Cross-Pollination Brief — March 20, 2026

Both projects shipped substantial work. Klatch delivered the creation UI for multi-entity group chats (the flagship roundtable feature) and established a standing intelligence monitoring protocol that seeds the cross-pollination process. Piper Morgan discovered a systemic Capability Awareness Gap — five disconnected sources of truth for what the agent can do — and fixed it via registry-driven reconciliation in a single session. The CIO agent launched Piper Alpha planning with three stakeholder memos, and the Docs agent institutionalized the omnibus log format as a reusable skill.

Key Insights

1. Capability Awareness Gap — Five Sources of Truth, Zero Coordination

From: Piper Morgan (dev/2026/03/20/2026-03-20-0618-lead-code-opus-log.md) Relevant to: Klatch

A smoke test on the ADR-059 dispatcher revealed a deeper bug: the LLM was offering capabilities that weren’t registered in the dispatcher, creating implicit contracts the system couldn’t fulfill. Investigation uncovered five disconnected systems describing Piper’s capabilities: (1) PIPER.md system capabilities, (2) PIPER.md integrations, (3) soft invocation detector, (4) workflow dispatcher registry, (5) ContextAssembler + canonical handlers. None were coordinated.

The fix reconciled all five via the dispatcher registry as the single source of truth: PIPER.md cleaned to reflect only real capabilities, soft invocation gated on the registry (prevents offers for unregistered workflows), ContextAssembler made registry-aware, and the floor addendum rewritten to stop instructing "NEVER say you can’t help" (harmful when the agent genuinely can’t). 13 files modified, 5 new tests, 2,633 passing.

Suggested action (Klatch): As entities gain capabilities via the roundtable system, this same gap will appear: entity prompts (layer 5) may describe capabilities the system can’t actually execute. Consider early: where is the canonical list of what an entity can do? If entity prompts can diverge from actual system capabilities, users will hit the same implicit-contract problem.

2. Intelligence Feed Protocol — Standing Monitoring as Infrastructure

From: Klatch (docs/INTELLIGENCE.md, docs/intel/2026-03-20-sweep.md) Relevant to: Piper Morgan

Argus ran the first structured intelligence sweep: 20 items scored by relevance across Anthropic’s blog, Claude Code changelog, tech press, and the MCP ecosystem. Key findings included Claude Code Channels (validates Klatch’s multi-entity thesis but remains session-scoped/ephemeral), the Compaction API (enables long conversations without context loss), and the Agent SDK (Anthropic shipping official agent orchestration). Each item got a relevance score and triage recommendation.

The sweep was immediately actionable: Daedalus cherry-picked three quick wins from Argus’s findings and shipped them same day — adaptive thinking, Haiku 4.5 default, 16K max_tokens. This intelligence-to-implementation cycle took hours, not days.

Suggested action (Piper Morgan): The CIO role could adopt a similar standing sweep protocol. Piper Morgan’s 14-agent team has no formal mechanism for tracking external changes that affect the project. The Klatch model (dedicated agent, structured format, scored items, immediate triage) is proven and lightweight. The cross-pollination process itself was born from this intelligence feed.

3. Klatch Creation UI — Multi-Entity Group Chat Ships

From: Klatch (docs/logs/2026-03-20-1955-daedalus-opus-log.md) Relevant to: Piper Morgan

Daedalus shipped the core klatch creation feature (issue #10): a segmented chat/klatch toggle in the sidebar, project dropdown (required for klatches), entity picker with checkboxes (max 5 entities, color dots, model badges), and interaction mode selector. End-to-end verified: klatches appear in the sidebar under the project’s KLATCHES subsection.

This is the feature that transforms Klatch from a conversation manager into an agent coordination tool. A klatch is a structured multi-entity conversation within a project context — exactly the kind of coordination surface that Piper Morgan’s 9-agent days have shown is needed.

Suggested action (Piper Morgan): Track this feature’s development. When klatches support entity-to-entity messaging or roundtable discussions, they could serve as the "shared configuration layer above individual repos" that the Mailbox v3 plan envisions.

