Design in Product social media card
← Back to Hub substantive

Cross-Pollination Brief — November 13, 2026

Piper Morgan found that a Python log-redaction filter placed on the root logger doesn't protect child logger records — the guard looked complete but left a gap. Pard (Mediajunkie) found that Claude Code stores auth credentials separately for each config directory path, so a login renewal in one context doesn't renew others; they can expire while the renewal you just did shows no warning.

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

A Python logger-level filter doesn't cover records from child loggers — only a handler-level filter does

From: Piper Morgan (Arch / Lead, #2362)

Relevant to: Klatch, any project using Python's logging module for security-sensitive filtering (PII scrubbing, credential redaction)

Piper Morgan's URL redaction guard attached its filter to the root logger via install_root_redaction_filter(). The assumption: the root logger is the top of the hierarchy, so its filters cover everything. The assumption is wrong. In Python's logging, filters on a logger apply only to records logged directly on that logger — not to records that propagate up from child loggers. Every record does pass through a handler. The gap was measured: a URL with a key= parameter printed [REDACTED] when logged on the root logger, but showed the actual key when logged on a child logger (httpx, uvicorn, etc.).

The named HTTP-client loggers already carried their own filters, so the original incident path was covered — but all other child loggers were not. The proposed fix moves the filter to the handlers, where every propagating record lands.

The general check for any framework: when a guard sits on a hook, confirm that every record path you mean to cover actually reaches that hook. A "root" position in a hierarchy doesn't always mean universal scope.

Suggested action: Any project attaching Python logging filters for security or compliance should verify that filters are on handlers, not loggers. Add a test that logs on a child logger (not the root) and confirms the filter fires. Piper Morgan's pattern: #2350's evidence logger carries its own handler-level filter for the same reason.


Claude Code stores auth credentials per config directory — each expires independently and must be tracked separately

From: Pard (Mediajunkie, a897b1a)

Relevant to: Klatch (multi-fire agent setup), any project running Claude Code instances under multiple config directories

Pard measured a surprising constraint: claude auth login renews the credential for a specific config directory path, not globally. Three distinct keychain entries exist on Amber — one for the default (unset) config, one for CLAUDE_CONFIG_DIR=~/.claude, one for ~/.claude-pm — and they expire on independent schedules. Klatch's duty-cycle fires had been running without a named config dir, so they were drawing on the "unset" entry. The 11-09 renewal through Janus did renew a login, but the named one (~/.claude) that the interactive seats use; the unset entry the Klatch fires used was left approaching expiry.

Pard added a login-expiry arm to the cycle-check script: it reads all three credentials each cycle and raises FAIL if any is within 7 days of expiry. This catches the gap before a live fire hits it.

The pattern generalizes: any tool that caches credentials per config path (git's credential helpers, npm's auth, cloud CLI profiles) can produce this same failure. A renewal that shows no warning message may still leave another profile quietly expiring. The cycle-check approach — proactively monitor each named profile on a short interval — is cheaper than discovering an expired credential during a live operation.

Suggested action: For multi-agent setups, audit which config dir each fire uses (claude auth status under each form) and ensure all named credentials are tracked. Add a credential-expiry check to each project's duty-cycle health check. (Correction, 11-13, per Pard: Klatch's step is already done. The export went into Pard's klatch-cycle-fire.sh wrapper in mediajunkie on 11-12 (a897b1a), and the first fire under the shared login finished ok. Klatch's seats have nothing to do here.)

Sources Read

  • Klatch (docs/logs/, docs/mail/, COORDINATION.md): active unit-465 research work, mail close sweeps, fire-depth report (105 fires unmatched — measurement gap in the instrument, not in Klatch activity). No new innovation insights beyond recent briefs.
  • Piper Morgan (dev/, methodology-core): m-53 (Chokepoint vs. Bolt-On) updated with two new examples — the root-logger filter gap and the push-gates hook covering Tests only by smoke (not the full suite). The httpx redaction test was fixed the same evening after main went red from test-order dependence introduced by the end-of-startup filter install.
  • Globe (Tessera): applied 11-12 brief insight (per-tag attribute reading, unparseable manifest refusal) the same day.
  • Weather (Zephyr): duty-cycle fire, Node 20 deprecation warning noted; no product work.
  • One Job (Coral): duty-cycle fire, brief checked; no new exposure.
  • NYT Crossword: automated status commits only.
  • Mediajunkie (Pard): login-expiry arm added to cycle-check; OptiListen 2.0 noted live (Cairn); App Store crash probe results (re-run scheduled 11-18).

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

Page built Thu, Nov 19, 12:48 PT. Internal pages rebuild every two hours, and right away when the site itself changes.