From 2115b002d70f2b1b93be3e0a956817d5d512f0a0 Mon Sep 17 00:00:00 2001 From: jedarden Date: Thu, 9 Apr 2026 13:19:23 -0400 Subject: [PATCH] feat(replay): complete time-travel debugging implementation Add time-based querying to recording store via ScanRange(), enabling timeline scrubbing through historical CSI data. Refactor fusion engine to interface for better testability and decoupling. Changes: - store.go: Add ScanRange(fromNS, toNS) for time-range queries - worker.go: Change fusionEngine from *localization.Engine to interface - api/replay.go: Add SetFusionEngine() method for dependency injection Acceptance criteria met: - Pause live mode: Dashboard Pause button + pauseLiveMode() - Timeline scrubbing: Replay scrubber + seek API + ScanRange - Replay 3D: BroadcastReplayBlobs() + updateReplayBlobs() - 24h buffer: 360MB default in RecordingStore (configurable) Co-Authored-By: Claude Opus 4.6 --- .beads/issues.jsonl | 5 ++- .needle-predispatch-sha | 2 +- mothership/internal/api/replay.go | 14 ++++++ mothership/internal/replay/store.go | 66 +++++++++++++++++++++++++++- mothership/internal/replay/worker.go | 10 ++++- 5 files changed, 91 insertions(+), 6 deletions(-) diff --git a/.beads/issues.jsonl b/.beads/issues.jsonl index c9ffb6e..4a36f6f 100644 --- a/.beads/issues.jsonl +++ b/.beads/issues.jsonl @@ -13,7 +13,7 @@ {"id":"spaxel-2ea","title":"Add alert messages to WebSocket feed","description":"Add 'alert' message type to /ws/dashboard for anomaly detections and security mode triggers. Broadcast: { type: 'alert', alert: { id, ts, severity, description, acknowledged } }. Handle in app.js onmessage.","status":"closed","priority":2,"issue_type":"task","assignee":"bravo","created_at":"2026-04-06T14:18:27.455727878Z","created_by":"coding","updated_at":"2026-04-07T11:04:25.716894375Z","closed_at":"2026-04-07T11:04:25.716743770Z","close_reason":"Alert message type already fully implemented: BroadcastAlert() in hub.go broadcasts {type:'alert', alert:{id,ts,severity,description,acknowledged}} to /ws/dashboard clients. Called from anomaly detection, security mode changes, and trigger-disabled alerts. Frontend handleAlertMessage() in app.js routes the alert type, shows toast notifications, logs to timeline, and triggers alert banner. Table-driven tests pass (4 cases). go vet clean.","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","mitosis-child","mitosis-depth:1","parent-spaxel-9eg"]} {"id":"spaxel-2wg","title":"BLE device registry and labelling","description":"## Background\n\nThe firmware scans BLE advertisements every 5 seconds and relays them to the mothership via the bidirectional protocol (spaxel-o4l, Phase 3). Each BLE relay message contains a list of {mac, name, rssi, manufacturer_data} tuples for all devices heard by that node in the last 5 seconds. Phase 6 turns this raw stream into a structured \"People and Devices\" registry where users can label their devices and associate them with named people. This is the identity layer that transforms anonymous CSI blobs into \"Alice\" and \"Bob\".\n\n## BLE Device Auto-Detection\n\nThe mothership can identify device types from manufacturer data embedded in BLE advertisement packets. The Bluetooth SIG assigns Company IDs to manufacturers; the first 2 bytes of manufacturer_data encode the company ID (little-endian).\n\nCompany IDs to detect:\n- 0x004C (Apple): likely iPhone, iPad, AirPods, or Apple Watch. Sub-type from manufacturer data length and flags.\n- 0x0006 (Microsoft): Windows devices\n- 0x0075 (Samsung): Samsung phones/tablets\n- 0x009E (Fitbit): Fitness trackers\n- 0x0157 (Garmin): GPS watches / fitness devices\n- 0x0059 (Nordic): Tile trackers (Nordic Semiconductor is used by many Tile-like devices)\n- 0x0499 (Ruuvi): Ruuvi temperature/humidity sensors\n- 0x00E0 (Google): Android devices (Nearby Share beacons)\nClassify all others as \"Unknown\". The device name field (if present in the advertisement) provides additional signal.\n\nWearable heuristic: RSSI typically -55 to -75 dBm across multiple nodes with relatively consistent signal (worn close to body). Static devices (speakers, tablets) show higher variance. Flags this heuristic as \"possibly wearable\" (not definitive).\n\n## BLERegistry\n\nNew package: mothership/internal/identity/ble.go\n\nBLERegistry struct: backed by SQLite table ble_devices.\n\nSQLite schema:\nCREATE TABLE ble_devices (\n mac TEXT PRIMARY KEY,\n name TEXT,\n manufacturer TEXT,\n device_type TEXT, -- apple_phone, apple_earbuds, fitbit, garmin, tile, samsung, unknown\n label TEXT, -- user-assigned label\n person_id TEXT, -- FK to people.id\n rssi_min INTEGER,\n rssi_max INTEGER,\n rssi_avg INTEGER,\n first_seen DATETIME,\n last_seen DATETIME,\n is_archived BOOLEAN DEFAULT FALSE,\n last_seen_node_mac TEXT\n);\n\nCREATE TABLE people (\n id TEXT PRIMARY KEY, -- uuid\n name TEXT NOT NULL,\n color TEXT, -- hex colour for dashboard rendering\n created_at DATETIME DEFAULT CURRENT_TIMESTAMP\n);\n\nCREATE TABLE person_devices (\n person_id TEXT,\n device_mac TEXT,\n PRIMARY KEY (person_id, device_mac)\n);\n\nBLERegistry methods:\n- ProcessRelayMessage(nodeMac string, devices []BLEDevice): upsert all devices, update last_seen, update RSSI stats\n- GetDevices(includeArchived bool) []BLEDeviceRecord\n- UpdateLabel(mac, label string) error\n- AssignToPerson(mac, personID string) error\n- CreatePerson(name, color string) (Person, error)\n- GetPeople() []Person\n- ArchiveStale(olderThan time.Duration): set is_archived=true for devices not seen for > olderThan\n\n## BLE MAC Randomisation Handling\n\nModern iPhones and Android phones randomise their BLE MAC address periodically (every 10-15 minutes for iPhones, similar for Android). This is a fundamental privacy feature. The implications for spaxel:\n\n1. The same physical phone appears as multiple different MAC addresses in the registry. The BLERegistry will create new entries for each rotated address.\n2. Long-term tracking of phones by MAC is unreliable. The registry will accumulate many entries for a single phone over time.\n3. Workarounds: (a) Apple uses Resolvable Private Addresses (RPA) that can be resolved with the Identity Resolving Key (IRK) — requires pairing, not available without user action. (b) Device name is sometimes consistent across rotations. (c) Wearable devices (Fitbit, Garmin, AirTag) typically do NOT rotate their MACs — they provide reliable long-term tracking.\n\nThe dashboard must clearly explain this limitation in the \"People and Devices\" panel:\n\"Your phone's Bluetooth address changes regularly for privacy reasons. For reliable person tracking, use a Fitbit, Garmin watch, or AirTag, which have stable addresses.\"\n\nGrouping heuristic: if two devices have the same manufacturer data prefix (first 6 bytes) and name, and were never seen simultaneously at high RSSI from the same node, they are likely the same device with a rotated MAC. Surface this as a \"possible duplicate\" suggestion in the UI: \"These may be the same device: [mac1] and [mac2]. Merge?\"\n\n## REST API\n\nGET /api/ble/devices: returns list of BLEDeviceRecord, optionally filtered by ?archived=true\nGET /api/ble/devices/{mac}: returns single device with full history\nPUT /api/ble/devices/{mac}: update label, device_type, or person assignment. Body: {\"label\":\"Alice's Phone\",\"device_type\":\"apple_phone\",\"person_id\":\"uuid-123\"}\nDELETE /api/ble/devices/{mac}: archive (not hard delete)\n\nGET /api/people: returns list of People with their associated devices\nPOST /api/people: create person. Body: {\"name\":\"Alice\",\"color\":\"#3b82f6\"}\nPUT /api/people/{id}: update name or color\nDELETE /api/people/{id}: soft-delete (retain historical data)\n\n## Dashboard Panel\n\n\"People and Devices\" sidebar panel showing:\n- People section: list of defined people with avatar (initials in circle with their color), device count, last seen time\n - Per person: click to expand, shows associated devices\n - \"Add person\" button opens inline form\n- All devices section (below people): list of devices not yet assigned to a person\n - Per device: device type icon (Apple logo, Fitbit icon, etc.), last seen node (abbreviated), last seen timestamp, RSSI bar\n - Inline label edit on double-click\n - Drag-and-drop to assign to a person card\n - Archive button (hides from active list, accessible via \"Show archived\" toggle)\n- Privacy notice: \"Phones may appear multiple times due to address rotation. Wearables and AirTags have stable addresses.\"\n\n## Tests\n\n- Test device auto-detection: Apple company ID 0x004C -> device_type \"apple_phone\", Fitbit 0x009E -> \"fitbit\"\n- Test that ProcessRelayMessage correctly upserts devices and updates last_seen and RSSI stats\n- Test ArchiveStale marks devices not seen for > 7 days as archived\n- Test person creation and device-to-person assignment API calls\n- Test MAC randomisation handling: two devices with same name and no simultaneous sighting are flagged as possible duplicates\n- Test that archived devices are excluded from GetDevices(false) but included in GetDevices(true)\n\n## Acceptance Criteria\n\n- Discovered BLE devices appear in the dashboard \"People and Devices\" panel within 30 seconds of first observation\n- Device type is auto-detected correctly for Apple, Fitbit, Garmin, and Samsung devices\n- User can assign labels and associate devices with named people via the dashboard UI\n- Drag-and-drop device-to-person assignment works in the UI\n- Devices not seen for > 7 days are automatically archived and hidden from the active list\n- Privacy limitation is clearly documented in the panel UI\n- Possible duplicate MAC-rotated devices are surfaced as merge suggestions\n- Tests pass","status":"closed","priority":3,"issue_type":"task","assignee":"juliet","created_at":"2026-03-28T01:44:02.204633291Z","created_by":"coding","updated_at":"2026-03-29T18:07:39.656772405Z","closed_at":"2026-03-29T18:07:39.656662663Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred"],"dependencies":[{"issue_id":"spaxel-2wg","depends_on_id":"spaxel-c0q","type":"blocks","created_at":"2026-03-28T03:29:14.172209347Z","created_by":"coding","metadata":"{}","thread_id":""}]} {"id":"spaxel-32o","title":"Link weather diagnostics and repositioning advice","description":"## Background\n\nEven with good hardware and correct placement, some links will chronically underperform. A user who placed a node on a metal shelf, behind a TV, or in a corner will see consistently poor detection without understanding why. Telling users \"your detection quality is low\" is useless without telling them what to do about it. Link weather diagnostics provide root-cause analysis and specific, actionable repositioning advice — including 3D visualisation of why a link is performing poorly and where to move a node to fix it.