Post-Cutoff Re-Sweep — Day 1: the cutoff did NOT land

Scheduled re-sweep (TROVE.md §6 "Re-sweep dates" → post-cutoff re-sweep after 2026-07-08 07:00 UTC; GAP 3 / task #47). Run 2026-07-08 15:50 UTC via two pinned Sonnet lanes — X/Twitter via agent-reach (OpenCLI backend) + web/official — plus a live pxpipe→Fable probe. Raw lane notes: `/.claude/jobs/1881b475/tmp/pc_x_findings.md, pc_web_findings.md`.

VERDICT — EXTENSION HOLDS. No cutoff landed on July 8. The re-sweep was scheduled on the assumption that "true post-cutoff data" would exist after 2026-07-08 07:00 UTC (spree synthesis § honest gaps). It does not: Anthropic extended subscription-included Fable 5 access to 2026-07-12, and the extension is still the latest official word from both @claudeai and @AnthropicAI. The post-cutoff gap stays OPEN — its true re-sweep moves to after 2026-07-12/13. This doc records what the re-sweep did verify (the corrected timeline, the real credit pricing, the field tactic) and re-arms the gap.

1. Corrected timeline (three date fixes vs. our own notes)

The verbatim extension tweet: "We're extending access to Claude Fable 5 on all paid plans through July 12."

2. The two facts that change the doctrine (not the date — these)

a) The reprice trigger is real and it flips Fable to PREMIUM. Live on platform.claude.com pricing docs: Fable 5's credit rate is $10 / MTok in, $50 / MTok out — above Opus 4.8. So the matrix's fable-5 · cost 1 is true only inside the subscription window. When the window closes (post-Jul-12), Fable does not become "credit-metered walk-away-only at some vague cost" — it becomes the single most expensive model on the roster. The eventual reprice must rank Fable's metered cost below Opus, not leave it implied.

b) The catch inside the extension. Next tweet in the same thread (Jul 7 17:36:59 UTC, status/2074548243971604641): "you can use up to 50% of your weekly usage limit on Claude Fable 5. After that, you can keep using Fable 5 with usage credits, or switch to another model." This is why the reply threads are angry despite "extension" being real — the fight is about the 50% weekly cap (and whether it resets), not about losing access. Field-relevant: heavy Fable orchestration can silently cross 50% of the weekly limit mid-window and start billing credits at the premium rate above.

3. The field tactic worth encoding — the "advisor" pattern

@ClaudeDevs (Jul 7 21:26 UTC, status/2074606058128224365), amplified same evening by steipete (i/status/2074638582418231495, who was still actively using Fable hours after the extension): "Use Fable 5 as an advisor. An executor (Sonnet 5) calls Fable 5 for guidance. Most tokens billed at the lower executor rate."

This is the credit-era version of our own orchestration-seat doctrine, but inverted for cost: instead of Fable as the orchestrator burning premium tokens on every turn, Fable becomes a consulted advisor — Sonnet executes and pays the cheap rate on the bulk tokens, Fable is invoked only for the hard judgment calls. Pairs exactly with pxpipe (already compresses Fable's input ~86%) and with the existing "have Sonnet tune the prompt before an expensive Fable run" rule. Fold into CLAUDE.md as the post-window default.

4. Method validation (lanes #1 and #4 of the task)

• agent-reach (X) — WORKS. agent-reach doctor --json → Twitter active_backend: OpenCLI (reuses desktop Chrome login). Direct-read verbs (thread / tweets / profile) fully reliable; pulled the verbatim tweet + thread from the primary account. The search verb degraded mid-run (2 ok → timeout → repeated HTTP 429 session rate-limits, 15–30 min cooldown), which capped broader named-voice coverage. Direct X fetch is now 402-paywalled, so tweet content was corroborated across Forbes / BleepingComputer / TechTimes quoting it verbatim. → the "use agent-reach for X" norm is validated; prefer direct-read verbs over search to dodge the rate limiter.
• pxpipe delegation — GREEN. Proxy up (:47821 HTTP 200); runner intact (pins the full claude-fable-5, not the Opus-fallback fable alias; prepends BOIL_THE_OCEAN). Live pxpipe-run.sh --no-boil round-trip returned PXPIPE_PROBE_OK cleanly — no billing/access error. Independent confirmation Fable is reachable and responding right now on this Max setup (an access probe, not a billing-mode probe).

5. Honest nulls (re-armed for the true post-cutoff sweep)

• Top-voices response is thin — by design, there's little to respond to yet. mattpocock's recent Fable posts all predate the Jul-7 announcement (hallucination/model-taxonomy, not pricing). swyx, IndyDevDan, simonw could not be checked on this specific topic before the X search endpoint rate-limited. Not chasing a 429'd endpoint for secondary color — deferred to the post-Jul-12 sweep, when there's an actual reaction to document.
• No newer reversal exists. Recent @claudeai + @AnthropicAI timelines checked; the Jul-12 extension is the most recent Fable-pricing statement on either account at fetch time.
• Billing mode inside the window (does heavy use auto-cross the 50% cap into credits for this account?) — unverified; would need a live usage-window read, out of scope for a read-only sweep.

6. Re-arm

Next true post-cutoff re-sweep: after 2026-07-12 (revert) / 2026-07-13 (first full metered day) — and treat even that as provisional, since Anthropic called the removal capacity-driven/temporary. Until then: Fable is subscription (cost 1) up to the 50% weekly cap; the advisor pattern is the pre-staged post-window default; the reprice values ($10/$50 MTok) are banked and ready to apply on the revert. TROVE §6 updated to match.