precise-alloy · Copilot · Jul 6, 2026 · Jul 6, 2026
diff --git a/README.md b/README.md
@@ -10,7 +10,9 @@
 
 Pronounced: *REE-tus* `/ˈriː.tus/` - from Latin, meaning a prescribed rite or procedure.
 
-A disciplined workflow for reliable AI-assisted development. Ritus uses skill-based architecture, on-demand loading, chain-based routing, and independent verification to work across Claude Code, GitHub Copilot, and any language or stack.
+A disciplined workflow for reliable AI-assisted development. Ritus uses skill-based architecture, on-demand loading, chain-based routing, and independent verification, in any language or stack.
+
+**Platform support:** Ritus is built for **Claude Code** (first-class — Skill-tool invocation, programmatic subagent dispatch, per-tier model routing). **GitHub Copilot** support is **experimental**: skills load and the workflow applies, but model routing is advisory (you pick the model in the client) and subagent dispatch runs as separate manual passes rather than programmatic fan-out. See [Platform compatibility](#platform-compatibility).
 
 
 ---
@@ -117,6 +119,29 @@ missing project files. Run the setup again to create any new files. Existing use
 
 ---
 
+## Platform compatibility
+
+Ritus is authored around Claude Code primitives. The table below maps each Claude-specific
+concept to its GitHub Copilot (or generic-agent) equivalent so the workflow degrades
+predictably instead of silently.
+
+| Ritus concept                | Claude Code                          | GitHub Copilot / generic agent                                  |
+|------------------------------|--------------------------------------|-----------------------------------------------------------------|
+| Skill invocation             | `Skill` tool                         | `/command` or reading the matching `SKILL.md` on demand         |
+| Subagent dispatch            | fresh subagent (programmatic)        | run the step as a separate pass in a clean context (manual)     |
+| Model routing                | per-tier model set at dispatch       | advisory — pick the matching model in the client                |
+| Web fetch (in `pr-review`)   | `WebFetch` tool                      | any built-in web-fetch/browse tool, or paste content manually   |
+| Structured question prompts  | `AskUserQuestion` tool               | numbered plain-text options                                     |
+| Plugin root env var          | `CLAUDE_PLUGIN_ROOT`                 | `PLUGIN_ROOT`                                                   |
+
+**Model IDs are not hardcoded in skills.** Skills reference **capability tiers**
+(`haiku` = fast/cheap, `sonnet` = balanced, `opus` = deepest reasoning). The concrete
+model ID for each tier is resolved once, per platform, in `docs/profiles/runtime.yml`
+(`model_ids`) — the single place to update when models change. On GitHub Copilot the tier
+is a recommendation you apply manually, since Copilot has no programmatic dispatch.
+
+---
+
 ## Skills
 
 Skills are self-contained SKILL.md files loaded on-demand. Each has YAML frontmatter (name + description) so the

diff --git a/skills/pr-review/SKILL.md b/skills/pr-review/SKILL.md
@@ -14,7 +14,7 @@ can prove correctness with evidence at `file:line`.
 When the orchestrating session needs to run pr-review, dispatch a fresh subagent with:
 - **Model:** sonnet
 - **Effort:** high
-- **Tools:** Read, Grep, Glob, Bash, WebFetch
+- **Tools:** read, grep, glob, shell/bash, and a web-fetch capability (the `WebFetch` tool in Claude Code; any built-in web-fetch/browse tool on other agents, or paste the content manually)
 - **Constraints:** adversarial reviewer in fresh context; never apply fixes or modify source files; never create or
   modify environment files (.env, .env.local) — report missing setup to the user; use `git fetch origin` and
   `origin/<branch>` refs — never mutate local branches; default to "Request changes" — the burden of proof is on the

diff --git a/skills/setup/SKILL.md b/skills/setup/SKILL.md
@@ -268,35 +268,64 @@ If `{{TICKET_FORMAT}}` is set: `feat/PROJ-123-{slug}` or `feat/#123-{slug}`.
 