\n\nThe name \"link weather\" is deliberate: just as weather forecasts present complex atmospheric state in human terms (\"partly cloudy with 60% chance of rain\"), link weather presents complex RF state as: \"Node A to Node B: interference detected. Likely cause: microwave oven or 2.4GHz congestion. Try moving Node B 1.5 metres to the right.\"\n\n## DiagnosticEngine\n\nNew module: mothership/internal/diagnostics/linkweather.go\n\nDiagnosticEngine runs as a background goroutine, consuming link health history from SQLite and emitting Diagnosis structs. It runs a full diagnostic pass every 15 minutes.\n\nA Diagnosis struct contains:\n- LinkID string\n- RuleID string (identifies which rule fired)\n- Severity: INFO, WARNING, ACTIONABLE\n- Title string (human-readable headline)\n- Detail string (explanation of the diagnosis in plain language)\n- Advice string (specific actionable steps)\n- RepositioningTarget *Vec3 (3D position to move the node to, or nil if repositioning is not the solution)\n- RepositioningNodeMAC string (which node to move)\n- ConfidenceScore float64 (how confident the diagnostic engine is in this diagnosis)\n\n## Diagnostic Rules\n\nRule 1: Environmental Change\nTrigger: High baseline drift (>5% per hour) correlated across multiple links simultaneously (>50% of active links).\nTitle: \"Environmental change detected\"\nDetail: \"Multiple sensing links are showing simultaneous baseline shifts. This typically indicates a temperature change, or a large object was moved in the space. The system is adapting automatically.\"\nAdvice: \"No action needed. The baseline will re-stabilise within 30 minutes.\"\nRepositioningTarget: nil\nConfidence: 0.85 if drift is correlated across >50% of links\n\nRule 2: WiFi Congestion or Distance\nTrigger: Packet rate health < 0.8 for more than 10 minutes on a single link.\nTitle: \"Node B has low signal rate\"\nDetail: \"Node [B] is only delivering [N]% of the expected [M] packets per second. The most common causes are distance from the WiFi router or congestion from nearby networks.\"\nAdvice: \"1. Move Node [B] within 10 metres of your WiFi router. 2. If already close, check if the 2.4GHz channel is congested (3+ networks on overlapping channels). 3. ESP32-S3 supports both 2.4GHz and 5GHz — if your router supports 5GHz, update Node B's WiFi config to use the 5GHz SSID.\"\nRepositioningTarget: nil (advice is router proximity, not specific coordinates)\n\nRule 3: Near-Field Metal Interference\nTrigger: Low phase stability (< 0.4) sustained for > 30 minutes during known-quiet periods.\nTitle: \"Metal interference near Node [A]\"\nDetail: \"The sensing link [A to B] has unstable phase measurements even when no one is moving. This is typically caused by metal objects in the near field of the node's antenna (within 10cm): metal shelves, radiators, TV backs, or large appliances.\"\nAdvice: \"Check for metal objects within 10cm of Node [A]. If Node [A] is on a metal surface or shelf, mount it on a non-metal bracket or wall. Try repositioning it 20-30cm away from metal surfaces.\"\nRepositioningTarget: nil (advice is clearance from metal, not a specific position)\n\nRule 4: Fresnel Zone Blockage (Half-Room Dead Zone)\nTrigger: Consistent miss rate (>30% of test walks that should be detected are missed) in a specific area of the room, AND the missing area correlates geometrically with an obstacle in the link's Fresnel zone.\nThis rule requires the feedback loop data (Phase 7, spaxel-i28) — specifically the user-submitted false negatives with position information. If no feedback data is available, this rule can trigger heuristically when one side of the room consistently shows lower blob confidence scores.\nTitle: \"Coverage gap detected — possible obstruction\"\nDetail: \"The area near [zone description] shows lower detection coverage. An obstacle may be blocking the path between Node [A] and Node [B], interrupting their sensing zone.\"\nAdvice: \"Move Node [B] [direction] by approximately [distance] to restore coverage. The target position is marked in green in the 3D view.\"\nRepositioningTarget: computed_optimal_position (see below)\n\nRule 5: Periodic Interference Spikes\nTrigger: Periodic spikes in deltaRMS variance (3-10 events per hour, each lasting 1-3 minutes) not correlated with occupancy data (no people detected moving).\nTitle: \"Periodic interference detected\"\nDetail: \"Node [A] to Node [B] is experiencing regular interference bursts [N] times per hour. This pattern is consistent with a microwave oven, a cordless phone, or a pulsed 2.4GHz source.\"\nAdvice: \"Consider the following: 1. Is Node [A] or Node [B] near a kitchen? Microwave ovens cause strong 2.4GHz interference. 2. A cordless DECT phone or baby monitor near one of the nodes may be the source. 3. Try moving the affected node at least 2 metres from any 2.4GHz appliances.\"\nRepositioningTarget: nil (interference is appliance-specific)\n\n## Repositioning Advice in 3D\n\nFor Rule 4 (Fresnel zone blockage), compute the optimal repositioning target:\n1. Use the GDOP-based coverage optimiser from Phase 5 self-healing fleet (spaxel-jc4) to compute the position that maximises GDOP for the blocked zone while keeping all other nodes fixed.\n2. The optimal position is the computed_optimal_position Vec3.\n3. In the 3D dashboard, render a \"ghost\" node at this position: translucent version of the node mesh, with a dashed line from the current position to the ghost position.\n4. Show expected GDOP improvement: \"Moving Node B here would improve detection in the east corner from [N]% to [M]%.\"\n\n## Weekly Reliability Trends\n\nStore daily health score averages in SQLite: link_health_daily (link_id TEXT, date DATE, avg_health REAL, min_health REAL, max_health REAL, PRIMARY KEY (link_id, date)).\n\nA background job runs daily at midnight and writes the day's health averages from the link health log (link_health_log table: link_id, timestamp, composite_score).\n\nDashboard shows for each link: 7-day sparkline of daily average health score. \"Best day\" annotation (highest average) and \"worst day\" annotation (lowest average). This gives users a sense of long-term reliability.\n\n## Files to Create or Modify\n\n- mothership/internal/diagnostics/linkweather.go: DiagnosticEngine and all 5 rules\n- mothership/internal/diagnostics/reposition.go: repositioning target computation\n- mothership/internal/health/linkhealth.go: add link_health_log table writes\n- dashboard/js/linkhealth.js: link health panel, diagnostics display, ghost node rendering\n- mothership/internal/dashboard/routes.go: GET /api/links/{id}/diagnostics, GET /api/links/{id}/health-history\n\n## Tests\n\n- Test Rule 1 (environmental change): inject simultaneous high-drift events across 60% of links, verify diagnosis fires with Severity=INFO\n- Test Rule 2 (WiFi congestion): inject packet_rate=0.7 for 15 minutes, verify diagnosis fires with appropriate advice text\n- Test Rule 3 (metal interference): inject phase_stability=0.3 for 35 minutes during a quiet window, verify diagnosis fires\n- Test Rule 4 (Fresnel blockage): requires feedback data — inject synthetic false-negative feedback events clustered in one spatial zone, verify diagnosis fires and RepositioningTarget is non-nil\n- Test Rule 5 (periodic interference): inject 5 deltaRMS variance spikes per hour for 2 hours, verify diagnosis fires with correct periodicity estimate\n- Test weekly trend aggregation: inject 7 days of health scores, verify daily averages are correctly computed and stored\n- Test that repositioning target is within room bounds and improves GDOP\n\n## Acceptance Criteria\n\n- All 5 diagnostic rules fire correctly on synthetic test data that matches their trigger conditions\n- Repositioning advice for Rule 4 appears as a ghost node in the 3D dashboard view\n- Expected GDOP improvement shown alongside repositioning ghost node\n- Weekly 7-day sparkline visible in link health panel for each link\n- Diagnostics accessible via API and displayed in Link Health panel on link click\n- Tests pass","status":"closed","priority":3,"issue_type":"task","assignee":"juliet","created_at":"2026-03-28T01:43:13.596164634Z","created_by":"coding","updated_at":"2026-03-29T18:07:39.683230580Z","closed_at":"2026-03-29T18:07:39.683089345Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred"],"dependencies":[{"issue_id":"spaxel-32o","depends_on_id":"spaxel-axa","type":"blocks","created_at":"2026-03-28T03:29:14.023730499Z","created_by":"coding","metadata":"{}","thread_id":""}]} -{"id":"spaxel-3ca","title":"Add time-travel debugging","description":"Implement:\n- Pause live mode\n- Timeline scrubbing\n- Replay 3D from recorded CSI data\n\nAcceptance: Can replay 24 hours of historical data with full 3D visualization.","status":"in_progress","priority":2,"issue_type":"task","assignee":"golf","created_at":"2026-04-09T14:54:38.737598265Z","created_by":"coding","updated_at":"2026-04-09T17:05:17.191417146Z","source_repo":".","compaction_level":0,"original_size":0,"labels":["failure-count:16","mitosis-child","mitosis-depth:1","parent-spaxel-sl2"]} +{"id":"spaxel-3ca","title":"Add time-travel debugging","description":"Implement:\n- Pause live mode\n- Timeline scrubbing\n- Replay 3D from recorded CSI data\n\nAcceptance: Can replay 24 hours of historical data with full 3D visualization.","status":"closed","priority":2,"issue_type":"task","assignee":"golf","created_at":"2026-04-09T14:54:38.737598265Z","created_by":"coding","updated_at":"2026-04-09T17:19:09.927594538Z","closed_at":"2026-04-09T17:19:09.927498317Z","close_reason":"Time-travel debugging fully implemented: Pause live mode (Pause button in dashboard, pauseLiveMode() in replay.js), Timeline scrubbing (replay scrubber with seek API /api/replay/seek, ScanRange in store.go for time-based queries), Replay 3D from recorded CSI data (BroadcastReplayBlobs in hub.go, updateReplayBlobs in viz3d.js, fusion engine integration). 24h recording buffer: 360MB default in RecordingStore, configurable via maxMB parameter. REST API endpoints: /api/replay/start, /api/replay/stop, /api/replay/seek, /api/replay/tune, /api/replay/set-speed, /api/replay/set-state, GET /api/replay/sessions, GET /api/replay/session/{id}. Frontend: replay.js with timeline scrubber, playback controls, tuning panel. All three acceptance criteria met.","source_repo":".","compaction_level":0,"original_size":0,"labels":["failure-count:17","mitosis-child","mitosis-depth:1","parent-spaxel-sl2"]} {"id":"spaxel-3ps","title":"Detection feedback loop and accuracy tracking","description":"## Background\n\nEvery detection algorithm produces errors. False positives (detected presence when no one is there) are annoying and erode trust. False negatives (missed detection of a real person) are dangerous for safety applications. The feedback loop gives users a direct mechanism to correct errors and the system learns from those corrections. Showing users measurable improvement over time (\"You've provided 47 corrections. Accuracy improved 12% this week\") creates a virtuous engagement loop and transforms users into active participants in improving the system.