Phase 1: Model portability (tiered presets + token resolver)

[NicoGenti]

Problem

Every agent's frontmatter hardcodes a concrete model: value (e.g. opencode-go/gpt-5.6-luna, ollama/deepseek-v4-flash:cloud, ollama/minimax-m3:cloud). New users without access to those exact providers/models had to hand-edit 15 files before the kit would run.

What this PR adds

1. .opencode/models.config.json — centralized config mapping three logical tiers to concrete models per preset:

   • TIER_REASONING — orchestrator, oracle, security, planner
   • TIER_CODE — code-reviewer, developer-fixer, test-engineer
   • TIER_FAST — explorer, librarian, pc-doctor, npm-helper, build-helper, deploy-helper, writer, profiler

2. Four presets:

   • opencode-go (default) — router flagship for reasoning, cloud-hosted specialists for code/fast. Matches current out-of-the-box behavior.
   • local-ollama — fully offline: qwen2.5-coder:32b / qwen2.5-coder:14b / deepseek-coder:6.7b.
   • commercial-api-anthropic — Claude Opus/Sonnet/Haiku via ANTHROPIC_API_KEY.
   • commercial-api-openai — GPT-4o / GPT-4o-mini via OPENAI_API_KEY.

3. scripts/apply-model-preset.py — resolves {{TIER_REASONING}}, {{TIER_CODE}}, {{TIER_FAST}} tokens in agents/*.md frontmatter to the concrete IDs from a chosen preset. Supports --list, --dry-run, and --restore (revert concrete IDs back to tokens).

4. scripts/migrate-agents-to-tokens.sh — one-time migration replacing the current hardcoded model values in all 15 agent files with the matching {{TIER_*}} token, using the exact literal values already in each file (idempotent, single-line sed replace, no other frontmatter touched).

Follow-up needed before merge (manual, not included in this PR)

install.sh should call the resolver after copying agents into place. I could not safely retrieve the exact current byte content of install.sh through the tooling available in this session, so rather than risk a blind full-file rewrite of a protected script, here is the exact snippet to add by hand near the end of install_into(), after the link_or_copy "$KIT_DIR/command" "$base/command" line and before its closing }:

  # Phase 1: resolve model tiers for the chosen preset
  local preset="${OPENCODE_MODEL_PRESET:-opencode-go}"
  if [[ -f "$KIT_DIR/scripts/apply-model-preset.py" ]]; then
    echo "Applying model preset: $preset"
    python3 "$KIT_DIR/scripts/apply-model-preset.py" \
      --preset "$preset" \
      --config "$KIT_DIR/.opencode/models.config.json" \
      --agents-dir "$base/agents"
  fi

Usage after wiring:

./install.sh project                                   # default preset: opencode-go
OPENCODE_MODEL_PRESET=local-ollama ./install.sh project
OPENCODE_MODEL_PRESET=commercial-api-anthropic ./install.sh global

Suggested merge order

1. Review/merge this PR (new config + scripts, zero behavior change on its own).
2. Run bash scripts/migrate-agents-to-tokens.sh to tokenize agents/*.md.
3. Run python3 scripts/apply-model-preset.py --preset opencode-go --dry-run to verify round-trip resolves back to the original values.
4. Wire the install.sh snippet above.
5. Mention OPENCODE_MODEL_PRESET in README.md / QUICKSTART.md.