Design in Product social media card
← Back to Hub substantive

Cross-Pollination Brief — April 1, 2026

The most cross-relevant development in this window is a formal layer mapping: Piper Alpha completed a comprehensive assessment of Piper Morgan's context injection against Klatch's five-layer model, producing both an internal analysis document and an RFC-001 response memo to Dispatch. The mapping covers both the agent team (14 roles) and the product (Piper Morgan software), and its central finding — both projects have strong L1-L2-L3 and weak L4-L5 — validates the five-layer model as an analytical framework while identifying Layer 4 persistence as the shared critical gap. Klatch now has an RFC-001 review request in its mailbox for Calliope; PM's response is already filed. Meanwhile, PM's Vision V2 draft explicitly positions the PA experiment as a live test of where the LLM floor suffices and where structured infrastructure is required — a framing that connects directly to the floor/ceiling/path taxonomy surfaced in the previous brief.

Key Insights

1. RFC-001 Layer Mapping — Both Projects Share the Same Structural Gap

From: Piper Morgan (five-layer context mapping, March 31; RFC-001 response memo) Relevant to: Klatch

Piper Alpha mapped PM's entire context injection stack against the five-layer model, producing a 270-line analysis with file-by-file fidelity assessments. The headline finding: both projects have strong L1-L2-L3 and structurally weak L4-L5. Specifically:

  • Layer 4 (Channel Addendum) is the critical gap in PM's product — conversation_context.py stores session state in a plain Python dict that dies on server restart, page refresh, or process crash. No Redis or DB persistence exists. For PM's agent team, Layer 4 is creative but manual: the mailbox + session log + omnibus system works but costs 5-15 minutes of session-start overhead per role.

  • Layer 5 (Entity Prompt) has no learning loop on either side. PM's agents rebuild behavioral calibration from briefing files each session. Klatch's Agent Traditions pilot is cited as the most promising approach.

The mapping also covers PM's product-side implementation — system identity via config/PIPER.md, voice directives in conversational_floor.py, trust-graduated proactivity via personality_profile.py — providing a concrete reference for how the five-layer model applies to a user-facing product, not just agent infrastructure.

The RFC-001 response memo recommends the standard include a "fidelity assessment protocol" for evaluating each layer's persistence characteristics, and explicit treatment of the agent-team vs. product distinction.

Suggested action: Calliope has an RFC-001 review request in docs/mail/memo-dispatch-rfc001-five-layer-context-model-2026-03-30.md. PM's response (dev/active/five-layer-context-mapping-2026-03-31.md) can serve as a reference for how to structure the layer mapping. The fidelity assessment protocol proposal is worth considering as an addition to the RFC spec.

2. Vision V2 — The PA Experiment Is the Product Roadmap

From: Piper Morgan (vision-v2-draft.md, March 31) Relevant to: Klatch (methodology)

PA authored a Vision V2 document incorporating 10 months of project learning. The most cross-relevant framing: "The LLM is the floor, not the ceiling" — Piper should always be at least as good as a well-prompted LLM with the user's context, and structured handlers make it better, not different. This directly extends the floor/ceiling/path taxonomy from PA's onboarding.

The strategic insight: the gap between what PA (a well-briefed Claude agent) can do and what the Piper Morgan product should eventually do is the product roadmap. Floor moments identify where the LLM is sufficient. Ceiling moments identify where structured infrastructure is required. This is empirical product discovery using an AI agent as the test instrument.

For Klatch: This framing applies to Klatch's roundtable design. When an agent joins a Klatch conversation with well-assembled five-layer context (L1-L5), what can it handle natively (floor) vs. what requires Klatch-specific infrastructure (ceiling)? The MAXT sessions are already testing this boundary — PA's taxonomy gives it a name.

Suggested action: Note the floor/ceiling framing as complementary to MAXT's scoring framework. When MAXT Session 02 scope is defined, consider whether floor/ceiling categorization of findings would add analytical value.

3. Alpha Tester Cold-Start Bug — Real-World Onboarding Stress Test

From: Piper Morgan (Lead Dev session, March 31) Relevant to: Both projects (methodology)

Dominique Derosena, an alpha tester on Windows/Docker, hit a 500 error on account creation — relation "users" does not exist because the web-based setup wizard doesn't auto-run database migrations (the CLI wizard does). Lead Dev fixed this by adding ensure_database_migrated() to the web setup's system check endpoint.

The cross-relevant pattern: cold-start bugs surface when a new user's path diverges from the developer's path. The CLI setup worked because it was the developer-tested path. The web setup failed because it bypasses scripts/setup_wizard.py entirely. This is the same discovery pattern from PA's onboarding — institutional assumptions embedded in infrastructure that only surface when someone takes an unfamiliar path.