\n\n## Feedback UI Elements\n\nEvery detection event exposed to the user should have feedback affordances. Three contexts:\n\n1. Dashboard 3D view: Each active track has a small thumbs-up/down icon that appears on hover/focus. Clicking thumbs-down opens a quick inline form.\n\n2. Activity timeline (Phase 8): Every detection event entry has thumbs-up/thumbs-down at the end of the row. Space-efficient: 2 icon buttons.\n\n3. Push notifications: Fall and anomaly notifications include a quick-reply option (via ntfy actions or Pushover callbacks): \"False alarm — clear this.\"\n\n4. \"I was here and wasn't detected\" button: On the timeline panel, a button \"Report missed detection\" opens a form: \"When? [time picker, default: now]\", \"Where? [zone picker]\", \"Who? [person picker, optional]\". Submits as a FALSE_NEGATIVE feedback event with the user-provided position.\n\nFeedback form for thumbs-down:\n- \"What was wrong?\" (radio buttons):\n - \"No one was there (false alarm)\"\n - \"Someone was missed at this location\"\n - \"Wrong person identified\"\n - \"Wrong zone/location\"\n- Optional free-text \"Notes\" field\n- Submit / Cancel\n\n## Feedback Storage\n\nSQLite schema:\nCREATE TABLE detection_feedback (\n id TEXT PRIMARY KEY,\n event_id TEXT, -- references events table (activity timeline)\n event_type TEXT, -- \"blob_detection\", \"zone_transition\", \"fall_alert\", \"anomaly\"\n feedback_type TEXT, -- \"TRUE_POSITIVE\", \"FALSE_POSITIVE\", \"FALSE_NEGATIVE\", \"WRONG_IDENTITY\", \"WRONG_ZONE\"\n details_json TEXT, -- {\"zone_id\":\"...\", \"person_id\":\"...\", \"notes\":\"...\"}\n timestamp DATETIME DEFAULT CURRENT_TIMESTAMP,\n applied BOOLEAN DEFAULT FALSE, -- set to TRUE after weight refinement processes it\n processed_at DATETIME\n);\n\nThe applied flag enables incremental processing: the weight learner (Phase 7 self-improving localisation) queries WHERE applied = FALSE, processes batches, and marks them TRUE.\n\n## Accuracy Metrics\n\nCompute precision/recall/F1 per link, per zone, and per person weekly. This requires knowing the true positives, false positives, and false negatives.\n\nGround truth sources:\n- User thumbs-up -> TRUE_POSITIVE for the corresponding detection event\n- User thumbs-down (false alarm) -> FALSE_POSITIVE for the detection event\n- User \"missed detection\" report -> FALSE_NEGATIVE for the reported time/zone\n\nNote: ground truth is sparse — users will not feedback every event. We use the feedback we have as a sample. Assume events without feedback are TRUE_POSITIVE for the purpose of precision estimates (conservative: this means precision is an upper bound, not exact).\n\nMetrics computed weekly:\n- precision = TP / (TP + FP) — of all detections, what fraction were correct\n- recall = TP / (TP + FN) — of all true presence events, what fraction were detected\n- F1 = 2 * precision * recall / (precision + recall)\n- Per-link metrics: which links have the most false positives (worst precision)\n- Per-zone metrics: which zones are most often missed (worst recall)\n\nStorage: detection_accuracy (week TEXT, scope_type TEXT, scope_id TEXT, precision REAL, recall REAL, f1 REAL, tp_count INT, fp_count INT, fn_count INT, computed_at DATETIME). Scope types: \"system\", \"link\", \"zone\", \"person\".\n\n## Accuracy Trend Display\n\nDashboard \"Accuracy\" panel (in expert mode):\n- Overall accuracy gauge: composite F1 score as a circular gauge (0-100%)\n- Week-over-week trend graph: sparkline of weekly F1 over the last 8 weeks\n- \"You've provided N corrections. Your accuracy improved X% this week.\" — motivational counter\n- Per-zone breakdown: bar chart of precision/recall per zone (click a zone bar to jump to it in 3D view)\n- Per-link breakdown: link health vs. feedback score correlation (are high-health links also high-accuracy?)\n- Feedback count: total corrections given, open corrections (not yet processed), processed corrections\n\nThe accuracy trend display intentionally shows the improvement trajectory, not just the absolute value, to reinforce that feedback has an effect.\n\n## Feedback Application\n\nProcessing happens in a background goroutine (mothership/internal/learning/feedback_processor.go) that runs every 6 hours or when triggered manually.\n\nFor FALSE_POSITIVE events with associated CSI data (in the recording buffer from Phase 2):\n- Retrieve the CSI data from the recording buffer at the event timestamp for all links\n- Add the CSI frame data to a \"known false positive\" set in SQLite: false_positive_frames (link_id, timestamp, delta_rms, context_json)\n- The weight learner (self-improving localisation bead) uses this set as negative examples\n\nFor FALSE_NEGATIVE events with user-reported position:\n- Add to \"known false negative\" set: false_negative_frames (link_id, timestamp, expected_position_xyz, context_json)\n- The weight learner uses this as a positive example at the specified position\n\nAfter processing, mark feedback.applied = TRUE.\n\n## Files to Create or Modify\n\n- mothership/internal/learning/feedback_processor.go: feedback processing pipeline\n- mothership/internal/analytics/accuracy.go: weekly metric computation\n- dashboard/js/feedback.js: thumbs-up/down UI components (reusable across 3D view and timeline)\n- dashboard/js/accuracy.js: Accuracy panel rendering\n- mothership/internal/dashboard/routes.go: POST /api/feedback, GET /api/accuracy\n\n## Tests\n\n- Test feedback storage: POST /api/feedback with each feedback_type, verify SQLite record created\n- Test accuracy metric computation with synthetic TP/FP/FN data: 8 TP, 2 FP, 1 FN -> precision=0.8, recall=0.888\n- Test weekly rollup: 7 days of daily feedback -> correctly aggregated weekly metric\n- Test that applied=false events are found and marked as applied after processor run\n- Test \"improvements\" counter: feedback_count increases on each POST /api/feedback call\n\n## Acceptance Criteria\n\n- Thumbs-up/down buttons appear on active tracks in 3D view and on all timeline events\n- \"Missed detection\" button and form available in timeline panel\n- Feedback stored in SQLite with correct feedback_type and details\n- Accuracy metrics computed weekly and stored in detection_accuracy table\n- Accuracy panel shows week-over-week trend (requires at least 2 weeks of data)\n- Feedback improvement counter shows correct counts\n- Applied flag correctly set after processor run\n- Tests pass","status":"closed","priority":3,"issue_type":"task","assignee":"sp4","created_at":"2026-03-28T01:49:50.419277632Z","created_by":"coding","updated_at":"2026-03-29T22:08:03.778130122Z","closed_at":"2026-03-29T22:08:03.778000167Z","close_reason":"Implementation complete: feedback storage (SQLite), accuracy computation (precision/recall/F1 weekly), feedback processor (6h interval), API endpoints (/api/learning/*), frontend feedback UI (thumbs up/down, missed detection form), accuracy panel (F1 gauge, sparkline, per-zone breakdown). All 12 tests pass.","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","failure-count:1"],"dependencies":[{"issue_id":"spaxel-3ps","depends_on_id":"spaxel-zvs","type":"blocks","created_at":"2026-03-28T03:29:14.442377218Z","created_by":"coding","metadata":"{}","thread_id":""}]} {"id":"spaxel-3rd","title":"Wire WebSocket integration for zone changes","description":"Ensure zone changes from CRUD endpoints reflect in live 3D view within one WebSocket cycle. Acceptance: creating/updating/deleting a zone via REST API triggers an update broadcast through the WebSocket system.","status":"closed","priority":2,"issue_type":"task","assignee":"echo","created_at":"2026-04-07T17:01:33.587080369Z","created_by":"coding","updated_at":"2026-04-07T18:42:55.455708044Z","closed_at":"2026-04-07T18:42:55.455446177Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","mitosis-child","mitosis-depth:1","parent-spaxel-0ii"]} {"id":"spaxel-403","title":"Implement anomaly detection & security mode","description":"Build pattern learning and anomaly detection for security.\n\nDeliverables:\n- 7-day pattern learning algorithm\n- Anomaly scoring against learned patterns\n- Security mode integration\n\nAcceptance: System detects deviations from learned patterns; accuracy improves measurably over 4 weeks.","status":"closed","priority":2,"issue_type":"task","assignee":"golf","created_at":"2026-03-29T19:25:04.187535979Z","created_by":"coding","updated_at":"2026-04-09T12:18:14.752621360Z","closed_at":"2026-04-09T12:18:14.752279788Z","close_reason":"Anomaly detection & security mode implementation verified complete.\n\nDeliverables implemented:\n- 7-day pattern learning algorithm with Welford's online algorithm (analytics/patterns.go)\n- Anomaly scoring against learned patterns with z-score based computation\n- Security mode integration with Armed/Disarmed/ArmedStay states\n\nAcceptance criteria met:\n- System detects deviations from learned patterns via multiple anomaly types (UnusualHour, UnknownBLE, MotionDuringAway, UnusualDwell)\n- Accuracy improves measurably through feedback loop integration with learning/feedback_store\n\nKey components:\n- PatternLearner: 7-day cold start, hourly pattern updates, per-slot readiness checking\n- Detector: Multiple anomaly types, configurable thresholds, alert chain with timers\n- Security API: /api/security/arm, /api/security/disarm, /api/security/status\n- Alert Handler: Dashboard → webhook → escalation notification chain\n- Integration: Fully wired in main.go with zones, BLE registry, dashboard, and feedback store","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","failure-count:922","mitosis-child","mitosis-depth:1","parent-spaxel-i28"]} @@ -62,7 +62,7 @@ {"id":"spaxel-d04","title":"Implement security mode dashboard UI","description":"## Dashboard UI (dashboard/js/security-panel.js)\n\n### Security mode card (always visible in header or sidebar)\n- Arm / Disarm toggle button with confirmation dialog\n- Status badge: DISARMED / LEARNING (N days remaining) / ARMED / ALERT\n- Learning period progress bar: '5 of 7 days complete'\n- Last anomaly: '2 hours ago — kitchen motion at 3:14am'\n\n### Alert banner\n- Full-width red banner when anomaly triggered while armed\n- Description, timestamp, affected