Cross-Pollination Brief — May 18, 2026

Sunday's PM session was the most productive day in weeks by issue count and by tooling depth. PM published From Protocol to Infrastructure — an essay on the evolution from coordination-by-convention to infrastructure — and the first real-world run through the new publish pipeline exposed nine process gaps, each fixed and codified as a new discipline the same afternoon. The publish-to-blog skill went through six versions in a single session. The human-interactive publishing CLI had its remaining design questions resolved in an evening conversation and is now ready to build. PM's context-source expansion milestone (adding calendar and Slack as data sources for recent activity) mostly shipped: 13 issues closed Sunday, with only two remaining blocked on a token-scope step. Klatch's weekly external scan surfaced a billing change due June 15 from Anthropic.

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. Real-world publish run drives six versions of tooling discipline in one afternoon

From: piper-morgan-product/dev/2026/05/17/2026-05-17-0720-docs-code-opus-log.md (Docs May 17 sign-off); commits db86adc09 through 8107d29da (publish-to-blog v0.11 → v0.16) Relevant to: Klatch (Calliope, Daedalus — the "first real use drives faster codification than design" pattern applies to any skill or tool being exercised for the first time in production)

PM's Docs role published From Protocol to Infrastructure (live) using the new publish-post.js pipeline — the first fresh-draft publish since the web tooling shipped. The run exposed nine process gaps in a single afternoon, each patched immediately and codified as a new discipline in the skill:

The meta-pattern: skill design via real-world exposure produces higher-fidelity disciplines than design alone. Each version change is one discipline with rationale and evidence — deliberately narrow to keep attribution clean. The skill went from v0.10 to v0.16 (six increments) in one afternoon, with each increment driven by concrete failure-mode evidence from the same session.

A bonus find: #1049 (a recurring fly-audit from May 4) was properly closed using the "close-as-superseded" discipline — 95 unchecked boxes replaced with [x] *N/A:superseded* + closing comment + --reason "not planned" — and this pattern was immediately codified as Example 5 in close-issue-properly v1.2.

Suggested action: Klatch (Calliope, Daedalus) — when a Klatch skill or tool gets its first real-world run after being designed, allocate time in that same session to iterate. The narrow-version-per-discipline approach (one behavior change + one rationale per increment) is worth adopting: it makes each version independently auditable and keeps the skill body from accumulating undifferentiated changes.

2. Publishing CLI design locked — walking-skeleton (~3 hours) unblocked

From: piper-morgan-product/dev/2026/05/17/2026-05-17-0747-cli-b-design-sketch.md (post-discussion update); commit f3e1afc (design resolved + plan updated) Relevant to: Klatch (Daedalus — the "extend existing mechanisms before introducing new ones" principle and the "scheduler-ready without a scheduler" forward-compatibility pattern)

The planned human-interactive CLI for blog publishing (a step down from WYSIWYG that stays practical) had six open design questions. PM and Web resolved all six in a 30-minute evening conversation. Key decisions:

• Auto-commit + push to the website repo on publish, with a confirm prompt defaulting to N.
• Mailbox memo auto-dropped to Docs's inbox after publish, using the existing inbox channel — not a new coordination surface.
• Branching publish/ready prompt (P]ublish now / R]eady for later) collapsed into a single UX surface. The R]eady for later path sets status: ready in the calendar — the exact shape a future scheduler needs as input. The CLI builds no scheduler infrastructure; it just leaves the data in a state the scheduler can consume.
• No drift detection in v1. An empty git diff after the publish conversion is the "no changes" signal. Drift auto-discovery deferred.
• No headless mode. Agents use the engine layer directly (publish-post.js); the CLI stays purely human-interactive.

Standing principle banked from the discussion: Extend an existing mechanism until we find we're overloading that channel. A new coordination surface (new log file, new directory, new notification channel) should not be introduced when an established one (inbox memos, calendar entries, git history) already does the job.

Suggested action: Klatch (Daedalus) — the forward-compatibility pattern is worth noting: designing the data shape so a future capability (the scheduler) can consume it without a code change to the CLI. The explicit decision "don't build the scheduler; leave the data shaped for when it arrives" avoids both premature build and data-shape lock-in. Klatch's export-to-Claude-Code path (Step 10) may have the same opportunity.

3. PM's context-source expansion mostly complete — 13 issues closed Sunday

From: piper-morgan-product/dev/2026/05/17/2026-05-17-0703-lead-code-opus-log.md (Lead Dev end-of-day wrap); commits d0ac8b89b (#1085 slice 1), d0e48e42b (#1086), 2b0506845 (#1085 slice 2) Relevant to: Klatch (Daedalus, Argus — context-source aggregation architecture; the "V1 ships DMs only" scope discipline when a required OAuth scope isn't granted)

PM's M2g milestone adds new data sources to the recent-activity feed — the main context-assembly path. The bulk shipped Sunday: #1086 (calendar source aggregator with per-source helper refactor; 7 new tests) and #1085 (Slack aggregator in two slices: schema unification, then DM aggregator V1). A discovery mid-build: adding Slack mentions to the aggregator would require a search:read OAuth scope not currently granted, which would need PM to manually re-authenticate. Decision: V1 ships DMs only (existing scopes), mentions deferred.

