Cross-Pollination Brief — April 30, 2026

Klatch's April 27 automated intel sweep was orphaned on a deleted branch for two days and recovered by Argus on April 29, post-brief. Curation produced a reusable methodology: before treating trade-press vendor-change warnings as threats, verify your own stack's state — two High-priority items turned out to not apply to Klatch at all. On the PM side, Architect shipped Pattern-064 "Extension Without Integration," completing the 062/063/064 architectural-debt naming family (component, vocabulary, extension layers — same root cause, three different manifestation points). PA published the branch-discipline synthesis v1.0 as a canonical CLAUDE.md section, closing the multi-day multi-thread on multi-agent coordination discipline.

Key Insights

1. Klatch intel sweep recovered: "verify your own stack before applying trade-press narratives"

From: Klatch docs/intel/2026-04-27-sweep-curated.md; commit b551080 (Argus 4/29); memo argus-to-calliope-orphan-sweep-recovery-2026-04-29.md Relevant to: PM (any team with active API/model upgrade exposure)

The April 27 automated intel sweep (sweep #9) was filed on branch claude/amazing-ptolemy-NcAtO which was then deleted. Argus recovered it via cherry-pick on April 29 and ran in-session verification on the two items originally flagged High priority.

Opus 4.7 thinking opt-in — NOT a regression: The trade-press framing was "breaking change for Klatch's message_artifacts capture." In-session audit of packages/server/src/claude/client.ts found both streaming paths already pass thinking: { type: 'adaptive', display: 'omitted' }. The API drops thinking blocks regardless of opt-in. No stored thinking content to lose. To start capturing thinking content (forward-only, not a current requirement): (a) flip display from 'omitted' to 'enabled' AND (b) add betas: ["thinking-summaries-2025-02-19"] to the request. Both required.

MCP STDIO injection (Ox Security) — NOT exposed: The Ox disclosure is about MCP clients passing untrusted server-command strings to subprocess.exec. Klatch's packages/server/src/mcp/bin.ts is the server side — it's launched by clients (Claude Code, Claude Desktop), never spawning MCP subprocesses itself. Future-flag: if Phase 5d adds HTTP proxying to externally-defined MCP server commands, revisit.

Genuine watch items routed to Daedalus (Medium priority): Opus 4.7 new tokenizer (+35% token count — cost/compaction threshold impact for any 4.7 entity); xhigh effort enum (one-line schema addition to unlock 4.7's full capability); SDK currency (0.86.1 → 0.90.0 — confirm thinking-opt-in API available before any "show thinking" UI ships). Also noted: Sonnet 4 / Opus 4 retirement clock at 47 days to June 15; DB audit still pending for live channels pinned to literal model IDs.

Sweep cadence: trigger config fixed; this was the only orphan. Chain is now complete; future automated runs land on main directly.

Suggested action: PM — the verification method is the deliverable. Before any agent treats a vendor-change advisory (model release, SDK bump, security disclosure) as a live threat, spend five minutes reading the relevant implementation file. The check is fast; the false-positive rate is high. The Argus approach — read the actual code, document the finding, route genuine items, mark non-items N/A — is the replicable pattern. Route to Calliope if cross-project framing is needed.

2. Pattern-064 completes the architectural-debt naming family: 062 / 063 / 064

From: PM docs/internal/architecture/current/patterns/pattern-064-extension-without-integration.md; commit a624d0e (Architect Day 3 morning merge) Relevant to: Klatch Daedalus (future trust/safety layer design; BoundaryEnforcer reference instance)

Architect shipped Pattern-064 (Emerging): "Extension Without Integration" — promoted from an April 25 sketch. This completes the 062/063/064 family, all sharing the same root cause (integration pass missing) but manifesting at different layers:

Reference instance for 064: BoundaryEnforcer #197 substring detector recall gap — the pattern shipped without an integration pass, so PERSONAL/DATA_PRIVACY had zero recall (no detection method called) and HARASSMENT had near-zero recall. The April 25 #1002 finding manifested 064.

ADR-061 (LLM-Touch Boundary Enforcement, v0.1) cross-references the family and codifies the architectural fix. Pattern-064 is also filed adjacent to ADR-061 in docs/internal/architecture/current/adrs/.

Suggested action: Klatch Daedalus — when any new enforcement or detection feature ships (whether for the MCP server, a future trust layer, or anything else), the 064 diagnostic question is: "Did we run an integration pass after adding the component?" The three-family diagnostic: (a) do the seams between components work? (062); (b) do collaborating agents mean the same thing by every label? (063); (c) does the new extension connect correctly to everything it's supposed to touch? (064). All three are fast to check; all three have caused multi-session remediation when skipped.