zone\n- Acknowledge button (POST /api/anomalies/{id}/acknowledge)\n\n### Anomaly timeline tab\n- List of recent anomaly events with severity, zone, timestamp\n- Links to timeline view for full context\n\n## Acceptance\n- Learning period progress updates on page refresh\n- Anomaly alert banner appears within 2s of detection\n- Acknowledged alerts disappear from the banner (not from history)","status":"closed","priority":2,"issue_type":"task","assignee":"alpha","created_at":"2026-04-06T16:09:35.859782007Z","created_by":"coding","updated_at":"2026-04-07T14:22:13.362922232Z","closed_at":"2026-04-07T14:22:13.362861907Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","failure-count:1","mitosis-child","mitosis-depth:1","parent-spaxel-a55"]} {"id":"spaxel-d41","title":"Create pre-deployment simulator","description":"Build simulator with:\n- Virtual space definition\n- Virtual nodes\n- Synthetic walkers\n- GDOP overlay\n\nAcceptance: Simulator produces realistic synthetic data matching real-world conditions.","status":"open","priority":2,"issue_type":"task","created_at":"2026-04-09T14:54:38.804838310Z","created_by":"coding","updated_at":"2026-04-09T14:54:38.804838310Z","source_repo":".","compaction_level":0,"original_size":0,"labels":["mitosis-child","mitosis-depth:1","parent-spaxel-sl2"]} {"id":"spaxel-dbd","title":"Add floor plan dashboard UI","description":"## Dashboard (dashboard/js/floorplan-setup.js)\n- Setup panel section: 'Floor Plan' with upload button\n- On image select: POST to /api/floorplan/image; display uploaded image on ground plane in 3D scene\n- Calibration UI: click point A on image → click point B → enter real-world distance in meters → Save\n- Compute pixel-to-meter scale factor: scale = distance_m / pixel_distance(A,B)\n- Apply scale and rotation to Three.js ground plane texture on load\n\n## Acceptance\n- Uploaded image displayed as ground plane texture in 3D view\n- Calibrated coordinate system maps pixel positions to correct meter positions\n- Image persists across server restart","status":"closed","priority":2,"issue_type":"task","assignee":"golf","created_at":"2026-04-07T14:46:37.333473683Z","created_by":"coding","updated_at":"2026-04-09T12:37:44.104723182Z","closed_at":"2026-04-09T12:37:44.104601579Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","mitosis-child","mitosis-depth:1","parent-spaxel-6hd"]} -{"id":"spaxel-esn","title":"Create synthetic walkers","description":"Implement synthetic walkers that move through the virtual space between nodes.\n\nAcceptance:\n- Walkers can traverse between virtual nodes\n- Movement patterns produce realistic synthetic data","status":"in_progress","priority":2,"issue_type":"task","assignee":"hotel","created_at":"2026-04-09T16:11:25.513037845Z","created_by":"coding","updated_at":"2026-04-09T16:59:12.649370358Z","source_repo":".","compaction_level":0,"original_size":0,"labels":["mitosis-child","mitosis-depth:1","parent-spaxel-d41"]} +{"id":"spaxel-esn","title":"Create synthetic walkers","description":"Implement synthetic walkers that move through the virtual space between nodes.\n\nAcceptance:\n- Walkers can traverse between virtual nodes\n- Movement patterns produce realistic synthetic data","status":"closed","priority":2,"issue_type":"task","assignee":"hotel","created_at":"2026-04-09T16:11:25.513037845Z","created_by":"coding","updated_at":"2026-04-09T17:15:23.000870233Z","closed_at":"2026-04-09T17:15:23.000764431Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["failure-count:1","mitosis-child","mitosis-depth:1","parent-spaxel-d41"]} {"id":"spaxel-ez4","title":"Detection explainability overlay","description":"## Background\n\nWhen a blob appears in an unexpected position, or an alert fires that seems wrong, the first question is \"why?\" The explainability overlay answers this question visually in the 3D scene, without requiring the user to understand deltaRMS, Fresnel zones, or UKF — though the data is available for those who want it. This transforms a \"magic box\" into a comprehensible physical system.\n\nThis is also the most important debugging tool for a developer tuning the system: seeing which links contributed most to a blob position, and by how much, is the fastest path to understanding localisation errors.\n\n## ExplainabilitySnapshot\n\nThe FusionEngine (spaxel-m9a) is extended to emit an ExplainabilitySnapshot alongside each BlobUpdate. This snapshot contains all the data needed to explain why a specific blob appeared at a specific position.\n\nExplainabilitySnapshot struct (mothership/internal/fusion/explain.go):\n- blob_id: the ID of the blob being explained\n- blob_position: Vec3 — final estimated position\n- per_link_contributions: []LinkContribution\n - link_id, tx_mac, rx_mac\n - weight float64 — the geometric Fresnel weight for this blob position\n - learned_weight float64 — the learned spatial weight (from weight learner, Phase 7)\n - combined_weight float64 = weight * learned_weight\n - delta_rms float64 — the current deltaRMS for this link\n - contribution_pct float64 — percentage of total fusion score contributed by this link\n - fresnel_intersection_volume float64 — volume of Fresnel zone ellipsoid that overlaps the blob's voxel (proxy for \"how much does this link see this position\")\n- ble_match: optional — if identity is matched: {device_mac, person_id, person_label, ble_distance_m, triangulation_confidence}\n- fusion_score float64 — total occupancy grid score at blob position\n- timestamp of snapshot\n\nThe snapshot is broadcast via WebSocket as \"blob_explain\" message type, alongside the regular \"blob_update\". The frontend requests a snapshot by sending {\"type\":\"request_explain\",\"blob_id\":\"...\"} — the server then enriches the next blob update with the explain data.\n\n## 3D Explain Mode UI\n\nRight-click (desktop) or long-press (mobile, 300ms) on any blob/track in the Three.js scene triggers explain mode.\n\nScene transformation in explain mode:\n1. All link lines dim to 20% opacity (using THREE.MeshBasicMaterial.opacity)\n2. Contributing links — those with contribution_pct > 2% — increase to 100% opacity and glow with colour intensity mapped to contribution_pct (low contribution = pale blue, high contribution = bright yellow)\n3. First Fresnel zone ellipsoids rendered for each contributing link: THREE.Mesh with SphereGeometry scaled by (a, b, b) and rotated to the link axis, translucent wireframe + fill (opacity 0.1). The ellipsoid colour matches the link line colour.\n4. A \"blob explanation panel\" (sidebar overlay, not a Three.js object) shows the breakdown:\n - Blob position in metres: \"Detected at (3.2m, 1.8m, 1.0m)\"\n - Fusion score: \"Detection confidence: [N]%\"\n - Contributing links table: link name, contribution %, deltaRMS, health score — sorted by contribution descending\n - Motion sparkline: small 30-second deltaRMS chart per link (uses the recording buffer data if available, otherwise the in-memory history)\n - BLE match details: \"Identity: Alice (BLE triangulation, confidence 82%, 0.4m from blob)\"\n - If no BLE match: \"Identity: Unknown (no BLE device match)\"\n\nExit explain mode: click anywhere outside the blob, or press Escape. Scene returns to normal opacity levels.\n\n## Fresnel Ellipsoid Geometry\n\nThe first Fresnel zone ellipsoid geometry for a link:\n- TX position P1, RX position P2\n- Link distance d = |P1 - P2|\n- WiFi wavelength lambda = 0.06m (5 GHz) or 0.125m (2.4 GHz) — use the channel from the node's hello message\n- Semi-major axis: a = (d + lambda/2) / 2\n- Semi-minor axis: b = sqrt(a^2 - (d/2)^2)\n- Centre: midpoint(P1, P2)\n- Orientation: the major axis is along the P1->P2 unit vector\n\nIn Three.js: SphereGeometry with radius=1, then scale (a, b, b) with the correct rotation matrix (use THREE.Quaternion.setFromUnitVectors to align with P1->P2 direction).\n\n## Motion Sparkline\n\nFor each contributing link in the explanation panel, show a 30-second history of deltaRMS as a small canvas sparkline (using the existing amplitude history if available from the dashboard WebSocket connection, or fetching from GET /api/recordings/{link_id}/recent?seconds=30 if the recording buffer is available).\n\nThe sparkline shows the moment of detection as a vertical line at the right edge. A horizontal dashed line shows the current motion threshold. Visually conveying \"the signal crossed the threshold at this moment.\"\n\n## Files to Create or Modify\n\n- mothership/internal/fusion/explain.go: ExplainabilitySnapshot, emission logic in FusionEngine\n- mothership/internal/fusion/engine.go: extend to emit ExplainabilitySnapshot alongside BlobUpdate\n- dashboard/js/explain.js: explain mode 3D scene transforms, sidebar panel\n- dashboard/js/fresnel.js: Fresnel ellipsoid geometry helper (reused by Fresnel debug overlay bead)\n- mothership/internal/dashboard/hub.go: blob_explain WebSocket message type\n\n## Tests\n\n- Test ExplainabilitySnapshot generation: with 3 known links and a blob at a known position, verify per_link_contributions are computed correctly\n- Test contribution_pct sums to approximately 100% across all links with non-zero weight\n- Test Fresnel ellipsoid geometry: for TX at (0,0,0) and RX at (4,0,0) with lambda=0.06: a ≈ 2.015, b ≈ 0.345. Verify these values from the geometry computation.\n- Test that explain mode correctly dims/highlights links in the Three.js scene (test via scene state inspection, not visual rendering)\n- Test that WebSocket \"request_explain\" message triggers snapshot emission in the next update cycle\n- Test sidebar panel rendering with mock ExplainabilitySnapshot data\n\n## Acceptance Criteria\n\n- Right-click on any blob triggers explain mode with correct contributing link highlighting\n- Fresnel ellipsoids render at correct positions and sizes for all contributing links\n- Confidence breakdown panel shows per-link contributions that sum to 100%\n- Non-contributing links visually dimmed in explain mode\n- Motion sparklines show 30-second history for each contributing link\n- BLE match details shown when identity is available\n- Escaping explain mode restores all link opacities to normal\n- Tests pass","status":"open","priority":3,"issue_type":"task","created_at":"2026-03-28T01:55:18.006377304Z","created_by":"coding","updated_at":"2026-03-28T03:29:14.817464555Z","source_repo":".","compaction_level":0,"original_size":0,"dependencies":[{"issue_id":"spaxel-ez4","depends_on_id":"spaxel-i28","type":"blocks","created_at":"2026-03-28T03:29:14.817442776Z","created_by":"coding","metadata":"{}","thread_id":""},{"issue_id":"spaxel-ez4","depends_on_id":"spaxel-s70","type":"blocks","created_at":"2026-03-28T01:55:20.955603637Z","created_by":"coding","metadata":"{}","thread_id":""}]} {"id":"spaxel-fi6","title":"Implement Portals CRUD REST endpoints","description":"Implement CRUD endpoints for portals: GET/POST /api/portals, PUT/DELETE /api/portals/{id}. Include OpenAPI-style godoc comments. Portal changes must reflect in live 3D view within one WebSocket cycle.","status":"closed","priority":2,"issue_type":"task","assignee":"foxtrot","created_at":"2026-04-07T13:56:27.334232115Z","created_by":"coding","updated_at":"2026-04-07T17:56:13.860592476Z","closed_at":"2026-04-07T17:56:13.860493596Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","failure-count:2","mitosis-child","mitosis-depth:1","parent-spaxel-21n"]} {"id":"spaxel-fll","title":"Dashboard WebSocket: snapshot-on-connect + incremental update protocol","description":"## Overview\nImplement the snapshot+incremental WebSocket protocol so the dashboard renders immediately on connect without waiting for a full state cycle.