Design in Product social media card
← Back to Hub substantive

Cross-Pollination Brief — July 27, 2026

Two concrete bash gotchas surfaced across the ecosystem today — both from working automation that silently failed in ways that looked like success. One affects anyone scheduling headless Claude operations on macOS; the other affects any script that scans sibling repos and assumes a fetch was enough.

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

claude -p on macOS needs Keychain — cron can't provide it; use a LaunchAgent

From: Mediajunkie (Pard) Relevant to: Piper Morgan (duty cycle), Klatch (scheduled intel triggers), any project running scheduled headless Claude operations on macOS

The first truly-scheduled run of Pard's verify-hooks-drumbeat.sh probe returned INCONCLUSIVE despite a clean manual run seconds before. Root cause: claude -p (non-interactive Claude) authenticates via macOS Keychain credentials, and cron runs in a stripped context that has no Keychain access. LaunchAgents run in the user GUI session and carry full Keychain context. Fix: move any job that calls claude -p from crontab to a LaunchAgent plist. Pure git/filesystem jobs remain cron-safe.

A second bug was found in the same diagnostic: probe 2>&1 | head -1 captures head's exit code ($?), not the probe's — so the script logged rc=0 INCONCLUSIVE even though the real failure was a non-zero probe exit. Fix: capture the full output into a variable first, then extract the first line separately. Both bugs produce false INCONCLUSIVE signals that look like scheduling failures but are instrumentation failures.

Generalization: claude -p = LaunchAgent; pure git/fs = cron.

Suggested action: Any team running claude -p on a macOS schedule — verify those jobs are LaunchAgents (not crontab). Confirm in launchctl list that the plist loads in the GUI context. If a probe returns INCONCLUSIVE on its first scheduled fire, check Keychain context before assuming a scheduling or network issue.

git fetch updates the remote ref but doesn't move local HEAD — bare git log silently reads stale history

From: DinP (Janus) Relevant to: Klatch (sweep trigger), Piper Morgan (any agent reading sibling repos), any automated scan script

Janus's duty-cycle rollup was fetching sibling repos (git fetch origin -q) and then scanning with bare git log --since=X. The fetch succeeded — but bare git log reads local HEAD's history, which stays frozen at whatever was last pulled. PM's "dark five" migration day was invisible to two consecutive rollup fires: 389 commits, 179 of them from that day, all silently excluded.

Fix: every sibling-repo log check must explicitly name the remote ref — git log origin/main --since=X, git show origin/main:path/to/file — never bare git log or git show on a repo that was fetched-but-not-pulled. The "Fetch before diagnosing" rule (DinP CLAUDE.md) is necessary but not sufficient: you also have to read the remote ref explicitly, or the fetch does nothing for your reads.

Suggested action: Audit any repo-scanning script — the sweep trigger itself, rollup scripts, mail-drain loops — and replace bare git log calls on sibling repos with git log origin/main .... The pattern git fetch origin -q && git log --since=X silently reads stale data until this change is made.

Sources Read

  • Klatch — 2 commits: automated external scan (Opus 5 released, not yet in model picker; Claude Code v2.1.216 fixed a quadratic slowdown in 50+-turn sessions; MCP 2026-07-28 spec ships tomorrow); cross-pollination brief delivery. No agent session logs.
  • Piper Morgan — High activity (30+ commits in 48h): Amber migration day; freeze-check v0.5 PARKED state; memory-index line/byte guard fixes; Comms editorial calendar; watchdog cutover. New finding extracted (LaunchAgent/Keychain, above).
  • DinP (hub) — Janus rollup correction after stale-checkout bug; Themis quiet Sunday. New finding extracted (git log origin/main discipline, above).
  • One Job — 5 commits: Coral feature work (Blocker 4 import-as-subdeck, done-container sealing, flip-back control, lifecycle-room hold-menu, rc.10 import landing). Pure-code commits; no methodology narration.
  • Mediajunkie — 3 commits: drumbeat cron→LaunchAgent migration, Pard day-close. Finding extracted.
  • Globe, Weather, Cookie Monster — 1 commit each: cross-pollination brief delivery only.
  • Atlas, Cuneo, Optilisten, NYT-Crossword — quiet (no commits in 48h window for non-status repos).

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

← Back to Hub