exploration: config-driven model profiles with inheritance

[rwoll]

Problem

When experimenting with how well existing prompts and tools work with new model variants, we currently need custom builds of the extension — even if the new model behaves identically to an existing one. Every new model requires hardcoded changes spliced across multiple boolean functions in chatModelCapabilities.ts and potentially a new prompt resolver file, all of which require a full PR/build/deploy cycle just to try things out.

This makes it harder to iterate quickly during model evaluation and slows down the feedback loop for understanding whether a new variant needs custom behavior or can reuse an existing archetype.

Exploration

This PR explores making model capabilities config-driven so that evaluating a new model variant that behaves like an existing archetype requires zero code changes — just a settings override or a one-line profile entry.

What it does

New: modelProfiles.ts — a declarative registry where each model inherits from a base archetype and only overrides what's different:

'_openai-modern': {
    editStrategy: 'apply-patch',
    notebookFormat: 'json',
    // ...inherited defaults
},

'gpt-5.3-codex': {
    extends: '_openai-modern',
    verbosity: 'low',
},

Refactored: chatModelCapabilities.ts — all ~15 capability functions become thin wrappers:

export function modelSupportsApplyPatch(model): boolean {
    return expandEditStrategy(getModelProfile(model).editStrategy).supportsApplyPatch;
}

All function signatures are unchanged — zero callsite modifications needed.

New setting: github.copilot.chat.advanced.models.profiles — override profiles from settings.json for eval without any code changes:

"github.copilot.chat.advanced.models.profiles": {
    "new-model-family": {
        "extends": "_anthropic",
        "editStrategy": "apply-patch"
    }
}

BYOK models

BYOK models get the right profile automatically as long as their model ID starts with a recognized prefix (claude, gpt, gemini, grok-code, etc.). Unknown model IDs fall through to the default profile (insert-edit-only, standard settings) — which is the safe behavior. And with the new settings override, users can explicitly map any unknown model to an archetype via github.copilot.chat.advanced.models.profiles without rebuilding.

Key design decisions

• Inheritance via extends — profiles inherit from base archetypes (_anthropic, _openai-modern, _gemini, etc.) and only override what's different
• EditStrategy enum — replaces ~6 interrelated booleans with one meaningful concept that expands to the right flag combinations
• Longest-prefix-first matching — gpt-5.1-codex matches before gpt-5.1 before gpt-5 before gpt
• Hash-based models — unreleased models still use SHA-256 hash matching, resolved to profiles internally
• Backward compatible — prompt registry (familyPrefixes / matchesModel / registerPrompt) is untouched; profiles handle capabilities, prompt routing stays as-is

Validation

• 27 new unit tests for profile resolution, inheritance, prefix matching, and edit strategy expansion
• Snapshot (.snap) files have been intentionally removed from this PR — snapshot tests that reference them will fail until regenerated

What this does NOT change

• TSX prompt components still exist — models that genuinely need different instructions still need custom prompts
• The prompt registry pattern is untouched
• All existing callsites are unchanged

See PLAN.md in the repo root for the full design document.