Pending decisions

2026-08-02 · Metered-spend / model-tier session wrap-up. Everything else from this effort is shipped or self-monitoring; these six are the only items that need Fede. Click an option — answers save for everyone on the site.

1 · Outbound-safety framework (PR #5285)

The "make it generic" refactor: pre-send safety becomes a policy registry; fair housing is the first (and only) registered policy, byte-identical behavior — same prompt, model, position, fail-closed direction. Full suite green. Collections legal-review and broadcast safety deliberately stay put (behavior-bearing migrations, own evals later). The PR is 6 files of code + tests only.

A — Approve, merge it. Zero behavior change today; a collections check becomes a one-file add later. Recommended B — Request changes (say what in the notes below). C — Drop it, keep the hard-coded gate.

2 · Second bearer for eval rotation

Bench traffic now fails closed instead of ever billing metered, and rotates across subscription bearers — but the rotation list has one entry (asher). All four accounts are banked in the Mac keychain, but those are short-lived session tokens the laptop auto-refreshes; prod SSM needs a long-lived setup-token bearer, and minting one is a one-time interactive browser approval.

A — Mint clara's long-lived token (~1 min: ! ccswitch use clara@propflowai.co && claude setup-token; the rest is handled). Recommended B — Stay single-bearer for now: if asher caps, evals skip + alert (never bill metered). Fine, just less resilient.

3 · Three Lambdas withheld from the subscription credential

agent-runtime, appfolio-sync, conversation-events-bridge bundle background-lane code but weren't granted the background token — widening a personal credential's footprint is an owner call. Their bench-path traffic is rare.

A — Keep withheld: smallest credential footprint; the drift guard documents the gap. Recommended B — Grant all three: closes the last theoretical bench-billing gap.

4 · Maintenance-nightly model eval

Honest answer to "don't we have enough info now": yes. Sonnet 4.6 is the incumbent that ran this surface for months; this eval only matters as a pre-flight for a future Sonnet 5 upgrade attempt.

A — Skip now; run only as the gate before any future Sonnet 5 re-attempt. Recommended B — Run it now anyway (needs Temporal + armed bench property + live Twilio, supervised).

5 · FAST-tier (Haiku) sweep

"Sweep for what": the same compare-models-on-evals exercise, but for the FAST tier — the Haiku lane behind quick classifiers (fair-housing gate, topic stamps, routing). Nothing is broken there and Haiku is already the cheapest tier, so a sweep answers a quality question, not a cost one.

A — Skip: no cost upside, no known quality complaint. Recommended B — Run it: 11 discriminating suites, subscription-powered, ~1 evening.

6 · Latency early-warning ("catch it sooner next time")

Reply latency grew +25% in 6 days before anyone noticed. Proposed stack, cheapest first: a slope alert on daily median reply time; tracing spans on context assembly; a synthetic canary conversation.

A — Slope alert only: one cron + one Slack line; catches the next drift in ~2 days instead of 6. Recommended B — Full stack (alert + assembly tracing + canary). C — Skip for now.