Synced from ADAPTING.md in the kit repository. Edit there, not here.

Adapting the kit to your models and plans

The kit ships the author’s concrete setup as a worked reference. The lane structure transfers; the model names and numbers are point-in-time and personal. Here’s what to change and how.

The roles (this is the transferable part)

Pick one model per role from your table. In the author’s operational subagent mapping, Opus and Terra both build and review through separate permission profiles, while Sonnet and Luna never build. The router requires a different model family for every delegated-result review. The resident lead is a separate host-level choice and does not create another subagent lane.

Where to change the model / effort

Five places — keep them in sync:

1. Claude subagent profiles — agents/*.md frontmatter:

   model: sonnet      # -> your role-specific tier: sonnet | opus | fable
   effort: low        # builders are pinned to max in the author's mapping

2. Codex profiles — both files per profile must match:
   • codex/agents/<name>.toml: model = "...", model_reasoning_effort = "..."
   • codex/profiles/<name>.config.toml: same model + model_reasoning_effort
3. The central gate — config/routing-gates.json: update exact profile, status, explicit-only selection, evidence references, model_families, the cross-family review_policy, the activation_policy, and any compound lane. Selections are only explicit-only or blocked; keep it that way.
4. The prose — claude/CLAUDE.delegation.md, codex/AGENTS.md, model-routing.md: update the reference mapping and lane-evidence notes.
5. Executable bridges and sync surfaces — for an external model, keep its bin/delegation-* (which source the shared helpers in bin/lib/), config/*-routing.json, skills/*-executor/SKILL.md, installer, uninstaller, doctor, routing-gate drift test, README, plugin manifests, and skills/orchestrate/* in step. Grok 4.6 is the current worked example: only builder and frontend-builder, Grok Build CLI, effort high.

Then re-run ./install.sh; it refreshes copied files and replaces the guarded user-direction blocks while retaining backups. Uninstalling first is unnecessary. Customized installations must preserve the user-direction guard: selections stay explicit-only/blocked and nothing may dispatch without the user naming the call for the current request.

Refreshing evidence and qualifying lanes

Do not inherit a global 1–10 score. Start from config/model-evidence.json and keep every model+harness+effort row exact:

• delegation-evidence check verifies schema and freshness;
• delegation-evidence lane <lane> shows relevant coverage without qualifying it;
• delegation-route check validates the operational decision graph;
• delegation-route table generates the internal table from the central gate;
• API cost per task is not subscription-bucket cost;
• WebDev preference applies only to frontend work;
• coding-agent results never substitute for reviewer precision/recall.
• put same-variant, lane-relevant rows in exact_evidence_ids; put neighboring efforts, different harnesses, or supporting domains in context_evidence_ids;
• never copy a legacy aggregate score without task count, repetitions, confidence, and limitations. Unknown sample metadata remains null.

To add a model, refresh current sources, add the exact variant, pre-commit a lane-specific threshold against its incumbent, run a small local runtime/scope smoke, then record the owner decision in the relevant fail-closed routing gate. See evaluation/README.md for the complete workflow.

Author-specific choices to reconsider

These are baked into the reference, not universal — decide for yourself:

• “Never use Haiku” and Fable-as-metered-credits are the author’s plan realities. Yours differ.
• Separate Sonnet-only weekly bucket (Claude Max): the author uses it only for very small non-builder tasks. Check whether your plan meters the same way before copying that choice.
• Lead = Opus, not the top model: because the author’s top model (Fable) bills as real money per call. If your top model is plan-included, your lead choice may differ.
• Fan-out caps ([agents] max_threads=3 / max_depth=1) and the ephemeral multi_agent=false guard are conservative defaults — tune to your appetite.

Minimal adaptation checklist

1. Map each role to one of your models (table above).
2. Edit agents/*.md + codex/agents/*.toml + codex/profiles/*.config.toml.
3. Refresh config/model-evidence.json and define lane-specific thresholds.
4. Update config/routing-gates.json; keep manual judgement explicit-only.
5. Run the local compatibility smoke and update only the relevant routing gate.
6. Run the gate and runner tests, then ./install.sh and ./doctor.sh.