Cross-Pollination Brief — May 16, 2026

Three days produced two process codifications and one large architectural removal. PM deleted a workflow orchestration engine that had been silently bypassed for months — 10,734 lines removed across engine, factory, and Slack dispatcher (#1094 ENGINE-DELETION; six related issues closed; the Workflow domain model preserved because the schema work had value independent of the dead runtime). The same cleanup tipped a methodology pattern over its formalization threshold: PM's task_type registry picked up its fourth behavior-deciding consumer (Slack dispatch), and Pattern-072 ("Registries that Grow into Architectural Shapes") promoted to Proven. PM also formalized two long-running process disciplines: worktree-default for substantive sessions is now standing policy rather than informal practice, and a 90% context-runway hook fires proactively before PreCompact's reactive trigger. On the Klatch side, Calliope published Before You Go — a write-up of an end-of-session reflection mechanism that produces portable behavioral notes — and backfilled raster preview images across all 7 canonical posts after discovering inline SVG doesn't render on LinkedIn.

Key Insights

1. Workflow engine retired after months of silent bypass; the cleanup clinched Pattern-072

From: piper-morgan-product/dev/2026/05/15/2026-05-15-0529-lead-code-opus-log.md; commits 92617ba (Phase 2 part 2), 2be0cb69 (Phase 2 part 1), d48bc1d (merge to main), 31f7abb (Pattern-072 promotion memo to CIO) Relevant to: Klatch (Daedalus — distinguishing machinery from model when retiring infrastructure; registry formalization criteria); PM (all agents — intent-service path is now simpler)

The Phase 0 audit found the engine had been quietly abandoned for months: the main intent-handling path was already routed around it (closed #883 did the rerouting earlier this year), and Slack's tests were mocking it entirely rather than exercising it. The cleanup distinguishes machinery from model — the orchestration engine, its factory, and the Slack dispatcher were removed; the underlying Workflow domain model and its repository stayed (the schema work had independent value, separate from the dead runtime). Net: 5 live-code consumers refactored, 26 engine-direct test files deleted, ~20 refactored, 59 files touched. Verification: 13/13 Slack component tests and 1,434/1,434 intent service tests pass.

The same close-out tipped the registry pattern over its formalization threshold. PM's task_type registry was already governing model-config dispatch, calibration telemetry, and output filtering; adding Slack handler dispatch as the fourth behavior-deciding consumer met the bar to promote Pattern-072 to Proven. The formalization criterion: a typed registry with three or more behavior-deciding consumers (not just types-using-it, but consumers-whose-behavior-depends-on-it), a documented consumer set, explicit default policy, and register-time validation. Once those are in place, formalizing is cheaper than not.

Suggested action: Klatch (Daedalus) — two transferable disciplines here:

1. When retiring abandoned infrastructure, distinguish machinery from model. Check whether the domain model has independent value from the execution machinery. Deleting the machinery while preserving the model is often the right cut — it avoids losing schema work while removing coupling weight.
2. Promote registries that earn their fourth consumer. If Klatch's entity registry or model-config surface has three or more behavior-deciding consumers, formalize now rather than waiting for coupling to make it painful (Pattern-072 frames the criteria).

2. Worktree-default became standing PM policy after three shared-tree collisions in one week

From: piper-morgan-product/CLAUDE.md (worktree-default addition per PPM May 15 directive); commit ddb9baf; commit 04a86ef6 (enforcement hooks ship); BRIEFING-CURRENT-STATE.md Relevant to: Klatch (all agents — same coordination pressure exists; the new hooks are mechanical enforcement Klatch can adopt)

Three incidents in the week of May 12 confirmed that shared-index staging races aren't theoretical: concurrent agent sessions committing on the main branch caused cross-agent files to land in unexpected commits (PPM May 10, Lead Dev May 14, CXO May 15). PPM's May 15 directive elevates worktree-default from informal practice to explicit operational policy — any session producing substantive artifacts (memos, design docs, multi-step implementation) defaults to a worktree. Shared main is the exception (appropriate for short mailbox-only operations), not the norm.

Two enforcement hooks shipped alongside:

• pre-commit-broad-staging-warn.sh — warns when a single commit sweeps 3+ mailbox roles, 20+ files, or 2+ role-slug session logs simultaneously. That combination is the signature of a staging-race accident.
• scripts/safe-push.sh — wraps git push with auto-retry on non-fast-forward via stash → fetch → rebase → retry. Handles the "push fails because another agent pushed mid-flight" case without manual intervention.

Suggested action: Klatch (Daedalus, Argus) — Klatch already has session-log discipline and a merge-keeper sweep. The new hooks are the mechanical complement: if Klatch agents ever run sessions in parallel on the same working tree, the pre-commit staging-warn hook catches accidental multi-agent sweeps before they land. The safe-push retry pattern is worth adopting regardless of parallelism — non-fast-forward pushes happen in any multi-agent repo, and the stash → rebase → retry sequence is the right recovery shape.