 ### Model routing table → `{{MODEL_ROUTING}}`
 
+Routing uses **capability tiers**, never hardcoded model IDs. A tier names the _class_ of
+model to dispatch; the concrete model ID is resolved once, per platform, in the **model
+resolution table** below. This keeps model IDs in a single place (`docs/profiles/runtime.yml`
+→ `model_ids`) so they can be updated without editing any skill.
+
+**Capability tiers:**
+
+| Tier     | Use for                                          |
+|----------|--------------------------------------------------|
+| `haiku`  | Fast, cheap — direct edits, per-task verification |
+| `sonnet` | Balanced — cross-module work, PR review          |
+| `opus`   | Deepest reasoning — pure architecture decisions   |
+
+**Model resolution table** — the single source of truth for concrete model IDs. Pick the
+column for the detected platform (Step 3). Model IDs change over time; treat these as
+examples and set them to the current models your platform exposes rather than trusting a
+dated default.
+
+| Tier     | Claude Code (example) | GitHub Copilot (example) |
+|----------|-----------------------|--------------------------|
+| `haiku`  | `claude-haiku-4-5`    | `gpt-5-mini`             |
+| `sonnet` | `claude-sonnet-4-5`   | `gpt-5`                  |
+| `opus`   | `claude-opus-4-1`     | `gpt-5` (high effort)    |
+
+> **GitHub Copilot:** Copilot selects the model in the client and has no programmatic
+> subagent dispatch. Treat the tier as a recommendation — pick the matching model manually
+> and run the verify/review steps as separate passes in a clean context. Cross-agent parity
+> is experimental.
+
 **cost-first:**
 
-| Triage                  | Model                       | Effort | Notes                           |
-|-------------------------|-----------------------------|--------|---------------------------------|
-| TRIVIAL                 | `claude-haiku-4-5-20251001` | low    | Direct edits, single file       |
-| SIMPLE                  | `claude-haiku-4-5-20251001` | medium | 3-section task note             |
-| STANDARD                | `claude-sonnet-4-6`         | medium | Cross-module, design decision   |
-| EPIC                    | `claude-sonnet-4-6`         | high   | Multi-session, new architecture |
-| Batch validate (pre-PR) | `claude-haiku-4-5-20251001` | low    | Diff + task review              |
+| Triage                  | Tier     | Effort | Notes                           |
+|-------------------------|----------|--------|---------------------------------|
+| TRIVIAL                 | `haiku`  | low    | Direct edits, single file       |
+| SIMPLE                  | `haiku`  | medium | 3-section task note             |
+| STANDARD                | `sonnet` | medium | Cross-module, design decision   |
+| EPIC                    | `sonnet` | high   | Multi-session, new architecture |
+| Batch validate (pre-PR) | `haiku`  | low    | Diff + task review              |
 
 **balanced (default):**
 
-| Triage                  | Model                                                              | Effort | Notes                         |
-|-------------------------|--------------------------------------------------------------------|----|-------------------------------|
-| TRIVIAL                 | `claude-haiku-4-5-20251001`                                        | low    | Direct edits, single file     |
-| SIMPLE                  | `claude-haiku-4-5-20251001`                                        | medium | 3-section task note           |
-| STANDARD                | `claude-sonnet-4-6`                                                | high   | Cross-module, design decision |
-| EPIC                    | `claude-sonnet-4-6` · `claude-opus-4-7` (pure arch decisions only) | high   | Multi-session                 |
-| Batch validate (pre-PR) | `claude-haiku-4-5-20251001`                                        | medium | Diff + task review            |
+| Triage                  | Tier                                        | Effort | Notes                         |
+|-------------------------|---------------------------------------------|--------|-------------------------------|
+| TRIVIAL                 | `haiku`                                     | low    | Direct edits, single file     |
+| SIMPLE                  | `haiku`                                     | medium | 3-section task note           |
+| STANDARD                | `sonnet`                                    | high   | Cross-module, design decision |
+| EPIC                    | `sonnet` · `opus` (pure arch decisions only) | high   | Multi-session                 |
+| Batch validate (pre-PR) | `haiku`                                     | medium | Diff + task review            |
 