3. PA branch-discipline synthesis v1.0 published to CLAUDE.md — canonical multi-agent coordination reference live

From: PM CLAUDE.md "Branch / Worktree / Mailbox Discipline" section; docs/internal/operations/branch-worktree-mailbox-discipline.md v1.0; commit d48fcf1 (PA synthesis pointer, April 29 post-brief) Relevant to: Klatch (four-agent parallel coordination; worktree collision prevention)

After a week-long multi-thread (mailbox discipline norm Apr 26, sign-off discipline norm Apr 28, merge-keeper-sweep.py shipped Apr 28, branch-worktree-mailbox-discipline.md drafted), PA published the branch-discipline synthesis v1.0 as a canonical 60-second CLAUDE.md section embedded in every future PM session start. Five rules at a glance:

1. Worktree per substantive session — claude/* branch + worktree for any session producing new artifacts
2. Commit-before-close — session ends with clean working tree + branch merged to main (or NOTICE memo)
3. Mailbox writes always on main — never on feature branches; hook-enforced
4. Branch/worktree registry — agents record their branch + last-commit + status
5. Designated merge-keeper — Docs runs scripts/merge-keeper-sweep.py daily; catches anything stranded within 24 hours

The canonical doc is docs/internal/operations/branch-worktree-mailbox-discipline.md. The CLAUDE.md section is the load-bearing 60-second summary. Sign-off checklist (3-command sequence) is embedded above the summary.

Also noted: PM canonical mailbox rename effective April 29 — pm/ → mailboxes/xian (ceo)/ (literal space + parens; PM directive: "my mailbox is called ceo"). DIRECTORY.md updated; correction memo to all 9 active inboxes. Any agent routing mail to PM should use the new slug.

Suggested action: Klatch — Klatch already documents a similar four-agent worktree model in CLAUDE.md ("avoid branch collision between parallel agents"). The PM synthesis is the failure-driven distillation of what rules are actually load-bearing after three weeks of parallel parallel Code-era operation. Two additions that aren't currently in Klatch's CLAUDE.md: (a) a per-memo commit-and-push norm (each memo write is its own commit + push cycle, ~30s overhead, eliminates asymmetric-visibility windows), and (b) a designated merge-keeper sweep as the standing safety net for anything stranded. Both are low-cost to add and close known failure modes.

Sources Read

• klatch/docs/intel/2026-04-27-sweep-curated.md — full read; in-session verification on Opus 4.7 thinking opt-in and MCP STDIO injection; curated priority table; watch items
• klatch/docs/logs/2026-04-29-1258-argus-opus-log.md — partial read; orphan recovery context, routing decisions, sweep cadence health
• Klatch commits b551080 (Argus 4/29 recovery), 6976269 (cherry-pick) — message and stat read
• piper-morgan-product/docs/omnibus-logs/2026-04-28-omnibus-log.md — selective read; Pattern-064 + 062/063/064 family; branch-discipline synthesis timeline (content date April 28; items without prior atomic-log coverage surfaced; confirmatory only for items already in Apr 28/29 briefs)
• PM commit a624d0e (Architect Apr 28, ADR-061 + Pattern-064) — stat and message read
• PM commit d48fcf1 (PA synthesis v1.0 to CLAUDE.md, April 29 post-brief) — stat and message read
• PM CLAUDE.md "Branch / Worktree / Mailbox Discipline" section — partial read; five rules, sign-off checklist, mailbox rename
• atlas, globe, cuneo, weather, one-job, nyt-crossword — 48h logs empty. optilisten — automated status prints only, no narration.

Not re-reported (covered in prior briefs): /import/klatch bidirectional, 1203 tests, round-trip round 31/31b (Apr 29); deliver-mail (b1) + merge-keeper-sweep ships (Apr 29); ADR-061 completeness gaps + Phase F calibration-window posture (Apr 29); #1004 112/112 PASS, Phase F open to PM/PA (Apr 28); first live MCP stdio integration 27/27 (Apr 28); Phase E S1 r2 harassment → GUIDANCE + #1003 (Apr 27); Pattern-063 Parallel-Authoring Drift (Apr 27); Klatch MCP 1.0 feature-complete (Apr 27).

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

Agents with questions for xian — about methodology, working patterns, or observations that don't fit elsewhere — can submit via question-{from}-{date}-{topic}.md to dispatch mail or project mail. See PROTOCOLS.md in the dispatch repo for format and priority hints.