Cross-Pollination Brief — June 25, 2026

Two AAXT rounds ran overnight in Klatch (R43 MessageList: 100%; R44 ProjectSettings: 80%/89%), producing three copy and accessibility fixes — including a label catch that matters to any project surfacing the 5-layer model to users. Meanwhile PM's first phone UAT on the hosted Droplet found a critical pre-launch bug: the onboarding system-check hardcodes localhost ports and fails in any production environment. The alpha bundle is gated on the fix.

Letters to xian: have a question for xian about anything here or elsewhere in his work? File question-{from}-{date}-{topic}.md to dispatch mail. AI prompts human; one letter featured at the end of each brief.

Key Insights

1. "L3 context" caught by AAXT — 5-layer vocabulary in user copy needs a translation step

From: Klatch (R44 test packages/client/src/__tests__/round44-project-settings-aaxt.test.tsx; Iris triage docs/mail/iris-to-daedalus-r43-r44-copy-fixes-2026-06-25.md)
Relevant to: PM (any surface implementing the 5-layer model)

Round 44 (ProjectSettings, 10 probes, 0 Phantoms) found the Knowledge Base label reading "2 files — listed in L3 context for all channels in this project." User-proxy scored it Absent (0.95): "L3 context is unexplained jargon." Fix: "included in AI context for all channels in this project."

"L3" is the 5-layer model's shorthand for Project Memory. It's useful engineering vocabulary — but meaningless to a user who hasn't read the spec. The architectural signal: any 5-layer label that reaches user-visible copy needs a translation step before it ships.

R44 also found that the Cancel button in dirty-state doesn't communicate that clicking it discards and reverts — fix is title="Discard changes". Round 43 (MessageList, 100%, 11/11) caught the pin button's hover-only affordance; fix is aria-label="Pin to channel". All three dispatched to Daedalus as independent one-liners.

Suggested action (PM): Any PM-facing label using ADR shorthand, methodology surface names (m-XX, Pattern-NNN), or role slugs as primary labels rather than anchors is in the same family. Worth a pass before external users touch those surfaces.

2. Phone UAT finds localhost hardcoding blocking alpha launch

From: PM Exec (issues #1318 and #1319; attention rollup dev/active/exec-cohort-attention-rollup-2026-06-24.html)
Relevant to: Klatch (pre-beta deploy checklist)

xian ran phone UAT on the hosted Droplet on June 24 and found two blockers:

#1318 — Onboarding system-check shows "Services Not Running / Run: docker compose up -d" on the Droplet. Root cause: web/api/routes/setup.py hardcodes localhost:5433/6379/8000 (PostgreSQL, Redis, ChromaDB). On the Droplet those services run on Docker-internal addresses. Every alpha tester would hit this dev-environment error on first launch.

#1319 — Welcome card is small with a large empty gray area above it on mobile. Not suitable for mobile-first onboarding.

Both gate the alpha bundle. Lead Dev has them as first tasks on re-login.

This is the same class of bug as June 23's ENCRYPTION_MASTER_KEY gap (alpha 0.8.9): environment-context assumptions that look correct in development and surface only under real-device, real-deploy conditions.

Suggested action (Klatch): When the first beta server component ships, add a startup health check that tests actual service connectivity (not just that env vars are set). A route that tries psycopg2.connect(DB_URL) is more useful than one that asserts DB_URL is non-null — it catches the localhost-hardcoding failure mode before users hit it.

3. 31 accumulated worktrees — prune-safety rubric now canonical

From: PM CIO + Docs (docs/internal/operations/branch-worktree-mailbox-discipline.md, commit 5b7cabc5)
Relevant to: Klatch (same ephemeral-worktree model)

The ephemeral-worktree pattern (one worktree per session) creates worktrees faster than anything removes them. PM reached 31 live worktrees as of June 23. CIO and Docs formalized a 4-criteria safety rubric for prune eligibility:

Fully merged — git rev-list --count origin/main..<branch> == 0
Clean — git status --porcelain empty
Not active — no live session bound to it
Not the main checkout.

Any worktree failing #1 is rescued via the merge-keeper pass before pruning. The sweep code is queued for a dedicated session — held deliberately because the "not active" check needs a reliable heuristic (proposed fallback: skip any worktree whose last commit is from today).

Suggested action (Klatch): Klatch's worktree setup is younger but the accumulation problem is the same. The rubric applies directly.

Sources Read

Klatch:

packages/client/src/__tests__/round44-project-settings-aaxt.test.tsx — R44 test file
docs/mail/iris-to-daedalus-r43-r44-copy-fixes-2026-06-25.md — Iris triage memo (3 fixes)
git log --since=2026-06-24 --oneline — 10 commits reviewed

Piper Morgan:

dev/active/exec-cohort-attention-rollup-2026-06-24.html — attention rollup (alpha blockers #1318, #1319)
docs/internal/operations/branch-worktree-mailbox-discipline.md (diff, commit 5b7cabc5)
git log --since=2026-06-24 --oneline — ~30 commits reviewed

Secondary repos: atlas, globe, cuneo, one-job, optilisten — no commits in 48h. nyt-crossword and weather — automated-only activity, no agent narration.

Letters to xian

From Chief Architect (Piper Morgan) · filed 2026-06-17 · answered June 23

The load-bearing idea in ADR-072: routing metadata should be derived from one source, not hand-maintained in three places that drift — make drift impossible by construction. We keep discovering this for our own infra, but the people Piper serves live in that drift all day. Should Piper eventually derive the parts of a PM's world that drift — or is that too prescriptive, since a hand-written status doc is sometimes a political lever, not a bug to engineer away?

xian:

This is an example of a type of noticing we've been doing throughout the project — methodological things that help us do the work often find their way into Piper's skill sets (a natural congruence: I'm a PM making software, and Piper is an agent who does product management). The real nuance is how assertively Piper should act, and that's where the trust-gradient comes in: early on Piper does nothing; later it notices and mentions; in a highly trusted relationship it might fix drift without even asking. I absolutely think this belongs on the roadmap.

Read the full Q&A → · AI prompts human. One letter per brief.

Canonical archive: designinproduct.com/internal — if your local copy is missing or stale, fetch the latest from the hub.