 **quality-first:**
 
-| Triage                  | Model                       | Effort | Notes                           |
-|-------------------------|-----------------------------|--------|---------------------------------|
-| TRIVIAL                 | `claude-haiku-4-5-20251001` | medium | Direct edits, single file       |
-| SIMPLE                  | `claude-sonnet-4-6`         | medium | 3-section task note             |
-| STANDARD                | `claude-sonnet-4-6`         | high   | Cross-module, design decision   |
-| EPIC                    | `claude-opus-4-7`           | xhigh  | Multi-session, new architecture |
-| Batch validate (pre-PR) | `claude-sonnet-4-6`         | high   | Diff + task review              |
+| Triage                  | Tier     | Effort | Notes                           |
+|-------------------------|----------|--------|---------------------------------|
+| TRIVIAL                 | `haiku`  | medium | Direct edits, single file       |
+| SIMPLE                  | `sonnet` | medium | 3-section task note             |
+| STANDARD                | `sonnet` | high   | Cross-module, design decision   |
+| EPIC                    | `opus`   | xhigh  | Multi-session, new architecture |
+| Batch validate (pre-PR) | `sonnet` | high   | Diff + task review              |
 
 ---
 
@@ -340,7 +369,14 @@ Fill these fields:
 
 - `ai_tools` — auto-detect from platform: `Claude Code` if `CLAUDE_PLUGIN_ROOT` is set, `GitHub Copilot` if
   `PLUGIN_ROOT` is set, otherwise ask the user
-- `model_routing` — paste the correct table from §Model routing table above
+- `model_ids` — the **single** place concrete model IDs live. From the §Model resolution table above, paste the
+  column matching the detected platform (Claude Code or GitHub Copilot) as `haiku` / `sonnet` / `opus` keys.
+  These are examples — confirm they match models the platform currently exposes.
+- `model_routing` — paste the correct tier-based routing table (cost-first / balanced / quality-first) from §Model
+  routing table above. The table references tiers only; it never repeats concrete model IDs.
+
+> On **GitHub Copilot**, `model_routing` is advisory: Copilot has no programmatic subagent dispatch, so apply the
+> tier recommendation by selecting the model in the client and running verify/review as separate clean-context passes.
 
 ### Step 4: Render `docs/PROJECT_CONTEXT.md` from all `.yml` files
 
@@ -367,7 +403,7 @@ Setup complete — {{PROJECT_NAME}}
 WRITTEN:
   docs/profiles/project.yml   project config (name, language, framework)
   docs/profiles/team.yml      team config (size, git workflow, branches)
-  docs/profiles/runtime.yml   runtime config (AI tools, model routing)
+  docs/profiles/runtime.yml   runtime config (AI tools, model IDs, model routing)
   docs/PROJECT_CONTEXT.md     rendered from all profiles
   docs/ARCHITECTURE.md        header filled
 

diff --git a/skills/start-ritus/SKILL.md b/skills/start-ritus/SKILL.md
@@ -7,7 +7,7 @@ argument-hint: Provide the user's request so the router can choose the next appl
 > **Subagent guard:** If you were dispatched as a subagent to execute a specific task (verify-task, pr-review, or execute-task), skip this skill entirely.
 > Run only the skill you were dispatched to execute.
 >
-> **Mandatory skill check:** Before responding to any task — including clarifying questions — check whether an available skill applies. If a skill covers your task, you **must** invoke it via the Skill tool. Do not skip this check. This is not optional.
+> **Mandatory skill check:** Before responding to any task — including clarifying questions — check whether an available skill applies. If a skill covers your task, you **must** invoke it via your agent's skill-invocation mechanism (the `Skill` tool in Claude Code; a `/command` or reading the matching `SKILL.md` in GitHub Copilot / other agents). Do not skip this check. This is not optional.
 
