1.9.0

Added

• fable5-fanout-guard.ps1, a PowerShell twin of the fanout guard, runs on native Windows without Git Bash, with no jq or curl dependency, and registers in exec form (interpreter as command).
• The guard's decision table, JSON parsing, and fail-open behavior are proven by a 32-case suite on PowerShell 7 in a Linux container; Windows-specific path handling and registration have not been exercised on a real Windows machine.
• WSL needs nothing new, since it already behaves as ordinary Linux.
• verify-reminder.sh and flight-deck-update-check.sh stay bash-only and are simply absent on native Windows without Git Bash; since neither gates a tool call, that costs a nudge rather than a safeguard.
• Installer and uninstaller prompts are now platform-aware, with a Windows/WSL section in the README and a platform matrix in hooks/HOOKS.md. Detecting the platform first resolves the config directory (honoring CLAUDE_CONFIG_DIR), copy-vs-symlink, chmod +x, and which hooks to offer.
• tools/doctor.sh now detects macOS, Linux, WSL, and native Windows under Git Bash, honors CLAUDE_CONFIG_DIR, and accepts either fanout-guard variant (.sh or .ps1) as a valid install. It flags a missing jq only when a bash hook is installed, since a PowerShell-only install needs none.
• design-fidelity gained a scout pass for the Claude Design import surface: run on a session's first handoff, or whenever the surface seems to have changed, it checks current capabilities and the exchange format against the real documentation and schema rather than memory.
• Phase-2 design-fidelity primitives are documented ahead of the work: the DesignSync tool and the /design-sync skill drive it, using one HTML preview per component.
• Two gotchas carry forward: get_file's 256 KiB cap, exceeded by previews with embedded fonts (regenerate from code rather than editing in place), and design systems shaped like an app, which fall outside the converter's envelope (write previews directly).

Changed

• The routing grid and CLAUDE.md activation block are recalibrated for the current model lineup: Haiku for zero-judgment mechanics, Sonnet as the default for real code, Opus for complex agentic work, Fable for the hardest, longest-horizon reasoning.
• A hard ceiling caps parallel agents at 20; past that, the user must have asked for it explicitly, since current frontier models reach for subagents on their own.
• Verifier subagents are demoted from a default to an exception: checking another agent's deliverable is the orchestrator's own job, and a verifier subagent spins up only when the check itself is heavy or an independent adversarial read is deliberately wanted.
• The effort ladder is restated accurately: high is already the default on the top tiers, so setting it no longer implies a change; low and medium are the real cost control on execution lots, and max is session-scope only, never a valid per-call value.
• A scope-discipline line is added throughout: the plan now states where a lot stops and holds it there, since current models widen a task or lot on their own without saying so.
• Close-out is hardened from guidance into a hard rule: the knowledge pass fires as the first action on any close-out signal, before any mechanical teardown. A single file touched earlier, or a not-yet-merged state, no longer discharges or defers it.