile info).
- PID browser + presets rebuild on profile switch; added Table view; raw-JSON
profile editor dialog (validates schema+formulas before saving).
Tests: profiles load+compile, formula sandbox rejects hostile input, decoders
still match real truck bytes, crank/derived/dead-PID/replay -- all pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
The initial registry was a curated core; add the rest of the PIDs the
ford-60-pid-hunt workflow surfaced, with honest confidence tags:
- doc: VGT duty (096D), Fan speed (099F)
- tentative: Injection timing (09CC, *10/64 not /10), PCM battery (1172),
fuel-pump duty (1672), fuel level (16C1, uncalibrated),
mass-fuel-desired (1411, raw only -- no verified GPH formula)
Add VGT to the driving preset. Tests still pass.
Tier reminder: 'verified' = multi-source/standard or truck-confirmed; the
on-truck-confirmed subset remains ICP/EBP/MAP/BARO/EOT/gear/TSS + FICM_M.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
Foundation for the PySide6 + pyqtgraph Windows GUI, shared with the terminal
tool. Pure data/IO -- no Qt, no curses.
obdcore/
link.py ElmLink -- ELM327 serial (Mode-01/22, ATRV, DTC read/clear)
mock.py MockLink -- synthetic crank for tests + GUI dev (no truck)
registry.py PidRegistry (verified Ford 6.0 PIDs + confidence) + DtcDatabase
scheduler.py PollScheduler -- prioritized round-robin polling, dead-PID park,
derived channels; tick() is fake-clock test-drivable
store.py TimeSeriesStore (ring buffers + min/max) + CsvRecorder/replay
Design centers on the ELM327 bandwidth limit (~7-15 reads/sec): the active
view subscribes PIDs at chosen rates; acquisition runs off the UI thread;
the GUI only reads the store. FICM_M (09D0) promoted to verified after the
2026-06-30 on-truck crank read (48.0V, intermittent).
tests/test_obdcore.py: decoders vs real truck bytes, crank ramp + peak,
derived BOOST, dead-PID park/revive, record/replay roundtrip -- all pass.
ARCHITECTURE.md: layers, data model, GUI plan, 6.0 stock-PID limits
(no EGT/oil-PSI), feature backlog, P0-P5 roadmap.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yT89n4zR4qbrySoSiEyZs
justin