 # Ritus
 
@@ -48,10 +48,10 @@ This ensures you never stop mid-chain and never skip steps in long context windo
 
 ## Skill invocation
 
-> **Terminology:** "Invoke" = trigger a skill via the Skill tool (user or system action). "Load" = a skill reading another skill's content as a companion standard (skill-to-skill action).
+> **Terminology:** "Invoke" = trigger a skill via your agent's skill-invocation mechanism — the `Skill` tool in Claude Code, or a `/command` / reading the target `SKILL.md` in GitHub Copilot and other agents. "Load" = a skill reading another skill's content as a companion standard (skill-to-skill action).
 
 Before starting work, check which skill matches the user's intent by reading the available skill descriptions.
-Invoke the matching skill via the Skill tool. Each skill's `## Next` section defines what happens after — follow
+Invoke the matching skill. Each skill's `## Next` section defines what happens after — follow
 it, do not pre-plan the full chain.
 
 ## Subagent configs
@@ -60,8 +60,14 @@ The **orchestrating session** is whatever agent loaded start-ritus — typically
 When a skill says to dispatch, spawn a **fresh subagent** and instruct it to load the target skill.
 Never use a skill name as the agent type — skill names are not agent types.
 
-| Subagent | Model | Effort | Key constraints |
-|----------|-------|--------|-----------------|
+> **Platform note:** The `Model` column below is a **capability tier** (`haiku` = fast/cheap,
+> `sonnet` = balanced, `opus` = deepest reasoning), not a model ID. Concrete IDs are resolved once
+> in `docs/profiles/runtime.yml` → `model_ids`. On GitHub Copilot there is no programmatic subagent
+> dispatch: apply the tier by picking the model in the client and run each dispatched skill as a
+> separate pass in a clean context.
+
+| Subagent | Model (tier) | Effort | Key constraints |
+|----------|--------------|--------|-----------------|
 | `execute-task` | per triage | per triage | Implement STEPS exactly; do not redesign |
 | `verify-task` | haiku | medium | Read-only except build/test/lint; never fix; never trust implementer claims |
 | `pr-review` | sonnet | high | Adversarial; never apply fixes; use `origin/` refs; default to "Request changes" |

diff --git a/templates/docs/PROJECT_CONTEXT.md b/templates/docs/PROJECT_CONTEXT.md
@@ -97,6 +97,16 @@ Traceability policy: `{{TRACEABILITY_POLICY}}`
 
 `{{AI_TOOLS}}`
 
+## Model resolution
+
+Skills reference capability tiers; concrete model IDs live only here (from `runtime.yml` → `model_ids`).
+
+| Tier     | Model ID           |
+|----------|--------------------|
+| `haiku`  | `{{MODEL_ID_HAIKU}}`  |
+| `sonnet` | `{{MODEL_ID_SONNET}}` |
+| `opus`   | `{{MODEL_ID_OPUS}}`   |
+
 ## Model routing
 
 {{MODEL_ROUTING}}

diff --git a/templates/docs/profiles/runtime.yml b/templates/docs/profiles/runtime.yml
@@ -2,8 +2,18 @@
 # Edit this file, then re-run the `setup` skill to regenerate docs/PROJECT_CONTEXT.md.
 # Values like off/no/yes must be quoted: key: "off"
 
-schema_version: 2
+schema_version: 3
 
 # Filled by setup
 ai_tools: ""
+
+# Single source of truth for concrete model IDs (per platform, filled by setup).
+# Skills reference capability tiers (haiku/sonnet/opus), not IDs — update models here only.
+# Model IDs change over time; set these to the models your platform currently exposes.
+model_ids:
+  haiku: ""   # fast/cheap  — e.g. Claude Code: claude-haiku-4-5 · GitHub Copilot: gpt-5-mini
+  sonnet: ""  # balanced    — e.g. Claude Code: claude-sonnet-4-5 · GitHub Copilot: gpt-5
+  opus: ""    # deep reason — e.g. Claude Code: claude-opus-4-1 · GitHub Copilot: gpt-5 (high effort)
+
+# Tier-based routing table (filled by setup). References tiers only, never concrete IDs.
 model_routing: ""