Failed and deferred approaches

Approaches that were tried and rejected, or deliberately deferred, with the reason. Read this before retrying a dead end — the next agent should not spend a turn rediscovering why something doesn’t work here. Update this page when an approach is abandoned or a deferred idea is revisited.

The durable status record is PROJECT_STATUS.md; this page is the curated “do not retry without new information” list.

Rejected: cloud anything as a default

Why rejected: Pace’s headline differentiator is fully-on-device operation (speed + zero operating cost + “0 bytes sent off this Mac”). Cloud LLM, cloud STT, cloud TTS, and cloud telemetry call paths have all been removed. Cloud bridge / Direct API / CLI direct-spawn exist only as opt-in tiers, each visibly indicated (amber capsule), audit-logged, and fail-loud. Making any cloud path the default contradicts the moat. Do not reintroduce a cloud default or a silent cloud fallback (the directAPIFallsBackToLocalOnCloudFailure opt-in is the one explicit, off-by-default exception).

Rejected: dual-agent prefetch (removed as dead code)

Why rejected: a partial-transcript-keyed RAG prewarm / dual-agent prefetch was removed as dead code — it added hot-path complexity for negligible, unmeasurable TTFSW upside. The expensive prewarm (VLM screen context) already runs at PTT press via PaceScreenContextService.prewarmScreenContext, and local retrieval (BM25 + in-memory episodic) is already sub-millisecond. Revisit only if a benchmark_ttfsw.sh run shows retrieval on the critical path. (Tracked in competitive/steal-catalog.md.)

Rejected: meetily’s meeting-notes shape (partially)

What was taken: the selectable meeting-note profile idea (general / standup / one-on-one, with a general byte-for-byte compat anchor).

Rejected as non-fits from meetily:

Its 7 markdown-table templates — too rigid; Pace uses one structured shape per profile.
VAD (voice-activity detection) — Pace’s PaceMeetingTurnSegmenter already covers segmentation (Accelerate RMS + hysteresis + echo trimming).
RMS ducking — two-track capture (mic + system as separate tracks) is better than ducking one mixed track.
Real-time transcription + diarization — separate large efforts, not in scope for the notes wedge.

Deferred: persistent KV planner backend

Why deferred: blocked on TinyGPT oMLX qualification. The in-process MLX planner (Qwen3-4B) is available behind the Settings → Models toggle but is not the default — fresh installs talk via Apple Foundation Models or LM Studio. Do not flip the bundled-MLX planner to default without a 4B-vs-30B eval-gate run on real hardware; per project memory, Pace-side model work is otherwise concluded.

Deferred: grammar-constrained v10 as runtime default

Why deferred: TinyGPT / eval gated. The shipping planner remains the current MLX / Qwen stack with response_format: json_schema decode constraining. The grammar-constrained gate is a shipping-decision harness, not the runtime path.

Deferred: real-app AX smokes in CI

Why deferred: TCC makes automated live-app tests fragile — terminal xcodebuild invalidates permissions. Live-app executor smokes (scripts/smoke-executor-surface.sh, scripts/smoke-real-apps.sh) are manual-only. The 3 hardware-bound tests are gated behind TEST_RUNNER_PACE_CI in CI and still run locally.

Deferred: hosted telemetry or accounts

Why deferred: contradicts the on-device moat. Analytics are local-only no-op/timing-safe hooks (PaceAnalytics.swift); no cloud analytics SDK is linked. Do not add accounts or hosted telemetry.

Deferred: speaking-time context prefetch (episodic/RAG)

Why deferred: the expensive prewarm already runs at PTT press and local retrieval is sub-millisecond, so a partial-transcript-keyed RAG prewarm adds hot-path complexity for negligible upside — and it mirrors the dual-agent prefetch already removed as dead code. Revisit only if a benchmark shows retrieval on the critical path.

Known live-testing follow-ups (not failures — unfinished edges)

These are not rejected approaches; they are known rough edges surfaced by the live gauntlet that should be fixed, not retried as-is:

Few-shot coordinate echo — the planner can echo coordinates from few-shot examples; needs a prompt/parser guard.
PacePlannerModelResolver silent brain swap — served gemma-3-12b when qwen unloaded; needs fail-loud + served-model in traces.
MCP fixture subprocess vs parallel test workers — known CI interaction; CI runs serial because parallel workers crash on the image.

Rejected: fixing the known non-blocking warnings

Why rejected: Swift 6 concurrency warnings and the deprecated onChange in OverlayWindow.swift are intentionally not fixed (per AGENTS.md). Do not attempt these unless explicitly asked — they are noise, not signal, and chasing them risks churn on stable paths.