v5.5.1

Fixed

• Routing hooks now deliver their deny reason to the model. The PreToolUse
  routing hooks (grep→Serena redirect, git add wildcard guard,
  phase-review-close gate) emitted only the legacy reason field, which current
  Claude Code does not read on a deny — so a blocked command arrived as a bare
  "denied" with no cause or remediation. They now emit permissionDecisionReason
  (the full remediation text, model-facing) plus a short top-level
  systemMessage that renders as a terse one-line transcript banner instead of
  the full remediation essay. deny(reason, summary=...) decouples the two
  audiences. [nexus-rpvqu]

Changed

• Tightened the grep→Serena redirect message (~38% smaller): the directive plus
  the exact Serena tool calls, with the human-doc justification prose removed.

Internal

• Substantial cc-validation harness overhaul (test infrastructure; not
  user-facing): keychain-based credential provisioning, deterministic
  --mcp-config launch with trust pre-seed, deferred-MCP-tool warmup, a
  scenario-isolation fix, and elimination of vacuous-pass / model-output-fragile
  assertions across the routing/MCP/permission scenarios.