[FEATURE] Model-fluid delegation profile — auto-tiered subagents with escalation

The problem

A session has one fixed model. Users who want cost-efficient operation must hand-build tiering: choosing which subagent model handles which sub-task, writing the conventions down, and enforcing them by prompt discipline. The model selection is a user chore when it could be a system capability.

What we built by hand (the evidence)

A standing workspace convention, enforced via CLAUDE.md across every project in the workspace:

• "Delegate down, review up": the main loop (Fable) sends simple/mechanical tasks to Sonnet subagents, bulk drafting/research to Opus, and reviews every subagent output before anything is reported or written
• Inline exception: tasks smaller than the handoff are done in the main loop
• Escalation ladder: a defect caught in review re-issues the task one tier up, with the failure passed as context (Sonnet's miss → Opus; Opus's miss → main loop); a task class that repeatedly graduates gets its default tier moved up permanently, recorded per-project
• Per-project refinements live in each project's CLAUDE.md

It works well — the fetch-tier agent interacts with CLAUDE.md context effectively and reports up cleanly. But it's all convention: nothing enforces it, nothing measures it, and every user who wants it must reinvent it.

The ask

A native delegation profile — per session or per workspace:

1. The thread has no single fixed model; an orchestrating tier plans and judges, while sub-tasks are auto-routed to the cheapest model adequate for the task class (fetch / make / judge)
2. Automatic escalation: verification failure re-runs the sub-task one tier up, carrying the failure context; repeated escalations recalibrate that task class's default tier
3. Transparency: the UI shows which tier did what and what escalated, so users can audit the routing
4. User-visible dials: "efficiency vs. quality" bias, and per-task-class overrides

Who benefits

Every usage-capped or budget-conscious user — and Anthropic, whose inference fleet spends premium-model tokens today on work a cheaper tier handles fine. The economics align: users get more work per usage window, Anthropic serves more users per GPU.

---

Distilled from a real working session on 2026-07-15 — operating a multi-project workspace under a hand-built tiering protocol. Happy to share the conventions files as a reference implementation. Companion request from the same session: machine-level Remote Control (filed separately).