name

form-sync-verification

description

Bounded, diagnostic-first procedure for verifying whether a pushed workout actually committed to FORM goggles. Use when: (1) testing changes to form_sync.py/full_sync.py BLE push logic, (2) investigating why transfers succeed (FILE_TRANSFER_SUCCESS) but workouts don't appear on the goggles, (3) any /goal loop over goggle sync. Hardware-in- the-loop: BLE is flaky and the device is irreplaceable mid-project, so this skill is as much about what NOT to do as what to do.

FORM Goggles Sync Verification

Known state (as of 2026-07)

Server API works: create workouts, save to library, fetch protobufs.
BLE transport works: goggles ACK every transfer (11/11 FILE_TRANSFER_SUCCESS).
The gap: transport success ≠ commit. Our pushes don't appear on the goggles; the official app's sync does. Root cause unknown — likely missing context files (DEVICE_SETTINGS_V2, USER_PROFILE_V2, REMOTE_CONFIG), or unanswered goggle-initiated requests after SYNC_START.
Premium entitlement blobs are cached in ~/.formgoggles_blobs/ for replay.

Success criterion — be specific or it's noise

Never use "a workout appears in the list" as the signal. Always:

Create the test workout with a unique, greppable name including a timestamp, e.g. SYNCTEST 20260706-1732, and record the workout id returned by the server API.
Verify by reading the goggles' own index, not by inference:
```
.venv/bin/python read_goggle_indexes.py --target all
```
(Always use .venv/bin/python — the system python3 lacks bleak/protobuf.) Success = that specific workout id/name present in the relevant index (imports or saved) that was absent in the pre-push baseline.
Always capture the baseline index BEFORE pushing, so "appeared" means a diff, not a guess.

Procedure (one attempt)

Preflight (no hardware needed): run with the project venv:
```
.venv/bin/python -c "import form_sync" \
&& .venv/bin/python -c "import json;c=json.load(open('$HOME/.formgoggles.json'));assert c.get('goggleMac') and c.get('accessToken')" \
&& ls ~/.formgoggles_blobs/entitlement.bin ~/.formgoggles_blobs/subscription.bin
```
Config keys are goggleMac / accessToken (camelCase). If anything is missing, stop and report — do not attempt BLE. Note: accessToken may be expired (tokenExpires field); server-API steps refresh via refreshToken, but cached-blob mode (--entitlement-mode cached) works without the server.
Baseline: goggles on and awake → .venv/bin/python read_goggle_indexes.py --target all. Save output to a timestamped file.
Push: one sync attempt (form_sync.py or full_sync.py depending on the hypothesis being tested), full output captured to a timestamped log.
Verify: re-read indexes. Diff against baseline for the specific test id.
Record: append hypothesis → change → transfer result → index diff → conclusion to SYNC_LOG.md in the repo root (create if missing). Every attempt gets an entry, including failures — the log IS the deliverable while the root cause is unknown.

Hard limits (hardware-in-the-loop guardrails)

These are non-negotiable in any autonomous or /goal context:

Max 3 turns per /goal loop; max 2 BLE sync attempts per session (BLE flakiness means attempt 3+ is noise, and each attempt drains/heats the device).
No destructive writes to the device: no file deletions, no factory/debug service writes (Services 2/3 are unprobed factory/debug — never write to them), no firmware operations, no --replace-id against a workout that wasn't created by this session.
Never clear device state or caches — not on the goggles, and not ~/.formgoggles_blobs/ (the premium entitlement blobs are irreplaceable after the subscription lapses).
No repeated official-app intervention without human confirmation. Asking Garrick to run the official app's sync is a consumable diagnostic (it changes device state and costs a comparison baseline) — at most once, and only with explicit confirmation.
Stop on repeated identical failure: if two attempts produce byte-identical transfer results and the same index diff (i.e., nothing), stop — the hypothesis is exhausted; more attempts are not evidence.
Timeouts are ambiguous, not failures: a BLE timeout may mean the goggles slept. One retry after confirming the device is awake; then treat as environment, not code.

If not solved: diagnostic report format

When a session/loop ends without the workout committing, produce a report (in SYNC_LOG.md) containing: hypothesis tested, exact command(s) run, transfer-level results per file, pre/post index diff (or "no diff"), any goggle-initiated messages observed after SYNC_START (from the notification handler log), and the single most informative next hypothesis. Do not restart with a new hypothesis in the same loop.

