Recording session time-profile — v2 onboarding demo (2026-07-03)

1. Time profile

Phase	Window	Wall clock	%	What happened
Dress rehearsal round 2 (full S1→live-flip proof)	07:28–07:40	12.6 min	16%	Restricted-Mode trust miss, "files deleted on server" dialog, live-flip finally proven
Between-segment vault prep (prep1–prep10)	07:48–07:57	9.6 min	12%	Rebuild vault from on-camera zip, plugin re-setup, click-by-click with screenshot checks
Rehearsal round 1 (Obsidian launch, login, coords)	07:07–07:15	8.3 min	11%	shot1–shot17; mis-clicks (search icon vs sign-in), pre-filled field, stray Firefox
Sync-on-save debugging	07:15–07:22	7.1 min	9%	BRAT installed plugin 0.5.0 without autoSyncOnSave; local 0.6.1 installed + Obsidian restart
seg3a bug-fix + retake prep	08:00–08:06	6.3 min	8%	Obsidian auto-list + wikilink popup ate Returns → vault3 content garbage; cleanup + record.mjs edits
Zero-state reset + docker DEV flap	07:22–07:28	5.3 min	7%	root-owned sqlite (rm: Permission denied), container Exited (137) after DEV=true recreation
Takes seg3a(retake)+seg3a2+seg3b + verify	08:06–08:11	5.1 min	7%	frame-extraction verify after each
Recon: spec + optB reuse + codebase research	06:58–07:02	4.2 min	5%	Found dev code is 111111 not 000000; DevMode config hunt across ~15 greps
Endcard + record.mjs authoring	07:40–07:44	3.9 min	5%	104s single Write of record.mjs
Takes seg1a + seg1b + verify	07:44–07:47	3.5 min	5%	Clean first takes
post.py compositing script write (cut-off)	08:11–08:14	3.4 min	4%	205s single Write; session snapshot ends here
seg2 take + frame checks	07:57–08:00	2.7 min	4%	Clean take; 8 frame extracts to verify
memcli + trip2g boot, sign-in probing	07:02–07:04	2.4 min	3%	Boot fine; sign-in code hunted in prior-session tool-result files
Vault/Obsidian config prep	07:05–07:07	1.4 min	2%
VO generation (ElevenLabs) + whisper check	07:04–07:05	1.2 min	2%	2 API iterations, v3 accepted; not a time sink
Total		77 min	100%

Grouped: rehearsal/verification 28 min (36%) · state reset/re-prep 15 min (19%) · recording+verify 11.3 min (15%) · script authoring 7.3 min (9%) · recon/boot/VO 9.2 min (12%) · bug-fix/retake 6.3 min (8%).

Longest single operations (gap to next tool call): 215s stall after a prep click (model latency digesting a screenshot), 205s Write post.py, 104s Write record.mjs, 103s docker DEV-wrapper flap (Exited (137)), 76s seg2 record+frame-extract, 67s Obsidian/Chromium teardown, 66s vault3 content restore.

The see-click loop itself is the diffuse killer: 66 screenshot Reads and 62 xdotool commands; gaps preceding screenshot Reads alone sum to ~16.5 min, and each cycle (click → import → Read PNG → think) runs 10–25s.

2. Retakes / retries / loops

• 1 true retake: seg3a — Obsidian's list auto-continuation + wikilink autocomplete popup swallowed Return keys, corrupting typed note content; content cleaned, record.mjs patched, re-rolled (~6 min lost). seg3a2 was a planned extra segment, not a retake.
• 4 Obsidian restarts: initial launch, plugin 0.5.0→0.6.1 upgrade, zero-state reset, mid-rehearsal pkill for vault2.
• 3 docker container cycles: initial boot, DEV=true recreation, recovery from Exited (137).
• 2 stray-window hunts: unexpected Chromium window (07:12) and a Firefox spawned by an off-script xdg-open (07:13) — PID forensics + targeted kill, ~2.5 min.
• ~6 coordinate mis-clicks corrected via screenshot loop (search toggle instead of sign-in icon, Restricted Mode instead of "Trust author", pre-filled email field needing ctrl+a, etc.).
• 2 full environment resets (rehearsal → zero state → dress rehearsal → zero state → takes). Each reset is click-by-click, ~5–10 min.

