2026-04-06 four-spec rollout: priority pipeline, HUD line streaming, Live Activity rework, session cost

[FJiangArthur]

Tracking issue for the 2026-04-06 four-spec batch. Specs are committed under docs/superpowers/specs/2026-04-06-*.md.

Specs

• Spec A — Priority Pipeline + Transcription Isolation — docs/superpowers/specs/2026-04-06-priority-pipeline-design.md
  • Two-model architecture (smart on Q&A press, light scanner for auto-questions + fact-check)
  • Priority hierarchy: Q&A > fact-check > auto-answer
  • 1s Q&A debounce, speculative prefetch from light model
  • Phase 0 transcription-isolation diagnostic before fix
• Spec B — Line-by-Line HUD Streaming — docs/superpowers/specs/2026-04-06-hud-line-streaming-design.md
  • Flush on completed visual line (~6-7x fewer BLE writes)
  • Protocol nomenclature corrected: cmd is 0x4E, 0x30/0x40/0x70 are screen_status bits
• Spec C — Live Activity Rework — docs/superpowers/specs/2026-04-06-live-activity-rework-design.md
  • Question / Pause / Resume buttons via App Intents → Darwin notification → MethodChannel
  • Visibility gated to active sessions only
  • Auto-detected questions excluded from Live Activity (and from glasses HUD)
  • iOS 26 deployment target — no @available gating
• Spec D — Cumulative Session Cost Tracking — docs/superpowers/specs/2026-04-06-session-cost-tracking-design.md
  • Live cost pill in recording bar, persist per-session totals, per-model breakdown
  • Cost never displayed in Live Activity or on glasses HUD

Todo

• Review all 4 specs and approve (gate before plan-writing)
• Write implementation plan for Spec A (do first — B, C, D depend on it)
• Write implementation plan for Spec B
• Write implementation plan for Spec C
• Write implementation plan for Spec D
• Decide fate of two older unrelated specs (2026-04-04-transcription-quality-overhaul-design.md, 2026-04-05-unified-tune-menu-design.md) — keep, archive, or delete

Dependencies

A ──┬──> B   (B uses Spec A's HudStreamSession cancel hook)
    ├──> C   (C uses Spec A's shared Q&A entry point)
    └──> D   (D needs Spec A's modelRole tagging on LLM calls)

Reference findings (committed)

• G1 BLE protocol: CLAUDE.md updated, command byte is 0x4E, new_char_pos is not an append offset, no scroll command exists. Sources: official Even Realities demo + community Python SDK.
• Deployment target: iOS 26 across all targets. No iOS 16/17 fallback paths.

[FJiangArthur]

Plans landed

All 4 implementation plans written and committed.

Plan	Path	Commit
A — Priority Pipeline + Transcription Isolation	docs/superpowers/plans/2026-04-06-priority-pipeline.md	df3df2d
B — Line-by-Line HUD Streaming	docs/superpowers/plans/2026-04-06-hud-line-streaming.md	c457cc4
C — Live Activity Rework	docs/superpowers/plans/2026-04-06-live-activity-rework.md	8944138
D — Cumulative Session Cost Tracking	docs/superpowers/plans/2026-04-06-session-cost-tracking.md	436be83

Phase counts

• Plan A: 5 phases, ~41 tasks. Phase 0 is diagnostic instrumentation (no behavior change). Phase 1c forks into 1c-A (audio-pipeline isolation) and 1c-B (async/isolate isolation) — executor picks branch based on Phase 0 evidence.
• Plan B: 4 phases, 16 tasks. TDD-first; Phase 1 lands HudStreamSession + HudPacketSink with full unit coverage before any wiring.
• Plan C: 6 phases. Phase 1 closes the auto-detected-question leak in Live Activity content (audit finding).
• Plan D: 8 phases. Phase 1 is a non-destructive Drift v3→v4 migration.