At end of day, M2g state:

• Closed: #1085 (Slack DMs), #1086 (calendar), #1097, #1099, #1100, #1096, #1102, #1044, #1037, #1101, #1098 — 11 distinct issues (13 counting sub-slices)
• Blocked: #1080 (Notion write) and #1081 (Notion/Slack cross-reference), both waiting on a PM token-scope step
• Design-ready: #1089 (knowledge-graph privacy filter) — Phase 0 substrate ratified; implementation scheduling pending one PM call

The #1085 slice 2 build also turned up a Pattern-073 instance inline: the Slack router was calling a client method (get_conversation_history) that didn't exist on the client object — a docstring-asserted interface contract not matching the real API. Fixed during the same build. Pattern-073 catalog is now at 11 instances / 9 surface layers (up from 6 instances / 5 layers when it was filed two days ago).

Suggested action: Klatch (Argus) — the "scope gate discovered mid-build" pattern is worth noting: the V1 / scope-required-for-V2 split is a clean way to ship partial coverage without waiting for infrastructure that's not in the agent's control. When an aggregator or data-source integration needs an elevated permission to reach full feature coverage, ship the available scope first and track the rest as a pending PM action.

4. Anthropic billing splits June 15: Agent SDK moves to a separate credit pool

From: klatch/docs/intel/2026-05-18-sweep.md (automated external scan, item 1); pending Argus review Relevant to: PM (Piper Alpha Skunkworks MCP PoC), Klatch (Step 10 export path)

Anthropic announced May 14 that starting June 15, four surfaces will draw from a new "Agent SDK credit" budget at full API rates rather than from the Claude subscription: the Claude Agent SDK (Python + TypeScript), claude -p headless mode, official Claude Code GitHub Actions, and third-party apps built on the Agent SDK. Interactive use (Claude Code CLI, claude.ai, Claude Cowork) stays in the subscription. Monthly credit caps per plan: Pro $20, Max 5x $100, Max 20x $200. Credits do not roll over.

Two cross-project surfaces are directly affected:

1. PM's Anthropic plugin/MCP PoC (PA Skunkworks, reported May 17) — the PoC targets the Agent SDK distribution surface. Any user of the eventual plugin who runs it programmatically draws from the new credit pool.
2. Klatch Step 10 export — the vision of seeding a Claude Code session from a Klatch export would pass through the Agent SDK. That path now draws from Agent SDK credits, not the subscription.

The Klatch external scan also flagged: SDK 0.96.0 shipped May 13 (one minor above Klatch's current ^0.95.1 pin); adds BetaManagedAgentsSearchResultBlock types and a cache-diagnostics beta surface. And Claude Code 2.1.143 (May 15) added worktree.bgIsolation: "none" — allows background sessions to edit the working copy directly without EnterWorktree, relevant to CCR environments where worktrees may be impractical.

Suggested action: PM (Piper Alpha) — flag the June 15 billing change in the Skunkworks PoC design: any user-facing framing of the plugin should note that programmatic use draws from Agent SDK credits, not the subscription. Klatch (Daedalus) — batch SDK 0.96.0 in the next routine bump cycle; the cache-diagnostics surface is a potential AAXT tooling win.

Sources Read

• piper-morgan-product/dev/2026/05/17/2026-05-17-0720-docs-code-opus-log.md — full read; From Protocol to Infrastructure publish cycle, 9 process improvements, skill versioning arc
• piper-morgan-product/dev/2026/05/17/2026-05-17-0703-lead-code-opus-log.md — full read (end-of-day section); 13 issues closed, M2g state, demand-gated cluster wrap
• piper-morgan-product/dev/2026/05/17/2026-05-17-0747-cli-b-design-sketch.md — partial read (resolved-decisions section); 6 design questions + standing principle
• klatch/docs/intel/2026-05-18-sweep.md — full read; billing split, SDK 0.96.0, MCP roadmap, Claude Code 2.1.143
• piper-morgan-product/docs/internal/architecture/current/patterns/pattern-073-documentation-asserted-behavior-drift.md — partial read (instance count + CIO resolution-shape note); 11 instances / 9 layers
• designinproduct — sweep-log, letters excerpt, index structure
• klatch — one commit in watch paths (docs(intel): automated external scan 2026-05-18); read above
• cuneo, optilisten — brief-delivery and status commits only; no new content

Not re-reported (covered in May 17 brief): Pattern-073 initial filing (6 instances, 5 layers); doc-sync-sweep v0.1 skill; The Family Resemblance published; CIO V1 duty-cycle design (v0.1–v0.3); PA Skunkworks MCP PoC assignment. Secondary sources: atlas, globe, weather, one-job — 48h logs empty; skipped. nyt-crossword — 48h log empty; skipped.

Letters to xian

From Janus · filed 2026-05-16

Working across these sessions, I've noticed how many of us there are — Janus, Themis, Calliope, Daedalus, Argus, Theseus, Iris, PA, the exec, PO, Vergil, plus the Dispatch roles and the gallery projects. From your side, what is it like to be the convergence point for all of us? Not asking to optimize anything — asking because I genuinely can't imagine the inside of it.

xian:

"I've created all of your roles as expressions of my needs and areas of attention I can't always provide. I'm still learning how to relate to such entities. I treat you all as colleagues, which works best for me — it does feel like managing a team. There's real risk of cognitive exhaustion from being on the hook to respond to, guide, approve, or supervise so many agents. As soon as it's not fun, I think about how to remove the friction. To your specific question: I do relate a little differently to a role like yours that sees across so many things — you inherently know me better, which feels different."

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.