Total loop/rework loss: ~15–18 min (~20% of session).

3. Top difficulties (ranked by time cost)

1. Unscripted UI state discovery (~28 min). Every button was found by blind-click + import -window root + Read PNG. Representative: 07:08:40 "I hit the search toggle instead. The sign-in icon is the leftmost one." Coordinates from the v1 run were not reusable/recorded, so both rehearsal rounds re-derived them.
2. Environment state not resettable by script (~15 min). Zero-state restore fought root-owned docker volumes (rm: cannot remove '.../local.sqlite3': Permission denied → alpine-container rm workaround), a DEV=true container recreation that OOM/exited (137), and manual vault re-download/re-setup between segments.
3. Plugin/app behavior surprises (~13 min). (a) BRAT delivered plugin 0.5.0 without autoSyncOnSave — the demo's core feature — vs local build 0.6.1; (b) "Trust author" click actually landed on Restricted Mode, silently disabling the plugin ("No push attempt at all — plugin may not have loaded"); (c) first sync of a downloaded vault raises a "Files deleted on server → Keep locally" dialog nobody planned for.
4. Obsidian typing semantics (~6 min + the only retake). Auto-continued lists and the [[ autocomplete popup eat Return keystrokes typed via xdotool.
5. Codebase archaeology mid-recording (~5 min). Dev sign-in code (111111 vs assumed 000000), DevMode env plumbing, title-extraction rules — all hunted with grep during the session instead of being in the scenario/harness docs.

4. Speed-up levers

1. Persist a coordinate/state map per stage layout. Both rehearsal rounds (20+ min) rediscovered the same buttons at 2560×1440. Save a coords.json (sign-in icon, trust button, community-plugins toggle, ribbon sync, hamburger→live-reload, download button) next to record.mjs and validate it with one screenshot diff at boot instead of 66 look-loops. This attacks the single biggest cost.
2. Script the reset, not just the take. Make reset-stage.sh do the full zero-state: docker run alpine rm for root-owned data (already proven in-session), container recreate with DEV=true baked into the memcli wrapper from the start, vault wipe/restore, Obsidian relaunch with pre-trusted vault + pre-enabled plugin (community-plugins.json + data.json written before launch — skips the Trust/Restricted-Mode dialog class entirely, which cost ~5 min and one silent failure).
3. Pre-bake the plugin and dialogs. Install the local 0.6.1 build into the vault template (never BRAT), and pre-answer the first-sync "files deleted on server" dialog off-camera in the harness (the session did this manually at 07:36 — encode it).
4. Kill the typing-retake class: paste note content via xdotool key ctrl+v from clipboard (or write the .md file on disk and let Obsidian pick it up, which the session already used at 07:40 for _index.md) instead of type-ing markdown that triggers list-continuation/wikilink popups. This removes the only retake.
5. Split content from capture. The 205s post.py and 104s record.mjs Writes plus VO are text-only artifacts — keep them as committed, parameterized scripts in scripts/hero/v2/ so a VO/endcard/timing tweak never re-triggers rehearsal or takes. VO is already cheap (~1 min) and cached per-beat; keep per-beat mp3s so one line change re-renders one beat.
6. Guard against off-script windows: the stray Firefox came from an xdg-open on :98. Set BROWSER=/bin/true/neutralize xdg-open in the stage env, and have record.mjs assert the exact window list before rolling — a black-frame/extra-window check is cheaper than PID forensics mid-session.

Caveat: compositing/encode/GIF cost is unknown — the snapshot ends before they ran. Profile that tail in the next session before optimizing it.