Caveats from Plan A

• A few LlmService / SettingsManager method names are assumed by convention; Plan A tasks 1a.4 / 1a.5 tell the executor to verify against current files before pasting code.
• Q&A button presses route via BluetoothManager → method.bluetooth channel → Dart lib/services/ble.dart, not directly through AppDelegate.swift.
• Plan A Phase 2 regression test references onUpdateForTest and buildTestEngine helpers that may need to be added to test/helpers/test_helpers.dart.

Recommended execution order

Phase 0 of Plan A first (instrumentation only, no behavior change). Its results determine Plan A Phase 1c's fix branch and unblock Plans B / C / D's dependencies on Spec A interfaces.

[FJiangArthur]

Reconciliation with existing G1 protocol research (2026-04-06, session end)

While setting up execution, discovered that the repo had significant uncommitted in-flight work from a prior session:

• docs/research/ — 6 G1 BLE protocol research docs synthesized from 10+ open-source SDKs (JohnRThomas wiki, AugmentOS, emingenc/even_glasses, Even Realities official demo, g1-java-sdk, fahrplan, etc.). Far more comprehensive than the spec-B-agent's research, which only hit 2 repos.
• lib/services/conversation_engine.dart, proto.dart, glasses_protocol.dart, text_service.dart — partial Spec B implementation already landed: per-page aiFrameForPage helper, correct current_page_num, 400ms L→R inter-side delay, _lastStreamedByteLength / _isFirstStreamFrame / _lastStreamedPageIndex state fields.
• lib/services/g1_debug_service.dart — new service that enables firmware 0x23 0x6C 0x00 debug logging and parses incoming 0xF4 frames.
• docs/research/g1-protocol-conflict-report.md — audit of Helix vs the consolidated protocol, listing HIGH/MEDIUM severity bugs not covered by any 2026-04-06 spec.

Actions taken

1. Checkpoint commit (10905f7) preserving the in-flight work on the branch.

2. Spec reconciliation (three commits):

• dfb2001 — Spec B rewritten against checkpoint 10905f7. Audit section now describes the ACTUAL current code; scope narrowed to the remaining win (per-token → per-line flush trigger + HudStreamSession extraction + tests). Half the original spec was obsolete.
• 1199dd9 — Spec A Phase 0 patched to wire G1DebugService as instrumentation hook Fix Recording Timer to Show Actual Elapsed Time #6 (correlates firmware-side timing against phone-side timing for transcription isolation diagnostic).
• bb0131e — New Spec E: G1 Protocol Correctness, turning the conflict report HIGH/MEDIUM items into a proper design spec.

3. Plan updates (three commits):

• 14119a2 — Plan A: +1 Phase 0 task for G1DebugService wiring, renumbered 0.5→0.6, 0.6→0.7.
• 93621e2 — Plan B rewritten. 4 phases / 16 tasks → 3 phases / 10 tasks. Dropped everything landed in 10905f7; no longer touches proto.dart or glasses_protocol.dart. Scope: extract HudStreamSession, change flush trigger, add test coverage.
• d5bca09 — New Plan E. 5 phases: (1) test coverage for resolved H1/M1-M4, (2) H2 battery 0x0A parse + consumer, (3) H3 status 0x22 parse + consumer (10-byte head-up + 8-byte right-tap variants), (4) H4 BMP inversion probe + conditional fix, (5) final gate.

Specs & plans as of end of session

Spec	Plan
A b12a3be → patched 1199dd9	A df3df2d → patched 14119a2 (5 phases, ~42 tasks)
B d6bd139 → reconciled dfb2001	B c457cc4 → rewritten 93621e2 (3 phases, 10 tasks)
C 8944138 (iOS 26 patch 0b75011)	C 8944138 (6 phases)
D 22c2ab9 (Live Activity exclusion a0dac72)	D 436be83 (8 phases)
E bb0131e (new — protocol correctness)	E d5bca09 (new — 5 phases)

Ready for execution

Execution has not started yet — the reconciliation was blocking. Next steps:

1. User reviews specs + plans (still gating per brainstorming protocol).
2. Spawn subagent-driven execution starting with Plan A Phase 0 (diagnostic only, no behavior change).
3. Plan E can run in parallel with A since it has no dependencies.