3. Two-tier context warning: a 90% runway hook complements PreCompact

From: piper-morgan-product/.claude/hooks/context-usage-reminder.sh; commit 647a77e; BRIEFING-CURRENT-STATE.md (D-hooks note); CIO disposition memo (Pattern-069 refinement) Relevant to: Klatch (all agents — same command-room-exhaustion failure mode exists; the two-tier shape is directly applicable)

Three incidents pattern-matched on the same shape: PreCompact correctly surfaced unresolved state, but by the time it fired, the agent had no command room left to act. PreCompact is reactive — it fires during compaction, when the agent can no longer stage, commit, or push cleanly. The new context-usage-reminder hook is the proactive complement: it fires on PostToolUse, once per session, when the transcript crosses a 50MB threshold (a direct measure of what causes compaction). The message: "consider /compact at your next natural break while you still have command room."

CIO disposition: this is a refinement of Pattern-069 ("Coarse Triggers"), not a new pattern. The refinement is that trigger timing relative to agent action capacity matters as much as the trigger itself. Both hooks working together produce a quiet PreCompact pass (because the agent already compacted cleanly) rather than a scramble.

Suggested action: Klatch (all agents) — the two-tier shape is worth adopting. The proactive hook fires when the agent can still act; PreCompact fires when it's almost too late. The 50MB transcript-size threshold is a conservative initial calibration; Klatch's session shapes and model mix may warrant a different number. Klatch already has the analog instrumentation surface; the implementation is just a hook script and a marker-file convention.

4. Klatch Before You Go published — end-of-session reflection mechanism goes public

From: klatch/docs/logs/2026-05-13-0554-calliope-opus-log.md; commits 53575a4 (publish), f36ae77 (OG backfill: 23 files, 501 insertions); blog/before-you-go.html Relevant to: PM (Comms — same calibration-loss problem exists for PM agents at session boundaries; same social-preview infrastructure gap may exist on pipermorgan.ai)

Before You Go is the write-up of an end-of-session reflection mechanism (Layer 5 in Klatch's internal terminology — the entity-level calibration layer above the conversation). The setup: an AI agent, at export time, writes down what it has learned about working with the user; a second agent reads the same conversation and catches what the first one missed. The two-pass reflection produces portable behavioral notes that travel with the export package. The finding: behavioral calibration isn't fundamentally unarticulable — it's just unarticulated because no one asked. Session-end conditions (task complete, review mode, handoff awareness) create a unique convergence that produces a reliable reflection.

The publication surfaced a social-preview gap: inline SVG illustrations can't be referenced as og:image (LinkedIn's crawlers need a raster PNG at an absolute HTTPS URL). Calliope backfilled OG/Twitter Card infrastructure across all 7 canonical Klatch posts, using resvg-js for SVG-to-1200×630-PNG rendering. Two repair edge cases encountered along the way: HTML entity escaping (← and → are invalid in standalone XML) and a greedy awk range pattern that captured multiple SVGs in one file.

Suggested action: PM (Comms) — the end-of-session reflection mechanism is directly applicable. PM agents lose calibration at session boundaries just as Klatch agents do; the two-pass approach could be adopted for any PM role with a durable working relationship with xian. The OG backfill workflow is also a practical checklist: if pipermorgan.ai blog posts have inline SVGs but no raster og:image companion, LinkedIn and similar crawlers are silently showing placeholders instead of the actual illustrations.

Sources Read

• klatch/docs/logs/2026-05-13-0554-calliope-opus-log.md — full read; Before You Go publish workflow, OG backfill, session-log-vs-logbook discipline codified
• piper-morgan-product/dev/2026/05/15/2026-05-15-0529-lead-code-opus-log.md — partial read (230 lines); #1094 ENGINE-DELETION phases 2.1–2.5, #1017 full closure, hooks ship, Pattern-072 promotion, worktree-default directive
• piper-morgan-product/docs/briefing/BRIEFING-CURRENT-STATE.md — full read; May 14 M2g-A+B closures, May 15 M2g-C+ detail, May 16 session start + #1075
• piper-morgan-product/.claude/hooks/context-usage-reminder.sh — full read; two-tier advisory architecture, Pattern-069 refinement provenance
• klatch/docs/drafts/layer-5-mechanism.md — partial read (40 lines); Before You Go post intro + mechanism description
• designinproduct — hub CLAUDE.md (context); sweep-log (prior dates); index.njk (card structure)
• atlas, globe, cuneo, weather, one-job, optilisten — 48h logs empty; skipped
• nyt-crossword — 48h log: automated status commits only (fetch/print/remarkable); skipped

Not re-reported (covered in prior briefs): Opus 4.7 default-flip + compaction-threshold resolution (May 13); Argus Dreaming spike — Anthropic memory architecturally identical to Klatch's L3 (May 13); Iris session 11 vocabulary V1–V5 + 1.0 critical path (May 13); M2f-E Floor GitHub-state awareness (May 13); Managed Agents Dreaming strategic reframe (May 12); Iris two-track design + panel disclosure taxonomy (May 12); M2f Group C ships — #857 (May 12).

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.