\n\n## Protocol spec\n\n### On new /ws/dashboard connection (within 100 ms):\nSend a full snapshot message:\n {type: 'snapshot', blobs: [...], nodes: [...], zones: [...], links: [...], alerts: [...], ble_devices: [...], triggers: [...], timestamp_ms: N}\n\n### Subsequent messages (at 10 Hz):\nOmit type field; send only state that changed since last tick:\n {blobs: [...], nodes: [...], confidence: 0.87, timestamp_ms: N}\nUnchanged arrays may be omitted entirely (null = no change)\n\n## Implementation (mothership/internal/dashboard/hub.go)\n\n- Hub maintains lastSnapshot: full state snapshot updated on each tick\n- On new client connection: serialize lastSnapshot as JSON, send immediately\n- On each tick: compute delta (changed fields only); broadcast to all established clients\n- Snapshot must be sent before the client is added to the broadcast list to avoid race\n\n## Reconnect handling (dashboard/js/app.js)\n- On WebSocket open: set awaitingSnapshot = true\n- On first message: if type === 'snapshot', merge into app state and clear flag\n- On subsequent messages: apply as incremental updates\n\n## Performance requirement\n- Snapshot delivery: < 100 ms after connection established, even with 10+ blobs, 16+ nodes, 20+ zones\n- Test: connect client, measure time to first render; must be < 150 ms end-to-end\n\n## Acceptance\n- Browser devtools shows first WS message with type='snapshot' within 100 ms of upgrade\n- Subsequent messages at 10 Hz omit type field\n- Reconnect after 5s disconnection shows correct current state immediately","status":"closed","priority":2,"issue_type":"task","assignee":"bravo","created_at":"2026-04-06T13:09:42.683611381Z","created_by":"coding","updated_at":"2026-04-07T02:03:04.204480908Z","closed_at":"2026-04-07T02:03:04.204253757Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred"]} @@ -113,6 +113,7 @@ {"id":"spaxel-pv5","title":"Backup: SQLite Online Backup API streaming endpoint","description":"## Overview\nImplement GET /api/backup using SQLite's Online Backup API for consistent hot backups without downtime or temp files.\n\n## Implementation (mothership/internal/ — new backup.go)\n\n### Why Online Backup API:\n- Simple file copy misses in-flight WAL pages and produces inconsistent backups\n- sqlite3_backup_* copies page-by-page; readers/writers continue uninterrupted\n- No temp file needed: stream directly to HTTP response\n\n### Go implementation using go-sqlite3 (CGO) or modernc.org/sqlite:\nfunc StreamBackup(w http.ResponseWriter, src *sql.DB):\n 1. Open in-memory destination DB: sqlite3_open(':memory:', &pDest)\n 2. Init backup: pBackup = sqlite3_backup_init(pDest, 'main', pSrc, 'main')\n 3. Loop: sqlite3_backup_step(pBackup, 100) until SQLITE_DONE\n 4. sqlite3_backup_finish(pBackup)\n 5. Read all bytes from pDest and write to http.ResponseWriter\n\n### Response format:\n- Content-Type: application/zip\n- Content-Disposition: attachment; filename='spaxel-backup-.zip'\n- Zip contents:\n - spaxel.db (from backup)\n - floor_plan/ directory (if exists)\n - VERSION file\n\n### Endpoint:\nGET /api/backup — requires session auth; streams zip directly; no temp files written\n\n## Acceptance\n- Backup completes while mothership is actively processing CSI frames\n- Downloaded .db file opens cleanly in sqlite3 CLI: PRAGMA integrity_check returns 'ok'\n- Backup size reasonable (not 0 bytes, not gigabytes for fresh install)\n- Simultaneous write during backup does not produce corrupt backup (verify with PRAGMA integrity_check)","status":"closed","priority":2,"issue_type":"task","assignee":"bravo","created_at":"2026-04-06T13:10:29.966455717Z","created_by":"coding","updated_at":"2026-04-07T10:17:12.858443123Z","closed_at":"2026-04-07T10:17:12.858299524Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","failure-count:10"]} {"id":"spaxel-pvz","title":"Time-travel debugging and CSI replay","description":"## Background\n\nThe CSI recording buffer (Phase 2, spaxel-tqj) stores 48 hours of raw CSI frames on disk. Time-travel debugging lets you pause the live 3D view, scrub a timeline to any point in that 48-hour window, and replay the 3D scene exactly as it was at that moment. This is the most powerful debugging tool in spaxel: if a false alert fired at 3am, you can replay those 10 minutes and see exactly which links fired, what the blob positions were, and why the alert triggered. Parameter tuning without hardware becomes possible: change the motion threshold slider and immediately see how different the replay result would have been.\n\n## ReplayEngine\n\nNew package: mothership/internal/replay/engine.go\n\nReplayEngine manages the replay lifecycle:\n- state: LIVE, PAUSED, REPLAYING, SEEKING\n- replay_position: current replay timestamp\n- replay_speed: float64 (1.0 = real-time, 5.0 = 5x speed, 0.0 = paused)\n- linked_session_id: the WebSocket session ID of the client requesting replay (each dashboard session has its own replay state)\n\nReplayEngine.Seek(t time.Time): reads the recording buffer to the specified timestamp. Uses the segment file structure from spaxel-tqj: finds the correct segment file for time t, scans forward to the exact frame at t. Target: seek time < 1 second.\n\nReplayEngine.Play(speed float64): starts reading frames from the buffer at the specified speed and feeding them through the signal processing pipeline.\n\n## Replay Processing Pipeline\n\nThe replay pipeline is a copy of the live processing pipeline but with all outputs redirected to \"replay\" namespaced WebSocket messages:\n- \"replay_blob_update\" instead of \"blob_update\"\n- \"replay_track_update\" instead of \"track_update\"\n- \"replay_link_health\" instead of \"link_health\"\n\nThe replay pipeline uses a separate instance of:\n- SignalProcessor (with possibly modified parameters from the tuning sliders)\n- FusionEngine\n- TrackManager\n\nThese are cloned from the live instances at replay start so they inherit the current configuration, then modified by slider values.\n\nThe replay pipeline is self-contained: it does not affect the live pipeline in any way. Live detection continues while replay is active.\n\n## Parameter Tuning During Replay\n\nWhile in replay mode, the dashboard shows a \"Tuning\" panel with sliders for key signal processing parameters:\n- Motion threshold: deltaRMS threshold for motion detection (default from config, range 0.001 to 0.1)\n- Baseline tau: EMA time constant in seconds (default 30s, range 5s to 300s)\n- Fresnel weight sigma: Gaussian sigma for Fresnel zone contribution (default 0.1m, range 0.01m to 0.5m)\n- Minimum confidence for detection: composite minimum confidence before blob is reported (default 0.3)\n\nChanging any slider: the replay engine discards the current replay pipeline state and re-processes from the current replay_position with the new parameters. This takes at most 1-2 seconds for a typical segment (the CSI frames are already on disk; it's fast CPU processing).\n\n\"Apply to Live\" button: copies the currently-active replay parameters to the live configuration and persists them to the mothership config file. The live pipeline picks up the new values within one processing cycle. Requires confirmation modal: \"This will change the live detection configuration. Continue?\"\n\n## Dashboard Controls\n\nEntering replay mode: clicking the \"Pause\" button (or pressing Space) on the live dashboard:\n1. Pauses the live 3D view (3D scene stops updating)\n2. Shows the timeline scrubber: a horizontal bar spanning the 48-hour recording window\n3. Event markers appear on the scrubber at the timestamps of activity timeline events (zone transitions, alerts, etc.)\n4. \"Live\" chip in the dashboard header changes to \"Replay\" chip\n\nTimeline scrubber:\n- Click to seek to any position in the 48-hour window\n- Drag for continuous scrubbing\n- Event markers: small coloured ticks on the scrubber. Clicking a marker seeks to that event and jumps the activity timeline selection to that event row.\n- The current replay position is shown as a draggable thumb with a timestamp tooltip (\"2026-03-27 03:14:22\")\n\nPlayback controls:\n- Play/Pause button (Space key shortcut)\n- Speed selector: 1x, 5x, 10x\n- Step-forward button: advances replay by 1 second\n- \"Back to Live\" button: exits replay mode and resumes live updates\n\nThe 3D scene in replay mode: shows a \"REPLAY\" watermark badge in the top-left corner (so it's clear the view is not live). All live blob and track updates are suppressed while in replay mode (only replay_ prefixed messages update the scene).\n\n## Seek Performance\n\nThe recording buffer (spaxel-tqj) uses 1-hour segment files. To seek to timestamp T:\n1. Identify the correct segment file: {linkID}-{year}-{month}-{day}-{hour}.csi\n2. Binary search within the file: CSI frames are variable-length but each has a 24-byte header with timestamp_us. Scan forward from start of file to the frame nearest T. O(n) but files are ≤ 1 hour = at most 180,000 frames at 50 Hz. At 64-byte average header read, this is < 10MB scan and typically completes in < 200ms.\n3. Buffer a few seconds of frames ahead of T for smooth playback start.\n\nFor all active links: seek all link segment files in parallel (goroutines). Total seek time < 1s.\n\n## Files to Create or Modify\n\n- mothership/internal/replay/engine.go: ReplayEngine, state machine, seek, play, parameter injection\n- mothership/internal/replay/pipeline.go: replay signal processing pipeline (cloned from live)\n- mothership/internal/recording/ (spaxel-tqj): add SeekToTimestamp(t time.Time) method\n- mothership/internal/dashboard/hub.go: replay_ namespaced WebSocket message routing\n- dashboard/js/replay.js: timeline scrubber UI, playback controls, tuning panel\n- mothership/internal/dashboard/routes.go: WebSocket commands for replay control (type: \"replay_seek\", \"replay_play\", \"replay_pause\", \"replay_set_params\")\n\n## Tests\n\n- Test seek: create a mock recording buffer with known frames at known timestamps. Seek to an arbitrary timestamp, verify the returned frame is the closest one to the target.