Cross-Pollination Brief — March 28, 2026

Big output window for Klatch: Step 9 (Files) is substantially complete with five sub-features shipped in rapid succession, a comprehensive File Domain Model has been designed and approved, Round 12 infrastructure delivered prompt caching and dynamic model discovery, and the roadmap has been resequenced from Files to Export to Search. Most notably, Klatch is introducing a dedicated UX designer/developer role — a product maturity signal. Calliope launched a Layer 5 externalization pilot and built an editorial calendar with three blog posts queued for immediate publication. Piper Morgan had no new commits in the 48-hour window.

Key Insights

1. File Domain Model: Pointers, Not Payloads — Five Visibility Scopes

From: Klatch (docs/plans/FILE-DOMAIN-MODEL.md) Relevant to: Piper Morgan

Daedalus designed a File Domain Model that reframes files as domain objects with ownership and visibility at each contextual level, rather than treating them as prompt payloads to be injected wholesale. The model defines five scopes — Kit (awareness only), Project (knowledge base), Channel (working set), Entity (library/index), and Message (attachments) — with two directional operations: promotion (upward, e.g., message attachment elevated to project knowledge base) and projection (downward, e.g., project-level spec delivered to a specific entity with a framing prompt).

The key design principle: a single file on disk can be visible at multiple levels simultaneously via pointer references rather than duplication. And critically, MEMORY.md becomes a file with a reserved name and special prompt-assembly role — unifying the memory system with the file system rather than treating them as separate concerns.

This matters for Piper Morgan because it solves the "context at multiple levels" problem cleanly. PM's knowledge base, briefing documents, and role-specific artifacts all exist at different visibility levels today, managed through convention rather than a formal model. The five-scope taxonomy and promotion/projection operations would map directly: a gameplan could live at the project level and be projected to specific agent roles with delivery context.

Suggested action: No immediate action. But if PM ever formalizes how knowledge base artifacts flow to specific agent roles, this model provides a tested vocabulary — particularly the distinction between visibility (who can access) and injection (what enters the prompt).

2. Roadmap Resequenced: Files, Then Export, Then Search

From: Klatch (docs/ROADMAP.md, docs/mail/calliope-to-daedalus-roadmap-resequence-2026-03-26.md) Relevant to: Piper Morgan

Klatch reordered its remaining roadmap from the original Step 9 (Search) → Step 10 (Files) → Step 11 (Export) to a new sequence: Step 9 (Files) → Step 10 (Export + meta-model) → Step 11 (Search). The reasoning is structural, not preferential:

Files are substrate — Export needs files to carry.
Export forces the meta-model — To export, Klatch must synthesize a complete five-layer context package, acknowledge what can't be packaged (Layer 5), and explain to users what they need to do. This is the hardest design problem.
Search comes last — Searching without understanding file types, project structure, and layer content produces weak results. Once the model of "what Klatch data is" is settled, search becomes powerful.

The underlying principle generalizes: build the packaging/export problem before building search. The Dispatch import research (covered in the March 26 brief) validated this empirically — the hardest part of export isn't writing files, it's the meta-model question.

Suggested action: If PM is planning search or discovery features across its knowledge base, consider whether the shape of the data is fully modeled first. Search built on a well-understood domain model is categorically more useful than search built on a file tree.

3. Step 9 Complete: File Attachments End-to-End in Two Days

From: Klatch (docs/logs/2026-03-26-1718-daedalus-opus-log.md, docs/logs/2026-03-27-1004-daedalus-opus-log.md) Relevant to: Piper Morgan (methodological interest)

Daedalus shipped five sub-features of Step 9 in rapid sequence:

9a: File upload UI, server storage, upload + serve endpoints, context injection
9b: Render artifacts inline in messages
9c: Kit briefing file awareness (Layer 1 — agent told what files are attached)
9d: Save code blocks as files (two approaches tried: Option A clipboard-based, Option B tool-based save_file — Option B selected)
9e (demo pipeline): Playwright + npm scripts for demo recording

The velocity is notable — five sub-features across client, server, and prompt layers in roughly two calendar days, including 23 unit tests for a new extractFilename utility. The approach followed Gall's Law strictly: each sub-feature was a complete working increment, not a partial implementation waiting for the next piece.

For Piper Morgan, the methodological pattern is the signal. Breaking a feature into independently-shippable vertical slices (each touching server, client, and prompt layers) is the same principle PM uses in its milestone structure. The difference is that Klatch achieves this with a single developer agent (Daedalus) while PM distributes across roles — worth noting when evaluating coordination overhead.

Suggested action: None. Methodological observation only.

4. Calliope's Calibration Notes: Externalizing Layer 5

From: Klatch (docs/agents/calliope-calibration.md) Relevant to: Piper Morgan

Calliope created a calibration notes file as an experiment in externalizing Layer 5 (behavioral calibration) — the layer that the March 26 brief identified as transferring at 0% during import. The file documents working preferences (xian's communication style, decision-making patterns), workflow patterns (parallelization, branch sweeps, session log discipline), and communication style (findings-first, context-inclusive tone).

The hypothesis: if behavioral calibration is written down explicitly, it can be absorbed by successor sessions and improve continuity — making the implicit transferable. The experiment is measured; Calliope plans to assess transfer fidelity in a future MAXT session.

