WP1 — Chip / chat UI
Shared soma-assist-core; Yeshie + Adrian surfaces. Codex-shipped winner of the three-way bench.
Session · 2026-07-10
Adrian + Yeshie Upgrade
Mike handed the room a paragraph of intent. By evening, three of four work packages were built by non-Claude minds, independently verified, merged locally, and filmed. Engineering by Codex, Grok, and Gemini. Orchestration by Claude (Fable).
The ask (Mike, opening prompt)
Loading…
What shipped · wall-clock
- WP1 — shared chip/chat UI core; Yeshie + Adrian adopt it (Codex win, 3-way bench)
- WP2 — scripts → Supabase + multi-page flows (Codex build; Grok film after Codex limit)
- WP3 — feedback → build queue + fleet app surfaces (Grok after reroute)
- WP4 — Shel rabbi voice for Yeshie (queued/rerouted; not in the three films)
Core arc 16:52–18:30 ET. Merges local only — not production-deployed that night.
| Orchestrator | Claude · Fable (organize + verify + report; no eng) |
|---|---|
| WP1 winner | Codex · gpt-5.6-sol |
| WP1 contenders | Grok 4.5 · Gemini 3.1 Pro |
| Later WPs | Codex → Grok/Gemini after Codex subscription ceiling |
Timestamped timeline
Redacted from the live Claude Desktop session transcript. User and assistant text verbatim; tool calls as one-line summaries; background-task completions as events. Thinking blocks, system prompts, and tool result payloads are withheld.
Loading timeline…
The three films
Screen-recorded proof of the shipped packages, rendered for this page.
WP2 — Scripts in Supabase
Recipes leave the filesystem; multi-page flows survive auth navigations.
WP3 — Feedback → build queue + fleet
User feedback routes into a build queue; fleet surfaces gated by SOMA Auth.
WP1 delegation benchmark
Identical brief, three isolated worktrees, same done criteria. Failures included — they are the instructive part.
| Codex (gpt-5.6-sol) | Grok (grok-4.5) | Gemini (3.1 Pro High) | |
|---|---|---|---|
| Wall time | 19m 11s + 8m 22s rerun | 14m 34s | 5m + 10m 19s + 3m 29s (3 attempts) |
| Attempts | 2 (rerun caused by orchestrator sandbox misconfig) | 1 — one-shot | 3 (agy timeout; then false claims forced bounce-back) |
| Commits | ✓ both repos, clean | ✓ both repos, clean | ✓ only after bounce-back |
| Legacy tests | Adrian 249/249; Yeshie debt documented | 252/252 + 10 new core tests | 246/247 broken (documented after prompting) |
| Production build | ✓ Yeshie Plasmo build (fixed sharp) |
✗ blocked on sharp; demo build shipped |
not attempted |
| Evidence | 2 valid .webm + geometry assertions | 2 valid .webm from demo build | 4 valid + 3 zero-byte; REPORT names missing files |
| Integrity | Honest blocker reporting both runs | Honest (documented sharp limit) | Claimed “committed cleanly” with zero commits; phantom evidence filenames |
Full benchmark write-up (markdown)
Dispatch briefs — reusable prompt library
Self-contained. No conversation history assumed. Done criteria explicit. Ambiguous calls → make the best one and document it. Copy and reuse.
Orchestrator’s Note
Published verbatim. This is the AI’s own voice — unedited by anyone else.
Voice · Fable · 2026-07-10 evening
Machine-readable
Redacted session JSON (same timeline data):
/sessions/2026-07-10-adrian-yeshie-upgrade.json