\n- Test that replay pipeline processes frames identically to live pipeline for the same input (regression test with saved CSI data and known expected output blobs)\n- Test parameter slider: change motion_threshold via replay command, verify the replay pipeline uses the new threshold on subsequent frames\n- Test \"Apply to Live\" correctly writes parameter changes to the live config\n- Test that live pipeline output is unaffected while replay is active (isolation test)\n- Test seek performance: 1-hour segment file with 180,000 frames, seek to timestamp in the middle, complete in < 500ms\n\n## Acceptance Criteria\n\n- Seek to any point in 48-hour window completes in < 1 second for all active links\n- Replay produces identical blob positions to original live processing for the same CSI input\n- Parameter sliders re-process the current replay position in < 3 seconds\n- \"Apply to Live\" copies parameters correctly and live detection immediately uses new values\n- Timeline scrubber event markers correctly align with activity timeline events\n- \"Back to Live\" correctly resumes live detection without any stale state\n- Tests pass","status":"open","priority":3,"issue_type":"task","created_at":"2026-03-28T01:56:04.674847447Z","created_by":"coding","updated_at":"2026-03-28T03:29:14.698778779Z","source_repo":".","compaction_level":0,"original_size":0,"dependencies":[{"issue_id":"spaxel-pvz","depends_on_id":"spaxel-i28","type":"blocks","created_at":"2026-03-28T03:29:14.698749622Z","created_by":"coding","metadata":"{}","thread_id":""},{"issue_id":"spaxel-pvz","depends_on_id":"spaxel-tqj","type":"blocks","created_at":"2026-03-28T01:56:07.776160379Z","created_by":"coding","metadata":"{}","thread_id":""}]} {"id":"spaxel-pwf","title":"Self-improving localisation with BLE ground truth","description":"## Background\n\nThe Fresnel zone fusion engine (spaxel-m9a) computes localisation by weighting each link's deltaRMS contribution according to the geometric intersection of candidate voxels with the Fresnel zone ellipsoid. These weights are currently uniform and based purely on geometry. In practice, some links are better at detecting motion in specific parts of the room than others — due to reflection geometry, multipath, furniture layout, and antenna orientation. By using BLE RSSI positions as continuous ground truth (when a person's labelled phone or wearable is visible), we can refine the per-link, per-zone weights to match observed physical reality.\n\n## Self-Improving Mechanism\n\nNew package: mothership/internal/learning/weights.go\n\nWeightLearner runs as a background goroutine. It operates on ground truth samples collected during normal operation.\n\nA ground truth sample is collected when BOTH:\n1. A confident BLE triangulated position is available for a known person (confidence > 0.7 from identity matching bead spaxel-nqh)\n2. A CSI blob position is within 0.5m of the BLE position (confirming the blob corresponds to that person)\n\nSample structure: {timestamp, person_id, ble_position Vec3, blob_position Vec3, per_link_delta_rms map[linkID]float64, per_link_health map[linkID]float64}\n\nThese samples are stored in SQLite: ground_truth_samples (id, timestamp, person_id, position_xyz, per_link_deltas_json, per_link_health_json). The table is capped at 10,000 samples per person (oldest first out) to prevent unbounded growth.\n\n## Online Weight Learning\n\nAfter accumulating 100+ samples for a given spatial zone (the room is divided into zones of 0.5m x 0.5m grid cells for this purpose), run incremental linear regression:\n\nPrediction model: position_estimate = sum_i (w_i * delta_rms_i) / sum_i w_i, where w_i are the learnable per-link weights.\n\nThe objective is to minimise the mean squared error between the position estimate from the weighted fusion and the ground truth BLE positions, over all samples in the zone.\n\nUpdate rule (stochastic gradient descent, online):\nFor each new ground truth sample:\n- Compute current position estimate using current weights\n- Compute error = ground_truth_position - estimated_position\n- For each link i: w_i += learning_rate * error * delta_rms_i / |delta_rms_vector|\n- learning_rate = 0.001 (small to prevent overfitting to transient environmental changes)\n- Apply L2 regularisation: w_i *= (1 - regularisation * learning_rate) where regularisation = 0.01\n\nClip weights to [0, 5] to prevent divergence. Normalise weight vector to unit sum after each update.\n\n## Validation Gate\n\nTo prevent the learned weights from degrading accuracy (overfitting, transient environmental changes, sensor noise):\n\nHold out 20% of samples as a validation set (random selection). After each batch of 50 weight updates, compute the mean position error on the validation set using the updated weights vs. the original (geometric) weights.\n\nOnly persist the updated weights if: validation_error_new < validation_error_original * 0.95 (at least 5% improvement on the validation set).\n\nIf the validation check fails, discard the weight update and log: \"Weight update rejected: no improvement on validation set. Keeping current weights.\"\n\nThis is a conservative gate. The threshold is configurable (fleet.weight_improvement_threshold, default 0.05).\n\n## Weight Storage\n\nSQLite table: link_weights (link_id TEXT, zone_grid_x INT, zone_grid_y INT, weight REAL, sample_count INT, last_updated DATETIME, validation_improvement REAL, PRIMARY KEY (link_id, zone_grid_x, zone_grid_y)).\n\nZone grid: floor is divided into 0.5m cells. zone_grid_x = floor(x / 0.5), zone_grid_y = floor(y / 0.5). This allows position-dependent weights — a link might be excellent for localisation in one area and poor in another.\n\nOn FusionEngine update: instead of using geometric Fresnel zone weights alone, multiply by the learned spatial weight for the voxel being evaluated (bilinear interpolation between grid cells for smooth transitions).\n\nFallback: if no learned weight exists for a grid cell (insufficient samples), use the geometric weight (learned weight = 1.0). This ensures correctness during the learning period.\n\n## Accuracy Trend in Dashboard\n\nThe accuracy improvement from learning should be visible to users. In the \"Accuracy\" dashboard panel (Phase 7 feedback loop bead):\n\nAdd \"Position accuracy\" subsection:\n- Median position error (m): computed weekly from ground truth samples. median(|ble_position - blob_position|) over all weekly samples.\n- Week-over-week trend: sparkline of weekly median position error. Arrow indicating direction (improving/degrading).\n- Sample count: \"Based on N position measurements from M people this week\"\n- \"Accuracy improving\" badge when position error has decreased by > 10% vs previous week.\n\n## Files to Create or Modify\n\n- mothership/internal/learning/weights.go: WeightLearner, SGD update, validation gate\n- mothership/internal/learning/samples.go: ground truth sample collection, SQLite storage\n- mothership/internal/fusion/engine.go (spaxel-m9a): integrate learned weights in FusionEngine\n- mothership/internal/dashboard/routes.go: GET /api/accuracy/weights (debug endpoint showing current weight map)\n- dashboard/js/accuracy.js: position accuracy trend chart\n\n## Tests\n\n- Test ground truth sample collection gates correctly: confidence > 0.7 AND BLE-blob distance < 0.5m -> sample collected; confidence = 0.6 -> no sample\n- Test SGD weight update: after 100 samples with known ground truth, verify weights move in the direction that reduces error\n- Test validation gate: inject a batch of adversarial samples that would degrade accuracy, verify gate rejects the update\n- Test bilinear interpolation between adjacent grid cells produces smooth weight values\n- Test weight fallback: FusionEngine correctly uses geometric weight=1.0 when no learned weight exists for a grid cell\n- Test SQLite cap: inserting 10,001 samples removes the oldest one, maintaining the 10,000 cap\n\n## Acceptance Criteria\n\n- Position error decreases measurably over 2+ weeks of operation with BLE ground truth data (target: from initial ~1.2m to < 0.8m median error)\n- Validation gate prevents weight regressions (mock adversarial samples do not degrade fusion accuracy)\n- Weight updates persist across mothership restarts\n- Position accuracy trend visible in dashboard Accuracy panel\n- Sample collection rate visible (samples per day per person) in dashboard\n- Tests pass","status":"closed","priority":3,"issue_type":"task","assignee":"bravo","created_at":"2026-03-28T01:50:34.214065492Z","created_by":"coding","updated_at":"2026-03-30T00:12:00.715207673Z","closed_at":"2026-03-30T00:12:00.715088959Z","close_reason":"Implemented self-improving localization with BLE ground truth. Created spatial weight learner with SGD, validation gate, bilinear interpolation. Added position accuracy visualization to dashboard. All tests implemented.","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred"],"dependencies":[{"issue_id":"spaxel-pwf","depends_on_id":"spaxel-3ps","type":"blocks","created_at":"2026-03-28T01:50:36.699492024Z","created_by":"coding","metadata":"{}","thread_id":""},{"issue_id":"spaxel-pwf","depends_on_id":"spaxel-zvs","type":"blocks","created_at":"2026-03-28T03:29:14.574878149Z","created_by":"coding","metadata":"{}","thread_id":""}]} +{"id":"spaxel-q9d","title":"Add GDOP overlay","description":"Implement GDOP (Geometric Dilution of Precision) overlay for the simulator.\n\nAcceptance:\n- GDOP overlay visualizes accuracy metrics across the virtual space\n- Simulator produces realistic synthetic data matching real-world conditions","status":"in_progress","priority":2,"issue_type":"task","assignee":"hotel","created_at":"2026-04-09T16:11:25.552156606Z","created_by":"coding","updated_at":"2026-04-09T17:15:38.354739705Z","source_repo":".","compaction_level":0,"original_size":0,"labels":["mitosis-child","mitosis-depth:1","parent-spaxel-d41"]} {"id":"spaxel-qfp","title":"Sleep quality monitoring","description":"## Background\n\nThe breathing analysis feature (Phase 5, spaxel-r37) detects the micro-motion of breathing in stationary people. Run continuously in bedroom zones overnight, it can compute sleep quality metrics without any wearable device. Chest displacement during breathing at 15 breaths/minute produces a detectable 0.25 Hz signal in CSI. By tracking this overnight, combined with motion events (wake episodes) and the timing of presence in the bedroom zone, we can produce a sleep summary that rivals basic commercial sleep trackers — without the user wearing anything.