Escalation order for the root cause (from prior analysis)

Respond to goggle-initiated requests after SYNC_START that we currently ignore (check notification handler logs for unanswered message types).
Add missing FormFileMessage types to the sync: DEVICE_SETTINGS_V2, USER_PROFILE_V2, REMOTE_CONFIG (full_sync.py already builds the first two — verify they're sent and ACKed in the right order relative to the app's sequence).
Capture official-app BLE traffic (ble_sniff.py / frida/hook_form.js) — requires human setup, propose it, don't do it autonomously.

FORM Goggles Sync Verification

Known state (as of 2026-07)

Server API works: create workouts, save to library, fetch protobufs.

BLE transport works: goggles ACK every transfer (11/11 FILE_TRANSFER_SUCCESS).

The gap: transport success ≠ commit. Our pushes don't appear on the goggles; the official app's sync does. Root cause unknown — likely missing context files (DEVICE_SETTINGS_V2, USER_PROFILE_V2, REMOTE_CONFIG), or unanswered goggle-initiated requests after SYNC_START.

Premium entitlement blobs are cached in ~/.formgoggles_blobs/ for replay.

Success criterion — be specific or it's noise

Never use "a workout appears in the list" as the signal. Always:

Create the test workout with a unique, greppable name including a timestamp, e.g. SYNCTEST 20260706-1732, and record the workout id returned by the server API.

Verify by reading the goggles' own index, not by inference:

.venv/bin/python read_goggle_indexes.py --target all

(Always use .venv/bin/python — the system python3 lacks bleak/protobuf.) Success = that specific workout id/name present in the relevant index (imports or saved) that was absent in the pre-push baseline.

Always capture the baseline index BEFORE pushing, so "appeared" means a diff, not a guess.

Procedure (one attempt)

Preflight (no hardware needed): run with the project venv:

.venv/bin/python -c "import form_sync" \
&& .venv/bin/python -c "import json;c=json.load(open('$HOME/.formgoggles.json'));assert c.get('goggleMac') and c.get('accessToken')" \
&& ls ~/.formgoggles_blobs/entitlement.bin ~/.formgoggles_blobs/subscription.bin

Config keys are goggleMac / accessToken (camelCase). If anything is missing, stop and report — do not attempt BLE. Note: accessToken may be expired (tokenExpires field); server-API steps refresh via refreshToken, but cached-blob mode (--entitlement-mode cached) works without the server.

Baseline: goggles on and awake → .venv/bin/python read_goggle_indexes.py --target all. Save output to a timestamped file.

Push: one sync attempt (form_sync.py or full_sync.py depending on the hypothesis being tested), full output captured to a timestamped log.

Verify: re-read indexes. Diff against baseline for the specific test id.

Record: append hypothesis → change → transfer result → index diff → conclusion to SYNC_LOG.md in the repo root (create if missing). Every attempt gets an entry, including failures — the log IS the deliverable while the root cause is unknown.

Hard limits (hardware-in-the-loop guardrails)

These are non-negotiable in any autonomous or /goal context:

Max 3 turns per /goal loop; max 2 BLE sync attempts per session (BLE flakiness means attempt 3+ is noise, and each attempt drains/heats the device).

No destructive writes to the device: no file deletions, no factory/debug service writes (Services 2/3 are unprobed factory/debug — never write to them), no firmware operations, no --replace-id against a workout that wasn't created by this session.

Never clear device state or caches — not on the goggles, and not ~/.formgoggles_blobs/ (the premium entitlement blobs are irreplaceable after the subscription lapses).

No repeated official-app intervention without human confirmation. Asking Garrick to run the official app's sync is a consumable diagnostic (it changes device state and costs a comparison baseline) — at most once, and only with explicit confirmation.

Stop on repeated identical failure: if two attempts produce byte-identical transfer results and the same index diff (i.e., nothing), stop — the hypothesis is exhausted; more attempts are not evidence.

Timeouts are ambiguous, not failures: a BLE timeout may mean the goggles slept. One retry after confirming the device is awake; then treat as environment, not code.

If not solved: diagnostic report format

Escalation order for the root cause (from prior analysis)

Respond to goggle-initiated requests after SYNC_START that we currently ignore (check notification handler logs for unanswered message types).

Add missing FormFileMessage types to the sync: DEVICE_SETTINGS_V2, USER_PROFILE_V2, REMOTE_CONFIG (full_sync.py already builds the first two — verify they're sent and ACKed in the right order relative to the app's sequence).

Capture official-app BLE traffic (ble_sniff.py / frida/hook_form.js) — requires human setup, propose it, don't do it autonomously.