Loki Mode v7.104.0

Default execution model is now Claude Sonnet 5

• Default execution model changed Opus -> Sonnet 5. The planning and
  development tiers now default to sonnet (providers/claude.sh), which the
  Claude CLI resolves to the latest Sonnet (Claude Sonnet 5). Builds are cheaper
  (~40% less per token than Opus) and faster, at near-frontier intelligence. This
  affects every user's default loki start/loki verify run. Completion is still
  council-verified. To restore Opus execution: set
  LOKI_CLAUDE_MODEL_DEVELOPMENT=opus (and _PLANNING=opus), or pick Opus from
  the dashboard. We do NOT claim identical accuracy (no controlled study); the
  quality gates are unchanged, so if it passes it passes and if it does not Loki
  tells you why.
• Opt-in strong advisor judge (LOKI_ADVISOR_MODEL). The trust-gate code
  reviewers still run on the account default by default (unchanged behavior). Set
  LOKI_ADVISOR_MODEL=opus (or the dashboard "Advisor" toggle) to pin the
  reviewers to a stronger, consistent Opus judge. haiku|sonnet|opus are honored;
  fable is refused (its safety classifiers can end a review turn with a refusal
  and break the council gate). Mirrored on both the bash and Bun routes.
• Dashboard model selection. A start-time model selector (Haiku / Sonnet /
  Opus, default Sonnet 5) and a mid-run switcher (applies at the next iteration)
  are available in the dashboard. The mid-run hot-switch was already supported by
  the engine (.loki/state/model-override, read each iteration); this exposes it
  in the UI. An explicit Opus pick dispatches Opus on every iteration.
• opus session-pin fix. With the Sonnet default, a LOKI_SESSION_MODEL=opus
  pin previously routed through the (now-Sonnet) planning tier and silently
  dispatched Sonnet. It now dispatches Opus on every iteration (still clamped by
  LOKI_MAX_TIER, so an operator cost ceiling still binds). Fixed consistently in
  the runner, the loki plan estimator, and the dashboard, and locked by the
  session-pin parity matrix.
• Model catalog + pricing corrections. providers/model_catalog.json refreshed
  to current IDs (claude-opus-4-8, claude-sonnet-5). Corrected the
  gpt-5.3-codex price to the real standard-tier $1.75/$14.00 per MTok (was
  $1.50/$12.00) in the cost tables. Added a display-only note for the Sonnet 5
  introductory price ($2/$10 per MTok through 2026-08-31); the estimator quotes the
  standing $3/$15 list price (over-estimating the display is the safe direction).