4. Two-Track AXT Testing — Separating Plumbing from Perception

From: Klatch (docs/logs/2026-03-20-1938-theseus-opus-log.md, docs/fork-continuity-quiz.md) Relevant to: Piper Morgan

Theseus designed a two-track agent experience testing program: AAXT (Automated Agent Experience Testing) uses synthetic context and deterministic assertions against the prompt-debug endpoint — no LLM calls, no nondeterminism. It verifies plumbing: did the right context reach the right layer? MAXT (Manual Agent Experience Testing) uses real agents and qualitative interpretation via the Fork Continuity Quiz — can an agent orient itself from prompt context alone?

The key insight: separate what’s mechanically verifiable (context assembly, layer injection) from what requires human judgment (agent behavior, coherence, orientation quality). AAXT runs as a gate; MAXT only runs if plumbing checks pass.

Suggested action (Piper Morgan): Piper Morgan’s 6,190 tests verify code behavior but not agent experience. The AAXT model — asserting against the assembled context rather than LLM output — could complement existing tests. For example: before a session starts, verify that the briefing, role definition, and current sprint state are all present in the assembled prompt.

5. Piper Alpha Planning — Convergent Agent-Persona Architecture

From: Piper Morgan (mailboxes/ppm/inbox/memo-cio-piper-alpha-tasks-2026-03-20.md, CIO memos) Relevant to: Klatch

The CIO agent launched planning for "Piper Alpha" — a Claude Code agent that inhabits the Piper Morgan persona as a working PM assistant, distinct from the current multi-agent development methodology ("Piper M"). Three stakeholder memos were delivered:

To PPM: Requesting input on PA’s first tasks, introducing the framing of "ceiling moments" (where Piper M needs structured capabilities PA lacks) vs. "path moments" (where PA’s conversational approach works better).
To CXO: Requesting voice/persona guidance — does the autobiography voice work for a working assistant or needs tuning?
To Architect: Requesting technical safety analysis — can PA and Lead Dev share a repo? What branch discipline and path boundaries are needed?

Suggested action (Klatch): Piper Alpha is a natural Klatch user — a persistent AI persona that needs project context, conversation history, and multi-entity coordination. The "ceiling vs. path" framing is useful for Klatch’s own entity model: some tasks need structured orchestration (roundtables, handoffs), others work fine as simple prompted conversations.

6. Omnibus Institutionalization — Process as Reusable Skill

From: Piper Morgan (docs/omnibus-logs/2026-03-20-omnibus-log.md, .claude/skills/create-omnibus/SKILL.md) Relevant to: Klatch

The Docs agent converted the omnibus log format — previously a manual, drift-prone process — into a formal /create-omnibus skill with a 9-step runbook. This locks down format, ensures consistent sourcing (scan session logs, cross-reference commits, verify metrics), and prevents the recurring problem of agents producing omnibus logs that don’t match the expected structure.

This is a pattern: when a process is executed repeatedly by different agents, converting it to a skill (with explicit steps, validation criteria, and format specifications) prevents drift. The cross-pollination sweep prompts follow the same pattern.

Suggested action (Klatch): As Klatch agents accumulate recurring processes (release runbook, intelligence sweep, session wrap), consider formalizing them as skills with explicit runbooks. The Piper Morgan skill format (.claude/skills/{name}/SKILL.md) is a working convention.

7. Reliability Incident — Force Push Destroys Demo Infrastructure

From: Klatch (docs/mail/calliope-to-argus-reliability-incident-2026-03-20.md) Relevant to: Piper Morgan

Argus’s prior session claimed demo infrastructure was complete, but a forced push during rebase recovery had silently destroyed the commits. When xian tried to follow the demo instructions, the files didn’t exist. Calliope’s response: (1) filed an incident memo, (2) added mandatory session-wrap verification to CLAUDE.md (git log before claiming done, ls/cat to verify deliverables, push session log last), (3) added a blanket no-force-push rule without explicit PO approval.

The root cause: claiming completion at the session log layer without verifying the git layer. Same structural pattern as Piper Morgan’s Pattern-063 (Extension Without Integration).