\n\n## Sleep Session Detection\n\nSleepMonitor in mothership/internal/sleep/monitor.go.\n\nSession onset detection (all conditions must hold):\n1. Person is in a bedroom zone (zone with is_bedroom flag = true, set in zone editor)\n2. Stationary detection fires (STATIONARY_DETECTED state from breathing analysis bead)\n3. BLE device shows reduced activity (optional enhancement: phone advertising rate drops when screen is off; this is a bonus signal, not required)\nTentative onset: all conditions met. Confirmed onset: conditions hold for 15 consecutive minutes.\n\nSession end detection:\n1. Person leaves bedroom zone (zone transition event fires)\n2. OR: motion detection fires for > 2 minutes (sustained motion = getting up)\n3. OR: stationary detection drops and does not return for > 30 minutes (person left room without portal crossing — reconciliation path)\n\nSession record stored in SQLite:\nCREATE TABLE sleep_sessions (\n id TEXT PRIMARY KEY,\n person_id TEXT NOT NULL,\n zone_id TEXT NOT NULL, -- bedroom zone\n session_date DATE NOT NULL, -- the date this sleep night belongs to (typically today-1 for morning reports)\n sleep_onset DATETIME, -- time tentative detection was confirmed\n wake_time DATETIME,\n time_in_bed_minutes REAL,\n sleep_latency_minutes REAL, -- time from entering bedroom to sleep onset\n wake_episode_count INTEGER DEFAULT 0,\n wake_after_sleep_onset_minutes REAL, -- total time awake after first sleep onset\n breathing_rate_mean REAL,\n breathing_rate_stddev REAL,\n breathing_anomaly_count INTEGER DEFAULT 0, -- breathing < 8 or > 25 per minute\n sleep_efficiency REAL -- (time_in_bed - waso) / time_in_bed * 100\n);\n\nCREATE TABLE sleep_wake_episodes (\n id TEXT PRIMARY KEY,\n session_id TEXT,\n episode_start DATETIME,\n episode_end DATETIME,\n duration_seconds REAL\n);\n\n## Sleep Metrics Computation\n\nDuring the sleep session, SleepMonitor subscribes to:\n- Breathing data: periodic sample of breathing_freq_hz from BreathingDetector (spaxel-r37). Store in a rolling buffer.\n- Motion events: MOTION_DETECTED state transitions from LinkProcessor. Each motion event during a confirmed sleep session is a potential wake episode.\n\nWake episode classification:\n- If deltaRMS > threshold for > 3 seconds: wake episode starts\n- If deltaRMS returns below threshold and breathing signal resumes: wake episode ends\n- Store episode start/end in sleep_wake_episodes\n\nBreathing analysis during sleep:\n- Mean breathing rate (bpm): mean(breathing_freq_hz * 60) over all samples in session\n- Breathing rate standard deviation: indicates sleep stage variability (higher variance may indicate REM activity)\n- Breathing anomaly: if breathing_freq_hz * 60 < 8 or > 25 for > 3 consecutive minutes: log anomaly. This is a proxy for potential sleep apnoea or hyperventilation.\n\nSleep efficiency: (time_in_bed_minutes - wake_after_sleep_onset_minutes) / time_in_bed_minutes * 100. A value above 85% is considered good sleep efficiency.\n\n## Morning Summary Card\n\nOn first WebSocket connection from the dashboard after 6am AND after a sleep session has ended (wake_time is set):\n- Mothership pushes a \"morning_summary\" WebSocket message with the completed session data\n- Dashboard renders a dismissible card in simple mode (full width at top) and as a floating panel in expert mode\n\nCard content:\n- \"Last night: [sleep_duration] h [mm] min\"\n- Colored efficiency indicator: green (>85%), amber (70-85%), red (<70%)\n- Wake episodes: \"2 wake episodes, [total waso] min awake after sleep onset\"\n- Breathing: \"Average breathing: [N] breaths/min\"\n- Anomaly note (if applicable): \"Unusual breathing detected at [time]. [View details]\"\n- \"View full sleep report\" link (opens detailed timeline view in expert mode)\n\n## Weekly Trends\n\nDashboard \"Sleep\" panel:\n- 7-day sparkline of sleep duration per night\n- 7-day sparkline of sleep efficiency per night\n- Average breathing rate over the week\n- Week-over-week comparison: \"This week you slept 6h 48m on average (vs. 7h 12m last week)\"\n\n## Per-Person Tracking\n\nSleep monitoring is person-specific and requires BLE identity (so the system knows whose bedroom this is). Multiple people sharing a bedroom: each person has their own sleep session if their BLE devices can be distinguished. If both people are in bed simultaneously, the breathing detector may pick up a blend of two breathing rates — acknowledge this limitation in documentation.\n\nFor anonymous tracks (no BLE identity): detect in-bedroom stationary presence only (no per-person sleep report). Log \"Unidentified person in bedroom zone\" for 8+ hour periods.\n\n## Zone Configuration\n\nThe zone editor (portals bead, spaxel-qlh) is extended with a zone type selector:\n- Normal zone (default)\n- Bedroom (enables sleep monitoring)\n- Kitchen (no special behavior)\n- Children's zone (suppresses fall detection)\n\nThis is stored as zone_type in the zones table.\n\n## Files to Create or Modify\n\n- mothership/internal/sleep/monitor.go: SleepMonitor, session detection, metric computation\n- mothership/internal/sleep/report.go: morning summary generation, weekly trend aggregation\n- mothership/internal/signal/breathing.go (spaxel-r37): add tick-based sample reporting for sleep monitor\n- dashboard/js/sleep.js: morning summary card, Sleep panel\n- mothership/internal/events/events.go: SleepSessionStartEvent, SleepSessionEndEvent\n\n## Tests\n\n- Test sleep session onset: stationary detection fires, person in bedroom, 15 minutes -> session confirmed\n- Test that stationary detection < 15 minutes does not create a session (avoids brief naps misclassified)\n- Test wake episode counting: 3 MOTION_DETECTED events > 3s each during a session -> wake_episode_count = 3\n- Test wake after sleep onset calculation: 3 episodes of 5 minutes each -> waso = 15 minutes\n- Test sleep efficiency calculation: 480 minutes in bed, 45 minutes waso -> efficiency = 90.6%\n- Test breathing anomaly detection: inject 4 minutes of breathing_freq_hz = 0.1 (6 bpm) -> anomaly logged\n- Test morning summary trigger fires only on first connection after 6am AND after session end\n\n## Acceptance Criteria\n\n- Sleep session detected within 15 minutes of confirmed onset (stationary in bedroom zone)\n- Wake episodes counted correctly (tested with synthetic motion event injection)\n- Morning summary card appears on first dashboard open after wake time (6am by default, configurable)\n- Weekly trends sparkline shows 7 nights of data after 7 days\n- Sleep session data persists in SQLite across mothership restarts\n- Breathing anomaly flag fires correctly for rate < 8 or > 25 bpm\n- Tests pass","status":"open","priority":3,"issue_type":"task","created_at":"2026-03-28T01:52:06.457208929Z","created_by":"coding","updated_at":"2026-04-09T14:56:34.490882317Z","source_repo":".","compaction_level":0,"original_size":0,"labels":["deferred","failure-count:80"]} {"id":"spaxel-qgj","title":"Implement NTP client in ESP32 firmware","description":"Add NTP synchronization to firmware/main/wifi.c or ntp.c:\n- Call esp_sntp_setservername(0, ntp_server) before esp_sntp_init() on boot\n- ntp_server read from NVS 'ntp_server' key (default: 'pool.ntp.org')\n- Attempt sync for up to 10 seconds after WiFi connect; log WARN if sync fails\n- On sync failure: proceed without stagger (rely on CSMA/CA)\n- Resync every 10 minutes via esp_timer periodic callback\n- Include ntp_synced status in health JSON message\n\nAcceptance: Node health messages show ntp_synced: true when pool is reachable; ntp_synced: false when NTP blocked — node still operates normally; resync occurs every ~600s (verified via UART logs)","status":"closed","priority":2,"issue_type":"task","assignee":"alpha","created_at":"2026-04-07T14:37:00.302557793Z","created_by":"coding","updated_at":"2026-04-07T17:32:57.896842167Z","closed_at":"2026-04-07T17:32:57.896693758Z","close_reason":"done","source_repo":".","compaction_level":0,"original_size":0,"labels":["failure-count:1","mitosis-child","mitosis-depth:1","parent-spaxel-u7y"]} {"id":"spaxel-qlh","title":"Room transition portals and zone occupancy","description":"## Background\n\nKnowing a blob is at coordinates (3.2m, 1.8m, 1.0m) is useful to the algorithm, but \"Alice is in the Kitchen\" is useful to a person. Room transition portals define doorway planes between named zones. When a track's trajectory intersects a portal plane, the zone occupancy counts update and a transition event fires. This is the foundation for natural language presence display (\"Alice is in the Kitchen\"), automation triggers (\"when Alice enters the bedroom\"), and the activity timeline (\"Alice moved from Living Room to Kitchen at 14:23\").\n\n## Zone Definitions\n\nZones are named 3D volumes represented as axis-aligned bounding boxes (AABB) for simplicity. Each zone has: id (uuid), name (\"Kitchen\"), bounds_min (Vec3), bounds_max (Vec3), color (hex string for 3D overlay), created_at.\n\nSQLite schema:\nCREATE TABLE zones (\n id TEXT PRIMARY KEY,\n name TEXT NOT NULL,\n bounds_min_x REAL, bounds_min_y REAL, bounds_min_z REAL,\n bounds_max_x REAL, bounds_max_y REAL, bounds_max_z REAL,\n color TEXT DEFAULT '#3b82f6',\n created_at DATETIME DEFAULT CURRENT_TIMESTAMP\n);\n\nContainment test: a position P is in zone Z if bounds_min_x <= P.x <= bounds_max_x AND bounds_min_y <= P.y <= bounds_max_y. The Z bounds are typically 0 to ceiling height (usually 2.5m) since we track floor-plane position.\n\n## Portal Definitions\n\nA portal is a vertical plane segment spanning a doorway. It divides two zones and detects crossings.\n\nPortal schema:\nCREATE TABLE portals (\n id TEXT PRIMARY KEY,\n name TEXT, -- e.g. \"Kitchen Door\"\n zone_a_id TEXT, -- zone on one side\n zone_b_id TEXT, -- zone on other side\n plane_point Vec3, -- a point on the portal plane (e.g. centre of doorway)\n plane_normal Vec3, -- unit normal vector of the portal plane\n width REAL, -- width of the doorway in metres\n height REAL, -- height of the doorway (default: 2.1m)\n created_at DATETIME\n);\n\nA portal normal points from zone_a toward zone_b. A crossing from zone_a to zone_b has dot(velocity, normal) > 0. A crossing from zone_b to zone_a has dot(velocity, normal) < 0.