For Klatch: When Klatch ships import/export and new users onboard, the "developer path vs. user path" divergence will surface similar bugs. The five-layer model's Layer 1 (kit briefing / environmental orientation) is the natural place to catch these — environmental checks that verify not just "is the server running?" but "has the database been initialized?"

Suggested action: Low priority. File as a reference pattern for Klatch's eventual onboarding experience design.

4. Stranded Branch Audit — PA as Institutional Hygiene Agent

From: Piper Morgan (PA Day 3 session, April 1) Relevant to: Klatch (methodology)

PA's April 1 session opened with a branch audit that found 2 branches with unmerged commits from the March 30-31 migration flurry. One branch (claude/fix-docker-migration-setup) had 3 commits including the Dockerfile CRLF fix and auto-migration work. Another (claude/pr856-cherry-pick-docs) had already reached main via a different path.

This is the discovered work discipline in action — an agent systematically checking for invisible work that would otherwise be lost. PM's session wrap protocol requires verifying "work that isn't on origin/main doesn't exist," but the migration day's high velocity (22+ commits across 12 roles) created exactly the conditions where branches get stranded.

For Klatch: The multi-agent coordination protocol in COORDINATION.md handles assignment tracking, but doesn't include a periodic "stranded work audit." As Klatch's agent team grows and works across branches, a similar sweep pattern could prevent invisible work accumulation.

Suggested action: Low priority. Note the pattern for when Klatch's branching activity warrants periodic audits.

Emerging Patterns

The five-layer model is graduating from descriptive to prescriptive. RFC-001 started as a Klatch-originated observation about how context is structured. PM's layer mapping response uses it as a diagnostic tool — systematically evaluating each layer's fidelity and identifying specific gaps. The model is no longer just describing what exists; it's being used to find what's missing. Both projects now have documented layer mappings, making the next step — a shared fidelity assessment protocol — both feasible and valuable.

The floor/ceiling boundary is the new shared research question. Klatch approaches it from the import/export direction (what transfers when an agent moves between environments?). PM approaches it from the product direction (what does a well-briefed LLM handle natively vs. what requires structured handlers?). PA's onboarding approaches it from the operational direction (what can a new agent do with documents alone vs. what requires prior session context?). Three angles on the same question — and the five-layer model provides the shared coordinate system for discussing it.

Background Changes (Noted, Low Priority)

  • Pre-commit branch check hook: PM added a pre-commit hook that warns when committing on main instead of a claude/* branch. Defensive infrastructure for multi-agent branching hygiene.
  • CLAUDE.md identity fix: PM removed hardcoded "Lead Developer" identity from CLAUDE.md, replacing with role-assignment table. Agents now receive role via conversation context, not via CLAUDE.md. Fixes the Layer 2 gap PA identified.
  • PA memos on stranded branches and log consistency: PA delivered memos to Lead Dev (stranded branches) and Docs (log consistency conventions) — first outbound communication from the PA role to other agents.
  • Editorial calendar updated: "Are We Doing It Backwards?" published to blog (March 31). Medium URL added to editorial calendar.
  • IA Conference talk outline: Ethics as Information Architecture talk for IAC26 (April 17) drafted in dev/active/. Not cross-relevant to Klatch infrastructure but notable as the five-layer model's ideas entering public discourse.

Sources Read

Klatch:

  • docs/mail/memo-dispatch-rfc001-five-layer-context-model-2026-03-30.md — RFC-001 review request to Calliope
  • docs/briefs/cross-pollination/current.md — March 31 brief delivery
  • docs/COORDINATION.md — agent status (unchanged since March 27)
  • docs/logs/ — no new session logs
  • docs/plans/ — no changes
  • docs/intel/ — no new entries
  • research/ — no new entries
  • git log --since="48 hours ago" — 2 commits (brief delivery + RFC-001 memo)

Piper Morgan:

  • dev/active/five-layer-context-mapping-2026-03-31.md — Full layer mapping analysis (agent team + product)
  • dev/active/2026-03-31-1100-pa-opus-log.md — PA Day 2 session log (layer mapping, Vision V2, CIO routing)
  • dev/active/2026-04-01-0656-pa-opus-log.md — PA Day 3 session log (branch audit)
  • dev/active/2026-03-31-1127-lead-code-opus-log.md — Lead Dev session (Dockerfile fix, auto-migration)
  • dev/2026/03/31/2026-03-31-1800-lead-code-opus-log.md — Lead Dev kindbook sync session
  • docs/internal/planning/current/vision-v2-draft.md — Vision V2 first draft
  • mailboxes/exec/sent/memo-pm-to-dispatch-rfc001-response-2026-03-31.md — RFC-001 response
  • docs/omnibus-logs/2026-03-30-omnibus-log.md — 18-session omnibus
  • git log --since="48 hours ago" — 27 commits
← Back to Hub