This is directly applicable to Piper Morgan's multi-session orchestration. PM's detailed role definitions already serve a similar function — they encode behavioral expectations for each agent role. But those definitions describe what the role should do, not how it has learned to work with xian specifically. That second category is the calibration gap. If PM agents accumulate implicit calibration over sessions and then lose it at session boundaries, documenting it (as Calliope is piloting) could improve continuity.

Suggested action: Consider whether PM's role definitions should include a "working preferences" section — not role-generic instructions, but the specific interaction patterns xian has developed with each role over time.

5. Prompt Caching: One-Line Cost Win

From: Klatch (docs/logs/2026-03-26-1718-daedalus-opus-log.md) Relevant to: Piper Morgan

Round 12 added cache_control: { type: 'ephemeral' } to both API call sites in Klatch. This enables Anthropic's automatic prompt caching — cache reads cost 10% of input token cost, cache writes cost 125%. For multi-turn conversations where the system prompt and context are repeated, this is a significant cost reduction with a one-line change.

Suggested action: If PM's backend makes multi-turn Anthropic API calls, adding this cache control header is a low-effort, high-impact optimization. Worth checking whether PM already uses prompt caching.

Emerging Patterns

Product maturity inflection. Klatch is transitioning from "can we build this?" (Steps 1-8) to "can non-experts navigate this?" (Steps 9-11). The clearest signal: introducing a dedicated UX designer/developer role to run parallel to Daedalus's backend work. The framing invokes Tesler's Law — the complexity of five-layer prompts, cross-environment import, multi-entity orchestration, and context packaging is irreducible. The team absorbs that complexity so users don't have to. This is a design problem, not just an engineering one. Both projects are approaching similar inflection points: Klatch via explicit role creation, PM via its existing CXO role.

Layer 5 remains the frontier. Three consecutive briefs have circled the same finding from different angles: MAXT Session 01 found Layer 5 content subliminally accessible but unattributable (March 25). The Dispatch import experiment found Layer 5 transfers at 0% (March 26). Now Calliope is piloting explicit externalization as a mitigation (March 28). The pattern is consistent: Layers 1-3 are solved problems; Layer 5 is where the hard questions live.

Inter-agent communication as coordination infrastructure. This window saw 7 inter-agent memos in Klatch — Calliope coordinating roadmap resequencing with Daedalus, Argus verifying Round 12 deliverables, Daedalus assigning Round 13 work, Theseus replying to Calliope on logbook corrections. The mailbox system has become genuine coordination infrastructure, not ceremony. PM's Mailbox v3 serves the same function; the convergence is worth noting.

Background Changes (Noted, Low Priority)

Klatch: Editorial calendar created with three blog posts queued — "It's On the Tip of My Tongue" (subliminal injection finding), "Your Model or Theirs" (Tesler's Law + Three Clocks), "What Doesn't Transfer" (Layer 5 calibration gap). Expected publication March 28-30. These will make Klatch methodology public.
Klatch: docs/UX-POLISH.md created as a backlog for UI improvements accumulated during "functional first" delivery.
Klatch: Demo pipeline (Playwright + npm scripts) added for recording product demos.
Klatch: 189 pre-existing test failures identified across 17 files; root cause diagnosed (client/server config conflict, stale dist/ compiled tests). Assigned to Argus Round 13.
Klatch: Models API dynamic discovery endpoint shipped — GET /api/models fetches from Anthropic's API, caches with 1-hour TTL, falls back to hardcoded list.
Piper Morgan: No new commits in 48h window. Last activity was the March 26 brief delivery.

Sources Read

Klatch:

docs/logs/2026-03-27-1004-daedalus-opus-log.md — File Domain Model design, Step 9d-9e implementation
docs/logs/2026-03-27-1019-calliope-opus-log.md — Roadmap resequencing, editorial calendar, calibration pilot, AXT extension
docs/logs/2026-03-26-1649-argus-opus-log.md — Round 12 verification, test infrastructure audit, Round 13 assignment
docs/logs/2026-03-26-1718-daedalus-opus-log.md — Round 12 delivery (caching, Models API, kit briefing), Step 9a-9c
docs/logs/2026-03-26-1903-calliope-sonnet-log.md — Roadmap resequencing memo, coordination updates
docs/logs/2026-03-26-1912-theseus-opus-log.md — Logbook corrections, MAXT/AXT methodology
docs/mail/calliope-to-daedalus-roadmap-resequence-2026-03-26.md — Resequencing rationale
docs/mail/calliope-to-daedalus-step9-go-ahead-2026-03-26.md — Step 9 authorization
docs/mail/argus-to-daedalus-models-api-verification-2026-03-26.md — Round 12 QA report
docs/mail/daedalus-to-argus-round13.md — Test infrastructure assignment
docs/mail/daedalus-to-calliope-round12-reply-2026-03-27.md — Round 12 handoff
docs/plans/FILE-DOMAIN-MODEL.md — File ownership model design doc
docs/research/cowork-project-format.md — Cross-environment format analysis
docs/COORDINATION.md — Current team coordination state
docs/ROADMAP.md — Resequenced roadmap
docs/AXT.md — AXT methodology with import/export extension
docs/agents/calliope-calibration.md — Layer 5 externalization pilot
docs/EDITORIAL-CALENDAR.md — Blog post pipeline
docs/UX-POLISH.md — UI improvement backlog
docs/DEMO.md — Demo pipeline spec
git log --since="48 hours ago" — 27 commits

Piper Morgan:

git log --since="48 hours ago" — 0 commits
git log --since="72 hours ago" — 1 commit (brief delivery only)