Fable phase 3 pre-spec — June 22 cutover, dispatch gate + spend log

Context (all measured/decided 2026-06-11)

• Anthropic pulls Fable 5 from ALL subscription plans June 22/23 → usage credits ($10/$50 per M; new tokenizer ≈ +30% tokens vs Opus). Sources: [[2026-06-10-stratechery-fable-5-anthropic-alignment-ai-tiers]], [[2026-06-10-every-how-to-get-the-most-out-of-fable-5]].
• Founder budget reality: ~$200 credits loaded + $200/mo console cap → max exposure $400/mo. He wants to "lean into this capability," loose ROI criteria.
• Phases 1+2 SHIPPED 2026-06-11 ([[2026-06-11-system-cron-runner]]): cron suite off the fat session → measured spend driver (70% cache reads from ~30 wakes/day) eliminated. Steady-state Fable exposure post-restructure is the always-on channels session + founder-directed work only.
• Founder deferred phase 3 "closer to the cutover date" (iMessage 06-11 08:08).

The flip (founder, ~10 min, June 21-22)

1. /model → claude-opus-4-8[1m], save as default. (Restart script has no --model flag by design — it inherits the saved default; verified 06-09.)
2. Confirm console: credit balance + the $200/mo cap still set as intended.
3. Greenlight the Fable dispatch gate below (or adjust trigger list).

Ray-side build (do in the 06-18→06-21 window)

1. Fable dispatch gate

Fable becomes a callable judgment seat, not the substrate. Two mechanics, both already supported by the harness:

• In-session: Agent tool accepts model: "fable" — dispatch judgment- heavy subagent seats on Fable while the main loop runs Opus 4.8.
• Headless: claude -p ... --model claude-fable-5 via the system runner for scheduled judgment work (manifest model column gains a fable mapping in cron-dispatcher.py MODEL_MAP — one line).

Earns Fable (adapted from Every's four-quality filter — organized context + defined goal + checkable done + cost-justified importance):

• Investing thesis synthesis / backtest interpretation (verify-strategic-output targets)
• CAF restructure design + architecture work (the main bet)
• Founder direct asks where he says "go deep" (or Ray judges depth-critical)
• Multi-hour autonomous builds' DESIGN phase (implementation seats stay Opus/Sonnet)

Does NOT earn Fable: cron cycles, triage, report formatting, mechanical edits, anything a precheck/template already shapes.

Hard exclusion (pending terms review): any phData client-adjacent content. Mythos-class traffic carries mandatory 30-day retention (Stratechery 06-10), and the 06-11 newsletter watch surfaced a possible NDA/data-retention conflict for client work (Every, Mike Taylor finding — vault: 06-reference/2026-06-11-every-ai-everywhere-all-at-once.md). The existing employer-boundary rule already keeps confidential phData artifacts out of the agent entirely; this extends it: even permissible client-adjacent prep stays OFF Fable until the retention terms are assessed. Opus 4.8 (subscription) is the ceiling for that work.

Budget shape: $200/mo ≈ $6.6/day ≈ 1-2 deep dispatches/day at the measured $3-8/dispatch. Console cap is the hard stop; the gate is for ROI, not enforcement.

2. Spend log

Every Fable dispatch appends one line to ~/.claude/state/fable-spend-log.md (via Bash — Write tool is denied on ~/.claude/** headless):

<iso> | <task one-liner> | <justification vs filter> | <cost from usage JSON>

• Headless runs: cron-job-runner.sh already captures total_cost_usd in run logs — a 5-line addition tees Fable-run costs into the spend log automatically.
• In-session dispatches: the dispatching turn logs it (convention, documented in CLAUDE.md-adjacent SOP at build time).
• Surfacing: /morning-prep gains one optional line (month-to-date Fable spend vs $200) — only shown when spend > $0 that month; monthly ROI review rides the existing /finance-pulse or a founder ask.

3. Pre-flip verification checklist

• Opus 4.8 main loop smoke: one restart on the saved default, channels live, rearm hook arms 2 session jobs.
• MODEL_MAP fable entry + one forced headless Fable run completes + logs cost.
• Spend-log tee verified on that run.
• Handoff updated: "main loop = Opus 4.8; Fable = credit-billed judgment seat, gate + log active."

Open items at flip time

• Whether 1M-context [1m] Opus variant stays the default (input-cost vs context-headroom tradeoff — main session rarely needs >200K post-restructure).
• Scheduled cloud agents pilot (EDGAR/smart-money) — independent, still parked.

Related

• [[2026-06-11-system-cron-runner]] — phases 1+2 (shipped)
• [[2026-06-09-fable5-harness-review]] — the original cost modeling
• [[2026-06-10-every-how-to-get-the-most-out-of-fable-5]] — the four-quality filter