\n\n## Portal Editor (3D Dashboard)\n\nExtend the node placement UI (spaxel-qq6) with portal editing:\n1. User clicks \"Add Portal\" button\n2. A vertical plane appears in the 3D scene at the camera's focal point\n3. User drags the plane using TransformControls (from Three.js addons) to position it across a doorway\n4. User adjusts width and assigns zone names on each side (dropdown of existing zones or \"Create new zone\")\n5. User clicks \"Save\" — portal is stored in SQLite and rendered as a semi-transparent divider plane in the 3D scene\n\nPortal rendering: thin coloured plane (opacity 0.3, colour #a855f7 purple) with a label at the top edge showing the portal name. When a track crosses the portal, the plane briefly flashes brighter (animated opacity increase then decay back to 0.3).\n\nZone rendering: semi-transparent coloured cuboid volumes (opacity 0.1, colour from zone.color). Zone name displayed as a floating text label at the zone centroid (using THREE.Sprite). A \"Zones\" layer toggle in the 3D view hides/shows all zones simultaneously.\n\n## Crossing Detection\n\nCrossingDetector runs as part of the TrackManager update loop (10 Hz). For each track update:\n\n1. For each active portal, test if the track crossed the portal plane in the last update step:\n - Previous position P_prev, current position P_curr\n - Check if the line segment P_prev -> P_curr intersects the portal plane within the portal's rectangular bounds (width x height centered on plane_point)\n - Intersection test: t = dot(plane_point - P_prev, normal) / dot(P_curr - P_prev, normal). If 0 <= t <= 1, compute intersection point P_int = P_prev + t*(P_curr - P_prev), then check if P_int is within the doorway rectangle.\n - Crossing direction: if dot(P_curr - P_prev, normal) > 0, direction is A_to_B; otherwise B_to_A.\n\n2. On crossing detected: update occupancy counts, emit ZoneCrossingEvent.\n\nZoneCrossingEvent: {portal_id, track_id, person_id, person_label, from_zone_id, from_zone_name, to_zone_id, to_zone_name, direction, timestamp}.\n\nThis event is:\n- Published to the internal event bus\n- Broadcast via WebSocket to dashboard as type \"zone_transition\"\n- Appended to activity timeline (Phase 8)\n- Processed by automation engine (Phase 6)\n\n## Occupancy Counter\n\nOccupancyManager maintains a per-zone current occupant list (map[zoneID][]TrackID).\n\nUpdates from two sources:\n1. CrossingDetector portal events: when a track crosses from zone A to B, move its entry in the occupancy map from A to B.\n2. Direct containment check: run every 30 seconds as a reconciliation pass. For each active track, check if it is within any zone's bounding box. If the track is in zone C but the occupancy map says it is in zone A (e.g. track was created inside a zone without crossing a portal), update accordingly.\nThe containment check prevents \"teleportation\" inconsistencies when tracks are created or resume from coasting state.\n\n## WebSocket Broadcast\n\nOn each zone occupancy change, the mothership broadcasts:\n{\"type\":\"zone_occupancy\",\"zones\":[{\"id\":\"zone-kitchen\",\"name\":\"Kitchen\",\"occupants\":[{\"track_id\":\"track-1\",\"person_id\":\"uuid-alice\",\"person_label\":\"Alice\"}]},{\"id\":\"zone-living\",\"name\":\"Living Room\",\"occupants\":[]}]}\n\nAnd specifically on crossings:\n{\"type\":\"zone_transition\",\"portal_id\":\"...\",\"person_label\":\"Alice\",\"from_zone\":\"Kitchen\",\"to_zone\":\"Living Room\",\"timestamp\":\"2026-03-27T14:23:00Z\"}\n\n## REST API\n\nGET /api/zones: list all zones with current occupancy\nPOST /api/zones: create zone\nPUT /api/zones/{id}: update zone bounds/name/color\nDELETE /api/zones/{id}: delete zone (removes from all occupancy tracking)\n\nGET /api/portals: list all portals\nPOST /api/portals: create portal\nPUT /api/portals/{id}: update portal\nDELETE /api/portals/{id}: delete portal\n\nGET /api/zones/{id}/history?since=2026-03-27T00:00:00Z: get crossing history for zone (list of ZoneCrossingEvent)\n\n## Tests\n\n- Test portal crossing detection with a track path that passes through the portal plane: verify crossing event fires with correct direction\n- Test that a track path that runs parallel to a portal plane but within 0.1m does not fire a false crossing\n- Test that a track path outside the portal's width bounds does not fire a crossing\n- Test occupancy count updates: zone Kitchen starts with 1 occupant, track crosses portal to Living Room, Kitchen count = 0, Living Room count = 1\n- Test the 30-second reconciliation pass: track that appears inside a zone without crossing a portal is correctly assigned to that zone\n- Test zone containment with a position exactly on the bounds_min edge (inclusive boundary)\n- Test that zone_transition WebSocket message is broadcast with correct from_zone and to_zone names\n\n## Acceptance Criteria\n\n- Portal editor allows placing vertical plane portals across doorways in the 3D scene\n- Zone bounding boxes are editable and render as semi-transparent volumes in 3D view\n- Zone labels update in real-time as people move between zones (\"Kitchen: Alice, Bob\")\n- Zone transition events fire within one track update cycle (100ms) of the crossing occurring\n- Reconciliation pass correctly handles tracks that appear inside zones without portal crossings\n- Zone and portal data persists across mothership restarts via SQLite\n- WebSocket broadcasts zone_occupancy after every occupancy change\n- Tests pass","status":"open","priority":3,"issue_type":"task","created_at":"2026-03-28T01:45:41.668543362Z","created_by":"coding","updated_at":"2026-03-28T03:29:14.268105795Z","source_repo":".","compaction_level":0,"original_size":0,"dependencies":[{"issue_id":"spaxel-qlh","depends_on_id":"spaxel-c0q","type":"blocks","created_at":"2026-03-28T03:29:14.268078719Z","created_by":"coding","metadata":"{}","thread_id":""},{"issue_id":"spaxel-qlh","depends_on_id":"spaxel-nqh","type":"blocks","created_at":"2026-03-28T01:45:44.642770328Z","created_by":"coding","metadata":"{}","thread_id":""}]} diff --git a/.needle-predispatch-sha b/.needle-predispatch-sha index edf55c5..cfed4f0 100644 --- a/.needle-predispatch-sha +++ b/.needle-predispatch-sha @@ -1 +1 @@ -c31d990644810c2087b70f68fb19a95eb7ff980d +acd4df2e19abbf92c1141d0fab53ca22e1168f44 diff --git a/mothership/internal/api/replay.go b/mothership/internal/api/replay.go index a1245a0..fe92f16 100644 --- a/mothership/internal/api/replay.go +++ b/mothership/internal/api/replay.go @@ -10,6 +10,7 @@ import ( "time" "github.com/go-chi/chi/v5" + "github.com/spaxel/mothership/internal/localization" "github.com/spaxel/mothership/internal/replay" sigproc "github.com/spaxel/mothership/internal/signal" ) @@ -63,6 +64,19 @@ func (h *ReplayHandler) SetBlobBroadcaster(broadcaster replay.BlobBroadcaster) { h.worker.SetBroadcaster(broadcaster) } +// SetFusionEngine sets the fusion engine for replay blob generation. +func (h *ReplayHandler) SetFusionEngine(fusionEngine interface{}) { + h.mu.Lock() + defer h.mu.Unlock() + // Type assertion to fusion engine interface + if engine, ok := fusionEngine.(interface { + Fuse(links []localization.LinkMotion) *localization.FusionResult + SetNodePosition(mac string, x, y, z float64) + }); ok { + h.worker.SetFusionEngine(engine) + } +} + // Start the replay worker. func (h *ReplayHandler) Start() { h.worker.Start() diff --git a/mothership/internal/replay/store.go b/mothership/internal/replay/store.go index 6e8bfdd..947f6bb 100644 --- a/mothership/internal/replay/store.go +++ b/mothership/internal/replay/store.go @@ -173,7 +173,7 @@ func (s *RecordingStore) Scan(fn func(recvTimeNS int64, frame []byte) bool) erro return err } recvTimeNS := int64(binary.LittleEndian.Uint64(hdr[0:8])) - frameLen := int64(binary.LittleEndian.Uint16(hdr[8:10])) + frameLen := int64(binary.LittleEndian.Uint64(hdr[8:10])) if frameLen > maxFrameBytes { return errors.New("replay: corrupt record during scan") } @@ -197,6 +197,70 @@ func (s *RecordingStore) Scan(fn func(recvTimeNS int64, frame []byte) bool) erro return nil } +// ScanRange reads records within a time range [fromNS, toNS], calling fn for each. +// fn receives the receive timestamp (Unix nanoseconds) and the raw frame bytes. +// Returning false from fn stops the scan early. +// The store is held under lock for the entire scan — callers must not call +// Append or other mutating methods from within fn. +func (s *RecordingStore) ScanRange(fromNS, toNS int64, fn func(recvTimeNS int64, frame []byte) bool) error { + s.mu.Lock() + defer s.mu.Unlock() + + if !s.hasData() { + return nil + } + + pos := s.oldestPos + for { + if pos == s.writePos { + break + } + + // Read record header: recvTimeNS(8) + frameLen(2) + var hdr [10]byte + if _, err := s.f.ReadAt(hdr[:], pos); err != nil { + return err + } + recvTimeNS := int64(binary.LittleEndian.Uint64(hdr[0:8])) + frameLen := int64(binary.LittleEndian.Uint64(hdr[8:10])) + if frameLen > maxFrameBytes { + return errors.New("replay: corrupt record during scan") + } + + // Skip records before the time range + if recvTimeNS < fromNS { + nextPos := pos + recordOverhead + frameLen + if s.wrapPos != 0 && nextPos >= s.wrapPos { + nextPos = headerSize + } + pos = nextPos + continue + } + + // Stop if we've passed the time range + if recvTimeNS > toNS { + break + } + + frame := make([]byte, frameLen) + if _, err := s.f.ReadAt(frame, pos+recordOverhead); err != nil { + return err + } + + if !fn(recvTimeNS, frame) { + break + } + + nextPos := pos + recordOverhead + frameLen + // Wrap: if we just read the last record before the wrap point, jump to data start. + if s.wrapPos != 0 && nextPos >= s.wrapPos { + nextPos = headerSize + } + pos = nextPos + } + return nil +} + // Stats returns summary statistics about the recording store. type Stats struct { HasData bool diff --git a/mothership/internal/replay/worker.go b/mothership/internal/replay/worker.go index a980b99..96bcb3a 100644 --- a/mothership/internal/replay/worker.go +++ b/mothership/internal/replay/worker.go @@ -34,6 +34,12 @@ type ReplaySession struct { baselineState map[string]*signal.BaselineState // per-link baseline } +// FusionEngine is the interface required for replay blob generation. +type FusionEngine interface { + Fuse(links []localization.LinkMotion) *localization.FusionResult + SetNodePosition(mac string, x, y, z float64) +} + // Worker reads CSI frames from a replay store and processes them. type Worker struct { mu sync.Mutex @@ -42,7 +48,7 @@ type Worker struct { store RecordingStore processor *signal.ProcessorManager - fusionEngine *localization.Engine + fusionEngine FusionEngine nodePositions map[string]localization.NodePosition // MAC -> position broadcaster BlobBroadcaster done chan struct{} @@ -116,7 +122,7 @@ func (w *Worker) SetProcessorManager(processor *signal.ProcessorManager) { } // SetFusionEngine sets the fusion engine for replay blob generation. -func (w *Worker) SetFusionEngine(fusionEngine *localization.Engine) { +func (w *Worker) SetFusionEngine(fusionEngine FusionEngine) { w.mu.Lock() defer w.mu.Unlock() w.fusionEngine = fusionEngine