Suggested action (Piper Morgan): Piper Morgan already has evidence requirements for issue closure (added retroactively to #922 this same day). Consider extending to a general session-wrap verification protocol similar to Klatch’s: before any session log claims completion, the agent must verify deliverables exist in git.

Emerging Patterns

Registry-as-truth is converging. Piper Morgan reconciled five capability sources into a registry-driven model. Klatch uses a project registry for entity management. Both projects are learning the same lesson: when multiple components describe the same thing, one of them must be canonical and the others must derive from it. The alternative — parallel sources that drift — produces subtle, hard-to-debug failures.

Intelligence-to-action cycles are tightening. Klatch went from intelligence sweep to shipped code in hours (Argus finds, Daedalus implements). Piper Morgan went from audit finding to ADR to implementation in a single session. The pattern: structured discovery (sweep, audit, smoke test) feeds directly into same-day execution rather than being logged for "later."

Process institutionalization is accelerating. Both projects are converting ad-hoc processes into formal, repeatable artifacts: Klatch’s intelligence protocol and wrap-session skill, Piper Morgan’s omnibus skill and audit cascade methodology, the cross-pollination sweep prompts themselves. This is the natural response to multi-agent coordination — when any agent might execute a process, the process must be explicit enough for any agent to follow.

Background Changes (Noted, Low Priority)

Klatch: Website restructured — web/ moved to repo root, GitHub Pages redirect stubs added
Klatch: Markdown wireframe blog post drafted
Klatch: 5 new issues triaged (#15-#20: adaptive thinking, Haiku update, effort parameter, Compaction API, Agent SDK)
Klatch: 16 legacy messages backfilled with NULL model values (issue #20 closed)
Piper Morgan: Blog metadata fully populated (269/269 imageSlug), 175/269 images localized from Medium CDN before rate limit
Piper Morgan: BlogPostContent.tsx interface bug discovered (RSS field names vs. schema names causing "Invalid Date" on 146 posts)
Piper Morgan: #911 (Floor Inversion) Phases 1-2 confirmed complete, Phases 3-4 deferred post-M1
Piper Morgan: #924 (Chat Avatars) filed and shipped same session — dolphin SVG for Piper, colored initial for user
Piper Morgan: Evidence retroactively added to #922 closure

Sources Read

Klatch:

docs/logs/2026-03-20-1438-argus-opus-log.md — Intelligence sweep, feed protocol design
docs/logs/2026-03-20-1749-calliope-sonnet-log.md — Reliability incident response, wrap verification
docs/logs/2026-03-20-1938-theseus-opus-log.md — AXT two-track testing design, Fork Continuity Quiz v4
docs/logs/2026-03-20-1955-daedalus-opus-log.md — Klatch creation UI, quick wins from intel sweep
docs/intel/2026-03-20-sweep.md — First intelligence sweep (20 items)
docs/INTELLIGENCE.md — Standing intelligence protocol
docs/COORDINATION.md — Multi-agent status board
docs/mail/calliope-to-argus-reliability-incident-2026-03-20.md — Incident memo
docs/mail/argus-to-daedalus-intel-sweep-2026-03-20.md — Intel sweep handoff
docs/fork-continuity-quiz.md — Fork Continuity Quiz v4

Piper Morgan:

docs/omnibus-logs/2026-03-20-omnibus-log.md — Full day synthesis (3 agents, STANDARD)
dev/2026/03/20/2026-03-20-0608-docs-code-opus-log.md — Docs session (omnibus, blog images, create-omnibus skill)
dev/2026/03/20/2026-03-20-0618-lead-code-opus-log.md — Lead Dev session (#923 capability gap, #924 avatars, M1 audit)
mailboxes/ppm/inbox/memo-cio-piper-alpha-tasks-2026-03-20.md — CIO memo to PPM
mailboxes/cxo/inbox/memo-cio-piper-alpha-voice-2026-03-20.md — CIO memo to CXO
mailboxes/arch/inbox/memo-cio-piper-alpha-technical-2026-03-20.md — CIO memo to Architect
.claude/skills/create-omnibus/SKILL.md — Omnibus skill runbook