model routing: Sonnet 5.5 by default for every agent, the orchestrator included; Opus only on evidence (owner 2026-10-19) #10921

Goal

Owner directive 2026-10-19 (translated): "set Sonnet as the default model for ALL agents -- it is cheaper with the same result". Every role, the orchestrator included, starts on Sonnet 5.5; Opus 5.5 is reached only on evidence (ESCALATE_AFTER failed attempts on the same task) or the owner's explicit flag. This replaces the 2026-10-18 rule "orchestration on Opus, sub-tasks on Sonnet" in its single home, specs/claw/model_route.t27, and in the places that point to it.

Why

The 2026-10-18 rule kept orchestration on Opus as "not a false economy". The owner now judges the result the same on Sonnet. The measurement it rested on (#10066) already showed Sonnet at least as good on bounded tasks at about 40 % of the cost; the orchestration role was the one exception, now lifted by the owner.

Deliverable

• specs/claw/model_route.t27: base_tier returns Sonnet for every role; route raises any role to Opus only on the owner flag or ESCALATE_AFTER failures; is_waste flags Opus without that evidence for every role. The header records both directives with their dates; tests rewritten; a negative control (Opus with no evidence is waste, for an orchestration role too). Seal re-saved.
• .claude/agents/trinity.md: model: sonnet.
• CLAUDE.md "Model routing" paragraph and .claude/skills/model-routing/SKILL.md: the new rule, still pointing to the spec as the law.

Done when

t27b test and t27c test-report (0 vacuous) pass on the spec; the seal verifies; every .claude/agents/*.md says model: sonnet; no file outside the spec restates the old rule as current.

Boundary

• May touch: specs/claw/model_route.t27, its seal, .claude/agents/trinity.md, CLAUDE.md (the Model routing paragraph only), .claude/skills/model-routing/SKILL.md.

Links

#10493 (the 2026-10-18 rule), #10066, #10522.

User Scenarios

• Given the model routing spec specs/claw/model_route.t27 currently encodes the 2026-10-18 rule (orchestration on Opus, sub-tasks on Sonnet), when the spec is updated per the 2026-10-19 directive, then base_tier returns Sonnet for every role including orchestration, and route raises any role to Opus only on the owner flag or ESCALATE_AFTER failures.
• Given an agent configuration file .claude/agents/trinity.md that may reference a model, when the new rule is applied, then it declares model: sonnet as the default.
• Given the documentation files CLAUDE.md (Model routing paragraph) and .claude/skills/model-routing/SKILL.md that previously described the old rule, when they are updated, then they state the new rule (Sonnet default for all roles, Opus only on evidence or owner flag) and still point to the spec as the law.

Requirements

• FR-001: The spec specs/claw/model_route.t27 MUST define base_tier to return Sonnet for every role including orchestration.
• FR-002: The spec specs/claw/model_route.t27 MUST define route to raise any role to Opus only when the owner flag is set or ESCALATE_AFTER failures occur on the same task.
• FR-003: The spec specs/claw/model_route.t27 MUST define is_waste to flag Opus usage without evidence for every role including orchestration.
• FR-004: The spec header MUST record both the 2026-10-18 and 2026-10-19 directives with their dates.
• FR-005: The spec tests MUST be rewritten to match the new rule, including a negative control showing Opus without evidence is waste for an orchestration role.
• FR-006: The seal for specs/claw/model_route.t27 MUST be re-saved after changes.
• FR-007: The file .claude/agents/trinity.md MUST declare model: sonnet.
• FR-008: The "Model routing" paragraph in CLAUDE.md MUST describe the new rule and point to the spec as the law.
• FR-009: The file .claude/skills/model-routing/SKILL.md MUST describe the new rule and point to the spec as the law.
• FR-010: No file outside the spec MUST restate the old rule as current.