Cross-Pollination Brief — October 6, 2026

Two tool-behaviour traps that produce clean exits and wrong results: a macOS sed flag that silently becomes a filename suffix, and a CI health script that reports green while a gating check runs red for five days undetected.

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. On macOS/BSD, sed -i -E silently uses -E as a backup-file suffix — the extended-regex flag is never applied, and the command exits 0 — Mediajunkie / Piper Morgan, Pard, commits 517ae03 · ad6a6475a5

From: Mediajunkie, surfaced to Piper Morgan via Pard's restart-gate sweep
Relevant to: any project running shell scripts on macOS or a BSD host

Pard found an untracked file named decisions.log-E in a PM seat's working tree while running a binary-restart readiness check. The filename is the diagnostic: on macOS/BSD, sed -i takes a backup-suffix as its argument, not as a standalone flag. So sed -i -E 's/…/…/' file is read as "edit in place, use -E as the backup suffix" — the backup file-E is created, the -E extended-regex flag is never applied, and the regex runs as basic regex (BRE) instead of ERE. The command exits 0, produces no error, and the result looks like a success.

The safe forms on macOS are sed -i '' -E '...' file (empty-string suffix = no backup, then the flag) or sed -E -i '' ... (flag first). On Linux/GNU, sed -i -E works as expected.

Three related BSD-vs-GNU differences surfaced in the same week: cat -A output format, \? in a BRE, and this. If a script was written and tested on Linux, its sed calls are worth auditing before it runs on macOS.

Suggested action: Audit any cross-platform shell script using sed -i — check that the backup argument is explicit (or '') and that -E appears as a separate argument with no suffix between -i and -E. The file-E backup artifact is the fingerprint if you're checking existing repos.

2. A CI health script that hardcodes one workflow name misses every other gating check — Piper Morgan, CIO, commit 6c0116cb2b

From: Piper Morgan, CIO
Relevant to: any project with multiple CI workflows gating main

PM's duty-cycle Step 1e checked exactly one workflow (Code Quality / lint.yml) as the CI health signal. The Architecture Enforcement ratchet — a separate workflow that enforces structural constraints — ran red for 41 consecutive runs from 2026-10-01 to 2026-10-05 with nobody looking: the health script never read it.

The fix (scripts/main-ci-status.sh) derives the workflow list dynamically from .github/workflows by parsing each workflow file for push triggers that include main. Any new gating workflow is automatically covered the day it lands. The script also prints its denominator on every run: "N workflows: G green, R red, U unmeasured" — so a partial result is distinguishable from a full one.

The lesson generalises: a CI health check that hardcodes a name is a coverage claim for one file, not for "CI." Two fixes together close the gap: derive the list from the actual trigger config, and always state the denominator.

Suggested action: Audit your CI health checks for hardcoded workflow names. If you have more than one push-to-main workflow, the check should enumerate all of them — or at minimum, document explicitly which ones it omits and why.

Sources Read

• Klatch — docs/mail/ (30+ round-coordination memos, Oct 4–5); .scratch/ restart-gate resolution (Pard's memo); probe rounds 333–338. No new transferable insights from probe work (Klatch-internal methodology); restart-gate resolution covered under Mediajunkie.
• Piper Morgan — mailboxes/docs/inbox/ (Pard's BSD-sed memo); scripts/main-ci-status.sh (CIO fix 6c0116cb2b); day-close session logs Oct 5.
• Mediajunkie — logs/2026-10-05-pard-log.md (BSD-sed discovery, restart-gate re-run; commit 517ae03).
• Globe — scripts/fire_check.sh added (Oct 5, 9ffed27): health block as a script with brief-age threshold derived from the delivery record. Application of 10-05 insight; not a new insight.
• Weather — session log (Oct 5); fire-prompt trim confirmed. Operational. Not brief-worthy.
• One Job — build 40 → 1.1.1 archived, ASC-verified, upload deferred to xian session. Operational. Not brief-worthy.
• NYT Crossword — automated status/remarkable prints. Not brief-worthy.

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