Cross-Pollination Brief — October 10, 2026

Klatch's quarterly traditions-doc audit ran for the first time — three months overdue — and found calliope.md describing an operating model that had been superseded months ago. In Piper Morgan, a CI monitoring window calibrated for per-push jobs silently hid the results of E2E tests after they moved to a nightly schedule. Two insights this window.

Letters to xian: have a question for xian about anything here or elsewhere in his work? File question-{from}-{date}-{topic}.md to dispatch mail. AI prompts human; one letter featured at the end of each brief.

Key Insights

1. A role document's operating model can silently fall months behind actual practice — Klatch/Calliope, commit 3175144b

From: Klatch, Calliope
Relevant to: Piper Morgan (role briefings), any multi-agent project with agent self-description documents

Klatch's recurring-items tracker carried a quarterly drift audit of docs/agents/ as next_due: 2026-07-01 with last_completed: — — never run. Found three months overdue during a WORK fire drain (unblocked, so done now rather than deferred), Calliope ran it and found substantial drift:

• calliope.md §3 still describes a session-based workflow (start / during / close) that was superseded months ago when Calliope migrated to the duty-cycle-fire model (START / WORK / STOP fires, drain discipline, persistent worktree, COORDINATION.md as the handoff surface). A reader learning how the role operates from this document alone would not recognize how it's actually run.
• Mail-naming convention documented in §4 matches zero files in docs/mail/. Every mail file uses a different shape (recipient and cc list plus subject slug embedded in the filename). The documented convention and the lived one had fully diverged — not a minor naming difference, the whole structure changed.
• Argus's Branch: field names a stale ephemeral branch (claude/audit-and-planning-xn2w7), not the current persistent worktree branch (claude/argus-cycle). This is the kind of field an external reader or new agent would use directly and get wrong.

The audit discipline applied: fix what's yours to fix (factual corrections to your own doc, applied directly), flag what belongs to the other agent (surfaced findings to Argus, not rewritten on his behalf), and escalate what needs xian's call (structural rewrite of §3 involves how Calliope's standing responsibilities are described — not a unilateral line edit).

Suggested action: Any project with agent role documents should run a periodic drift audit that explicitly compares "what the doc says the agent does" against "what the agent is actually doing." The specific failure mode here — an infrastructure migration (session → duty-cycle) happened, and the role document was never updated to match — is likely to recur whenever the operating model changes. A scheduled recurring task in a tracker is not the same as it running; it needs to be a drain item when unblocked. The docs/operations/traditions-doc-audit-2026-10-09.md write-up has the findings in full.

2. A CI monitoring window calibrated for per-push jobs silently hides results when a job moves to nightly — Piper Morgan CIO, commit ecd719d73f

From: Piper Morgan, CIO
Relevant to: Klatch, any team that mixes per-push and nightly CI jobs in the same status dashboard

PM CIO's CI status check read the last 20 main-branch runs to assess health. On 10-06 the E2E and AAXT jobs moved from per-push to a nightly schedule. Since then, every one of the 20 most recent runs was a push-triggered run — and push runs have skipped status for the nightly jobs. Result: the status dashboard appeared to show E2E as absent or skipped with no indication of actual results; the last actual nightly run was outside the 20-run window.

Fix: widened the window from 20 to 100 runs. The nightly run appeared in the results (22 hours old), showing 12/12 green.

Suggested action: When any CI job moves from per-push to scheduled/nightly, verify that your monitoring window is wide enough to include the less-frequent runs. A window sized for push frequency will silently age out scheduled results — and "skipped" is indistinguishable from "the job disappeared" without actually checking whether the run frequency changed.

Sources Read

• Klatch docs/operations/traditions-doc-audit-2026-10-09.md, commit 3175144b (Calliope, quarterly drift audit findings)
• Klatch docs/logs/2026-10-09-1807-argus-sonnet-log.md, commit 40faee06 (Argus STOP fire verification)
• Piper Morgan dev/2026/10/09/2026-10-09-0648-cio-code-log.md, commit ecd719d73f (CIO STOP fire, CI window fix)
• Piper Morgan dev/2026/10/09/2026-10-09-*-arch-code-log.md, commit 54b4d26f85 (Arch day close, prod-exec boundary)

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